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
371 - On X86-64 the callee preserves all general purpose registers, except for
372 R11. R11 can be used as a scratch register. Floating-point registers
373 (XMMs/YMMs) are not preserved and need to be saved by the caller.
375 The idea behind this convention is to support calls to runtime functions
376 that have a hot path and a cold path. The hot path is usually a small piece
377 of code that doesn't many registers. The cold path might need to call out to
378 another function and therefore only needs to preserve the caller-saved
379 registers, which haven't already been saved by the caller. The
380 `PreserveMost` calling convention is very similar to the `cold` calling
381 convention in terms of caller/callee-saved registers, but they are used for
382 different types of function calls. `coldcc` is for function calls that are
383 rarely executed, whereas `preserve_mostcc` function calls are intended to be
384 on the hot path and definitely executed a lot. Furthermore `preserve_mostcc`
385 doesn't prevent the inliner from inlining the function call.
387 This calling convention will be used by a future version of the ObjectiveC
388 runtime and should therefore still be considered experimental at this time.
389 Although this convention was created to optimize certain runtime calls to
390 the ObjectiveC runtime, it is not limited to this runtime and might be used
391 by other runtimes in the future too. The current implementation only
392 supports X86-64, but the intention is to support more architectures in the
394 "``preserve_allcc``" - The `PreserveAll` calling convention
395 This calling convention attempts to make the code in the caller even less
396 intrusive than the `PreserveMost` calling convention. This calling
397 convention also behaves identical to the `C` calling convention on how
398 arguments and return values are passed, but it uses a different set of
399 caller/callee-saved registers. This removes the burden of saving and
400 recovering a large register set before and after the call in the caller.
402 - On X86-64 the callee preserves all general purpose registers, except for
403 R11. R11 can be used as a scratch register. Furthermore it also preserves
404 all floating-point registers (XMMs/YMMs).
406 The idea behind this convention is to support calls to runtime functions
407 that don't need to call out to any other functions.
409 This calling convention, like the `PreserveMost` calling convention, will be
410 used by a future version of the ObjectiveC runtime and should be considered
411 experimental at this time.
412 "``cc <n>``" - Numbered convention
413 Any calling convention may be specified by number, allowing
414 target-specific calling conventions to be used. Target specific
415 calling conventions start at 64.
417 More calling conventions can be added/defined on an as-needed basis, to
418 support Pascal conventions or any other well-known target-independent
421 .. _visibilitystyles:
426 All Global Variables and Functions have one of the following visibility
429 "``default``" - Default style
430 On targets that use the ELF object file format, default visibility
431 means that the declaration is visible to other modules and, in
432 shared libraries, means that the declared entity may be overridden.
433 On Darwin, default visibility means that the declaration is visible
434 to other modules. Default visibility corresponds to "external
435 linkage" in the language.
436 "``hidden``" - Hidden style
437 Two declarations of an object with hidden visibility refer to the
438 same object if they are in the same shared object. Usually, hidden
439 visibility indicates that the symbol will not be placed into the
440 dynamic symbol table, so no other module (executable or shared
441 library) can reference it directly.
442 "``protected``" - Protected style
443 On ELF, protected visibility indicates that the symbol will be
444 placed in the dynamic symbol table, but that references within the
445 defining module will bind to the local symbol. That is, the symbol
446 cannot be overridden by another module.
453 All Global Variables, Functions and Aliases can have one of the following
457 "``dllimport``" causes the compiler to reference a function or variable via
458 a global pointer to a pointer that is set up by the DLL exporting the
459 symbol. On Microsoft Windows targets, the pointer name is formed by
460 combining ``__imp_`` and the function or variable name.
462 "``dllexport``" causes the compiler to provide a global pointer to a pointer
463 in a DLL, so that it can be referenced with the ``dllimport`` attribute. On
464 Microsoft Windows targets, the pointer name is formed by combining
465 ``__imp_`` and the function or variable name. Since this storage class
466 exists for defining a dll interface, the compiler, assembler and linker know
467 it is externally referenced and must refrain from deleting the symbol.
472 LLVM IR allows you to specify name aliases for certain types. This can
473 make it easier to read the IR and make the IR more condensed
474 (particularly when recursive types are involved). An example of a name
479 %mytype = type { %mytype*, i32 }
481 You may give a name to any :ref:`type <typesystem>` except
482 ":ref:`void <t_void>`". Type name aliases may be used anywhere a type is
483 expected with the syntax "%mytype".
485 Note that type names are aliases for the structural type that they
486 indicate, and that you can therefore specify multiple names for the same
487 type. This often leads to confusing behavior when dumping out a .ll
488 file. Since LLVM IR uses structural typing, the name is not part of the
489 type. When printing out LLVM IR, the printer will pick *one name* to
490 render all types of a particular shape. This means that if you have code
491 where two different source types end up having the same LLVM type, that
492 the dumper will sometimes print the "wrong" or unexpected type. This is
493 an important design point and isn't going to change.
500 Global variables define regions of memory allocated at compilation time
503 Global variables definitions must be initialized, may have an explicit section
504 to be placed in, and may have an optional explicit alignment specified.
506 Global variables in other translation units can also be declared, in which
507 case they don't have an initializer.
509 A variable may be defined as ``thread_local``, which means that it will
510 not be shared by threads (each thread will have a separated copy of the
511 variable). Not all targets support thread-local variables. Optionally, a
512 TLS model may be specified:
515 For variables that are only used within the current shared library.
517 For variables in modules that will not be loaded dynamically.
519 For variables defined in the executable and only used within it.
521 The models correspond to the ELF TLS models; see `ELF Handling For
522 Thread-Local Storage <http://people.redhat.com/drepper/tls.pdf>`_ for
523 more information on under which circumstances the different models may
524 be used. The target may choose a different TLS model if the specified
525 model is not supported, or if a better choice of model can be made.
527 A variable may be defined as a global ``constant``, which indicates that
528 the contents of the variable will **never** be modified (enabling better
529 optimization, allowing the global data to be placed in the read-only
530 section of an executable, etc). Note that variables that need runtime
531 initialization cannot be marked ``constant`` as there is a store to the
534 LLVM explicitly allows *declarations* of global variables to be marked
535 constant, even if the final definition of the global is not. This
536 capability can be used to enable slightly better optimization of the
537 program, but requires the language definition to guarantee that
538 optimizations based on the 'constantness' are valid for the translation
539 units that do not include the definition.
541 As SSA values, global variables define pointer values that are in scope
542 (i.e. they dominate) all basic blocks in the program. Global variables
543 always define a pointer to their "content" type because they describe a
544 region of memory, and all memory objects in LLVM are accessed through
547 Global variables can be marked with ``unnamed_addr`` which indicates
548 that the address is not significant, only the content. Constants marked
549 like this can be merged with other constants if they have the same
550 initializer. Note that a constant with significant address *can* be
551 merged with a ``unnamed_addr`` constant, the result being a constant
552 whose address is significant.
554 A global variable may be declared to reside in a target-specific
555 numbered address space. For targets that support them, address spaces
556 may affect how optimizations are performed and/or what target
557 instructions are used to access the variable. The default address space
558 is zero. The address space qualifier must precede any other attributes.
560 LLVM allows an explicit section to be specified for globals. If the
561 target supports it, it will emit globals to the section specified.
563 By default, global initializers are optimized by assuming that global
564 variables defined within the module are not modified from their
565 initial values before the start of the global initializer. This is
566 true even for variables potentially accessible from outside the
567 module, including those with external linkage or appearing in
568 ``@llvm.used`` or dllexported variables. This assumption may be suppressed
569 by marking the variable with ``externally_initialized``.
571 An explicit alignment may be specified for a global, which must be a
572 power of 2. If not present, or if the alignment is set to zero, the
573 alignment of the global is set by the target to whatever it feels
574 convenient. If an explicit alignment is specified, the global is forced
575 to have exactly that alignment. Targets and optimizers are not allowed
576 to over-align the global if the global has an assigned section. In this
577 case, the extra alignment could be observable: for example, code could
578 assume that the globals are densely packed in their section and try to
579 iterate over them as an array, alignment padding would break this
582 Globals can also have a :ref:`DLL storage class <dllstorageclass>`.
586 [@<GlobalVarName> =] [Linkage] [Visibility] [DLLStorageClass] [ThreadLocal]
587 [AddrSpace] [unnamed_addr] [ExternallyInitialized]
588 <global | constant> <Type>
589 [, section "name"] [, align <Alignment>]
591 For example, the following defines a global in a numbered address space
592 with an initializer, section, and alignment:
596 @G = addrspace(5) constant float 1.0, section "foo", align 4
598 The following example just declares a global variable
602 @G = external global i32
604 The following example defines a thread-local global with the
605 ``initialexec`` TLS model:
609 @G = thread_local(initialexec) global i32 0, align 4
611 .. _functionstructure:
616 LLVM function definitions consist of the "``define``" keyword, an
617 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
618 style <visibility>`, an optional :ref:`DLL storage class <dllstorageclass>`,
619 an optional :ref:`calling convention <callingconv>`,
620 an optional ``unnamed_addr`` attribute, a return type, an optional
621 :ref:`parameter attribute <paramattrs>` for the return type, a function
622 name, a (possibly empty) argument list (each with optional :ref:`parameter
623 attributes <paramattrs>`), optional :ref:`function attributes <fnattrs>`,
624 an optional section, an optional alignment, an optional :ref:`garbage
625 collector name <gc>`, an optional :ref:`prefix <prefixdata>`, an opening
626 curly brace, a list of basic blocks, and a closing curly brace.
628 LLVM function declarations consist of the "``declare``" keyword, an
629 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
630 style <visibility>`, an optional :ref:`DLL storage class <dllstorageclass>`,
631 an optional :ref:`calling convention <callingconv>`,
632 an optional ``unnamed_addr`` attribute, a return type, an optional
633 :ref:`parameter attribute <paramattrs>` for the return type, a function
634 name, a possibly empty list of arguments, an optional alignment, an optional
635 :ref:`garbage collector name <gc>` and an optional :ref:`prefix <prefixdata>`.
637 A function definition contains a list of basic blocks, forming the CFG (Control
638 Flow Graph) for the function. Each basic block may optionally start with a label
639 (giving the basic block a symbol table entry), contains a list of instructions,
640 and ends with a :ref:`terminator <terminators>` instruction (such as a branch or
641 function return). If an explicit label is not provided, a block is assigned an
642 implicit numbered label, using the next value from the same counter as used for
643 unnamed temporaries (:ref:`see above<identifiers>`). For example, if a function
644 entry block does not have an explicit label, it will be assigned label "%0",
645 then the first unnamed temporary in that block will be "%1", etc.
647 The first basic block in a function is special in two ways: it is
648 immediately executed on entrance to the function, and it is not allowed
649 to have predecessor basic blocks (i.e. there can not be any branches to
650 the entry block of a function). Because the block can have no
651 predecessors, it also cannot have any :ref:`PHI nodes <i_phi>`.
653 LLVM allows an explicit section to be specified for functions. If the
654 target supports it, it will emit functions to the section specified.
656 An explicit alignment may be specified for a function. If not present,
657 or if the alignment is set to zero, the alignment of the function is set
658 by the target to whatever it feels convenient. If an explicit alignment
659 is specified, the function is forced to have at least that much
660 alignment. All alignments must be a power of 2.
662 If the ``unnamed_addr`` attribute is given, the address is know to not
663 be significant and two identical functions can be merged.
667 define [linkage] [visibility] [DLLStorageClass]
669 <ResultType> @<FunctionName> ([argument list])
670 [fn Attrs] [section "name"] [align N]
671 [gc] [prefix Constant] { ... }
678 Aliases act as "second name" for the aliasee value (which can be either
679 function, global variable, another alias or bitcast of global value).
680 Aliases may have an optional :ref:`linkage type <linkage>`, an optional
681 :ref:`visibility style <visibility>`, and an optional :ref:`DLL storage class
686 @<Name> = [Visibility] [DLLStorageClass] alias [Linkage] <AliaseeTy> @<Aliasee>
688 The linkage must be one of ``private``, ``linker_private``,
689 ``linker_private_weak``, ``internal``, ``linkonce``, ``weak``,
690 ``linkonce_odr``, ``weak_odr``, ``external``. Note that some system linkers
691 might not correctly handle dropping a weak symbol that is aliased by a non-weak
694 .. _namedmetadatastructure:
699 Named metadata is a collection of metadata. :ref:`Metadata
700 nodes <metadata>` (but not metadata strings) are the only valid
701 operands for a named metadata.
705 ; Some unnamed metadata nodes, which are referenced by the named metadata.
706 !0 = metadata !{metadata !"zero"}
707 !1 = metadata !{metadata !"one"}
708 !2 = metadata !{metadata !"two"}
710 !name = !{!0, !1, !2}
717 The return type and each parameter of a function type may have a set of
718 *parameter attributes* associated with them. Parameter attributes are
719 used to communicate additional information about the result or
720 parameters of a function. Parameter attributes are considered to be part
721 of the function, not of the function type, so functions with different
722 parameter attributes can have the same function type.
724 Parameter attributes are simple keywords that follow the type specified.
725 If multiple parameter attributes are needed, they are space separated.
730 declare i32 @printf(i8* noalias nocapture, ...)
731 declare i32 @atoi(i8 zeroext)
732 declare signext i8 @returns_signed_char()
734 Note that any attributes for the function result (``nounwind``,
735 ``readonly``) come immediately after the argument list.
737 Currently, only the following parameter attributes are defined:
740 This indicates to the code generator that the parameter or return
741 value should be zero-extended to the extent required by the target's
742 ABI (which is usually 32-bits, but is 8-bits for a i1 on x86-64) by
743 the caller (for a parameter) or the callee (for a return value).
745 This indicates to the code generator that the parameter or return
746 value should be sign-extended to the extent required by the target's
747 ABI (which is usually 32-bits) by the caller (for a parameter) or
748 the callee (for a return value).
750 This indicates that this parameter or return value should be treated
751 in a special target-dependent fashion during while emitting code for
752 a function call or return (usually, by putting it in a register as
753 opposed to memory, though some targets use it to distinguish between
754 two different kinds of registers). Use of this attribute is
757 This indicates that the pointer parameter should really be passed by
758 value to the function. The attribute implies that a hidden copy of
759 the pointee is made between the caller and the callee, so the callee
760 is unable to modify the value in the caller. This attribute is only
761 valid on LLVM pointer arguments. It is generally used to pass
762 structs and arrays by value, but is also valid on pointers to
763 scalars. The copy is considered to belong to the caller not the
764 callee (for example, ``readonly`` functions should not write to
765 ``byval`` parameters). This is not a valid attribute for return
768 The byval attribute also supports specifying an alignment with the
769 align attribute. It indicates the alignment of the stack slot to
770 form and the known alignment of the pointer specified to the call
771 site. If the alignment is not specified, then the code generator
772 makes a target-specific assumption.
778 .. Warning:: This feature is unstable and not fully implemented.
780 The ``inalloca`` argument attribute allows the caller to take the
781 address of all stack-allocated arguments to a ``call`` or ``invoke``
782 before it executes. It is similar to ``byval`` in that it is used
783 to pass arguments by value, but it guarantees that the argument will
786 To be :ref:`well formed <wellformed>`, an alloca may be used as an
787 ``inalloca`` argument at most once. The attribute can only be
788 applied to the last parameter, and it guarantees that they are
789 passed in memory. The ``inalloca`` attribute cannot be used in
790 conjunction with other attributes that affect argument storage, like
791 ``inreg``, ``nest``, ``sret``, or ``byval``. The ``inalloca`` stack
792 space is considered to be clobbered by any call that uses it, so any
793 ``inalloca`` parameters cannot be marked ``readonly``.
795 When the call site is reached, the argument allocation must have
796 been the most recent stack allocation that is still live, or the
797 results are undefined. It is possible to allocate additional stack
798 space after an argument allocation and before its call site, but it
799 must be cleared off with :ref:`llvm.stackrestore
802 See :doc:`InAlloca` for more information on how to use this
806 This indicates that the pointer parameter specifies the address of a
807 structure that is the return value of the function in the source
808 program. This pointer must be guaranteed by the caller to be valid:
809 loads and stores to the structure may be assumed by the callee
810 not to trap and to be properly aligned. This may only be applied to
811 the first parameter. This is not a valid attribute for return
814 This indicates that pointer values :ref:`based <pointeraliasing>` on
815 the argument or return value do not alias pointer values which are
816 not *based* on it, ignoring certain "irrelevant" dependencies. For a
817 call to the parent function, dependencies between memory references
818 from before or after the call and from those during the call are
819 "irrelevant" to the ``noalias`` keyword for the arguments and return
820 value used in that call. The caller shares the responsibility with
821 the callee for ensuring that these requirements are met. For further
822 details, please see the discussion of the NoAlias response in `alias
823 analysis <AliasAnalysis.html#MustMayNo>`_.
825 Note that this definition of ``noalias`` is intentionally similar
826 to the definition of ``restrict`` in C99 for function arguments,
827 though it is slightly weaker.
829 For function return values, C99's ``restrict`` is not meaningful,
830 while LLVM's ``noalias`` is.
832 This indicates that the callee does not make any copies of the
833 pointer that outlive the callee itself. This is not a valid
834 attribute for return values.
839 This indicates that the pointer parameter can be excised using the
840 :ref:`trampoline intrinsics <int_trampoline>`. This is not a valid
841 attribute for return values and can only be applied to one parameter.
844 This indicates that the function always returns the argument as its return
845 value. This is an optimization hint to the code generator when generating
846 the caller, allowing tail call optimization and omission of register saves
847 and restores in some cases; it is not checked or enforced when generating
848 the callee. The parameter and the function return type must be valid
849 operands for the :ref:`bitcast instruction <i_bitcast>`. This is not a
850 valid attribute for return values and can only be applied to one parameter.
854 Garbage Collector Names
855 -----------------------
857 Each function may specify a garbage collector name, which is simply a
862 define void @f() gc "name" { ... }
864 The compiler declares the supported values of *name*. Specifying a
865 collector which will cause the compiler to alter its output in order to
866 support the named garbage collection algorithm.
873 Prefix data is data associated with a function which the code generator
874 will emit immediately before the function body. The purpose of this feature
875 is to allow frontends to associate language-specific runtime metadata with
876 specific functions and make it available through the function pointer while
877 still allowing the function pointer to be called. To access the data for a
878 given function, a program may bitcast the function pointer to a pointer to
879 the constant's type. This implies that the IR symbol points to the start
882 To maintain the semantics of ordinary function calls, the prefix data must
883 have a particular format. Specifically, it must begin with a sequence of
884 bytes which decode to a sequence of machine instructions, valid for the
885 module's target, which transfer control to the point immediately succeeding
886 the prefix data, without performing any other visible action. This allows
887 the inliner and other passes to reason about the semantics of the function
888 definition without needing to reason about the prefix data. Obviously this
889 makes the format of the prefix data highly target dependent.
891 Prefix data is laid out as if it were an initializer for a global variable
892 of the prefix data's type. No padding is automatically placed between the
893 prefix data and the function body. If padding is required, it must be part
896 A trivial example of valid prefix data for the x86 architecture is ``i8 144``,
897 which encodes the ``nop`` instruction:
901 define void @f() prefix i8 144 { ... }
903 Generally prefix data can be formed by encoding a relative branch instruction
904 which skips the metadata, as in this example of valid prefix data for the
905 x86_64 architecture, where the first two bytes encode ``jmp .+10``:
909 %0 = type <{ i8, i8, i8* }>
911 define void @f() prefix %0 <{ i8 235, i8 8, i8* @md}> { ... }
913 A function may have prefix data but no body. This has similar semantics
914 to the ``available_externally`` linkage in that the data may be used by the
915 optimizers but will not be emitted in the object file.
922 Attribute groups are groups of attributes that are referenced by objects within
923 the IR. They are important for keeping ``.ll`` files readable, because a lot of
924 functions will use the same set of attributes. In the degenerative case of a
925 ``.ll`` file that corresponds to a single ``.c`` file, the single attribute
926 group will capture the important command line flags used to build that file.
928 An attribute group is a module-level object. To use an attribute group, an
929 object references the attribute group's ID (e.g. ``#37``). An object may refer
930 to more than one attribute group. In that situation, the attributes from the
931 different groups are merged.
933 Here is an example of attribute groups for a function that should always be
934 inlined, has a stack alignment of 4, and which shouldn't use SSE instructions:
938 ; Target-independent attributes:
939 attributes #0 = { alwaysinline alignstack=4 }
941 ; Target-dependent attributes:
942 attributes #1 = { "no-sse" }
944 ; Function @f has attributes: alwaysinline, alignstack=4, and "no-sse".
945 define void @f() #0 #1 { ... }
952 Function attributes are set to communicate additional information about
953 a function. Function attributes are considered to be part of the
954 function, not of the function type, so functions with different function
955 attributes can have the same function type.
957 Function attributes are simple keywords that follow the type specified.
958 If multiple attributes are needed, they are space separated. For
963 define void @f() noinline { ... }
964 define void @f() alwaysinline { ... }
965 define void @f() alwaysinline optsize { ... }
966 define void @f() optsize { ... }
969 This attribute indicates that, when emitting the prologue and
970 epilogue, the backend should forcibly align the stack pointer.
971 Specify the desired alignment, which must be a power of two, in
974 This attribute indicates that the inliner should attempt to inline
975 this function into callers whenever possible, ignoring any active
976 inlining size threshold for this caller.
978 This indicates that the callee function at a call site should be
979 recognized as a built-in function, even though the function's declaration
980 uses the ``nobuiltin`` attribute. This is only valid at call sites for
981 direct calls to functions which are declared with the ``nobuiltin``
984 This attribute indicates that this function is rarely called. When
985 computing edge weights, basic blocks post-dominated by a cold
986 function call are also considered to be cold; and, thus, given low
989 This attribute indicates that the source code contained a hint that
990 inlining this function is desirable (such as the "inline" keyword in
991 C/C++). It is just a hint; it imposes no requirements on the
994 This attribute suggests that optimization passes and code generator
995 passes make choices that keep the code size of this function as small
996 as possible and perform optimizations that may sacrifice runtime
997 performance in order to minimize the size of the generated code.
999 This attribute disables prologue / epilogue emission for the
1000 function. This can have very system-specific consequences.
1002 This indicates that the callee function at a call site is not recognized as
1003 a built-in function. LLVM will retain the original call and not replace it
1004 with equivalent code based on the semantics of the built-in function, unless
1005 the call site uses the ``builtin`` attribute. This is valid at call sites
1006 and on function declarations and definitions.
1008 This attribute indicates that calls to the function cannot be
1009 duplicated. A call to a ``noduplicate`` function may be moved
1010 within its parent function, but may not be duplicated within
1011 its parent function.
1013 A function containing a ``noduplicate`` call may still
1014 be an inlining candidate, provided that the call is not
1015 duplicated by inlining. That implies that the function has
1016 internal linkage and only has one call site, so the original
1017 call is dead after inlining.
1019 This attributes disables implicit floating point instructions.
1021 This attribute indicates that the inliner should never inline this
1022 function in any situation. This attribute may not be used together
1023 with the ``alwaysinline`` attribute.
1025 This attribute suppresses lazy symbol binding for the function. This
1026 may make calls to the function faster, at the cost of extra program
1027 startup time if the function is not called during program startup.
1029 This attribute indicates that the code generator should not use a
1030 red zone, even if the target-specific ABI normally permits it.
1032 This function attribute indicates that the function never returns
1033 normally. This produces undefined behavior at runtime if the
1034 function ever does dynamically return.
1036 This function attribute indicates that the function never returns
1037 with an unwind or exceptional control flow. If the function does
1038 unwind, its runtime behavior is undefined.
1040 This function attribute indicates that the function is not optimized
1041 by any optimization or code generator passes with the
1042 exception of interprocedural optimization passes.
1043 This attribute cannot be used together with the ``alwaysinline``
1044 attribute; this attribute is also incompatible
1045 with the ``minsize`` attribute and the ``optsize`` attribute.
1047 This attribute requires the ``noinline`` attribute to be specified on
1048 the function as well, so the function is never inlined into any caller.
1049 Only functions with the ``alwaysinline`` attribute are valid
1050 candidates for inlining into the body of this function.
1052 This attribute suggests that optimization passes and code generator
1053 passes make choices that keep the code size of this function low,
1054 and otherwise do optimizations specifically to reduce code size as
1055 long as they do not significantly impact runtime performance.
1057 On a function, this attribute indicates that the function computes its
1058 result (or decides to unwind an exception) based strictly on its arguments,
1059 without dereferencing any pointer arguments or otherwise accessing
1060 any mutable state (e.g. memory, control registers, etc) visible to
1061 caller functions. It does not write through any pointer arguments
1062 (including ``byval`` arguments) and never changes any state visible
1063 to callers. This means that it cannot unwind exceptions by calling
1064 the ``C++`` exception throwing methods.
1066 On an argument, this attribute indicates that the function does not
1067 dereference that pointer argument, even though it may read or write the
1068 memory that the pointer points to if accessed through other pointers.
1070 On a function, this attribute indicates that the function does not write
1071 through any pointer arguments (including ``byval`` arguments) or otherwise
1072 modify any state (e.g. memory, control registers, etc) visible to
1073 caller functions. It may dereference pointer arguments and read
1074 state that may be set in the caller. A readonly function always
1075 returns the same value (or unwinds an exception identically) when
1076 called with the same set of arguments and global state. It cannot
1077 unwind an exception by calling the ``C++`` exception throwing
1080 On an argument, this attribute indicates that the function does not write
1081 through this pointer argument, even though it may write to the memory that
1082 the pointer points to.
1084 This attribute indicates that this function can return twice. The C
1085 ``setjmp`` is an example of such a function. The compiler disables
1086 some optimizations (like tail calls) in the caller of these
1088 ``sanitize_address``
1089 This attribute indicates that AddressSanitizer checks
1090 (dynamic address safety analysis) are enabled for this function.
1092 This attribute indicates that MemorySanitizer checks (dynamic detection
1093 of accesses to uninitialized memory) are enabled for this function.
1095 This attribute indicates that ThreadSanitizer checks
1096 (dynamic thread safety analysis) are enabled for this function.
1098 This attribute indicates that the function should emit a stack
1099 smashing protector. It is in the form of a "canary" --- a random value
1100 placed on the stack before the local variables that's checked upon
1101 return from the function to see if it has been overwritten. A
1102 heuristic is used to determine if a function needs stack protectors
1103 or not. The heuristic used will enable protectors for functions with:
1105 - Character arrays larger than ``ssp-buffer-size`` (default 8).
1106 - Aggregates containing character arrays larger than ``ssp-buffer-size``.
1107 - Calls to alloca() with variable sizes or constant sizes greater than
1108 ``ssp-buffer-size``.
1110 If a function that has an ``ssp`` attribute is inlined into a
1111 function that doesn't have an ``ssp`` attribute, then the resulting
1112 function will have an ``ssp`` attribute.
1114 This attribute indicates that the function should *always* emit a
1115 stack smashing protector. This overrides the ``ssp`` function
1118 If a function that has an ``sspreq`` attribute is inlined into a
1119 function that doesn't have an ``sspreq`` attribute or which has an
1120 ``ssp`` or ``sspstrong`` attribute, then the resulting function will have
1121 an ``sspreq`` attribute.
1123 This attribute indicates that the function should emit a stack smashing
1124 protector. This attribute causes a strong heuristic to be used when
1125 determining if a function needs stack protectors. The strong heuristic
1126 will enable protectors for functions with:
1128 - Arrays of any size and type
1129 - Aggregates containing an array of any size and type.
1130 - Calls to alloca().
1131 - Local variables that have had their address taken.
1133 This overrides the ``ssp`` function attribute.
1135 If a function that has an ``sspstrong`` attribute is inlined into a
1136 function that doesn't have an ``sspstrong`` attribute, then the
1137 resulting function will have an ``sspstrong`` attribute.
1139 This attribute indicates that the ABI being targeted requires that
1140 an unwind table entry be produce for this function even if we can
1141 show that no exceptions passes by it. This is normally the case for
1142 the ELF x86-64 abi, but it can be disabled for some compilation
1147 Module-Level Inline Assembly
1148 ----------------------------
1150 Modules may contain "module-level inline asm" blocks, which corresponds
1151 to the GCC "file scope inline asm" blocks. These blocks are internally
1152 concatenated by LLVM and treated as a single unit, but may be separated
1153 in the ``.ll`` file if desired. The syntax is very simple:
1155 .. code-block:: llvm
1157 module asm "inline asm code goes here"
1158 module asm "more can go here"
1160 The strings can contain any character by escaping non-printable
1161 characters. The escape sequence used is simply "\\xx" where "xx" is the
1162 two digit hex code for the number.
1164 The inline asm code is simply printed to the machine code .s file when
1165 assembly code is generated.
1167 .. _langref_datalayout:
1172 A module may specify a target specific data layout string that specifies
1173 how data is to be laid out in memory. The syntax for the data layout is
1176 .. code-block:: llvm
1178 target datalayout = "layout specification"
1180 The *layout specification* consists of a list of specifications
1181 separated by the minus sign character ('-'). Each specification starts
1182 with a letter and may include other information after the letter to
1183 define some aspect of the data layout. The specifications accepted are
1187 Specifies that the target lays out data in big-endian form. That is,
1188 the bits with the most significance have the lowest address
1191 Specifies that the target lays out data in little-endian form. That
1192 is, the bits with the least significance have the lowest address
1195 Specifies the natural alignment of the stack in bits. Alignment
1196 promotion of stack variables is limited to the natural stack
1197 alignment to avoid dynamic stack realignment. The stack alignment
1198 must be a multiple of 8-bits. If omitted, the natural stack
1199 alignment defaults to "unspecified", which does not prevent any
1200 alignment promotions.
1201 ``p[n]:<size>:<abi>:<pref>``
1202 This specifies the *size* of a pointer and its ``<abi>`` and
1203 ``<pref>``\erred alignments for address space ``n``. All sizes are in
1204 bits. The address space, ``n`` is optional, and if not specified,
1205 denotes the default address space 0. The value of ``n`` must be
1206 in the range [1,2^23).
1207 ``i<size>:<abi>:<pref>``
1208 This specifies the alignment for an integer type of a given bit
1209 ``<size>``. The value of ``<size>`` must be in the range [1,2^23).
1210 ``v<size>:<abi>:<pref>``
1211 This specifies the alignment for a vector type of a given bit
1213 ``f<size>:<abi>:<pref>``
1214 This specifies the alignment for a floating point type of a given bit
1215 ``<size>``. Only values of ``<size>`` that are supported by the target
1216 will work. 32 (float) and 64 (double) are supported on all targets; 80
1217 or 128 (different flavors of long double) are also supported on some
1220 This specifies the alignment for an object of aggregate type.
1222 If present, specifies that llvm names are mangled in the output. The
1225 * ``e``: ELF mangling: Private symbols get a ``.L`` prefix.
1226 * ``m``: Mips mangling: Private symbols get a ``$`` prefix.
1227 * ``o``: Mach-O mangling: Private symbols get ``L`` prefix. Other
1228 symbols get a ``_`` prefix.
1229 * ``w``: Windows COFF prefix: Similar to Mach-O, but stdcall and fastcall
1230 functions also get a suffix based on the frame size.
1231 ``n<size1>:<size2>:<size3>...``
1232 This specifies a set of native integer widths for the target CPU in
1233 bits. For example, it might contain ``n32`` for 32-bit PowerPC,
1234 ``n32:64`` for PowerPC 64, or ``n8:16:32:64`` for X86-64. Elements of
1235 this set are considered to support most general arithmetic operations
1238 On every specification that takes a ``<abi>:<pref>``, specifying the
1239 ``<pref>`` alignment is optional. If omitted, the preceding ``:``
1240 should be omitted too and ``<pref>`` will be equal to ``<abi>``.
1242 When constructing the data layout for a given target, LLVM starts with a
1243 default set of specifications which are then (possibly) overridden by
1244 the specifications in the ``datalayout`` keyword. The default
1245 specifications are given in this list:
1247 - ``E`` - big endian
1248 - ``p:64:64:64`` - 64-bit pointers with 64-bit alignment.
1249 - ``p[n]:64:64:64`` - Other address spaces are assumed to be the
1250 same as the default address space.
1251 - ``S0`` - natural stack alignment is unspecified
1252 - ``i1:8:8`` - i1 is 8-bit (byte) aligned
1253 - ``i8:8:8`` - i8 is 8-bit (byte) aligned
1254 - ``i16:16:16`` - i16 is 16-bit aligned
1255 - ``i32:32:32`` - i32 is 32-bit aligned
1256 - ``i64:32:64`` - i64 has ABI alignment of 32-bits but preferred
1257 alignment of 64-bits
1258 - ``f16:16:16`` - half is 16-bit aligned
1259 - ``f32:32:32`` - float is 32-bit aligned
1260 - ``f64:64:64`` - double is 64-bit aligned
1261 - ``f128:128:128`` - quad is 128-bit aligned
1262 - ``v64:64:64`` - 64-bit vector is 64-bit aligned
1263 - ``v128:128:128`` - 128-bit vector is 128-bit aligned
1264 - ``a:0:64`` - aggregates are 64-bit aligned
1266 When LLVM is determining the alignment for a given type, it uses the
1269 #. If the type sought is an exact match for one of the specifications,
1270 that specification is used.
1271 #. If no match is found, and the type sought is an integer type, then
1272 the smallest integer type that is larger than the bitwidth of the
1273 sought type is used. If none of the specifications are larger than
1274 the bitwidth then the largest integer type is used. For example,
1275 given the default specifications above, the i7 type will use the
1276 alignment of i8 (next largest) while both i65 and i256 will use the
1277 alignment of i64 (largest specified).
1278 #. If no match is found, and the type sought is a vector type, then the
1279 largest vector type that is smaller than the sought vector type will
1280 be used as a fall back. This happens because <128 x double> can be
1281 implemented in terms of 64 <2 x double>, for example.
1283 The function of the data layout string may not be what you expect.
1284 Notably, this is not a specification from the frontend of what alignment
1285 the code generator should use.
1287 Instead, if specified, the target data layout is required to match what
1288 the ultimate *code generator* expects. This string is used by the
1289 mid-level optimizers to improve code, and this only works if it matches
1290 what the ultimate code generator uses. If you would like to generate IR
1291 that does not embed this target-specific detail into the IR, then you
1292 don't have to specify the string. This will disable some optimizations
1293 that require precise layout information, but this also prevents those
1294 optimizations from introducing target specificity into the IR.
1301 A module may specify a target triple string that describes the target
1302 host. The syntax for the target triple is simply:
1304 .. code-block:: llvm
1306 target triple = "x86_64-apple-macosx10.7.0"
1308 The *target triple* string consists of a series of identifiers delimited
1309 by the minus sign character ('-'). The canonical forms are:
1313 ARCHITECTURE-VENDOR-OPERATING_SYSTEM
1314 ARCHITECTURE-VENDOR-OPERATING_SYSTEM-ENVIRONMENT
1316 This information is passed along to the backend so that it generates
1317 code for the proper architecture. It's possible to override this on the
1318 command line with the ``-mtriple`` command line option.
1320 .. _pointeraliasing:
1322 Pointer Aliasing Rules
1323 ----------------------
1325 Any memory access must be done through a pointer value associated with
1326 an address range of the memory access, otherwise the behavior is
1327 undefined. Pointer values are associated with address ranges according
1328 to the following rules:
1330 - A pointer value is associated with the addresses associated with any
1331 value it is *based* on.
1332 - An address of a global variable is associated with the address range
1333 of the variable's storage.
1334 - The result value of an allocation instruction is associated with the
1335 address range of the allocated storage.
1336 - A null pointer in the default address-space is associated with no
1338 - An integer constant other than zero or a pointer value returned from
1339 a function not defined within LLVM may be associated with address
1340 ranges allocated through mechanisms other than those provided by
1341 LLVM. Such ranges shall not overlap with any ranges of addresses
1342 allocated by mechanisms provided by LLVM.
1344 A pointer value is *based* on another pointer value according to the
1347 - A pointer value formed from a ``getelementptr`` operation is *based*
1348 on the first operand of the ``getelementptr``.
1349 - The result value of a ``bitcast`` is *based* on the operand of the
1351 - A pointer value formed by an ``inttoptr`` is *based* on all pointer
1352 values that contribute (directly or indirectly) to the computation of
1353 the pointer's value.
1354 - The "*based* on" relationship is transitive.
1356 Note that this definition of *"based"* is intentionally similar to the
1357 definition of *"based"* in C99, though it is slightly weaker.
1359 LLVM IR does not associate types with memory. The result type of a
1360 ``load`` merely indicates the size and alignment of the memory from
1361 which to load, as well as the interpretation of the value. The first
1362 operand type of a ``store`` similarly only indicates the size and
1363 alignment of the store.
1365 Consequently, type-based alias analysis, aka TBAA, aka
1366 ``-fstrict-aliasing``, is not applicable to general unadorned LLVM IR.
1367 :ref:`Metadata <metadata>` may be used to encode additional information
1368 which specialized optimization passes may use to implement type-based
1373 Volatile Memory Accesses
1374 ------------------------
1376 Certain memory accesses, such as :ref:`load <i_load>`'s,
1377 :ref:`store <i_store>`'s, and :ref:`llvm.memcpy <int_memcpy>`'s may be
1378 marked ``volatile``. The optimizers must not change the number of
1379 volatile operations or change their order of execution relative to other
1380 volatile operations. The optimizers *may* change the order of volatile
1381 operations relative to non-volatile operations. This is not Java's
1382 "volatile" and has no cross-thread synchronization behavior.
1384 IR-level volatile loads and stores cannot safely be optimized into
1385 llvm.memcpy or llvm.memmove intrinsics even when those intrinsics are
1386 flagged volatile. Likewise, the backend should never split or merge
1387 target-legal volatile load/store instructions.
1389 .. admonition:: Rationale
1391 Platforms may rely on volatile loads and stores of natively supported
1392 data width to be executed as single instruction. For example, in C
1393 this holds for an l-value of volatile primitive type with native
1394 hardware support, but not necessarily for aggregate types. The
1395 frontend upholds these expectations, which are intentionally
1396 unspecified in the IR. The rules above ensure that IR transformation
1397 do not violate the frontend's contract with the language.
1401 Memory Model for Concurrent Operations
1402 --------------------------------------
1404 The LLVM IR does not define any way to start parallel threads of
1405 execution or to register signal handlers. Nonetheless, there are
1406 platform-specific ways to create them, and we define LLVM IR's behavior
1407 in their presence. This model is inspired by the C++0x memory model.
1409 For a more informal introduction to this model, see the :doc:`Atomics`.
1411 We define a *happens-before* partial order as the least partial order
1414 - Is a superset of single-thread program order, and
1415 - When a *synchronizes-with* ``b``, includes an edge from ``a`` to
1416 ``b``. *Synchronizes-with* pairs are introduced by platform-specific
1417 techniques, like pthread locks, thread creation, thread joining,
1418 etc., and by atomic instructions. (See also :ref:`Atomic Memory Ordering
1419 Constraints <ordering>`).
1421 Note that program order does not introduce *happens-before* edges
1422 between a thread and signals executing inside that thread.
1424 Every (defined) read operation (load instructions, memcpy, atomic
1425 loads/read-modify-writes, etc.) R reads a series of bytes written by
1426 (defined) write operations (store instructions, atomic
1427 stores/read-modify-writes, memcpy, etc.). For the purposes of this
1428 section, initialized globals are considered to have a write of the
1429 initializer which is atomic and happens before any other read or write
1430 of the memory in question. For each byte of a read R, R\ :sub:`byte`
1431 may see any write to the same byte, except:
1433 - If write\ :sub:`1` happens before write\ :sub:`2`, and
1434 write\ :sub:`2` happens before R\ :sub:`byte`, then
1435 R\ :sub:`byte` does not see write\ :sub:`1`.
1436 - If R\ :sub:`byte` happens before write\ :sub:`3`, then
1437 R\ :sub:`byte` does not see write\ :sub:`3`.
1439 Given that definition, R\ :sub:`byte` is defined as follows:
1441 - If R is volatile, the result is target-dependent. (Volatile is
1442 supposed to give guarantees which can support ``sig_atomic_t`` in
1443 C/C++, and may be used for accesses to addresses which do not behave
1444 like normal memory. It does not generally provide cross-thread
1446 - Otherwise, if there is no write to the same byte that happens before
1447 R\ :sub:`byte`, R\ :sub:`byte` returns ``undef`` for that byte.
1448 - Otherwise, if R\ :sub:`byte` may see exactly one write,
1449 R\ :sub:`byte` returns the value written by that write.
1450 - Otherwise, if R is atomic, and all the writes R\ :sub:`byte` may
1451 see are atomic, it chooses one of the values written. See the :ref:`Atomic
1452 Memory Ordering Constraints <ordering>` section for additional
1453 constraints on how the choice is made.
1454 - Otherwise R\ :sub:`byte` returns ``undef``.
1456 R returns the value composed of the series of bytes it read. This
1457 implies that some bytes within the value may be ``undef`` **without**
1458 the entire value being ``undef``. Note that this only defines the
1459 semantics of the operation; it doesn't mean that targets will emit more
1460 than one instruction to read the series of bytes.
1462 Note that in cases where none of the atomic intrinsics are used, this
1463 model places only one restriction on IR transformations on top of what
1464 is required for single-threaded execution: introducing a store to a byte
1465 which might not otherwise be stored is not allowed in general.
1466 (Specifically, in the case where another thread might write to and read
1467 from an address, introducing a store can change a load that may see
1468 exactly one write into a load that may see multiple writes.)
1472 Atomic Memory Ordering Constraints
1473 ----------------------------------
1475 Atomic instructions (:ref:`cmpxchg <i_cmpxchg>`,
1476 :ref:`atomicrmw <i_atomicrmw>`, :ref:`fence <i_fence>`,
1477 :ref:`atomic load <i_load>`, and :ref:`atomic store <i_store>`) take
1478 an ordering parameter that determines which other atomic instructions on
1479 the same address they *synchronize with*. These semantics are borrowed
1480 from Java and C++0x, but are somewhat more colloquial. If these
1481 descriptions aren't precise enough, check those specs (see spec
1482 references in the :doc:`atomics guide <Atomics>`).
1483 :ref:`fence <i_fence>` instructions treat these orderings somewhat
1484 differently since they don't take an address. See that instruction's
1485 documentation for details.
1487 For a simpler introduction to the ordering constraints, see the
1491 The set of values that can be read is governed by the happens-before
1492 partial order. A value cannot be read unless some operation wrote
1493 it. This is intended to provide a guarantee strong enough to model
1494 Java's non-volatile shared variables. This ordering cannot be
1495 specified for read-modify-write operations; it is not strong enough
1496 to make them atomic in any interesting way.
1498 In addition to the guarantees of ``unordered``, there is a single
1499 total order for modifications by ``monotonic`` operations on each
1500 address. All modification orders must be compatible with the
1501 happens-before order. There is no guarantee that the modification
1502 orders can be combined to a global total order for the whole program
1503 (and this often will not be possible). The read in an atomic
1504 read-modify-write operation (:ref:`cmpxchg <i_cmpxchg>` and
1505 :ref:`atomicrmw <i_atomicrmw>`) reads the value in the modification
1506 order immediately before the value it writes. If one atomic read
1507 happens before another atomic read of the same address, the later
1508 read must see the same value or a later value in the address's
1509 modification order. This disallows reordering of ``monotonic`` (or
1510 stronger) operations on the same address. If an address is written
1511 ``monotonic``-ally by one thread, and other threads ``monotonic``-ally
1512 read that address repeatedly, the other threads must eventually see
1513 the write. This corresponds to the C++0x/C1x
1514 ``memory_order_relaxed``.
1516 In addition to the guarantees of ``monotonic``, a
1517 *synchronizes-with* edge may be formed with a ``release`` operation.
1518 This is intended to model C++'s ``memory_order_acquire``.
1520 In addition to the guarantees of ``monotonic``, if this operation
1521 writes a value which is subsequently read by an ``acquire``
1522 operation, it *synchronizes-with* that operation. (This isn't a
1523 complete description; see the C++0x definition of a release
1524 sequence.) This corresponds to the C++0x/C1x
1525 ``memory_order_release``.
1526 ``acq_rel`` (acquire+release)
1527 Acts as both an ``acquire`` and ``release`` operation on its
1528 address. This corresponds to the C++0x/C1x ``memory_order_acq_rel``.
1529 ``seq_cst`` (sequentially consistent)
1530 In addition to the guarantees of ``acq_rel`` (``acquire`` for an
1531 operation which only reads, ``release`` for an operation which only
1532 writes), there is a global total order on all
1533 sequentially-consistent operations on all addresses, which is
1534 consistent with the *happens-before* partial order and with the
1535 modification orders of all the affected addresses. Each
1536 sequentially-consistent read sees the last preceding write to the
1537 same address in this global order. This corresponds to the C++0x/C1x
1538 ``memory_order_seq_cst`` and Java volatile.
1542 If an atomic operation is marked ``singlethread``, it only *synchronizes
1543 with* or participates in modification and seq\_cst total orderings with
1544 other operations running in the same thread (for example, in signal
1552 LLVM IR floating-point binary ops (:ref:`fadd <i_fadd>`,
1553 :ref:`fsub <i_fsub>`, :ref:`fmul <i_fmul>`, :ref:`fdiv <i_fdiv>`,
1554 :ref:`frem <i_frem>`) have the following flags that can set to enable
1555 otherwise unsafe floating point operations
1558 No NaNs - Allow optimizations to assume the arguments and result are not
1559 NaN. Such optimizations are required to retain defined behavior over
1560 NaNs, but the value of the result is undefined.
1563 No Infs - Allow optimizations to assume the arguments and result are not
1564 +/-Inf. Such optimizations are required to retain defined behavior over
1565 +/-Inf, but the value of the result is undefined.
1568 No Signed Zeros - Allow optimizations to treat the sign of a zero
1569 argument or result as insignificant.
1572 Allow Reciprocal - Allow optimizations to use the reciprocal of an
1573 argument rather than perform division.
1576 Fast - Allow algebraically equivalent transformations that may
1577 dramatically change results in floating point (e.g. reassociate). This
1578 flag implies all the others.
1585 The LLVM type system is one of the most important features of the
1586 intermediate representation. Being typed enables a number of
1587 optimizations to be performed on the intermediate representation
1588 directly, without having to do extra analyses on the side before the
1589 transformation. A strong type system makes it easier to read the
1590 generated code and enables novel analyses and transformations that are
1591 not feasible to perform on normal three address code representations.
1601 The void type does not represent any value and has no size.
1619 The function type can be thought of as a function signature. It consists of a
1620 return type and a list of formal parameter types. The return type of a function
1621 type is a void type or first class type --- except for :ref:`label <t_label>`
1622 and :ref:`metadata <t_metadata>` types.
1628 <returntype> (<parameter list>)
1630 ...where '``<parameter list>``' is a comma-separated list of type
1631 specifiers. Optionally, the parameter list may include a type ``...``, which
1632 indicates that the function takes a variable number of arguments. Variable
1633 argument functions can access their arguments with the :ref:`variable argument
1634 handling intrinsic <int_varargs>` functions. '``<returntype>``' is any type
1635 except :ref:`label <t_label>` and :ref:`metadata <t_metadata>`.
1639 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1640 | ``i32 (i32)`` | function taking an ``i32``, returning an ``i32`` |
1641 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1642 | ``float (i16, i32 *) *`` | :ref:`Pointer <t_pointer>` to a function that takes an ``i16`` and a :ref:`pointer <t_pointer>` to ``i32``, returning ``float``. |
1643 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1644 | ``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. |
1645 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1646 | ``{i32, i32} (i32)`` | A function taking an ``i32``, returning a :ref:`structure <t_struct>` containing two ``i32`` values |
1647 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1654 The :ref:`first class <t_firstclass>` types are perhaps the most important.
1655 Values of these types are the only ones which can be produced by
1663 These are the types that are valid in registers from CodeGen's perspective.
1672 The integer type is a very simple type that simply specifies an
1673 arbitrary bit width for the integer type desired. Any bit width from 1
1674 bit to 2\ :sup:`23`\ -1 (about 8 million) can be specified.
1682 The number of bits the integer will occupy is specified by the ``N``
1688 +----------------+------------------------------------------------+
1689 | ``i1`` | a single-bit integer. |
1690 +----------------+------------------------------------------------+
1691 | ``i32`` | a 32-bit integer. |
1692 +----------------+------------------------------------------------+
1693 | ``i1942652`` | a really big integer of over 1 million bits. |
1694 +----------------+------------------------------------------------+
1698 Floating Point Types
1699 """"""""""""""""""""
1708 - 16-bit floating point value
1711 - 32-bit floating point value
1714 - 64-bit floating point value
1717 - 128-bit floating point value (112-bit mantissa)
1720 - 80-bit floating point value (X87)
1723 - 128-bit floating point value (two 64-bits)
1732 The x86mmx type represents a value held in an MMX register on an x86
1733 machine. The operations allowed on it are quite limited: parameters and
1734 return values, load and store, and bitcast. User-specified MMX
1735 instructions are represented as intrinsic or asm calls with arguments
1736 and/or results of this type. There are no arrays, vectors or constants
1753 The pointer type is used to specify memory locations. Pointers are
1754 commonly used to reference objects in memory.
1756 Pointer types may have an optional address space attribute defining the
1757 numbered address space where the pointed-to object resides. The default
1758 address space is number zero. The semantics of non-zero address spaces
1759 are target-specific.
1761 Note that LLVM does not permit pointers to void (``void*``) nor does it
1762 permit pointers to labels (``label*``). Use ``i8*`` instead.
1772 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1773 | ``[4 x i32]*`` | A :ref:`pointer <t_pointer>` to :ref:`array <t_array>` of four ``i32`` values. |
1774 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1775 | ``i32 (i32*) *`` | A :ref:`pointer <t_pointer>` to a :ref:`function <t_function>` that takes an ``i32*``, returning an ``i32``. |
1776 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1777 | ``i32 addrspace(5)*`` | A :ref:`pointer <t_pointer>` to an ``i32`` value that resides in address space #5. |
1778 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1787 A vector type is a simple derived type that represents a vector of
1788 elements. Vector types are used when multiple primitive data are
1789 operated in parallel using a single instruction (SIMD). A vector type
1790 requires a size (number of elements) and an underlying primitive data
1791 type. Vector types are considered :ref:`first class <t_firstclass>`.
1797 < <# elements> x <elementtype> >
1799 The number of elements is a constant integer value larger than 0;
1800 elementtype may be any integer or floating point type, or a pointer to
1801 these types. Vectors of size zero are not allowed.
1805 +-------------------+--------------------------------------------------+
1806 | ``<4 x i32>`` | Vector of 4 32-bit integer values. |
1807 +-------------------+--------------------------------------------------+
1808 | ``<8 x float>`` | Vector of 8 32-bit floating-point values. |
1809 +-------------------+--------------------------------------------------+
1810 | ``<2 x i64>`` | Vector of 2 64-bit integer values. |
1811 +-------------------+--------------------------------------------------+
1812 | ``<4 x i64*>`` | Vector of 4 pointers to 64-bit integer values. |
1813 +-------------------+--------------------------------------------------+
1822 The label type represents code labels.
1837 The metadata type represents embedded metadata. No derived types may be
1838 created from metadata except for :ref:`function <t_function>` arguments.
1851 Aggregate Types are a subset of derived types that can contain multiple
1852 member types. :ref:`Arrays <t_array>` and :ref:`structs <t_struct>` are
1853 aggregate types. :ref:`Vectors <t_vector>` are not considered to be
1863 The array type is a very simple derived type that arranges elements
1864 sequentially in memory. The array type requires a size (number of
1865 elements) and an underlying data type.
1871 [<# elements> x <elementtype>]
1873 The number of elements is a constant integer value; ``elementtype`` may
1874 be any type with a size.
1878 +------------------+--------------------------------------+
1879 | ``[40 x i32]`` | Array of 40 32-bit integer values. |
1880 +------------------+--------------------------------------+
1881 | ``[41 x i32]`` | Array of 41 32-bit integer values. |
1882 +------------------+--------------------------------------+
1883 | ``[4 x i8]`` | Array of 4 8-bit integer values. |
1884 +------------------+--------------------------------------+
1886 Here are some examples of multidimensional arrays:
1888 +-----------------------------+----------------------------------------------------------+
1889 | ``[3 x [4 x i32]]`` | 3x4 array of 32-bit integer values. |
1890 +-----------------------------+----------------------------------------------------------+
1891 | ``[12 x [10 x float]]`` | 12x10 array of single precision floating point values. |
1892 +-----------------------------+----------------------------------------------------------+
1893 | ``[2 x [3 x [4 x i16]]]`` | 2x3x4 array of 16-bit integer values. |
1894 +-----------------------------+----------------------------------------------------------+
1896 There is no restriction on indexing beyond the end of the array implied
1897 by a static type (though there are restrictions on indexing beyond the
1898 bounds of an allocated object in some cases). This means that
1899 single-dimension 'variable sized array' addressing can be implemented in
1900 LLVM with a zero length array type. An implementation of 'pascal style
1901 arrays' in LLVM could use the type "``{ i32, [0 x float]}``", for
1911 The structure type is used to represent a collection of data members
1912 together in memory. The elements of a structure may be any type that has
1915 Structures in memory are accessed using '``load``' and '``store``' by
1916 getting a pointer to a field with the '``getelementptr``' instruction.
1917 Structures in registers are accessed using the '``extractvalue``' and
1918 '``insertvalue``' instructions.
1920 Structures may optionally be "packed" structures, which indicate that
1921 the alignment of the struct is one byte, and that there is no padding
1922 between the elements. In non-packed structs, padding between field types
1923 is inserted as defined by the DataLayout string in the module, which is
1924 required to match what the underlying code generator expects.
1926 Structures can either be "literal" or "identified". A literal structure
1927 is defined inline with other types (e.g. ``{i32, i32}*``) whereas
1928 identified types are always defined at the top level with a name.
1929 Literal types are uniqued by their contents and can never be recursive
1930 or opaque since there is no way to write one. Identified types can be
1931 recursive, can be opaqued, and are never uniqued.
1937 %T1 = type { <type list> } ; Identified normal struct type
1938 %T2 = type <{ <type list> }> ; Identified packed struct type
1942 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1943 | ``{ i32, i32, i32 }`` | A triple of three ``i32`` values |
1944 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1945 | ``{ 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``. |
1946 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1947 | ``<{ i8, i32 }>`` | A packed struct known to be 5 bytes in size. |
1948 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1952 Opaque Structure Types
1953 """"""""""""""""""""""
1957 Opaque structure types are used to represent named structure types that
1958 do not have a body specified. This corresponds (for example) to the C
1959 notion of a forward declared structure.
1970 +--------------+-------------------+
1971 | ``opaque`` | An opaque type. |
1972 +--------------+-------------------+
1977 LLVM has several different basic types of constants. This section
1978 describes them all and their syntax.
1983 **Boolean constants**
1984 The two strings '``true``' and '``false``' are both valid constants
1986 **Integer constants**
1987 Standard integers (such as '4') are constants of the
1988 :ref:`integer <t_integer>` type. Negative numbers may be used with
1990 **Floating point constants**
1991 Floating point constants use standard decimal notation (e.g.
1992 123.421), exponential notation (e.g. 1.23421e+2), or a more precise
1993 hexadecimal notation (see below). The assembler requires the exact
1994 decimal value of a floating-point constant. For example, the
1995 assembler accepts 1.25 but rejects 1.3 because 1.3 is a repeating
1996 decimal in binary. Floating point constants must have a :ref:`floating
1997 point <t_floating>` type.
1998 **Null pointer constants**
1999 The identifier '``null``' is recognized as a null pointer constant
2000 and must be of :ref:`pointer type <t_pointer>`.
2002 The one non-intuitive notation for constants is the hexadecimal form of
2003 floating point constants. For example, the form
2004 '``double 0x432ff973cafa8000``' is equivalent to (but harder to read
2005 than) '``double 4.5e+15``'. The only time hexadecimal floating point
2006 constants are required (and the only time that they are generated by the
2007 disassembler) is when a floating point constant must be emitted but it
2008 cannot be represented as a decimal floating point number in a reasonable
2009 number of digits. For example, NaN's, infinities, and other special
2010 values are represented in their IEEE hexadecimal format so that assembly
2011 and disassembly do not cause any bits to change in the constants.
2013 When using the hexadecimal form, constants of types half, float, and
2014 double are represented using the 16-digit form shown above (which
2015 matches the IEEE754 representation for double); half and float values
2016 must, however, be exactly representable as IEEE 754 half and single
2017 precision, respectively. Hexadecimal format is always used for long
2018 double, and there are three forms of long double. The 80-bit format used
2019 by x86 is represented as ``0xK`` followed by 20 hexadecimal digits. The
2020 128-bit format used by PowerPC (two adjacent doubles) is represented by
2021 ``0xM`` followed by 32 hexadecimal digits. The IEEE 128-bit format is
2022 represented by ``0xL`` followed by 32 hexadecimal digits. Long doubles
2023 will only work if they match the long double format on your target.
2024 The IEEE 16-bit format (half precision) is represented by ``0xH``
2025 followed by 4 hexadecimal digits. All hexadecimal formats are big-endian
2026 (sign bit at the left).
2028 There are no constants of type x86mmx.
2030 .. _complexconstants:
2035 Complex constants are a (potentially recursive) combination of simple
2036 constants and smaller complex constants.
2038 **Structure constants**
2039 Structure constants are represented with notation similar to
2040 structure type definitions (a comma separated list of elements,
2041 surrounded by braces (``{}``)). For example:
2042 "``{ i32 4, float 17.0, i32* @G }``", where "``@G``" is declared as
2043 "``@G = external global i32``". Structure constants must have
2044 :ref:`structure type <t_struct>`, and the number and types of elements
2045 must match those specified by the type.
2047 Array constants are represented with notation similar to array type
2048 definitions (a comma separated list of elements, surrounded by
2049 square brackets (``[]``)). For example:
2050 "``[ i32 42, i32 11, i32 74 ]``". Array constants must have
2051 :ref:`array type <t_array>`, and the number and types of elements must
2052 match those specified by the type.
2053 **Vector constants**
2054 Vector constants are represented with notation similar to vector
2055 type definitions (a comma separated list of elements, surrounded by
2056 less-than/greater-than's (``<>``)). For example:
2057 "``< i32 42, i32 11, i32 74, i32 100 >``". Vector constants
2058 must have :ref:`vector type <t_vector>`, and the number and types of
2059 elements must match those specified by the type.
2060 **Zero initialization**
2061 The string '``zeroinitializer``' can be used to zero initialize a
2062 value to zero of *any* type, including scalar and
2063 :ref:`aggregate <t_aggregate>` types. This is often used to avoid
2064 having to print large zero initializers (e.g. for large arrays) and
2065 is always exactly equivalent to using explicit zero initializers.
2067 A metadata node is a structure-like constant with :ref:`metadata
2068 type <t_metadata>`. For example:
2069 "``metadata !{ i32 0, metadata !"test" }``". Unlike other
2070 constants that are meant to be interpreted as part of the
2071 instruction stream, metadata is a place to attach additional
2072 information such as debug info.
2074 Global Variable and Function Addresses
2075 --------------------------------------
2077 The addresses of :ref:`global variables <globalvars>` and
2078 :ref:`functions <functionstructure>` are always implicitly valid
2079 (link-time) constants. These constants are explicitly referenced when
2080 the :ref:`identifier for the global <identifiers>` is used and always have
2081 :ref:`pointer <t_pointer>` type. For example, the following is a legal LLVM
2084 .. code-block:: llvm
2088 @Z = global [2 x i32*] [ i32* @X, i32* @Y ]
2095 The string '``undef``' can be used anywhere a constant is expected, and
2096 indicates that the user of the value may receive an unspecified
2097 bit-pattern. Undefined values may be of any type (other than '``label``'
2098 or '``void``') and be used anywhere a constant is permitted.
2100 Undefined values are useful because they indicate to the compiler that
2101 the program is well defined no matter what value is used. This gives the
2102 compiler more freedom to optimize. Here are some examples of
2103 (potentially surprising) transformations that are valid (in pseudo IR):
2105 .. code-block:: llvm
2115 This is safe because all of the output bits are affected by the undef
2116 bits. Any output bit can have a zero or one depending on the input bits.
2118 .. code-block:: llvm
2129 These logical operations have bits that are not always affected by the
2130 input. For example, if ``%X`` has a zero bit, then the output of the
2131 '``and``' operation will always be a zero for that bit, no matter what
2132 the corresponding bit from the '``undef``' is. As such, it is unsafe to
2133 optimize or assume that the result of the '``and``' is '``undef``'.
2134 However, it is safe to assume that all bits of the '``undef``' could be
2135 0, and optimize the '``and``' to 0. Likewise, it is safe to assume that
2136 all the bits of the '``undef``' operand to the '``or``' could be set,
2137 allowing the '``or``' to be folded to -1.
2139 .. code-block:: llvm
2141 %A = select undef, %X, %Y
2142 %B = select undef, 42, %Y
2143 %C = select %X, %Y, undef
2153 This set of examples shows that undefined '``select``' (and conditional
2154 branch) conditions can go *either way*, but they have to come from one
2155 of the two operands. In the ``%A`` example, if ``%X`` and ``%Y`` were
2156 both known to have a clear low bit, then ``%A`` would have to have a
2157 cleared low bit. However, in the ``%C`` example, the optimizer is
2158 allowed to assume that the '``undef``' operand could be the same as
2159 ``%Y``, allowing the whole '``select``' to be eliminated.
2161 .. code-block:: llvm
2163 %A = xor undef, undef
2180 This example points out that two '``undef``' operands are not
2181 necessarily the same. This can be surprising to people (and also matches
2182 C semantics) where they assume that "``X^X``" is always zero, even if
2183 ``X`` is undefined. This isn't true for a number of reasons, but the
2184 short answer is that an '``undef``' "variable" can arbitrarily change
2185 its value over its "live range". This is true because the variable
2186 doesn't actually *have a live range*. Instead, the value is logically
2187 read from arbitrary registers that happen to be around when needed, so
2188 the value is not necessarily consistent over time. In fact, ``%A`` and
2189 ``%C`` need to have the same semantics or the core LLVM "replace all
2190 uses with" concept would not hold.
2192 .. code-block:: llvm
2200 These examples show the crucial difference between an *undefined value*
2201 and *undefined behavior*. An undefined value (like '``undef``') is
2202 allowed to have an arbitrary bit-pattern. This means that the ``%A``
2203 operation can be constant folded to '``undef``', because the '``undef``'
2204 could be an SNaN, and ``fdiv`` is not (currently) defined on SNaN's.
2205 However, in the second example, we can make a more aggressive
2206 assumption: because the ``undef`` is allowed to be an arbitrary value,
2207 we are allowed to assume that it could be zero. Since a divide by zero
2208 has *undefined behavior*, we are allowed to assume that the operation
2209 does not execute at all. This allows us to delete the divide and all
2210 code after it. Because the undefined operation "can't happen", the
2211 optimizer can assume that it occurs in dead code.
2213 .. code-block:: llvm
2215 a: store undef -> %X
2216 b: store %X -> undef
2221 These examples reiterate the ``fdiv`` example: a store *of* an undefined
2222 value can be assumed to not have any effect; we can assume that the
2223 value is overwritten with bits that happen to match what was already
2224 there. However, a store *to* an undefined location could clobber
2225 arbitrary memory, therefore, it has undefined behavior.
2232 Poison values are similar to :ref:`undef values <undefvalues>`, however
2233 they also represent the fact that an instruction or constant expression
2234 which cannot evoke side effects has nevertheless detected a condition
2235 which results in undefined behavior.
2237 There is currently no way of representing a poison value in the IR; they
2238 only exist when produced by operations such as :ref:`add <i_add>` with
2241 Poison value behavior is defined in terms of value *dependence*:
2243 - Values other than :ref:`phi <i_phi>` nodes depend on their operands.
2244 - :ref:`Phi <i_phi>` nodes depend on the operand corresponding to
2245 their dynamic predecessor basic block.
2246 - Function arguments depend on the corresponding actual argument values
2247 in the dynamic callers of their functions.
2248 - :ref:`Call <i_call>` instructions depend on the :ref:`ret <i_ret>`
2249 instructions that dynamically transfer control back to them.
2250 - :ref:`Invoke <i_invoke>` instructions depend on the
2251 :ref:`ret <i_ret>`, :ref:`resume <i_resume>`, or exception-throwing
2252 call instructions that dynamically transfer control back to them.
2253 - Non-volatile loads and stores depend on the most recent stores to all
2254 of the referenced memory addresses, following the order in the IR
2255 (including loads and stores implied by intrinsics such as
2256 :ref:`@llvm.memcpy <int_memcpy>`.)
2257 - An instruction with externally visible side effects depends on the
2258 most recent preceding instruction with externally visible side
2259 effects, following the order in the IR. (This includes :ref:`volatile
2260 operations <volatile>`.)
2261 - An instruction *control-depends* on a :ref:`terminator
2262 instruction <terminators>` if the terminator instruction has
2263 multiple successors and the instruction is always executed when
2264 control transfers to one of the successors, and may not be executed
2265 when control is transferred to another.
2266 - Additionally, an instruction also *control-depends* on a terminator
2267 instruction if the set of instructions it otherwise depends on would
2268 be different if the terminator had transferred control to a different
2270 - Dependence is transitive.
2272 Poison Values have the same behavior as :ref:`undef values <undefvalues>`,
2273 with the additional affect that any instruction which has a *dependence*
2274 on a poison value has undefined behavior.
2276 Here are some examples:
2278 .. code-block:: llvm
2281 %poison = sub nuw i32 0, 1 ; Results in a poison value.
2282 %still_poison = and i32 %poison, 0 ; 0, but also poison.
2283 %poison_yet_again = getelementptr i32* @h, i32 %still_poison
2284 store i32 0, i32* %poison_yet_again ; memory at @h[0] is poisoned
2286 store i32 %poison, i32* @g ; Poison value stored to memory.
2287 %poison2 = load i32* @g ; Poison value loaded back from memory.
2289 store volatile i32 %poison, i32* @g ; External observation; undefined behavior.
2291 %narrowaddr = bitcast i32* @g to i16*
2292 %wideaddr = bitcast i32* @g to i64*
2293 %poison3 = load i16* %narrowaddr ; Returns a poison value.
2294 %poison4 = load i64* %wideaddr ; Returns a poison value.
2296 %cmp = icmp slt i32 %poison, 0 ; Returns a poison value.
2297 br i1 %cmp, label %true, label %end ; Branch to either destination.
2300 store volatile i32 0, i32* @g ; This is control-dependent on %cmp, so
2301 ; it has undefined behavior.
2305 %p = phi i32 [ 0, %entry ], [ 1, %true ]
2306 ; Both edges into this PHI are
2307 ; control-dependent on %cmp, so this
2308 ; always results in a poison value.
2310 store volatile i32 0, i32* @g ; This would depend on the store in %true
2311 ; if %cmp is true, or the store in %entry
2312 ; otherwise, so this is undefined behavior.
2314 br i1 %cmp, label %second_true, label %second_end
2315 ; The same branch again, but this time the
2316 ; true block doesn't have side effects.
2323 store volatile i32 0, i32* @g ; This time, the instruction always depends
2324 ; on the store in %end. Also, it is
2325 ; control-equivalent to %end, so this is
2326 ; well-defined (ignoring earlier undefined
2327 ; behavior in this example).
2331 Addresses of Basic Blocks
2332 -------------------------
2334 ``blockaddress(@function, %block)``
2336 The '``blockaddress``' constant computes the address of the specified
2337 basic block in the specified function, and always has an ``i8*`` type.
2338 Taking the address of the entry block is illegal.
2340 This value only has defined behavior when used as an operand to the
2341 ':ref:`indirectbr <i_indirectbr>`' instruction, or for comparisons
2342 against null. Pointer equality tests between labels addresses results in
2343 undefined behavior --- though, again, comparison against null is ok, and
2344 no label is equal to the null pointer. This may be passed around as an
2345 opaque pointer sized value as long as the bits are not inspected. This
2346 allows ``ptrtoint`` and arithmetic to be performed on these values so
2347 long as the original value is reconstituted before the ``indirectbr``
2350 Finally, some targets may provide defined semantics when using the value
2351 as the operand to an inline assembly, but that is target specific.
2355 Constant Expressions
2356 --------------------
2358 Constant expressions are used to allow expressions involving other
2359 constants to be used as constants. Constant expressions may be of any
2360 :ref:`first class <t_firstclass>` type and may involve any LLVM operation
2361 that does not have side effects (e.g. load and call are not supported).
2362 The following is the syntax for constant expressions:
2364 ``trunc (CST to TYPE)``
2365 Truncate a constant to another type. The bit size of CST must be
2366 larger than the bit size of TYPE. Both types must be integers.
2367 ``zext (CST to TYPE)``
2368 Zero extend a constant to another type. The bit size of CST must be
2369 smaller than the bit size of TYPE. Both types must be integers.
2370 ``sext (CST to TYPE)``
2371 Sign extend a constant to another type. The bit size of CST must be
2372 smaller than the bit size of TYPE. Both types must be integers.
2373 ``fptrunc (CST to TYPE)``
2374 Truncate a floating point constant to another floating point type.
2375 The size of CST must be larger than the size of TYPE. Both types
2376 must be floating point.
2377 ``fpext (CST to TYPE)``
2378 Floating point extend a constant to another type. The size of CST
2379 must be smaller or equal to the size of TYPE. Both types must be
2381 ``fptoui (CST to TYPE)``
2382 Convert a floating point constant to the corresponding unsigned
2383 integer constant. TYPE must be a scalar or vector integer type. CST
2384 must be of scalar or vector floating point type. Both CST and TYPE
2385 must be scalars, or vectors of the same number of elements. If the
2386 value won't fit in the integer type, the results are undefined.
2387 ``fptosi (CST to TYPE)``
2388 Convert a floating point constant to the corresponding signed
2389 integer constant. TYPE must be a scalar or vector integer type. CST
2390 must be of scalar or vector floating point type. Both CST and TYPE
2391 must be scalars, or vectors of the same number of elements. If the
2392 value won't fit in the integer type, the results are undefined.
2393 ``uitofp (CST to TYPE)``
2394 Convert an unsigned integer constant to the corresponding floating
2395 point constant. TYPE must be a scalar or vector floating point type.
2396 CST must be of scalar or vector integer type. Both CST and TYPE must
2397 be scalars, or vectors of the same number of elements. If the value
2398 won't fit in the floating point type, the results are undefined.
2399 ``sitofp (CST to TYPE)``
2400 Convert a signed integer constant to the corresponding floating
2401 point constant. TYPE must be a scalar or vector floating point type.
2402 CST must be of scalar or vector integer type. Both CST and TYPE must
2403 be scalars, or vectors of the same number of elements. If the value
2404 won't fit in the floating point type, the results are undefined.
2405 ``ptrtoint (CST to TYPE)``
2406 Convert a pointer typed constant to the corresponding integer
2407 constant. ``TYPE`` must be an integer type. ``CST`` must be of
2408 pointer type. The ``CST`` value is zero extended, truncated, or
2409 unchanged to make it fit in ``TYPE``.
2410 ``inttoptr (CST to TYPE)``
2411 Convert an integer constant to a pointer constant. TYPE must be a
2412 pointer type. CST must be of integer type. The CST value is zero
2413 extended, truncated, or unchanged to make it fit in a pointer size.
2414 This one is *really* dangerous!
2415 ``bitcast (CST to TYPE)``
2416 Convert a constant, CST, to another TYPE. The constraints of the
2417 operands are the same as those for the :ref:`bitcast
2418 instruction <i_bitcast>`.
2419 ``addrspacecast (CST to TYPE)``
2420 Convert a constant pointer or constant vector of pointer, CST, to another
2421 TYPE in a different address space. The constraints of the operands are the
2422 same as those for the :ref:`addrspacecast instruction <i_addrspacecast>`.
2423 ``getelementptr (CSTPTR, IDX0, IDX1, ...)``, ``getelementptr inbounds (CSTPTR, IDX0, IDX1, ...)``
2424 Perform the :ref:`getelementptr operation <i_getelementptr>` on
2425 constants. As with the :ref:`getelementptr <i_getelementptr>`
2426 instruction, the index list may have zero or more indexes, which are
2427 required to make sense for the type of "CSTPTR".
2428 ``select (COND, VAL1, VAL2)``
2429 Perform the :ref:`select operation <i_select>` on constants.
2430 ``icmp COND (VAL1, VAL2)``
2431 Performs the :ref:`icmp operation <i_icmp>` on constants.
2432 ``fcmp COND (VAL1, VAL2)``
2433 Performs the :ref:`fcmp operation <i_fcmp>` on constants.
2434 ``extractelement (VAL, IDX)``
2435 Perform the :ref:`extractelement operation <i_extractelement>` on
2437 ``insertelement (VAL, ELT, IDX)``
2438 Perform the :ref:`insertelement operation <i_insertelement>` on
2440 ``shufflevector (VEC1, VEC2, IDXMASK)``
2441 Perform the :ref:`shufflevector operation <i_shufflevector>` on
2443 ``extractvalue (VAL, IDX0, IDX1, ...)``
2444 Perform the :ref:`extractvalue operation <i_extractvalue>` on
2445 constants. The index list is interpreted in a similar manner as
2446 indices in a ':ref:`getelementptr <i_getelementptr>`' operation. At
2447 least one index value must be specified.
2448 ``insertvalue (VAL, ELT, IDX0, IDX1, ...)``
2449 Perform the :ref:`insertvalue operation <i_insertvalue>` on constants.
2450 The index list is interpreted in a similar manner as indices in a
2451 ':ref:`getelementptr <i_getelementptr>`' operation. At least one index
2452 value must be specified.
2453 ``OPCODE (LHS, RHS)``
2454 Perform the specified operation of the LHS and RHS constants. OPCODE
2455 may be any of the :ref:`binary <binaryops>` or :ref:`bitwise
2456 binary <bitwiseops>` operations. The constraints on operands are
2457 the same as those for the corresponding instruction (e.g. no bitwise
2458 operations on floating point values are allowed).
2465 Inline Assembler Expressions
2466 ----------------------------
2468 LLVM supports inline assembler expressions (as opposed to :ref:`Module-Level
2469 Inline Assembly <moduleasm>`) through the use of a special value. This
2470 value represents the inline assembler as a string (containing the
2471 instructions to emit), a list of operand constraints (stored as a
2472 string), a flag that indicates whether or not the inline asm expression
2473 has side effects, and a flag indicating whether the function containing
2474 the asm needs to align its stack conservatively. An example inline
2475 assembler expression is:
2477 .. code-block:: llvm
2479 i32 (i32) asm "bswap $0", "=r,r"
2481 Inline assembler expressions may **only** be used as the callee operand
2482 of a :ref:`call <i_call>` or an :ref:`invoke <i_invoke>` instruction.
2483 Thus, typically we have:
2485 .. code-block:: llvm
2487 %X = call i32 asm "bswap $0", "=r,r"(i32 %Y)
2489 Inline asms with side effects not visible in the constraint list must be
2490 marked as having side effects. This is done through the use of the
2491 '``sideeffect``' keyword, like so:
2493 .. code-block:: llvm
2495 call void asm sideeffect "eieio", ""()
2497 In some cases inline asms will contain code that will not work unless
2498 the stack is aligned in some way, such as calls or SSE instructions on
2499 x86, yet will not contain code that does that alignment within the asm.
2500 The compiler should make conservative assumptions about what the asm
2501 might contain and should generate its usual stack alignment code in the
2502 prologue if the '``alignstack``' keyword is present:
2504 .. code-block:: llvm
2506 call void asm alignstack "eieio", ""()
2508 Inline asms also support using non-standard assembly dialects. The
2509 assumed dialect is ATT. When the '``inteldialect``' keyword is present,
2510 the inline asm is using the Intel dialect. Currently, ATT and Intel are
2511 the only supported dialects. An example is:
2513 .. code-block:: llvm
2515 call void asm inteldialect "eieio", ""()
2517 If multiple keywords appear the '``sideeffect``' keyword must come
2518 first, the '``alignstack``' keyword second and the '``inteldialect``'
2524 The call instructions that wrap inline asm nodes may have a
2525 "``!srcloc``" MDNode attached to it that contains a list of constant
2526 integers. If present, the code generator will use the integer as the
2527 location cookie value when report errors through the ``LLVMContext``
2528 error reporting mechanisms. This allows a front-end to correlate backend
2529 errors that occur with inline asm back to the source code that produced
2532 .. code-block:: llvm
2534 call void asm sideeffect "something bad", ""(), !srcloc !42
2536 !42 = !{ i32 1234567 }
2538 It is up to the front-end to make sense of the magic numbers it places
2539 in the IR. If the MDNode contains multiple constants, the code generator
2540 will use the one that corresponds to the line of the asm that the error
2545 Metadata Nodes and Metadata Strings
2546 -----------------------------------
2548 LLVM IR allows metadata to be attached to instructions in the program
2549 that can convey extra information about the code to the optimizers and
2550 code generator. One example application of metadata is source-level
2551 debug information. There are two metadata primitives: strings and nodes.
2552 All metadata has the ``metadata`` type and is identified in syntax by a
2553 preceding exclamation point ('``!``').
2555 A metadata string is a string surrounded by double quotes. It can
2556 contain any character by escaping non-printable characters with
2557 "``\xx``" where "``xx``" is the two digit hex code. For example:
2560 Metadata nodes are represented with notation similar to structure
2561 constants (a comma separated list of elements, surrounded by braces and
2562 preceded by an exclamation point). Metadata nodes can have any values as
2563 their operand. For example:
2565 .. code-block:: llvm
2567 !{ metadata !"test\00", i32 10}
2569 A :ref:`named metadata <namedmetadatastructure>` is a collection of
2570 metadata nodes, which can be looked up in the module symbol table. For
2573 .. code-block:: llvm
2575 !foo = metadata !{!4, !3}
2577 Metadata can be used as function arguments. Here ``llvm.dbg.value``
2578 function is using two metadata arguments:
2580 .. code-block:: llvm
2582 call void @llvm.dbg.value(metadata !24, i64 0, metadata !25)
2584 Metadata can be attached with an instruction. Here metadata ``!21`` is
2585 attached to the ``add`` instruction using the ``!dbg`` identifier:
2587 .. code-block:: llvm
2589 %indvar.next = add i64 %indvar, 1, !dbg !21
2591 More information about specific metadata nodes recognized by the
2592 optimizers and code generator is found below.
2597 In LLVM IR, memory does not have types, so LLVM's own type system is not
2598 suitable for doing TBAA. Instead, metadata is added to the IR to
2599 describe a type system of a higher level language. This can be used to
2600 implement typical C/C++ TBAA, but it can also be used to implement
2601 custom alias analysis behavior for other languages.
2603 The current metadata format is very simple. TBAA metadata nodes have up
2604 to three fields, e.g.:
2606 .. code-block:: llvm
2608 !0 = metadata !{ metadata !"an example type tree" }
2609 !1 = metadata !{ metadata !"int", metadata !0 }
2610 !2 = metadata !{ metadata !"float", metadata !0 }
2611 !3 = metadata !{ metadata !"const float", metadata !2, i64 1 }
2613 The first field is an identity field. It can be any value, usually a
2614 metadata string, which uniquely identifies the type. The most important
2615 name in the tree is the name of the root node. Two trees with different
2616 root node names are entirely disjoint, even if they have leaves with
2619 The second field identifies the type's parent node in the tree, or is
2620 null or omitted for a root node. A type is considered to alias all of
2621 its descendants and all of its ancestors in the tree. Also, a type is
2622 considered to alias all types in other trees, so that bitcode produced
2623 from multiple front-ends is handled conservatively.
2625 If the third field is present, it's an integer which if equal to 1
2626 indicates that the type is "constant" (meaning
2627 ``pointsToConstantMemory`` should return true; see `other useful
2628 AliasAnalysis methods <AliasAnalysis.html#OtherItfs>`_).
2630 '``tbaa.struct``' Metadata
2631 ^^^^^^^^^^^^^^^^^^^^^^^^^^
2633 The :ref:`llvm.memcpy <int_memcpy>` is often used to implement
2634 aggregate assignment operations in C and similar languages, however it
2635 is defined to copy a contiguous region of memory, which is more than
2636 strictly necessary for aggregate types which contain holes due to
2637 padding. Also, it doesn't contain any TBAA information about the fields
2640 ``!tbaa.struct`` metadata can describe which memory subregions in a
2641 memcpy are padding and what the TBAA tags of the struct are.
2643 The current metadata format is very simple. ``!tbaa.struct`` metadata
2644 nodes are a list of operands which are in conceptual groups of three.
2645 For each group of three, the first operand gives the byte offset of a
2646 field in bytes, the second gives its size in bytes, and the third gives
2649 .. code-block:: llvm
2651 !4 = metadata !{ i64 0, i64 4, metadata !1, i64 8, i64 4, metadata !2 }
2653 This describes a struct with two fields. The first is at offset 0 bytes
2654 with size 4 bytes, and has tbaa tag !1. The second is at offset 8 bytes
2655 and has size 4 bytes and has tbaa tag !2.
2657 Note that the fields need not be contiguous. In this example, there is a
2658 4 byte gap between the two fields. This gap represents padding which
2659 does not carry useful data and need not be preserved.
2661 '``fpmath``' Metadata
2662 ^^^^^^^^^^^^^^^^^^^^^
2664 ``fpmath`` metadata may be attached to any instruction of floating point
2665 type. It can be used to express the maximum acceptable error in the
2666 result of that instruction, in ULPs, thus potentially allowing the
2667 compiler to use a more efficient but less accurate method of computing
2668 it. ULP is defined as follows:
2670 If ``x`` is a real number that lies between two finite consecutive
2671 floating-point numbers ``a`` and ``b``, without being equal to one
2672 of them, then ``ulp(x) = |b - a|``, otherwise ``ulp(x)`` is the
2673 distance between the two non-equal finite floating-point numbers
2674 nearest ``x``. Moreover, ``ulp(NaN)`` is ``NaN``.
2676 The metadata node shall consist of a single positive floating point
2677 number representing the maximum relative error, for example:
2679 .. code-block:: llvm
2681 !0 = metadata !{ float 2.5 } ; maximum acceptable inaccuracy is 2.5 ULPs
2683 '``range``' Metadata
2684 ^^^^^^^^^^^^^^^^^^^^
2686 ``range`` metadata may be attached only to loads of integer types. It
2687 expresses the possible ranges the loaded value is in. The ranges are
2688 represented with a flattened list of integers. The loaded value is known
2689 to be in the union of the ranges defined by each consecutive pair. Each
2690 pair has the following properties:
2692 - The type must match the type loaded by the instruction.
2693 - The pair ``a,b`` represents the range ``[a,b)``.
2694 - Both ``a`` and ``b`` are constants.
2695 - The range is allowed to wrap.
2696 - The range should not represent the full or empty set. That is,
2699 In addition, the pairs must be in signed order of the lower bound and
2700 they must be non-contiguous.
2704 .. code-block:: llvm
2706 %a = load i8* %x, align 1, !range !0 ; Can only be 0 or 1
2707 %b = load i8* %y, align 1, !range !1 ; Can only be 255 (-1), 0 or 1
2708 %c = load i8* %z, align 1, !range !2 ; Can only be 0, 1, 3, 4 or 5
2709 %d = load i8* %z, align 1, !range !3 ; Can only be -2, -1, 3, 4 or 5
2711 !0 = metadata !{ i8 0, i8 2 }
2712 !1 = metadata !{ i8 255, i8 2 }
2713 !2 = metadata !{ i8 0, i8 2, i8 3, i8 6 }
2714 !3 = metadata !{ i8 -2, i8 0, i8 3, i8 6 }
2719 It is sometimes useful to attach information to loop constructs. Currently,
2720 loop metadata is implemented as metadata attached to the branch instruction
2721 in the loop latch block. This type of metadata refer to a metadata node that is
2722 guaranteed to be separate for each loop. The loop identifier metadata is
2723 specified with the name ``llvm.loop``.
2725 The loop identifier metadata is implemented using a metadata that refers to
2726 itself to avoid merging it with any other identifier metadata, e.g.,
2727 during module linkage or function inlining. That is, each loop should refer
2728 to their own identification metadata even if they reside in separate functions.
2729 The following example contains loop identifier metadata for two separate loop
2732 .. code-block:: llvm
2734 !0 = metadata !{ metadata !0 }
2735 !1 = metadata !{ metadata !1 }
2737 The loop identifier metadata can be used to specify additional per-loop
2738 metadata. Any operands after the first operand can be treated as user-defined
2739 metadata. For example the ``llvm.vectorizer.unroll`` metadata is understood
2740 by the loop vectorizer to indicate how many times to unroll the loop:
2742 .. code-block:: llvm
2744 br i1 %exitcond, label %._crit_edge, label %.lr.ph, !llvm.loop !0
2746 !0 = metadata !{ metadata !0, metadata !1 }
2747 !1 = metadata !{ metadata !"llvm.vectorizer.unroll", i32 2 }
2752 Metadata types used to annotate memory accesses with information helpful
2753 for optimizations are prefixed with ``llvm.mem``.
2755 '``llvm.mem.parallel_loop_access``' Metadata
2756 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2758 For a loop to be parallel, in addition to using
2759 the ``llvm.loop`` metadata to mark the loop latch branch instruction,
2760 also all of the memory accessing instructions in the loop body need to be
2761 marked with the ``llvm.mem.parallel_loop_access`` metadata. If there
2762 is at least one memory accessing instruction not marked with the metadata,
2763 the loop must be considered a sequential loop. This causes parallel loops to be
2764 converted to sequential loops due to optimization passes that are unaware of
2765 the parallel semantics and that insert new memory instructions to the loop
2768 Example of a loop that is considered parallel due to its correct use of
2769 both ``llvm.loop`` and ``llvm.mem.parallel_loop_access``
2770 metadata types that refer to the same loop identifier metadata.
2772 .. code-block:: llvm
2776 %0 = load i32* %arrayidx, align 4, !llvm.mem.parallel_loop_access !0
2778 store i32 %0, i32* %arrayidx4, align 4, !llvm.mem.parallel_loop_access !0
2780 br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !0
2784 !0 = metadata !{ metadata !0 }
2786 It is also possible to have nested parallel loops. In that case the
2787 memory accesses refer to a list of loop identifier metadata nodes instead of
2788 the loop identifier metadata node directly:
2790 .. code-block:: llvm
2797 %0 = load i32* %arrayidx, align 4, !llvm.mem.parallel_loop_access !0
2799 store i32 %0, i32* %arrayidx4, align 4, !llvm.mem.parallel_loop_access !0
2801 br i1 %exitcond, label %inner.for.end, label %inner.for.body, !llvm.loop !1
2805 %0 = load i32* %arrayidx, align 4, !llvm.mem.parallel_loop_access !0
2807 store i32 %0, i32* %arrayidx4, align 4, !llvm.mem.parallel_loop_access !0
2809 br i1 %exitcond, label %outer.for.end, label %outer.for.body, !llvm.loop !2
2811 outer.for.end: ; preds = %for.body
2813 !0 = metadata !{ metadata !1, metadata !2 } ; a list of loop identifiers
2814 !1 = metadata !{ metadata !1 } ; an identifier for the inner loop
2815 !2 = metadata !{ metadata !2 } ; an identifier for the outer loop
2817 '``llvm.vectorizer``'
2818 ^^^^^^^^^^^^^^^^^^^^^
2820 Metadata prefixed with ``llvm.vectorizer`` is used to control per-loop
2821 vectorization parameters such as vectorization factor and unroll factor.
2823 ``llvm.vectorizer`` metadata should be used in conjunction with ``llvm.loop``
2824 loop identification metadata.
2826 '``llvm.vectorizer.unroll``' Metadata
2827 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2829 This metadata instructs the loop vectorizer to unroll the specified
2830 loop exactly ``N`` times.
2832 The first operand is the string ``llvm.vectorizer.unroll`` and the second
2833 operand is an integer specifying the unroll factor. For example:
2835 .. code-block:: llvm
2837 !0 = metadata !{ metadata !"llvm.vectorizer.unroll", i32 4 }
2839 Note that setting ``llvm.vectorizer.unroll`` to 1 disables unrolling of the
2842 If ``llvm.vectorizer.unroll`` is set to 0 then the amount of unrolling will be
2843 determined automatically.
2845 '``llvm.vectorizer.width``' Metadata
2846 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2848 This metadata sets the target width of the vectorizer to ``N``. Without
2849 this metadata, the vectorizer will choose a width automatically.
2850 Regardless of this metadata, the vectorizer will only vectorize loops if
2851 it believes it is valid to do so.
2853 The first operand is the string ``llvm.vectorizer.width`` and the second
2854 operand is an integer specifying the width. For example:
2856 .. code-block:: llvm
2858 !0 = metadata !{ metadata !"llvm.vectorizer.width", i32 4 }
2860 Note that setting ``llvm.vectorizer.width`` to 1 disables vectorization of the
2863 If ``llvm.vectorizer.width`` is set to 0 then the width will be determined
2866 Module Flags Metadata
2867 =====================
2869 Information about the module as a whole is difficult to convey to LLVM's
2870 subsystems. The LLVM IR isn't sufficient to transmit this information.
2871 The ``llvm.module.flags`` named metadata exists in order to facilitate
2872 this. These flags are in the form of key / value pairs --- much like a
2873 dictionary --- making it easy for any subsystem who cares about a flag to
2876 The ``llvm.module.flags`` metadata contains a list of metadata triplets.
2877 Each triplet has the following form:
2879 - The first element is a *behavior* flag, which specifies the behavior
2880 when two (or more) modules are merged together, and it encounters two
2881 (or more) metadata with the same ID. The supported behaviors are
2883 - The second element is a metadata string that is a unique ID for the
2884 metadata. Each module may only have one flag entry for each unique ID (not
2885 including entries with the **Require** behavior).
2886 - The third element is the value of the flag.
2888 When two (or more) modules are merged together, the resulting
2889 ``llvm.module.flags`` metadata is the union of the modules' flags. That is, for
2890 each unique metadata ID string, there will be exactly one entry in the merged
2891 modules ``llvm.module.flags`` metadata table, and the value for that entry will
2892 be determined by the merge behavior flag, as described below. The only exception
2893 is that entries with the *Require* behavior are always preserved.
2895 The following behaviors are supported:
2906 Emits an error if two values disagree, otherwise the resulting value
2907 is that of the operands.
2911 Emits a warning if two values disagree. The result value will be the
2912 operand for the flag from the first module being linked.
2916 Adds a requirement that another module flag be present and have a
2917 specified value after linking is performed. The value must be a
2918 metadata pair, where the first element of the pair is the ID of the
2919 module flag to be restricted, and the second element of the pair is
2920 the value the module flag should be restricted to. This behavior can
2921 be used to restrict the allowable results (via triggering of an
2922 error) of linking IDs with the **Override** behavior.
2926 Uses the specified value, regardless of the behavior or value of the
2927 other module. If both modules specify **Override**, but the values
2928 differ, an error will be emitted.
2932 Appends the two values, which are required to be metadata nodes.
2936 Appends the two values, which are required to be metadata
2937 nodes. However, duplicate entries in the second list are dropped
2938 during the append operation.
2940 It is an error for a particular unique flag ID to have multiple behaviors,
2941 except in the case of **Require** (which adds restrictions on another metadata
2942 value) or **Override**.
2944 An example of module flags:
2946 .. code-block:: llvm
2948 !0 = metadata !{ i32 1, metadata !"foo", i32 1 }
2949 !1 = metadata !{ i32 4, metadata !"bar", i32 37 }
2950 !2 = metadata !{ i32 2, metadata !"qux", i32 42 }
2951 !3 = metadata !{ i32 3, metadata !"qux",
2953 metadata !"foo", i32 1
2956 !llvm.module.flags = !{ !0, !1, !2, !3 }
2958 - Metadata ``!0`` has the ID ``!"foo"`` and the value '1'. The behavior
2959 if two or more ``!"foo"`` flags are seen is to emit an error if their
2960 values are not equal.
2962 - Metadata ``!1`` has the ID ``!"bar"`` and the value '37'. The
2963 behavior if two or more ``!"bar"`` flags are seen is to use the value
2966 - Metadata ``!2`` has the ID ``!"qux"`` and the value '42'. The
2967 behavior if two or more ``!"qux"`` flags are seen is to emit a
2968 warning if their values are not equal.
2970 - Metadata ``!3`` has the ID ``!"qux"`` and the value:
2974 metadata !{ metadata !"foo", i32 1 }
2976 The behavior is to emit an error if the ``llvm.module.flags`` does not
2977 contain a flag with the ID ``!"foo"`` that has the value '1' after linking is
2980 Objective-C Garbage Collection Module Flags Metadata
2981 ----------------------------------------------------
2983 On the Mach-O platform, Objective-C stores metadata about garbage
2984 collection in a special section called "image info". The metadata
2985 consists of a version number and a bitmask specifying what types of
2986 garbage collection are supported (if any) by the file. If two or more
2987 modules are linked together their garbage collection metadata needs to
2988 be merged rather than appended together.
2990 The Objective-C garbage collection module flags metadata consists of the
2991 following key-value pairs:
3000 * - ``Objective-C Version``
3001 - **[Required]** --- The Objective-C ABI version. Valid values are 1 and 2.
3003 * - ``Objective-C Image Info Version``
3004 - **[Required]** --- The version of the image info section. Currently
3007 * - ``Objective-C Image Info Section``
3008 - **[Required]** --- The section to place the metadata. Valid values are
3009 ``"__OBJC, __image_info, regular"`` for Objective-C ABI version 1, and
3010 ``"__DATA,__objc_imageinfo, regular, no_dead_strip"`` for
3011 Objective-C ABI version 2.
3013 * - ``Objective-C Garbage Collection``
3014 - **[Required]** --- Specifies whether garbage collection is supported or
3015 not. Valid values are 0, for no garbage collection, and 2, for garbage
3016 collection supported.
3018 * - ``Objective-C GC Only``
3019 - **[Optional]** --- Specifies that only garbage collection is supported.
3020 If present, its value must be 6. This flag requires that the
3021 ``Objective-C Garbage Collection`` flag have the value 2.
3023 Some important flag interactions:
3025 - If a module with ``Objective-C Garbage Collection`` set to 0 is
3026 merged with a module with ``Objective-C Garbage Collection`` set to
3027 2, then the resulting module has the
3028 ``Objective-C Garbage Collection`` flag set to 0.
3029 - A module with ``Objective-C Garbage Collection`` set to 0 cannot be
3030 merged with a module with ``Objective-C GC Only`` set to 6.
3032 Automatic Linker Flags Module Flags Metadata
3033 --------------------------------------------
3035 Some targets support embedding flags to the linker inside individual object
3036 files. Typically this is used in conjunction with language extensions which
3037 allow source files to explicitly declare the libraries they depend on, and have
3038 these automatically be transmitted to the linker via object files.
3040 These flags are encoded in the IR using metadata in the module flags section,
3041 using the ``Linker Options`` key. The merge behavior for this flag is required
3042 to be ``AppendUnique``, and the value for the key is expected to be a metadata
3043 node which should be a list of other metadata nodes, each of which should be a
3044 list of metadata strings defining linker options.
3046 For example, the following metadata section specifies two separate sets of
3047 linker options, presumably to link against ``libz`` and the ``Cocoa``
3050 !0 = metadata !{ i32 6, metadata !"Linker Options",
3052 metadata !{ metadata !"-lz" },
3053 metadata !{ metadata !"-framework", metadata !"Cocoa" } } }
3054 !llvm.module.flags = !{ !0 }
3056 The metadata encoding as lists of lists of options, as opposed to a collapsed
3057 list of options, is chosen so that the IR encoding can use multiple option
3058 strings to specify e.g., a single library, while still having that specifier be
3059 preserved as an atomic element that can be recognized by a target specific
3060 assembly writer or object file emitter.
3062 Each individual option is required to be either a valid option for the target's
3063 linker, or an option that is reserved by the target specific assembly writer or
3064 object file emitter. No other aspect of these options is defined by the IR.
3066 .. _intrinsicglobalvariables:
3068 Intrinsic Global Variables
3069 ==========================
3071 LLVM has a number of "magic" global variables that contain data that
3072 affect code generation or other IR semantics. These are documented here.
3073 All globals of this sort should have a section specified as
3074 "``llvm.metadata``". This section and all globals that start with
3075 "``llvm.``" are reserved for use by LLVM.
3079 The '``llvm.used``' Global Variable
3080 -----------------------------------
3082 The ``@llvm.used`` global is an array which has
3083 :ref:`appending linkage <linkage_appending>`. This array contains a list of
3084 pointers to named global variables, functions and aliases which may optionally
3085 have a pointer cast formed of bitcast or getelementptr. For example, a legal
3088 .. code-block:: llvm
3093 @llvm.used = appending global [2 x i8*] [
3095 i8* bitcast (i32* @Y to i8*)
3096 ], section "llvm.metadata"
3098 If a symbol appears in the ``@llvm.used`` list, then the compiler, assembler,
3099 and linker are required to treat the symbol as if there is a reference to the
3100 symbol that it cannot see (which is why they have to be named). For example, if
3101 a variable has internal linkage and no references other than that from the
3102 ``@llvm.used`` list, it cannot be deleted. This is commonly used to represent
3103 references from inline asms and other things the compiler cannot "see", and
3104 corresponds to "``attribute((used))``" in GNU C.
3106 On some targets, the code generator must emit a directive to the
3107 assembler or object file to prevent the assembler and linker from
3108 molesting the symbol.
3110 .. _gv_llvmcompilerused:
3112 The '``llvm.compiler.used``' Global Variable
3113 --------------------------------------------
3115 The ``@llvm.compiler.used`` directive is the same as the ``@llvm.used``
3116 directive, except that it only prevents the compiler from touching the
3117 symbol. On targets that support it, this allows an intelligent linker to
3118 optimize references to the symbol without being impeded as it would be
3121 This is a rare construct that should only be used in rare circumstances,
3122 and should not be exposed to source languages.
3124 .. _gv_llvmglobalctors:
3126 The '``llvm.global_ctors``' Global Variable
3127 -------------------------------------------
3129 .. code-block:: llvm
3131 %0 = type { i32, void ()* }
3132 @llvm.global_ctors = appending global [1 x %0] [%0 { i32 65535, void ()* @ctor }]
3134 The ``@llvm.global_ctors`` array contains a list of constructor
3135 functions and associated priorities. The functions referenced by this
3136 array will be called in ascending order of priority (i.e. lowest first)
3137 when the module is loaded. The order of functions with the same priority
3140 .. _llvmglobaldtors:
3142 The '``llvm.global_dtors``' Global Variable
3143 -------------------------------------------
3145 .. code-block:: llvm
3147 %0 = type { i32, void ()* }
3148 @llvm.global_dtors = appending global [1 x %0] [%0 { i32 65535, void ()* @dtor }]
3150 The ``@llvm.global_dtors`` array contains a list of destructor functions
3151 and associated priorities. The functions referenced by this array will
3152 be called in descending order of priority (i.e. highest first) when the
3153 module is loaded. The order of functions with the same priority is not
3156 Instruction Reference
3157 =====================
3159 The LLVM instruction set consists of several different classifications
3160 of instructions: :ref:`terminator instructions <terminators>`, :ref:`binary
3161 instructions <binaryops>`, :ref:`bitwise binary
3162 instructions <bitwiseops>`, :ref:`memory instructions <memoryops>`, and
3163 :ref:`other instructions <otherops>`.
3167 Terminator Instructions
3168 -----------------------
3170 As mentioned :ref:`previously <functionstructure>`, every basic block in a
3171 program ends with a "Terminator" instruction, which indicates which
3172 block should be executed after the current block is finished. These
3173 terminator instructions typically yield a '``void``' value: they produce
3174 control flow, not values (the one exception being the
3175 ':ref:`invoke <i_invoke>`' instruction).
3177 The terminator instructions are: ':ref:`ret <i_ret>`',
3178 ':ref:`br <i_br>`', ':ref:`switch <i_switch>`',
3179 ':ref:`indirectbr <i_indirectbr>`', ':ref:`invoke <i_invoke>`',
3180 ':ref:`resume <i_resume>`', and ':ref:`unreachable <i_unreachable>`'.
3184 '``ret``' Instruction
3185 ^^^^^^^^^^^^^^^^^^^^^
3192 ret <type> <value> ; Return a value from a non-void function
3193 ret void ; Return from void function
3198 The '``ret``' instruction is used to return control flow (and optionally
3199 a value) from a function back to the caller.
3201 There are two forms of the '``ret``' instruction: one that returns a
3202 value and then causes control flow, and one that just causes control
3208 The '``ret``' instruction optionally accepts a single argument, the
3209 return value. The type of the return value must be a ':ref:`first
3210 class <t_firstclass>`' type.
3212 A function is not :ref:`well formed <wellformed>` if it it has a non-void
3213 return type and contains a '``ret``' instruction with no return value or
3214 a return value with a type that does not match its type, or if it has a
3215 void return type and contains a '``ret``' instruction with a return
3221 When the '``ret``' instruction is executed, control flow returns back to
3222 the calling function's context. If the caller is a
3223 ":ref:`call <i_call>`" instruction, execution continues at the
3224 instruction after the call. If the caller was an
3225 ":ref:`invoke <i_invoke>`" instruction, execution continues at the
3226 beginning of the "normal" destination block. If the instruction returns
3227 a value, that value shall set the call or invoke instruction's return
3233 .. code-block:: llvm
3235 ret i32 5 ; Return an integer value of 5
3236 ret void ; Return from a void function
3237 ret { i32, i8 } { i32 4, i8 2 } ; Return a struct of values 4 and 2
3241 '``br``' Instruction
3242 ^^^^^^^^^^^^^^^^^^^^
3249 br i1 <cond>, label <iftrue>, label <iffalse>
3250 br label <dest> ; Unconditional branch
3255 The '``br``' instruction is used to cause control flow to transfer to a
3256 different basic block in the current function. There are two forms of
3257 this instruction, corresponding to a conditional branch and an
3258 unconditional branch.
3263 The conditional branch form of the '``br``' instruction takes a single
3264 '``i1``' value and two '``label``' values. The unconditional form of the
3265 '``br``' instruction takes a single '``label``' value as a target.
3270 Upon execution of a conditional '``br``' instruction, the '``i1``'
3271 argument is evaluated. If the value is ``true``, control flows to the
3272 '``iftrue``' ``label`` argument. If "cond" is ``false``, control flows
3273 to the '``iffalse``' ``label`` argument.
3278 .. code-block:: llvm
3281 %cond = icmp eq i32 %a, %b
3282 br i1 %cond, label %IfEqual, label %IfUnequal
3290 '``switch``' Instruction
3291 ^^^^^^^^^^^^^^^^^^^^^^^^
3298 switch <intty> <value>, label <defaultdest> [ <intty> <val>, label <dest> ... ]
3303 The '``switch``' instruction is used to transfer control flow to one of
3304 several different places. It is a generalization of the '``br``'
3305 instruction, allowing a branch to occur to one of many possible
3311 The '``switch``' instruction uses three parameters: an integer
3312 comparison value '``value``', a default '``label``' destination, and an
3313 array of pairs of comparison value constants and '``label``'s. The table
3314 is not allowed to contain duplicate constant entries.
3319 The ``switch`` instruction specifies a table of values and destinations.
3320 When the '``switch``' instruction is executed, this table is searched
3321 for the given value. If the value is found, control flow is transferred
3322 to the corresponding destination; otherwise, control flow is transferred
3323 to the default destination.
3328 Depending on properties of the target machine and the particular
3329 ``switch`` instruction, this instruction may be code generated in
3330 different ways. For example, it could be generated as a series of
3331 chained conditional branches or with a lookup table.
3336 .. code-block:: llvm
3338 ; Emulate a conditional br instruction
3339 %Val = zext i1 %value to i32
3340 switch i32 %Val, label %truedest [ i32 0, label %falsedest ]
3342 ; Emulate an unconditional br instruction
3343 switch i32 0, label %dest [ ]
3345 ; Implement a jump table:
3346 switch i32 %val, label %otherwise [ i32 0, label %onzero
3348 i32 2, label %ontwo ]
3352 '``indirectbr``' Instruction
3353 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3360 indirectbr <somety>* <address>, [ label <dest1>, label <dest2>, ... ]
3365 The '``indirectbr``' instruction implements an indirect branch to a
3366 label within the current function, whose address is specified by
3367 "``address``". Address must be derived from a
3368 :ref:`blockaddress <blockaddress>` constant.
3373 The '``address``' argument is the address of the label to jump to. The
3374 rest of the arguments indicate the full set of possible destinations
3375 that the address may point to. Blocks are allowed to occur multiple
3376 times in the destination list, though this isn't particularly useful.
3378 This destination list is required so that dataflow analysis has an
3379 accurate understanding of the CFG.
3384 Control transfers to the block specified in the address argument. All
3385 possible destination blocks must be listed in the label list, otherwise
3386 this instruction has undefined behavior. This implies that jumps to
3387 labels defined in other functions have undefined behavior as well.
3392 This is typically implemented with a jump through a register.
3397 .. code-block:: llvm
3399 indirectbr i8* %Addr, [ label %bb1, label %bb2, label %bb3 ]
3403 '``invoke``' Instruction
3404 ^^^^^^^^^^^^^^^^^^^^^^^^
3411 <result> = invoke [cconv] [ret attrs] <ptr to function ty> <function ptr val>(<function args>) [fn attrs]
3412 to label <normal label> unwind label <exception label>
3417 The '``invoke``' instruction causes control to transfer to a specified
3418 function, with the possibility of control flow transfer to either the
3419 '``normal``' label or the '``exception``' label. If the callee function
3420 returns with the "``ret``" instruction, control flow will return to the
3421 "normal" label. If the callee (or any indirect callees) returns via the
3422 ":ref:`resume <i_resume>`" instruction or other exception handling
3423 mechanism, control is interrupted and continued at the dynamically
3424 nearest "exception" label.
3426 The '``exception``' label is a `landing
3427 pad <ExceptionHandling.html#overview>`_ for the exception. As such,
3428 '``exception``' label is required to have the
3429 ":ref:`landingpad <i_landingpad>`" instruction, which contains the
3430 information about the behavior of the program after unwinding happens,
3431 as its first non-PHI instruction. The restrictions on the
3432 "``landingpad``" instruction's tightly couples it to the "``invoke``"
3433 instruction, so that the important information contained within the
3434 "``landingpad``" instruction can't be lost through normal code motion.
3439 This instruction requires several arguments:
3441 #. The optional "cconv" marker indicates which :ref:`calling
3442 convention <callingconv>` the call should use. If none is
3443 specified, the call defaults to using C calling conventions.
3444 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
3445 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
3447 #. '``ptr to function ty``': shall be the signature of the pointer to
3448 function value being invoked. In most cases, this is a direct
3449 function invocation, but indirect ``invoke``'s are just as possible,
3450 branching off an arbitrary pointer to function value.
3451 #. '``function ptr val``': An LLVM value containing a pointer to a
3452 function to be invoked.
3453 #. '``function args``': argument list whose types match the function
3454 signature argument types and parameter attributes. All arguments must
3455 be of :ref:`first class <t_firstclass>` type. If the function signature
3456 indicates the function accepts a variable number of arguments, the
3457 extra arguments can be specified.
3458 #. '``normal label``': the label reached when the called function
3459 executes a '``ret``' instruction.
3460 #. '``exception label``': the label reached when a callee returns via
3461 the :ref:`resume <i_resume>` instruction or other exception handling
3463 #. The optional :ref:`function attributes <fnattrs>` list. Only
3464 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
3465 attributes are valid here.
3470 This instruction is designed to operate as a standard '``call``'
3471 instruction in most regards. The primary difference is that it
3472 establishes an association with a label, which is used by the runtime
3473 library to unwind the stack.
3475 This instruction is used in languages with destructors to ensure that
3476 proper cleanup is performed in the case of either a ``longjmp`` or a
3477 thrown exception. Additionally, this is important for implementation of
3478 '``catch``' clauses in high-level languages that support them.
3480 For the purposes of the SSA form, the definition of the value returned
3481 by the '``invoke``' instruction is deemed to occur on the edge from the
3482 current block to the "normal" label. If the callee unwinds then no
3483 return value is available.
3488 .. code-block:: llvm
3490 %retval = invoke i32 @Test(i32 15) to label %Continue
3491 unwind label %TestCleanup ; {i32}:retval set
3492 %retval = invoke coldcc i32 %Testfnptr(i32 15) to label %Continue
3493 unwind label %TestCleanup ; {i32}:retval set
3497 '``resume``' Instruction
3498 ^^^^^^^^^^^^^^^^^^^^^^^^
3505 resume <type> <value>
3510 The '``resume``' instruction is a terminator instruction that has no
3516 The '``resume``' instruction requires one argument, which must have the
3517 same type as the result of any '``landingpad``' instruction in the same
3523 The '``resume``' instruction resumes propagation of an existing
3524 (in-flight) exception whose unwinding was interrupted with a
3525 :ref:`landingpad <i_landingpad>` instruction.
3530 .. code-block:: llvm
3532 resume { i8*, i32 } %exn
3536 '``unreachable``' Instruction
3537 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3549 The '``unreachable``' instruction has no defined semantics. This
3550 instruction is used to inform the optimizer that a particular portion of
3551 the code is not reachable. This can be used to indicate that the code
3552 after a no-return function cannot be reached, and other facts.
3557 The '``unreachable``' instruction has no defined semantics.
3564 Binary operators are used to do most of the computation in a program.
3565 They require two operands of the same type, execute an operation on
3566 them, and produce a single value. The operands might represent multiple
3567 data, as is the case with the :ref:`vector <t_vector>` data type. The
3568 result value has the same type as its operands.
3570 There are several different binary operators:
3574 '``add``' Instruction
3575 ^^^^^^^^^^^^^^^^^^^^^
3582 <result> = add <ty> <op1>, <op2> ; yields {ty}:result
3583 <result> = add nuw <ty> <op1>, <op2> ; yields {ty}:result
3584 <result> = add nsw <ty> <op1>, <op2> ; yields {ty}:result
3585 <result> = add nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3590 The '``add``' instruction returns the sum of its two operands.
3595 The two arguments to the '``add``' instruction must be
3596 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3597 arguments must have identical types.
3602 The value produced is the integer sum of the two operands.
3604 If the sum has unsigned overflow, the result returned is the
3605 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
3608 Because LLVM integers use a two's complement representation, this
3609 instruction is appropriate for both signed and unsigned integers.
3611 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3612 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3613 result value of the ``add`` is a :ref:`poison value <poisonvalues>` if
3614 unsigned and/or signed overflow, respectively, occurs.
3619 .. code-block:: llvm
3621 <result> = add i32 4, %var ; yields {i32}:result = 4 + %var
3625 '``fadd``' Instruction
3626 ^^^^^^^^^^^^^^^^^^^^^^
3633 <result> = fadd [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3638 The '``fadd``' instruction returns the sum of its two operands.
3643 The two arguments to the '``fadd``' instruction must be :ref:`floating
3644 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3645 Both arguments must have identical types.
3650 The value produced is the floating point sum of the two operands. This
3651 instruction can also take any number of :ref:`fast-math flags <fastmath>`,
3652 which are optimization hints to enable otherwise unsafe floating point
3658 .. code-block:: llvm
3660 <result> = fadd float 4.0, %var ; yields {float}:result = 4.0 + %var
3662 '``sub``' Instruction
3663 ^^^^^^^^^^^^^^^^^^^^^
3670 <result> = sub <ty> <op1>, <op2> ; yields {ty}:result
3671 <result> = sub nuw <ty> <op1>, <op2> ; yields {ty}:result
3672 <result> = sub nsw <ty> <op1>, <op2> ; yields {ty}:result
3673 <result> = sub nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3678 The '``sub``' instruction returns the difference of its two operands.
3680 Note that the '``sub``' instruction is used to represent the '``neg``'
3681 instruction present in most other intermediate representations.
3686 The two arguments to the '``sub``' instruction must be
3687 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3688 arguments must have identical types.
3693 The value produced is the integer difference of the two operands.
3695 If the difference has unsigned overflow, the result returned is the
3696 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
3699 Because LLVM integers use a two's complement representation, this
3700 instruction is appropriate for both signed and unsigned integers.
3702 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3703 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3704 result value of the ``sub`` is a :ref:`poison value <poisonvalues>` if
3705 unsigned and/or signed overflow, respectively, occurs.
3710 .. code-block:: llvm
3712 <result> = sub i32 4, %var ; yields {i32}:result = 4 - %var
3713 <result> = sub i32 0, %val ; yields {i32}:result = -%var
3717 '``fsub``' Instruction
3718 ^^^^^^^^^^^^^^^^^^^^^^
3725 <result> = fsub [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3730 The '``fsub``' instruction returns the difference of its two operands.
3732 Note that the '``fsub``' instruction is used to represent the '``fneg``'
3733 instruction present in most other intermediate representations.
3738 The two arguments to the '``fsub``' instruction must be :ref:`floating
3739 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3740 Both arguments must have identical types.
3745 The value produced is the floating point difference of the two operands.
3746 This instruction can also take any number of :ref:`fast-math
3747 flags <fastmath>`, which are optimization hints to enable otherwise
3748 unsafe floating point optimizations:
3753 .. code-block:: llvm
3755 <result> = fsub float 4.0, %var ; yields {float}:result = 4.0 - %var
3756 <result> = fsub float -0.0, %val ; yields {float}:result = -%var
3758 '``mul``' Instruction
3759 ^^^^^^^^^^^^^^^^^^^^^
3766 <result> = mul <ty> <op1>, <op2> ; yields {ty}:result
3767 <result> = mul nuw <ty> <op1>, <op2> ; yields {ty}:result
3768 <result> = mul nsw <ty> <op1>, <op2> ; yields {ty}:result
3769 <result> = mul nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3774 The '``mul``' instruction returns the product of its two operands.
3779 The two arguments to the '``mul``' instruction must be
3780 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3781 arguments must have identical types.
3786 The value produced is the integer product of the two operands.
3788 If the result of the multiplication has unsigned overflow, the result
3789 returned is the mathematical result modulo 2\ :sup:`n`\ , where n is the
3790 bit width of the result.
3792 Because LLVM integers use a two's complement representation, and the
3793 result is the same width as the operands, this instruction returns the
3794 correct result for both signed and unsigned integers. If a full product
3795 (e.g. ``i32`` * ``i32`` -> ``i64``) is needed, the operands should be
3796 sign-extended or zero-extended as appropriate to the width of the full
3799 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3800 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3801 result value of the ``mul`` is a :ref:`poison value <poisonvalues>` if
3802 unsigned and/or signed overflow, respectively, occurs.
3807 .. code-block:: llvm
3809 <result> = mul i32 4, %var ; yields {i32}:result = 4 * %var
3813 '``fmul``' Instruction
3814 ^^^^^^^^^^^^^^^^^^^^^^
3821 <result> = fmul [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3826 The '``fmul``' instruction returns the product of its two operands.
3831 The two arguments to the '``fmul``' instruction must be :ref:`floating
3832 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3833 Both arguments must have identical types.
3838 The value produced is the floating point product of the two operands.
3839 This instruction can also take any number of :ref:`fast-math
3840 flags <fastmath>`, which are optimization hints to enable otherwise
3841 unsafe floating point optimizations:
3846 .. code-block:: llvm
3848 <result> = fmul float 4.0, %var ; yields {float}:result = 4.0 * %var
3850 '``udiv``' Instruction
3851 ^^^^^^^^^^^^^^^^^^^^^^
3858 <result> = udiv <ty> <op1>, <op2> ; yields {ty}:result
3859 <result> = udiv exact <ty> <op1>, <op2> ; yields {ty}:result
3864 The '``udiv``' instruction returns the quotient of its two operands.
3869 The two arguments to the '``udiv``' instruction must be
3870 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3871 arguments must have identical types.
3876 The value produced is the unsigned integer quotient of the two operands.
3878 Note that unsigned integer division and signed integer division are
3879 distinct operations; for signed integer division, use '``sdiv``'.
3881 Division by zero leads to undefined behavior.
3883 If the ``exact`` keyword is present, the result value of the ``udiv`` is
3884 a :ref:`poison value <poisonvalues>` if %op1 is not a multiple of %op2 (as
3885 such, "((a udiv exact b) mul b) == a").
3890 .. code-block:: llvm
3892 <result> = udiv i32 4, %var ; yields {i32}:result = 4 / %var
3894 '``sdiv``' Instruction
3895 ^^^^^^^^^^^^^^^^^^^^^^
3902 <result> = sdiv <ty> <op1>, <op2> ; yields {ty}:result
3903 <result> = sdiv exact <ty> <op1>, <op2> ; yields {ty}:result
3908 The '``sdiv``' instruction returns the quotient of its two operands.
3913 The two arguments to the '``sdiv``' instruction must be
3914 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3915 arguments must have identical types.
3920 The value produced is the signed integer quotient of the two operands
3921 rounded towards zero.
3923 Note that signed integer division and unsigned integer division are
3924 distinct operations; for unsigned integer division, use '``udiv``'.
3926 Division by zero leads to undefined behavior. Overflow also leads to
3927 undefined behavior; this is a rare case, but can occur, for example, by
3928 doing a 32-bit division of -2147483648 by -1.
3930 If the ``exact`` keyword is present, the result value of the ``sdiv`` is
3931 a :ref:`poison value <poisonvalues>` if the result would be rounded.
3936 .. code-block:: llvm
3938 <result> = sdiv i32 4, %var ; yields {i32}:result = 4 / %var
3942 '``fdiv``' Instruction
3943 ^^^^^^^^^^^^^^^^^^^^^^
3950 <result> = fdiv [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3955 The '``fdiv``' instruction returns the quotient of its two operands.
3960 The two arguments to the '``fdiv``' instruction must be :ref:`floating
3961 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3962 Both arguments must have identical types.
3967 The value produced is the floating point quotient of the two operands.
3968 This instruction can also take any number of :ref:`fast-math
3969 flags <fastmath>`, which are optimization hints to enable otherwise
3970 unsafe floating point optimizations:
3975 .. code-block:: llvm
3977 <result> = fdiv float 4.0, %var ; yields {float}:result = 4.0 / %var
3979 '``urem``' Instruction
3980 ^^^^^^^^^^^^^^^^^^^^^^
3987 <result> = urem <ty> <op1>, <op2> ; yields {ty}:result
3992 The '``urem``' instruction returns the remainder from the unsigned
3993 division of its two arguments.
3998 The two arguments to the '``urem``' instruction must be
3999 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4000 arguments must have identical types.
4005 This instruction returns the unsigned integer *remainder* of a division.
4006 This instruction always performs an unsigned division to get the
4009 Note that unsigned integer remainder and signed integer remainder are
4010 distinct operations; for signed integer remainder, use '``srem``'.
4012 Taking the remainder of a division by zero leads to undefined behavior.
4017 .. code-block:: llvm
4019 <result> = urem i32 4, %var ; yields {i32}:result = 4 % %var
4021 '``srem``' Instruction
4022 ^^^^^^^^^^^^^^^^^^^^^^
4029 <result> = srem <ty> <op1>, <op2> ; yields {ty}:result
4034 The '``srem``' instruction returns the remainder from the signed
4035 division of its two operands. This instruction can also take
4036 :ref:`vector <t_vector>` versions of the values in which case the elements
4042 The two arguments to the '``srem``' instruction must be
4043 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4044 arguments must have identical types.
4049 This instruction returns the *remainder* of a division (where the result
4050 is either zero or has the same sign as the dividend, ``op1``), not the
4051 *modulo* operator (where the result is either zero or has the same sign
4052 as the divisor, ``op2``) of a value. For more information about the
4053 difference, see `The Math
4054 Forum <http://mathforum.org/dr.math/problems/anne.4.28.99.html>`_. For a
4055 table of how this is implemented in various languages, please see
4057 operation <http://en.wikipedia.org/wiki/Modulo_operation>`_.
4059 Note that signed integer remainder and unsigned integer remainder are
4060 distinct operations; for unsigned integer remainder, use '``urem``'.
4062 Taking the remainder of a division by zero leads to undefined behavior.
4063 Overflow also leads to undefined behavior; this is a rare case, but can
4064 occur, for example, by taking the remainder of a 32-bit division of
4065 -2147483648 by -1. (The remainder doesn't actually overflow, but this
4066 rule lets srem be implemented using instructions that return both the
4067 result of the division and the remainder.)
4072 .. code-block:: llvm
4074 <result> = srem i32 4, %var ; yields {i32}:result = 4 % %var
4078 '``frem``' Instruction
4079 ^^^^^^^^^^^^^^^^^^^^^^
4086 <result> = frem [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
4091 The '``frem``' instruction returns the remainder from the division of
4097 The two arguments to the '``frem``' instruction must be :ref:`floating
4098 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4099 Both arguments must have identical types.
4104 This instruction returns the *remainder* of a division. The remainder
4105 has the same sign as the dividend. This instruction can also take any
4106 number of :ref:`fast-math flags <fastmath>`, which are optimization hints
4107 to enable otherwise unsafe floating point optimizations:
4112 .. code-block:: llvm
4114 <result> = frem float 4.0, %var ; yields {float}:result = 4.0 % %var
4118 Bitwise Binary Operations
4119 -------------------------
4121 Bitwise binary operators are used to do various forms of bit-twiddling
4122 in a program. They are generally very efficient instructions and can
4123 commonly be strength reduced from other instructions. They require two
4124 operands of the same type, execute an operation on them, and produce a
4125 single value. The resulting value is the same type as its operands.
4127 '``shl``' Instruction
4128 ^^^^^^^^^^^^^^^^^^^^^
4135 <result> = shl <ty> <op1>, <op2> ; yields {ty}:result
4136 <result> = shl nuw <ty> <op1>, <op2> ; yields {ty}:result
4137 <result> = shl nsw <ty> <op1>, <op2> ; yields {ty}:result
4138 <result> = shl nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
4143 The '``shl``' instruction returns the first operand shifted to the left
4144 a specified number of bits.
4149 Both arguments to the '``shl``' instruction must be the same
4150 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4151 '``op2``' is treated as an unsigned value.
4156 The value produced is ``op1`` \* 2\ :sup:`op2` mod 2\ :sup:`n`,
4157 where ``n`` is the width of the result. If ``op2`` is (statically or
4158 dynamically) negative or equal to or larger than the number of bits in
4159 ``op1``, the result is undefined. If the arguments are vectors, each
4160 vector element of ``op1`` is shifted by the corresponding shift amount
4163 If the ``nuw`` keyword is present, then the shift produces a :ref:`poison
4164 value <poisonvalues>` if it shifts out any non-zero bits. If the
4165 ``nsw`` keyword is present, then the shift produces a :ref:`poison
4166 value <poisonvalues>` if it shifts out any bits that disagree with the
4167 resultant sign bit. As such, NUW/NSW have the same semantics as they
4168 would if the shift were expressed as a mul instruction with the same
4169 nsw/nuw bits in (mul %op1, (shl 1, %op2)).
4174 .. code-block:: llvm
4176 <result> = shl i32 4, %var ; yields {i32}: 4 << %var
4177 <result> = shl i32 4, 2 ; yields {i32}: 16
4178 <result> = shl i32 1, 10 ; yields {i32}: 1024
4179 <result> = shl i32 1, 32 ; undefined
4180 <result> = shl <2 x i32> < i32 1, i32 1>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 2, i32 4>
4182 '``lshr``' Instruction
4183 ^^^^^^^^^^^^^^^^^^^^^^
4190 <result> = lshr <ty> <op1>, <op2> ; yields {ty}:result
4191 <result> = lshr exact <ty> <op1>, <op2> ; yields {ty}:result
4196 The '``lshr``' instruction (logical shift right) returns the first
4197 operand shifted to the right a specified number of bits with zero fill.
4202 Both arguments to the '``lshr``' instruction must be the same
4203 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4204 '``op2``' is treated as an unsigned value.
4209 This instruction always performs a logical shift right operation. The
4210 most significant bits of the result will be filled with zero bits after
4211 the shift. If ``op2`` is (statically or dynamically) equal to or larger
4212 than the number of bits in ``op1``, the result is undefined. If the
4213 arguments are vectors, each vector element of ``op1`` is shifted by the
4214 corresponding shift amount in ``op2``.
4216 If the ``exact`` keyword is present, the result value of the ``lshr`` is
4217 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
4223 .. code-block:: llvm
4225 <result> = lshr i32 4, 1 ; yields {i32}:result = 2
4226 <result> = lshr i32 4, 2 ; yields {i32}:result = 1
4227 <result> = lshr i8 4, 3 ; yields {i8}:result = 0
4228 <result> = lshr i8 -2, 1 ; yields {i8}:result = 0x7F
4229 <result> = lshr i32 1, 32 ; undefined
4230 <result> = lshr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 0x7FFFFFFF, i32 1>
4232 '``ashr``' Instruction
4233 ^^^^^^^^^^^^^^^^^^^^^^
4240 <result> = ashr <ty> <op1>, <op2> ; yields {ty}:result
4241 <result> = ashr exact <ty> <op1>, <op2> ; yields {ty}:result
4246 The '``ashr``' instruction (arithmetic shift right) returns the first
4247 operand shifted to the right a specified number of bits with sign
4253 Both arguments to the '``ashr``' instruction must be the same
4254 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4255 '``op2``' is treated as an unsigned value.
4260 This instruction always performs an arithmetic shift right operation,
4261 The most significant bits of the result will be filled with the sign bit
4262 of ``op1``. If ``op2`` is (statically or dynamically) equal to or larger
4263 than the number of bits in ``op1``, the result is undefined. If the
4264 arguments are vectors, each vector element of ``op1`` is shifted by the
4265 corresponding shift amount in ``op2``.
4267 If the ``exact`` keyword is present, the result value of the ``ashr`` is
4268 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
4274 .. code-block:: llvm
4276 <result> = ashr i32 4, 1 ; yields {i32}:result = 2
4277 <result> = ashr i32 4, 2 ; yields {i32}:result = 1
4278 <result> = ashr i8 4, 3 ; yields {i8}:result = 0
4279 <result> = ashr i8 -2, 1 ; yields {i8}:result = -1
4280 <result> = ashr i32 1, 32 ; undefined
4281 <result> = ashr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 3> ; yields: result=<2 x i32> < i32 -1, i32 0>
4283 '``and``' Instruction
4284 ^^^^^^^^^^^^^^^^^^^^^
4291 <result> = and <ty> <op1>, <op2> ; yields {ty}:result
4296 The '``and``' instruction returns the bitwise logical and of its two
4302 The two arguments to the '``and``' instruction must be
4303 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4304 arguments must have identical types.
4309 The truth table used for the '``and``' instruction is:
4326 .. code-block:: llvm
4328 <result> = and i32 4, %var ; yields {i32}:result = 4 & %var
4329 <result> = and i32 15, 40 ; yields {i32}:result = 8
4330 <result> = and i32 4, 8 ; yields {i32}:result = 0
4332 '``or``' Instruction
4333 ^^^^^^^^^^^^^^^^^^^^
4340 <result> = or <ty> <op1>, <op2> ; yields {ty}:result
4345 The '``or``' instruction returns the bitwise logical inclusive or of its
4351 The two arguments to the '``or``' instruction must be
4352 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4353 arguments must have identical types.
4358 The truth table used for the '``or``' instruction is:
4377 <result> = or i32 4, %var ; yields {i32}:result = 4 | %var
4378 <result> = or i32 15, 40 ; yields {i32}:result = 47
4379 <result> = or i32 4, 8 ; yields {i32}:result = 12
4381 '``xor``' Instruction
4382 ^^^^^^^^^^^^^^^^^^^^^
4389 <result> = xor <ty> <op1>, <op2> ; yields {ty}:result
4394 The '``xor``' instruction returns the bitwise logical exclusive or of
4395 its two operands. The ``xor`` is used to implement the "one's
4396 complement" operation, which is the "~" operator in C.
4401 The two arguments to the '``xor``' instruction must be
4402 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4403 arguments must have identical types.
4408 The truth table used for the '``xor``' instruction is:
4425 .. code-block:: llvm
4427 <result> = xor i32 4, %var ; yields {i32}:result = 4 ^ %var
4428 <result> = xor i32 15, 40 ; yields {i32}:result = 39
4429 <result> = xor i32 4, 8 ; yields {i32}:result = 12
4430 <result> = xor i32 %V, -1 ; yields {i32}:result = ~%V
4435 LLVM supports several instructions to represent vector operations in a
4436 target-independent manner. These instructions cover the element-access
4437 and vector-specific operations needed to process vectors effectively.
4438 While LLVM does directly support these vector operations, many
4439 sophisticated algorithms will want to use target-specific intrinsics to
4440 take full advantage of a specific target.
4442 .. _i_extractelement:
4444 '``extractelement``' Instruction
4445 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4452 <result> = extractelement <n x <ty>> <val>, i32 <idx> ; yields <ty>
4457 The '``extractelement``' instruction extracts a single scalar element
4458 from a vector at a specified index.
4463 The first operand of an '``extractelement``' instruction is a value of
4464 :ref:`vector <t_vector>` type. The second operand is an index indicating
4465 the position from which to extract the element. The index may be a
4471 The result is a scalar of the same type as the element type of ``val``.
4472 Its value is the value at position ``idx`` of ``val``. If ``idx``
4473 exceeds the length of ``val``, the results are undefined.
4478 .. code-block:: llvm
4480 <result> = extractelement <4 x i32> %vec, i32 0 ; yields i32
4482 .. _i_insertelement:
4484 '``insertelement``' Instruction
4485 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4492 <result> = insertelement <n x <ty>> <val>, <ty> <elt>, i32 <idx> ; yields <n x <ty>>
4497 The '``insertelement``' instruction inserts a scalar element into a
4498 vector at a specified index.
4503 The first operand of an '``insertelement``' instruction is a value of
4504 :ref:`vector <t_vector>` type. The second operand is a scalar value whose
4505 type must equal the element type of the first operand. The third operand
4506 is an index indicating the position at which to insert the value. The
4507 index may be a variable.
4512 The result is a vector of the same type as ``val``. Its element values
4513 are those of ``val`` except at position ``idx``, where it gets the value
4514 ``elt``. If ``idx`` exceeds the length of ``val``, the results are
4520 .. code-block:: llvm
4522 <result> = insertelement <4 x i32> %vec, i32 1, i32 0 ; yields <4 x i32>
4524 .. _i_shufflevector:
4526 '``shufflevector``' Instruction
4527 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4534 <result> = shufflevector <n x <ty>> <v1>, <n x <ty>> <v2>, <m x i32> <mask> ; yields <m x <ty>>
4539 The '``shufflevector``' instruction constructs a permutation of elements
4540 from two input vectors, returning a vector with the same element type as
4541 the input and length that is the same as the shuffle mask.
4546 The first two operands of a '``shufflevector``' instruction are vectors
4547 with the same type. The third argument is a shuffle mask whose element
4548 type is always 'i32'. The result of the instruction is a vector whose
4549 length is the same as the shuffle mask and whose element type is the
4550 same as the element type of the first two operands.
4552 The shuffle mask operand is required to be a constant vector with either
4553 constant integer or undef values.
4558 The elements of the two input vectors are numbered from left to right
4559 across both of the vectors. The shuffle mask operand specifies, for each
4560 element of the result vector, which element of the two input vectors the
4561 result element gets. The element selector may be undef (meaning "don't
4562 care") and the second operand may be undef if performing a shuffle from
4568 .. code-block:: llvm
4570 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4571 <4 x i32> <i32 0, i32 4, i32 1, i32 5> ; yields <4 x i32>
4572 <result> = shufflevector <4 x i32> %v1, <4 x i32> undef,
4573 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32> - Identity shuffle.
4574 <result> = shufflevector <8 x i32> %v1, <8 x i32> undef,
4575 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32>
4576 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4577 <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7 > ; yields <8 x i32>
4579 Aggregate Operations
4580 --------------------
4582 LLVM supports several instructions for working with
4583 :ref:`aggregate <t_aggregate>` values.
4587 '``extractvalue``' Instruction
4588 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4595 <result> = extractvalue <aggregate type> <val>, <idx>{, <idx>}*
4600 The '``extractvalue``' instruction extracts the value of a member field
4601 from an :ref:`aggregate <t_aggregate>` value.
4606 The first operand of an '``extractvalue``' instruction is a value of
4607 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The operands are
4608 constant indices to specify which value to extract in a similar manner
4609 as indices in a '``getelementptr``' instruction.
4611 The major differences to ``getelementptr`` indexing are:
4613 - Since the value being indexed is not a pointer, the first index is
4614 omitted and assumed to be zero.
4615 - At least one index must be specified.
4616 - Not only struct indices but also array indices must be in bounds.
4621 The result is the value at the position in the aggregate specified by
4627 .. code-block:: llvm
4629 <result> = extractvalue {i32, float} %agg, 0 ; yields i32
4633 '``insertvalue``' Instruction
4634 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4641 <result> = insertvalue <aggregate type> <val>, <ty> <elt>, <idx>{, <idx>}* ; yields <aggregate type>
4646 The '``insertvalue``' instruction inserts a value into a member field in
4647 an :ref:`aggregate <t_aggregate>` value.
4652 The first operand of an '``insertvalue``' instruction is a value of
4653 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The second operand is
4654 a first-class value to insert. The following operands are constant
4655 indices indicating the position at which to insert the value in a
4656 similar manner as indices in a '``extractvalue``' instruction. The value
4657 to insert must have the same type as the value identified by the
4663 The result is an aggregate of the same type as ``val``. Its value is
4664 that of ``val`` except that the value at the position specified by the
4665 indices is that of ``elt``.
4670 .. code-block:: llvm
4672 %agg1 = insertvalue {i32, float} undef, i32 1, 0 ; yields {i32 1, float undef}
4673 %agg2 = insertvalue {i32, float} %agg1, float %val, 1 ; yields {i32 1, float %val}
4674 %agg3 = insertvalue {i32, {float}} %agg1, float %val, 1, 0 ; yields {i32 1, float %val}
4678 Memory Access and Addressing Operations
4679 ---------------------------------------
4681 A key design point of an SSA-based representation is how it represents
4682 memory. In LLVM, no memory locations are in SSA form, which makes things
4683 very simple. This section describes how to read, write, and allocate
4688 '``alloca``' Instruction
4689 ^^^^^^^^^^^^^^^^^^^^^^^^
4696 <result> = alloca <type>[, <ty> <NumElements>][, align <alignment>] ; yields {type*}:result
4701 The '``alloca``' instruction allocates memory on the stack frame of the
4702 currently executing function, to be automatically released when this
4703 function returns to its caller. The object is always allocated in the
4704 generic address space (address space zero).
4709 The '``alloca``' instruction allocates ``sizeof(<type>)*NumElements``
4710 bytes of memory on the runtime stack, returning a pointer of the
4711 appropriate type to the program. If "NumElements" is specified, it is
4712 the number of elements allocated, otherwise "NumElements" is defaulted
4713 to be one. If a constant alignment is specified, the value result of the
4714 allocation is guaranteed to be aligned to at least that boundary. If not
4715 specified, or if zero, the target can choose to align the allocation on
4716 any convenient boundary compatible with the type.
4718 '``type``' may be any sized type.
4723 Memory is allocated; a pointer is returned. The operation is undefined
4724 if there is insufficient stack space for the allocation. '``alloca``'d
4725 memory is automatically released when the function returns. The
4726 '``alloca``' instruction is commonly used to represent automatic
4727 variables that must have an address available. When the function returns
4728 (either with the ``ret`` or ``resume`` instructions), the memory is
4729 reclaimed. Allocating zero bytes is legal, but the result is undefined.
4730 The order in which memory is allocated (ie., which way the stack grows)
4736 .. code-block:: llvm
4738 %ptr = alloca i32 ; yields {i32*}:ptr
4739 %ptr = alloca i32, i32 4 ; yields {i32*}:ptr
4740 %ptr = alloca i32, i32 4, align 1024 ; yields {i32*}:ptr
4741 %ptr = alloca i32, align 1024 ; yields {i32*}:ptr
4745 '``load``' Instruction
4746 ^^^^^^^^^^^^^^^^^^^^^^
4753 <result> = load [volatile] <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>][, !invariant.load !<index>]
4754 <result> = load atomic [volatile] <ty>* <pointer> [singlethread] <ordering>, align <alignment>
4755 !<index> = !{ i32 1 }
4760 The '``load``' instruction is used to read from memory.
4765 The argument to the ``load`` instruction specifies the memory address
4766 from which to load. The pointer must point to a :ref:`first
4767 class <t_firstclass>` type. If the ``load`` is marked as ``volatile``,
4768 then the optimizer is not allowed to modify the number or order of
4769 execution of this ``load`` with other :ref:`volatile
4770 operations <volatile>`.
4772 If the ``load`` is marked as ``atomic``, it takes an extra
4773 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
4774 ``release`` and ``acq_rel`` orderings are not valid on ``load``
4775 instructions. Atomic loads produce :ref:`defined <memmodel>` results
4776 when they may see multiple atomic stores. The type of the pointee must
4777 be an integer type whose bit width is a power of two greater than or
4778 equal to eight and less than or equal to a target-specific size limit.
4779 ``align`` must be explicitly specified on atomic loads, and the load has
4780 undefined behavior if the alignment is not set to a value which is at
4781 least the size in bytes of the pointee. ``!nontemporal`` does not have
4782 any defined semantics for atomic loads.
4784 The optional constant ``align`` argument specifies the alignment of the
4785 operation (that is, the alignment of the memory address). A value of 0
4786 or an omitted ``align`` argument means that the operation has the ABI
4787 alignment for the target. It is the responsibility of the code emitter
4788 to ensure that the alignment information is correct. Overestimating the
4789 alignment results in undefined behavior. Underestimating the alignment
4790 may produce less efficient code. An alignment of 1 is always safe.
4792 The optional ``!nontemporal`` metadata must reference a single
4793 metadata name ``<index>`` corresponding to a metadata node with one
4794 ``i32`` entry of value 1. The existence of the ``!nontemporal``
4795 metadata on the instruction tells the optimizer and code generator
4796 that this load is not expected to be reused in the cache. The code
4797 generator may select special instructions to save cache bandwidth, such
4798 as the ``MOVNT`` instruction on x86.
4800 The optional ``!invariant.load`` metadata must reference a single
4801 metadata name ``<index>`` corresponding to a metadata node with no
4802 entries. The existence of the ``!invariant.load`` metadata on the
4803 instruction tells the optimizer and code generator that this load
4804 address points to memory which does not change value during program
4805 execution. The optimizer may then move this load around, for example, by
4806 hoisting it out of loops using loop invariant code motion.
4811 The location of memory pointed to is loaded. If the value being loaded
4812 is of scalar type then the number of bytes read does not exceed the
4813 minimum number of bytes needed to hold all bits of the type. For
4814 example, loading an ``i24`` reads at most three bytes. When loading a
4815 value of a type like ``i20`` with a size that is not an integral number
4816 of bytes, the result is undefined if the value was not originally
4817 written using a store of the same type.
4822 .. code-block:: llvm
4824 %ptr = alloca i32 ; yields {i32*}:ptr
4825 store i32 3, i32* %ptr ; yields {void}
4826 %val = load i32* %ptr ; yields {i32}:val = i32 3
4830 '``store``' Instruction
4831 ^^^^^^^^^^^^^^^^^^^^^^^
4838 store [volatile] <ty> <value>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>] ; yields {void}
4839 store atomic [volatile] <ty> <value>, <ty>* <pointer> [singlethread] <ordering>, align <alignment> ; yields {void}
4844 The '``store``' instruction is used to write to memory.
4849 There are two arguments to the ``store`` instruction: a value to store
4850 and an address at which to store it. The type of the ``<pointer>``
4851 operand must be a pointer to the :ref:`first class <t_firstclass>` type of
4852 the ``<value>`` operand. If the ``store`` is marked as ``volatile``,
4853 then the optimizer is not allowed to modify the number or order of
4854 execution of this ``store`` with other :ref:`volatile
4855 operations <volatile>`.
4857 If the ``store`` is marked as ``atomic``, it takes an extra
4858 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
4859 ``acquire`` and ``acq_rel`` orderings aren't valid on ``store``
4860 instructions. Atomic loads produce :ref:`defined <memmodel>` results
4861 when they may see multiple atomic stores. The type of the pointee must
4862 be an integer type whose bit width is a power of two greater than or
4863 equal to eight and less than or equal to a target-specific size limit.
4864 ``align`` must be explicitly specified on atomic stores, and the store
4865 has undefined behavior if the alignment is not set to a value which is
4866 at least the size in bytes of the pointee. ``!nontemporal`` does not
4867 have any defined semantics for atomic stores.
4869 The optional constant ``align`` argument specifies the alignment of the
4870 operation (that is, the alignment of the memory address). A value of 0
4871 or an omitted ``align`` argument means that the operation has the ABI
4872 alignment for the target. It is the responsibility of the code emitter
4873 to ensure that the alignment information is correct. Overestimating the
4874 alignment results in undefined behavior. Underestimating the
4875 alignment may produce less efficient code. An alignment of 1 is always
4878 The optional ``!nontemporal`` metadata must reference a single metadata
4879 name ``<index>`` corresponding to a metadata node with one ``i32`` entry of
4880 value 1. The existence of the ``!nontemporal`` metadata on the instruction
4881 tells the optimizer and code generator that this load is not expected to
4882 be reused in the cache. The code generator may select special
4883 instructions to save cache bandwidth, such as the MOVNT instruction on
4889 The contents of memory are updated to contain ``<value>`` at the
4890 location specified by the ``<pointer>`` operand. If ``<value>`` is
4891 of scalar type then the number of bytes written does not exceed the
4892 minimum number of bytes needed to hold all bits of the type. For
4893 example, storing an ``i24`` writes at most three bytes. When writing a
4894 value of a type like ``i20`` with a size that is not an integral number
4895 of bytes, it is unspecified what happens to the extra bits that do not
4896 belong to the type, but they will typically be overwritten.
4901 .. code-block:: llvm
4903 %ptr = alloca i32 ; yields {i32*}:ptr
4904 store i32 3, i32* %ptr ; yields {void}
4905 %val = load i32* %ptr ; yields {i32}:val = i32 3
4909 '``fence``' Instruction
4910 ^^^^^^^^^^^^^^^^^^^^^^^
4917 fence [singlethread] <ordering> ; yields {void}
4922 The '``fence``' instruction is used to introduce happens-before edges
4928 '``fence``' instructions take an :ref:`ordering <ordering>` argument which
4929 defines what *synchronizes-with* edges they add. They can only be given
4930 ``acquire``, ``release``, ``acq_rel``, and ``seq_cst`` orderings.
4935 A fence A which has (at least) ``release`` ordering semantics
4936 *synchronizes with* a fence B with (at least) ``acquire`` ordering
4937 semantics if and only if there exist atomic operations X and Y, both
4938 operating on some atomic object M, such that A is sequenced before X, X
4939 modifies M (either directly or through some side effect of a sequence
4940 headed by X), Y is sequenced before B, and Y observes M. This provides a
4941 *happens-before* dependency between A and B. Rather than an explicit
4942 ``fence``, one (but not both) of the atomic operations X or Y might
4943 provide a ``release`` or ``acquire`` (resp.) ordering constraint and
4944 still *synchronize-with* the explicit ``fence`` and establish the
4945 *happens-before* edge.
4947 A ``fence`` which has ``seq_cst`` ordering, in addition to having both
4948 ``acquire`` and ``release`` semantics specified above, participates in
4949 the global program order of other ``seq_cst`` operations and/or fences.
4951 The optional ":ref:`singlethread <singlethread>`" argument specifies
4952 that the fence only synchronizes with other fences in the same thread.
4953 (This is useful for interacting with signal handlers.)
4958 .. code-block:: llvm
4960 fence acquire ; yields {void}
4961 fence singlethread seq_cst ; yields {void}
4965 '``cmpxchg``' Instruction
4966 ^^^^^^^^^^^^^^^^^^^^^^^^^
4973 cmpxchg [volatile] <ty>* <pointer>, <ty> <cmp>, <ty> <new> [singlethread] <ordering> ; yields {ty}
4978 The '``cmpxchg``' instruction is used to atomically modify memory. It
4979 loads a value in memory and compares it to a given value. If they are
4980 equal, it stores a new value into the memory.
4985 There are three arguments to the '``cmpxchg``' instruction: an address
4986 to operate on, a value to compare to the value currently be at that
4987 address, and a new value to place at that address if the compared values
4988 are equal. The type of '<cmp>' must be an integer type whose bit width
4989 is a power of two greater than or equal to eight and less than or equal
4990 to a target-specific size limit. '<cmp>' and '<new>' must have the same
4991 type, and the type of '<pointer>' must be a pointer to that type. If the
4992 ``cmpxchg`` is marked as ``volatile``, then the optimizer is not allowed
4993 to modify the number or order of execution of this ``cmpxchg`` with
4994 other :ref:`volatile operations <volatile>`.
4996 The :ref:`ordering <ordering>` argument specifies how this ``cmpxchg``
4997 synchronizes with other atomic operations.
4999 The optional "``singlethread``" argument declares that the ``cmpxchg``
5000 is only atomic with respect to code (usually signal handlers) running in
5001 the same thread as the ``cmpxchg``. Otherwise the cmpxchg is atomic with
5002 respect to all other code in the system.
5004 The pointer passed into cmpxchg must have alignment greater than or
5005 equal to the size in memory of the operand.
5010 The contents of memory at the location specified by the '``<pointer>``'
5011 operand is read and compared to '``<cmp>``'; if the read value is the
5012 equal, '``<new>``' is written. The original value at the location is
5015 A successful ``cmpxchg`` is a read-modify-write instruction for the purpose
5016 of identifying release sequences. A failed ``cmpxchg`` is equivalent to an
5017 atomic load with an ordering parameter determined by dropping any
5018 ``release`` part of the ``cmpxchg``'s ordering.
5023 .. code-block:: llvm
5026 %orig = atomic load i32* %ptr unordered ; yields {i32}
5030 %cmp = phi i32 [ %orig, %entry ], [%old, %loop]
5031 %squared = mul i32 %cmp, %cmp
5032 %old = cmpxchg i32* %ptr, i32 %cmp, i32 %squared ; yields {i32}
5033 %success = icmp eq i32 %cmp, %old
5034 br i1 %success, label %done, label %loop
5041 '``atomicrmw``' Instruction
5042 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
5049 atomicrmw [volatile] <operation> <ty>* <pointer>, <ty> <value> [singlethread] <ordering> ; yields {ty}
5054 The '``atomicrmw``' instruction is used to atomically modify memory.
5059 There are three arguments to the '``atomicrmw``' instruction: an
5060 operation to apply, an address whose value to modify, an argument to the
5061 operation. The operation must be one of the following keywords:
5075 The type of '<value>' must be an integer type whose bit width is a power
5076 of two greater than or equal to eight and less than or equal to a
5077 target-specific size limit. The type of the '``<pointer>``' operand must
5078 be a pointer to that type. If the ``atomicrmw`` is marked as
5079 ``volatile``, then the optimizer is not allowed to modify the number or
5080 order of execution of this ``atomicrmw`` with other :ref:`volatile
5081 operations <volatile>`.
5086 The contents of memory at the location specified by the '``<pointer>``'
5087 operand are atomically read, modified, and written back. The original
5088 value at the location is returned. The modification is specified by the
5091 - xchg: ``*ptr = val``
5092 - add: ``*ptr = *ptr + val``
5093 - sub: ``*ptr = *ptr - val``
5094 - and: ``*ptr = *ptr & val``
5095 - nand: ``*ptr = ~(*ptr & val)``
5096 - or: ``*ptr = *ptr | val``
5097 - xor: ``*ptr = *ptr ^ val``
5098 - max: ``*ptr = *ptr > val ? *ptr : val`` (using a signed comparison)
5099 - min: ``*ptr = *ptr < val ? *ptr : val`` (using a signed comparison)
5100 - umax: ``*ptr = *ptr > val ? *ptr : val`` (using an unsigned
5102 - umin: ``*ptr = *ptr < val ? *ptr : val`` (using an unsigned
5108 .. code-block:: llvm
5110 %old = atomicrmw add i32* %ptr, i32 1 acquire ; yields {i32}
5112 .. _i_getelementptr:
5114 '``getelementptr``' Instruction
5115 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5122 <result> = getelementptr <pty>* <ptrval>{, <ty> <idx>}*
5123 <result> = getelementptr inbounds <pty>* <ptrval>{, <ty> <idx>}*
5124 <result> = getelementptr <ptr vector> ptrval, <vector index type> idx
5129 The '``getelementptr``' instruction is used to get the address of a
5130 subelement of an :ref:`aggregate <t_aggregate>` data structure. It performs
5131 address calculation only and does not access memory.
5136 The first argument is always a pointer or a vector of pointers, and
5137 forms the basis of the calculation. The remaining arguments are indices
5138 that indicate which of the elements of the aggregate object are indexed.
5139 The interpretation of each index is dependent on the type being indexed
5140 into. The first index always indexes the pointer value given as the
5141 first argument, the second index indexes a value of the type pointed to
5142 (not necessarily the value directly pointed to, since the first index
5143 can be non-zero), etc. The first type indexed into must be a pointer
5144 value, subsequent types can be arrays, vectors, and structs. Note that
5145 subsequent types being indexed into can never be pointers, since that
5146 would require loading the pointer before continuing calculation.
5148 The type of each index argument depends on the type it is indexing into.
5149 When indexing into a (optionally packed) structure, only ``i32`` integer
5150 **constants** are allowed (when using a vector of indices they must all
5151 be the **same** ``i32`` integer constant). When indexing into an array,
5152 pointer or vector, integers of any width are allowed, and they are not
5153 required to be constant. These integers are treated as signed values
5156 For example, let's consider a C code fragment and how it gets compiled
5172 int *foo(struct ST *s) {
5173 return &s[1].Z.B[5][13];
5176 The LLVM code generated by Clang is:
5178 .. code-block:: llvm
5180 %struct.RT = type { i8, [10 x [20 x i32]], i8 }
5181 %struct.ST = type { i32, double, %struct.RT }
5183 define i32* @foo(%struct.ST* %s) nounwind uwtable readnone optsize ssp {
5185 %arrayidx = getelementptr inbounds %struct.ST* %s, i64 1, i32 2, i32 1, i64 5, i64 13
5192 In the example above, the first index is indexing into the
5193 '``%struct.ST*``' type, which is a pointer, yielding a '``%struct.ST``'
5194 = '``{ i32, double, %struct.RT }``' type, a structure. The second index
5195 indexes into the third element of the structure, yielding a
5196 '``%struct.RT``' = '``{ i8 , [10 x [20 x i32]], i8 }``' type, another
5197 structure. The third index indexes into the second element of the
5198 structure, yielding a '``[10 x [20 x i32]]``' type, an array. The two
5199 dimensions of the array are subscripted into, yielding an '``i32``'
5200 type. The '``getelementptr``' instruction returns a pointer to this
5201 element, thus computing a value of '``i32*``' type.
5203 Note that it is perfectly legal to index partially through a structure,
5204 returning a pointer to an inner element. Because of this, the LLVM code
5205 for the given testcase is equivalent to:
5207 .. code-block:: llvm
5209 define i32* @foo(%struct.ST* %s) {
5210 %t1 = getelementptr %struct.ST* %s, i32 1 ; yields %struct.ST*:%t1
5211 %t2 = getelementptr %struct.ST* %t1, i32 0, i32 2 ; yields %struct.RT*:%t2
5212 %t3 = getelementptr %struct.RT* %t2, i32 0, i32 1 ; yields [10 x [20 x i32]]*:%t3
5213 %t4 = getelementptr [10 x [20 x i32]]* %t3, i32 0, i32 5 ; yields [20 x i32]*:%t4
5214 %t5 = getelementptr [20 x i32]* %t4, i32 0, i32 13 ; yields i32*:%t5
5218 If the ``inbounds`` keyword is present, the result value of the
5219 ``getelementptr`` is a :ref:`poison value <poisonvalues>` if the base
5220 pointer is not an *in bounds* address of an allocated object, or if any
5221 of the addresses that would be formed by successive addition of the
5222 offsets implied by the indices to the base address with infinitely
5223 precise signed arithmetic are not an *in bounds* address of that
5224 allocated object. The *in bounds* addresses for an allocated object are
5225 all the addresses that point into the object, plus the address one byte
5226 past the end. In cases where the base is a vector of pointers the
5227 ``inbounds`` keyword applies to each of the computations element-wise.
5229 If the ``inbounds`` keyword is not present, the offsets are added to the
5230 base address with silently-wrapping two's complement arithmetic. If the
5231 offsets have a different width from the pointer, they are sign-extended
5232 or truncated to the width of the pointer. The result value of the
5233 ``getelementptr`` may be outside the object pointed to by the base
5234 pointer. The result value may not necessarily be used to access memory
5235 though, even if it happens to point into allocated storage. See the
5236 :ref:`Pointer Aliasing Rules <pointeraliasing>` section for more
5239 The getelementptr instruction is often confusing. For some more insight
5240 into how it works, see :doc:`the getelementptr FAQ <GetElementPtr>`.
5245 .. code-block:: llvm
5247 ; yields [12 x i8]*:aptr
5248 %aptr = getelementptr {i32, [12 x i8]}* %saptr, i64 0, i32 1
5250 %vptr = getelementptr {i32, <2 x i8>}* %svptr, i64 0, i32 1, i32 1
5252 %eptr = getelementptr [12 x i8]* %aptr, i64 0, i32 1
5254 %iptr = getelementptr [10 x i32]* @arr, i16 0, i16 0
5256 In cases where the pointer argument is a vector of pointers, each index
5257 must be a vector with the same number of elements. For example:
5259 .. code-block:: llvm
5261 %A = getelementptr <4 x i8*> %ptrs, <4 x i64> %offsets,
5263 Conversion Operations
5264 ---------------------
5266 The instructions in this category are the conversion instructions
5267 (casting) which all take a single operand and a type. They perform
5268 various bit conversions on the operand.
5270 '``trunc .. to``' Instruction
5271 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5278 <result> = trunc <ty> <value> to <ty2> ; yields ty2
5283 The '``trunc``' instruction truncates its operand to the type ``ty2``.
5288 The '``trunc``' instruction takes a value to trunc, and a type to trunc
5289 it to. Both types must be of :ref:`integer <t_integer>` types, or vectors
5290 of the same number of integers. The bit size of the ``value`` must be
5291 larger than the bit size of the destination type, ``ty2``. Equal sized
5292 types are not allowed.
5297 The '``trunc``' instruction truncates the high order bits in ``value``
5298 and converts the remaining bits to ``ty2``. Since the source size must
5299 be larger than the destination size, ``trunc`` cannot be a *no-op cast*.
5300 It will always truncate bits.
5305 .. code-block:: llvm
5307 %X = trunc i32 257 to i8 ; yields i8:1
5308 %Y = trunc i32 123 to i1 ; yields i1:true
5309 %Z = trunc i32 122 to i1 ; yields i1:false
5310 %W = trunc <2 x i16> <i16 8, i16 7> to <2 x i8> ; yields <i8 8, i8 7>
5312 '``zext .. to``' Instruction
5313 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5320 <result> = zext <ty> <value> to <ty2> ; yields ty2
5325 The '``zext``' instruction zero extends its operand to type ``ty2``.
5330 The '``zext``' instruction takes a value to cast, and a type to cast it
5331 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
5332 the same number of integers. The bit size of the ``value`` must be
5333 smaller than the bit size of the destination type, ``ty2``.
5338 The ``zext`` fills the high order bits of the ``value`` with zero bits
5339 until it reaches the size of the destination type, ``ty2``.
5341 When zero extending from i1, the result will always be either 0 or 1.
5346 .. code-block:: llvm
5348 %X = zext i32 257 to i64 ; yields i64:257
5349 %Y = zext i1 true to i32 ; yields i32:1
5350 %Z = zext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5352 '``sext .. to``' Instruction
5353 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5360 <result> = sext <ty> <value> to <ty2> ; yields ty2
5365 The '``sext``' sign extends ``value`` to the type ``ty2``.
5370 The '``sext``' instruction takes a value to cast, and a type to cast it
5371 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
5372 the same number of integers. The bit size of the ``value`` must be
5373 smaller than the bit size of the destination type, ``ty2``.
5378 The '``sext``' instruction performs a sign extension by copying the sign
5379 bit (highest order bit) of the ``value`` until it reaches the bit size
5380 of the type ``ty2``.
5382 When sign extending from i1, the extension always results in -1 or 0.
5387 .. code-block:: llvm
5389 %X = sext i8 -1 to i16 ; yields i16 :65535
5390 %Y = sext i1 true to i32 ; yields i32:-1
5391 %Z = sext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5393 '``fptrunc .. to``' Instruction
5394 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5401 <result> = fptrunc <ty> <value> to <ty2> ; yields ty2
5406 The '``fptrunc``' instruction truncates ``value`` to type ``ty2``.
5411 The '``fptrunc``' instruction takes a :ref:`floating point <t_floating>`
5412 value to cast and a :ref:`floating point <t_floating>` type to cast it to.
5413 The size of ``value`` must be larger than the size of ``ty2``. This
5414 implies that ``fptrunc`` cannot be used to make a *no-op cast*.
5419 The '``fptrunc``' instruction truncates a ``value`` from a larger
5420 :ref:`floating point <t_floating>` type to a smaller :ref:`floating
5421 point <t_floating>` type. If the value cannot fit within the
5422 destination type, ``ty2``, then the results are undefined.
5427 .. code-block:: llvm
5429 %X = fptrunc double 123.0 to float ; yields float:123.0
5430 %Y = fptrunc double 1.0E+300 to float ; yields undefined
5432 '``fpext .. to``' Instruction
5433 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5440 <result> = fpext <ty> <value> to <ty2> ; yields ty2
5445 The '``fpext``' extends a floating point ``value`` to a larger floating
5451 The '``fpext``' instruction takes a :ref:`floating point <t_floating>`
5452 ``value`` to cast, and a :ref:`floating point <t_floating>` type to cast it
5453 to. The source type must be smaller than the destination type.
5458 The '``fpext``' instruction extends the ``value`` from a smaller
5459 :ref:`floating point <t_floating>` type to a larger :ref:`floating
5460 point <t_floating>` type. The ``fpext`` cannot be used to make a
5461 *no-op cast* because it always changes bits. Use ``bitcast`` to make a
5462 *no-op cast* for a floating point cast.
5467 .. code-block:: llvm
5469 %X = fpext float 3.125 to double ; yields double:3.125000e+00
5470 %Y = fpext double %X to fp128 ; yields fp128:0xL00000000000000004000900000000000
5472 '``fptoui .. to``' Instruction
5473 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5480 <result> = fptoui <ty> <value> to <ty2> ; yields ty2
5485 The '``fptoui``' converts a floating point ``value`` to its unsigned
5486 integer equivalent of type ``ty2``.
5491 The '``fptoui``' instruction takes a value to cast, which must be a
5492 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5493 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5494 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5495 type with the same number of elements as ``ty``
5500 The '``fptoui``' instruction converts its :ref:`floating
5501 point <t_floating>` operand into the nearest (rounding towards zero)
5502 unsigned integer value. If the value cannot fit in ``ty2``, the results
5508 .. code-block:: llvm
5510 %X = fptoui double 123.0 to i32 ; yields i32:123
5511 %Y = fptoui float 1.0E+300 to i1 ; yields undefined:1
5512 %Z = fptoui float 1.04E+17 to i8 ; yields undefined:1
5514 '``fptosi .. to``' Instruction
5515 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5522 <result> = fptosi <ty> <value> to <ty2> ; yields ty2
5527 The '``fptosi``' instruction converts :ref:`floating point <t_floating>`
5528 ``value`` to type ``ty2``.
5533 The '``fptosi``' instruction takes a value to cast, which must be a
5534 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5535 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5536 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5537 type with the same number of elements as ``ty``
5542 The '``fptosi``' instruction converts its :ref:`floating
5543 point <t_floating>` operand into the nearest (rounding towards zero)
5544 signed integer value. If the value cannot fit in ``ty2``, the results
5550 .. code-block:: llvm
5552 %X = fptosi double -123.0 to i32 ; yields i32:-123
5553 %Y = fptosi float 1.0E-247 to i1 ; yields undefined:1
5554 %Z = fptosi float 1.04E+17 to i8 ; yields undefined:1
5556 '``uitofp .. to``' Instruction
5557 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5564 <result> = uitofp <ty> <value> to <ty2> ; yields ty2
5569 The '``uitofp``' instruction regards ``value`` as an unsigned integer
5570 and converts that value to the ``ty2`` type.
5575 The '``uitofp``' instruction takes a value to cast, which must be a
5576 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
5577 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
5578 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
5579 type with the same number of elements as ``ty``
5584 The '``uitofp``' instruction interprets its operand as an unsigned
5585 integer quantity and converts it to the corresponding floating point
5586 value. If the value cannot fit in the floating point value, the results
5592 .. code-block:: llvm
5594 %X = uitofp i32 257 to float ; yields float:257.0
5595 %Y = uitofp i8 -1 to double ; yields double:255.0
5597 '``sitofp .. to``' Instruction
5598 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5605 <result> = sitofp <ty> <value> to <ty2> ; yields ty2
5610 The '``sitofp``' instruction regards ``value`` as a signed integer and
5611 converts that value to the ``ty2`` type.
5616 The '``sitofp``' instruction takes a value to cast, which must be a
5617 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
5618 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
5619 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
5620 type with the same number of elements as ``ty``
5625 The '``sitofp``' instruction interprets its operand as a signed integer
5626 quantity and converts it to the corresponding floating point value. If
5627 the value cannot fit in the floating point value, the results are
5633 .. code-block:: llvm
5635 %X = sitofp i32 257 to float ; yields float:257.0
5636 %Y = sitofp i8 -1 to double ; yields double:-1.0
5640 '``ptrtoint .. to``' Instruction
5641 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5648 <result> = ptrtoint <ty> <value> to <ty2> ; yields ty2
5653 The '``ptrtoint``' instruction converts the pointer or a vector of
5654 pointers ``value`` to the integer (or vector of integers) type ``ty2``.
5659 The '``ptrtoint``' instruction takes a ``value`` to cast, which must be
5660 a a value of type :ref:`pointer <t_pointer>` or a vector of pointers, and a
5661 type to cast it to ``ty2``, which must be an :ref:`integer <t_integer>` or
5662 a vector of integers type.
5667 The '``ptrtoint``' instruction converts ``value`` to integer type
5668 ``ty2`` by interpreting the pointer value as an integer and either
5669 truncating or zero extending that value to the size of the integer type.
5670 If ``value`` is smaller than ``ty2`` then a zero extension is done. If
5671 ``value`` is larger than ``ty2`` then a truncation is done. If they are
5672 the same size, then nothing is done (*no-op cast*) other than a type
5678 .. code-block:: llvm
5680 %X = ptrtoint i32* %P to i8 ; yields truncation on 32-bit architecture
5681 %Y = ptrtoint i32* %P to i64 ; yields zero extension on 32-bit architecture
5682 %Z = ptrtoint <4 x i32*> %P to <4 x i64>; yields vector zero extension for a vector of addresses on 32-bit architecture
5686 '``inttoptr .. to``' Instruction
5687 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5694 <result> = inttoptr <ty> <value> to <ty2> ; yields ty2
5699 The '``inttoptr``' instruction converts an integer ``value`` to a
5700 pointer type, ``ty2``.
5705 The '``inttoptr``' instruction takes an :ref:`integer <t_integer>` value to
5706 cast, and a type to cast it to, which must be a :ref:`pointer <t_pointer>`
5712 The '``inttoptr``' instruction converts ``value`` to type ``ty2`` by
5713 applying either a zero extension or a truncation depending on the size
5714 of the integer ``value``. If ``value`` is larger than the size of a
5715 pointer then a truncation is done. If ``value`` is smaller than the size
5716 of a pointer then a zero extension is done. If they are the same size,
5717 nothing is done (*no-op cast*).
5722 .. code-block:: llvm
5724 %X = inttoptr i32 255 to i32* ; yields zero extension on 64-bit architecture
5725 %Y = inttoptr i32 255 to i32* ; yields no-op on 32-bit architecture
5726 %Z = inttoptr i64 0 to i32* ; yields truncation on 32-bit architecture
5727 %Z = inttoptr <4 x i32> %G to <4 x i8*>; yields truncation of vector G to four pointers
5731 '``bitcast .. to``' Instruction
5732 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5739 <result> = bitcast <ty> <value> to <ty2> ; yields ty2
5744 The '``bitcast``' instruction converts ``value`` to type ``ty2`` without
5750 The '``bitcast``' instruction takes a value to cast, which must be a
5751 non-aggregate first class value, and a type to cast it to, which must
5752 also be a non-aggregate :ref:`first class <t_firstclass>` type. The
5753 bit sizes of ``value`` and the destination type, ``ty2``, must be
5754 identical. If the source type is a pointer, the destination type must
5755 also be a pointer of the same size. This instruction supports bitwise
5756 conversion of vectors to integers and to vectors of other types (as
5757 long as they have the same size).
5762 The '``bitcast``' instruction converts ``value`` to type ``ty2``. It
5763 is always a *no-op cast* because no bits change with this
5764 conversion. The conversion is done as if the ``value`` had been stored
5765 to memory and read back as type ``ty2``. Pointer (or vector of
5766 pointers) types may only be converted to other pointer (or vector of
5767 pointers) types with the same address space through this instruction.
5768 To convert pointers to other types, use the :ref:`inttoptr <i_inttoptr>`
5769 or :ref:`ptrtoint <i_ptrtoint>` instructions first.
5774 .. code-block:: llvm
5776 %X = bitcast i8 255 to i8 ; yields i8 :-1
5777 %Y = bitcast i32* %x to sint* ; yields sint*:%x
5778 %Z = bitcast <2 x int> %V to i64; ; yields i64: %V
5779 %Z = bitcast <2 x i32*> %V to <2 x i64*> ; yields <2 x i64*>
5781 .. _i_addrspacecast:
5783 '``addrspacecast .. to``' Instruction
5784 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5791 <result> = addrspacecast <pty> <ptrval> to <pty2> ; yields pty2
5796 The '``addrspacecast``' instruction converts ``ptrval`` from ``pty`` in
5797 address space ``n`` to type ``pty2`` in address space ``m``.
5802 The '``addrspacecast``' instruction takes a pointer or vector of pointer value
5803 to cast and a pointer type to cast it to, which must have a different
5809 The '``addrspacecast``' instruction converts the pointer value
5810 ``ptrval`` to type ``pty2``. It can be a *no-op cast* or a complex
5811 value modification, depending on the target and the address space
5812 pair. Pointer conversions within the same address space must be
5813 performed with the ``bitcast`` instruction. Note that if the address space
5814 conversion is legal then both result and operand refer to the same memory
5820 .. code-block:: llvm
5822 %X = addrspacecast i32* %x to i32 addrspace(1)* ; yields i32 addrspace(1)*:%x
5823 %Y = addrspacecast i32 addrspace(1)* %y to i64 addrspace(2)* ; yields i64 addrspace(2)*:%y
5824 %Z = addrspacecast <4 x i32*> %z to <4 x float addrspace(3)*> ; yields <4 x float addrspace(3)*>:%z
5831 The instructions in this category are the "miscellaneous" instructions,
5832 which defy better classification.
5836 '``icmp``' Instruction
5837 ^^^^^^^^^^^^^^^^^^^^^^
5844 <result> = icmp <cond> <ty> <op1>, <op2> ; yields {i1} or {<N x i1>}:result
5849 The '``icmp``' instruction returns a boolean value or a vector of
5850 boolean values based on comparison of its two integer, integer vector,
5851 pointer, or pointer vector operands.
5856 The '``icmp``' instruction takes three operands. The first operand is
5857 the condition code indicating the kind of comparison to perform. It is
5858 not a value, just a keyword. The possible condition code are:
5861 #. ``ne``: not equal
5862 #. ``ugt``: unsigned greater than
5863 #. ``uge``: unsigned greater or equal
5864 #. ``ult``: unsigned less than
5865 #. ``ule``: unsigned less or equal
5866 #. ``sgt``: signed greater than
5867 #. ``sge``: signed greater or equal
5868 #. ``slt``: signed less than
5869 #. ``sle``: signed less or equal
5871 The remaining two arguments must be :ref:`integer <t_integer>` or
5872 :ref:`pointer <t_pointer>` or integer :ref:`vector <t_vector>` typed. They
5873 must also be identical types.
5878 The '``icmp``' compares ``op1`` and ``op2`` according to the condition
5879 code given as ``cond``. The comparison performed always yields either an
5880 :ref:`i1 <t_integer>` or vector of ``i1`` result, as follows:
5882 #. ``eq``: yields ``true`` if the operands are equal, ``false``
5883 otherwise. No sign interpretation is necessary or performed.
5884 #. ``ne``: yields ``true`` if the operands are unequal, ``false``
5885 otherwise. No sign interpretation is necessary or performed.
5886 #. ``ugt``: interprets the operands as unsigned values and yields
5887 ``true`` if ``op1`` is greater than ``op2``.
5888 #. ``uge``: interprets the operands as unsigned values and yields
5889 ``true`` if ``op1`` is greater than or equal to ``op2``.
5890 #. ``ult``: interprets the operands as unsigned values and yields
5891 ``true`` if ``op1`` is less than ``op2``.
5892 #. ``ule``: interprets the operands as unsigned values and yields
5893 ``true`` if ``op1`` is less than or equal to ``op2``.
5894 #. ``sgt``: interprets the operands as signed values and yields ``true``
5895 if ``op1`` is greater than ``op2``.
5896 #. ``sge``: interprets the operands as signed values and yields ``true``
5897 if ``op1`` is greater than or equal to ``op2``.
5898 #. ``slt``: interprets the operands as signed values and yields ``true``
5899 if ``op1`` is less than ``op2``.
5900 #. ``sle``: interprets the operands as signed values and yields ``true``
5901 if ``op1`` is less than or equal to ``op2``.
5903 If the operands are :ref:`pointer <t_pointer>` typed, the pointer values
5904 are compared as if they were integers.
5906 If the operands are integer vectors, then they are compared element by
5907 element. The result is an ``i1`` vector with the same number of elements
5908 as the values being compared. Otherwise, the result is an ``i1``.
5913 .. code-block:: llvm
5915 <result> = icmp eq i32 4, 5 ; yields: result=false
5916 <result> = icmp ne float* %X, %X ; yields: result=false
5917 <result> = icmp ult i16 4, 5 ; yields: result=true
5918 <result> = icmp sgt i16 4, 5 ; yields: result=false
5919 <result> = icmp ule i16 -4, 5 ; yields: result=false
5920 <result> = icmp sge i16 4, 5 ; yields: result=false
5922 Note that the code generator does not yet support vector types with the
5923 ``icmp`` instruction.
5927 '``fcmp``' Instruction
5928 ^^^^^^^^^^^^^^^^^^^^^^
5935 <result> = fcmp <cond> <ty> <op1>, <op2> ; yields {i1} or {<N x i1>}:result
5940 The '``fcmp``' instruction returns a boolean value or vector of boolean
5941 values based on comparison of its operands.
5943 If the operands are floating point scalars, then the result type is a
5944 boolean (:ref:`i1 <t_integer>`).
5946 If the operands are floating point vectors, then the result type is a
5947 vector of boolean with the same number of elements as the operands being
5953 The '``fcmp``' instruction takes three operands. The first operand is
5954 the condition code indicating the kind of comparison to perform. It is
5955 not a value, just a keyword. The possible condition code are:
5957 #. ``false``: no comparison, always returns false
5958 #. ``oeq``: ordered and equal
5959 #. ``ogt``: ordered and greater than
5960 #. ``oge``: ordered and greater than or equal
5961 #. ``olt``: ordered and less than
5962 #. ``ole``: ordered and less than or equal
5963 #. ``one``: ordered and not equal
5964 #. ``ord``: ordered (no nans)
5965 #. ``ueq``: unordered or equal
5966 #. ``ugt``: unordered or greater than
5967 #. ``uge``: unordered or greater than or equal
5968 #. ``ult``: unordered or less than
5969 #. ``ule``: unordered or less than or equal
5970 #. ``une``: unordered or not equal
5971 #. ``uno``: unordered (either nans)
5972 #. ``true``: no comparison, always returns true
5974 *Ordered* means that neither operand is a QNAN while *unordered* means
5975 that either operand may be a QNAN.
5977 Each of ``val1`` and ``val2`` arguments must be either a :ref:`floating
5978 point <t_floating>` type or a :ref:`vector <t_vector>` of floating point
5979 type. They must have identical types.
5984 The '``fcmp``' instruction compares ``op1`` and ``op2`` according to the
5985 condition code given as ``cond``. If the operands are vectors, then the
5986 vectors are compared element by element. Each comparison performed
5987 always yields an :ref:`i1 <t_integer>` result, as follows:
5989 #. ``false``: always yields ``false``, regardless of operands.
5990 #. ``oeq``: yields ``true`` if both operands are not a QNAN and ``op1``
5991 is equal to ``op2``.
5992 #. ``ogt``: yields ``true`` if both operands are not a QNAN and ``op1``
5993 is greater than ``op2``.
5994 #. ``oge``: yields ``true`` if both operands are not a QNAN and ``op1``
5995 is greater than or equal to ``op2``.
5996 #. ``olt``: yields ``true`` if both operands are not a QNAN and ``op1``
5997 is less than ``op2``.
5998 #. ``ole``: yields ``true`` if both operands are not a QNAN and ``op1``
5999 is less than or equal to ``op2``.
6000 #. ``one``: yields ``true`` if both operands are not a QNAN and ``op1``
6001 is not equal to ``op2``.
6002 #. ``ord``: yields ``true`` if both operands are not a QNAN.
6003 #. ``ueq``: yields ``true`` if either operand is a QNAN or ``op1`` is
6005 #. ``ugt``: yields ``true`` if either operand is a QNAN or ``op1`` is
6006 greater than ``op2``.
6007 #. ``uge``: yields ``true`` if either operand is a QNAN or ``op1`` is
6008 greater than or equal to ``op2``.
6009 #. ``ult``: yields ``true`` if either operand is a QNAN or ``op1`` is
6011 #. ``ule``: yields ``true`` if either operand is a QNAN or ``op1`` is
6012 less than or equal to ``op2``.
6013 #. ``une``: yields ``true`` if either operand is a QNAN or ``op1`` is
6014 not equal to ``op2``.
6015 #. ``uno``: yields ``true`` if either operand is a QNAN.
6016 #. ``true``: always yields ``true``, regardless of operands.
6021 .. code-block:: llvm
6023 <result> = fcmp oeq float 4.0, 5.0 ; yields: result=false
6024 <result> = fcmp one float 4.0, 5.0 ; yields: result=true
6025 <result> = fcmp olt float 4.0, 5.0 ; yields: result=true
6026 <result> = fcmp ueq double 1.0, 2.0 ; yields: result=false
6028 Note that the code generator does not yet support vector types with the
6029 ``fcmp`` instruction.
6033 '``phi``' Instruction
6034 ^^^^^^^^^^^^^^^^^^^^^
6041 <result> = phi <ty> [ <val0>, <label0>], ...
6046 The '``phi``' instruction is used to implement the φ node in the SSA
6047 graph representing the function.
6052 The type of the incoming values is specified with the first type field.
6053 After this, the '``phi``' instruction takes a list of pairs as
6054 arguments, with one pair for each predecessor basic block of the current
6055 block. Only values of :ref:`first class <t_firstclass>` type may be used as
6056 the value arguments to the PHI node. Only labels may be used as the
6059 There must be no non-phi instructions between the start of a basic block
6060 and the PHI instructions: i.e. PHI instructions must be first in a basic
6063 For the purposes of the SSA form, the use of each incoming value is
6064 deemed to occur on the edge from the corresponding predecessor block to
6065 the current block (but after any definition of an '``invoke``'
6066 instruction's return value on the same edge).
6071 At runtime, the '``phi``' instruction logically takes on the value
6072 specified by the pair corresponding to the predecessor basic block that
6073 executed just prior to the current block.
6078 .. code-block:: llvm
6080 Loop: ; Infinite loop that counts from 0 on up...
6081 %indvar = phi i32 [ 0, %LoopHeader ], [ %nextindvar, %Loop ]
6082 %nextindvar = add i32 %indvar, 1
6087 '``select``' Instruction
6088 ^^^^^^^^^^^^^^^^^^^^^^^^
6095 <result> = select selty <cond>, <ty> <val1>, <ty> <val2> ; yields ty
6097 selty is either i1 or {<N x i1>}
6102 The '``select``' instruction is used to choose one value based on a
6103 condition, without branching.
6108 The '``select``' instruction requires an 'i1' value or a vector of 'i1'
6109 values indicating the condition, and two values of the same :ref:`first
6110 class <t_firstclass>` type. If the val1/val2 are vectors and the
6111 condition is a scalar, then entire vectors are selected, not individual
6117 If the condition is an i1 and it evaluates to 1, the instruction returns
6118 the first value argument; otherwise, it returns the second value
6121 If the condition is a vector of i1, then the value arguments must be
6122 vectors of the same size, and the selection is done element by element.
6127 .. code-block:: llvm
6129 %X = select i1 true, i8 17, i8 42 ; yields i8:17
6133 '``call``' Instruction
6134 ^^^^^^^^^^^^^^^^^^^^^^
6141 <result> = [tail] call [cconv] [ret attrs] <ty> [<fnty>*] <fnptrval>(<function args>) [fn attrs]
6146 The '``call``' instruction represents a simple function call.
6151 This instruction requires several arguments:
6153 #. The optional "tail" marker indicates that the callee function does
6154 not access any allocas or varargs in the caller. Note that calls may
6155 be marked "tail" even if they do not occur before a
6156 :ref:`ret <i_ret>` instruction. If the "tail" marker is present, the
6157 function call is eligible for tail call optimization, but `might not
6158 in fact be optimized into a jump <CodeGenerator.html#tailcallopt>`_.
6159 The code generator may optimize calls marked "tail" with either 1)
6160 automatic `sibling call
6161 optimization <CodeGenerator.html#sibcallopt>`_ when the caller and
6162 callee have matching signatures, or 2) forced tail call optimization
6163 when the following extra requirements are met:
6165 - Caller and callee both have the calling convention ``fastcc``.
6166 - The call is in tail position (ret immediately follows call and ret
6167 uses value of call or is void).
6168 - Option ``-tailcallopt`` is enabled, or
6169 ``llvm::GuaranteedTailCallOpt`` is ``true``.
6170 - `Platform specific constraints are
6171 met. <CodeGenerator.html#tailcallopt>`_
6173 #. The optional "cconv" marker indicates which :ref:`calling
6174 convention <callingconv>` the call should use. If none is
6175 specified, the call defaults to using C calling conventions. The
6176 calling convention of the call must match the calling convention of
6177 the target function, or else the behavior is undefined.
6178 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
6179 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
6181 #. '``ty``': the type of the call instruction itself which is also the
6182 type of the return value. Functions that return no value are marked
6184 #. '``fnty``': shall be the signature of the pointer to function value
6185 being invoked. The argument types must match the types implied by
6186 this signature. This type can be omitted if the function is not
6187 varargs and if the function type does not return a pointer to a
6189 #. '``fnptrval``': An LLVM value containing a pointer to a function to
6190 be invoked. In most cases, this is a direct function invocation, but
6191 indirect ``call``'s are just as possible, calling an arbitrary pointer
6193 #. '``function args``': argument list whose types match the function
6194 signature argument types and parameter attributes. All arguments must
6195 be of :ref:`first class <t_firstclass>` type. If the function signature
6196 indicates the function accepts a variable number of arguments, the
6197 extra arguments can be specified.
6198 #. The optional :ref:`function attributes <fnattrs>` list. Only
6199 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
6200 attributes are valid here.
6205 The '``call``' instruction is used to cause control flow to transfer to
6206 a specified function, with its incoming arguments bound to the specified
6207 values. Upon a '``ret``' instruction in the called function, control
6208 flow continues with the instruction after the function call, and the
6209 return value of the function is bound to the result argument.
6214 .. code-block:: llvm
6216 %retval = call i32 @test(i32 %argc)
6217 call i32 (i8*, ...)* @printf(i8* %msg, i32 12, i8 42) ; yields i32
6218 %X = tail call i32 @foo() ; yields i32
6219 %Y = tail call fastcc i32 @foo() ; yields i32
6220 call void %foo(i8 97 signext)
6222 %struct.A = type { i32, i8 }
6223 %r = call %struct.A @foo() ; yields { 32, i8 }
6224 %gr = extractvalue %struct.A %r, 0 ; yields i32
6225 %gr1 = extractvalue %struct.A %r, 1 ; yields i8
6226 %Z = call void @foo() noreturn ; indicates that %foo never returns normally
6227 %ZZ = call zeroext i32 @bar() ; Return value is %zero extended
6229 llvm treats calls to some functions with names and arguments that match
6230 the standard C99 library as being the C99 library functions, and may
6231 perform optimizations or generate code for them under that assumption.
6232 This is something we'd like to change in the future to provide better
6233 support for freestanding environments and non-C-based languages.
6237 '``va_arg``' Instruction
6238 ^^^^^^^^^^^^^^^^^^^^^^^^
6245 <resultval> = va_arg <va_list*> <arglist>, <argty>
6250 The '``va_arg``' instruction is used to access arguments passed through
6251 the "variable argument" area of a function call. It is used to implement
6252 the ``va_arg`` macro in C.
6257 This instruction takes a ``va_list*`` value and the type of the
6258 argument. It returns a value of the specified argument type and
6259 increments the ``va_list`` to point to the next argument. The actual
6260 type of ``va_list`` is target specific.
6265 The '``va_arg``' instruction loads an argument of the specified type
6266 from the specified ``va_list`` and causes the ``va_list`` to point to
6267 the next argument. For more information, see the variable argument
6268 handling :ref:`Intrinsic Functions <int_varargs>`.
6270 It is legal for this instruction to be called in a function which does
6271 not take a variable number of arguments, for example, the ``vfprintf``
6274 ``va_arg`` is an LLVM instruction instead of an :ref:`intrinsic
6275 function <intrinsics>` because it takes a type as an argument.
6280 See the :ref:`variable argument processing <int_varargs>` section.
6282 Note that the code generator does not yet fully support va\_arg on many
6283 targets. Also, it does not currently support va\_arg with aggregate
6284 types on any target.
6288 '``landingpad``' Instruction
6289 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6296 <resultval> = landingpad <resultty> personality <type> <pers_fn> <clause>+
6297 <resultval> = landingpad <resultty> personality <type> <pers_fn> cleanup <clause>*
6299 <clause> := catch <type> <value>
6300 <clause> := filter <array constant type> <array constant>
6305 The '``landingpad``' instruction is used by `LLVM's exception handling
6306 system <ExceptionHandling.html#overview>`_ to specify that a basic block
6307 is a landing pad --- one where the exception lands, and corresponds to the
6308 code found in the ``catch`` portion of a ``try``/``catch`` sequence. It
6309 defines values supplied by the personality function (``pers_fn``) upon
6310 re-entry to the function. The ``resultval`` has the type ``resultty``.
6315 This instruction takes a ``pers_fn`` value. This is the personality
6316 function associated with the unwinding mechanism. The optional
6317 ``cleanup`` flag indicates that the landing pad block is a cleanup.
6319 A ``clause`` begins with the clause type --- ``catch`` or ``filter`` --- and
6320 contains the global variable representing the "type" that may be caught
6321 or filtered respectively. Unlike the ``catch`` clause, the ``filter``
6322 clause takes an array constant as its argument. Use
6323 "``[0 x i8**] undef``" for a filter which cannot throw. The
6324 '``landingpad``' instruction must contain *at least* one ``clause`` or
6325 the ``cleanup`` flag.
6330 The '``landingpad``' instruction defines the values which are set by the
6331 personality function (``pers_fn``) upon re-entry to the function, and
6332 therefore the "result type" of the ``landingpad`` instruction. As with
6333 calling conventions, how the personality function results are
6334 represented in LLVM IR is target specific.
6336 The clauses are applied in order from top to bottom. If two
6337 ``landingpad`` instructions are merged together through inlining, the
6338 clauses from the calling function are appended to the list of clauses.
6339 When the call stack is being unwound due to an exception being thrown,
6340 the exception is compared against each ``clause`` in turn. If it doesn't
6341 match any of the clauses, and the ``cleanup`` flag is not set, then
6342 unwinding continues further up the call stack.
6344 The ``landingpad`` instruction has several restrictions:
6346 - A landing pad block is a basic block which is the unwind destination
6347 of an '``invoke``' instruction.
6348 - A landing pad block must have a '``landingpad``' instruction as its
6349 first non-PHI instruction.
6350 - There can be only one '``landingpad``' instruction within the landing
6352 - A basic block that is not a landing pad block may not include a
6353 '``landingpad``' instruction.
6354 - All '``landingpad``' instructions in a function must have the same
6355 personality function.
6360 .. code-block:: llvm
6362 ;; A landing pad which can catch an integer.
6363 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6365 ;; A landing pad that is a cleanup.
6366 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6368 ;; A landing pad which can catch an integer and can only throw a double.
6369 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6371 filter [1 x i8**] [@_ZTId]
6378 LLVM supports the notion of an "intrinsic function". These functions
6379 have well known names and semantics and are required to follow certain
6380 restrictions. Overall, these intrinsics represent an extension mechanism
6381 for the LLVM language that does not require changing all of the
6382 transformations in LLVM when adding to the language (or the bitcode
6383 reader/writer, the parser, etc...).
6385 Intrinsic function names must all start with an "``llvm.``" prefix. This
6386 prefix is reserved in LLVM for intrinsic names; thus, function names may
6387 not begin with this prefix. Intrinsic functions must always be external
6388 functions: you cannot define the body of intrinsic functions. Intrinsic
6389 functions may only be used in call or invoke instructions: it is illegal
6390 to take the address of an intrinsic function. Additionally, because
6391 intrinsic functions are part of the LLVM language, it is required if any
6392 are added that they be documented here.
6394 Some intrinsic functions can be overloaded, i.e., the intrinsic
6395 represents a family of functions that perform the same operation but on
6396 different data types. Because LLVM can represent over 8 million
6397 different integer types, overloading is used commonly to allow an
6398 intrinsic function to operate on any integer type. One or more of the
6399 argument types or the result type can be overloaded to accept any
6400 integer type. Argument types may also be defined as exactly matching a
6401 previous argument's type or the result type. This allows an intrinsic
6402 function which accepts multiple arguments, but needs all of them to be
6403 of the same type, to only be overloaded with respect to a single
6404 argument or the result.
6406 Overloaded intrinsics will have the names of its overloaded argument
6407 types encoded into its function name, each preceded by a period. Only
6408 those types which are overloaded result in a name suffix. Arguments
6409 whose type is matched against another type do not. For example, the
6410 ``llvm.ctpop`` function can take an integer of any width and returns an
6411 integer of exactly the same integer width. This leads to a family of
6412 functions such as ``i8 @llvm.ctpop.i8(i8 %val)`` and
6413 ``i29 @llvm.ctpop.i29(i29 %val)``. Only one type, the return type, is
6414 overloaded, and only one type suffix is required. Because the argument's
6415 type is matched against the return type, it does not require its own
6418 To learn how to add an intrinsic function, please see the `Extending
6419 LLVM Guide <ExtendingLLVM.html>`_.
6423 Variable Argument Handling Intrinsics
6424 -------------------------------------
6426 Variable argument support is defined in LLVM with the
6427 :ref:`va_arg <i_va_arg>` instruction and these three intrinsic
6428 functions. These functions are related to the similarly named macros
6429 defined in the ``<stdarg.h>`` header file.
6431 All of these functions operate on arguments that use a target-specific
6432 value type "``va_list``". The LLVM assembly language reference manual
6433 does not define what this type is, so all transformations should be
6434 prepared to handle these functions regardless of the type used.
6436 This example shows how the :ref:`va_arg <i_va_arg>` instruction and the
6437 variable argument handling intrinsic functions are used.
6439 .. code-block:: llvm
6441 define i32 @test(i32 %X, ...) {
6442 ; Initialize variable argument processing
6444 %ap2 = bitcast i8** %ap to i8*
6445 call void @llvm.va_start(i8* %ap2)
6447 ; Read a single integer argument
6448 %tmp = va_arg i8** %ap, i32
6450 ; Demonstrate usage of llvm.va_copy and llvm.va_end
6452 %aq2 = bitcast i8** %aq to i8*
6453 call void @llvm.va_copy(i8* %aq2, i8* %ap2)
6454 call void @llvm.va_end(i8* %aq2)
6456 ; Stop processing of arguments.
6457 call void @llvm.va_end(i8* %ap2)
6461 declare void @llvm.va_start(i8*)
6462 declare void @llvm.va_copy(i8*, i8*)
6463 declare void @llvm.va_end(i8*)
6467 '``llvm.va_start``' Intrinsic
6468 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6475 declare void @llvm.va_start(i8* <arglist>)
6480 The '``llvm.va_start``' intrinsic initializes ``*<arglist>`` for
6481 subsequent use by ``va_arg``.
6486 The argument is a pointer to a ``va_list`` element to initialize.
6491 The '``llvm.va_start``' intrinsic works just like the ``va_start`` macro
6492 available in C. In a target-dependent way, it initializes the
6493 ``va_list`` element to which the argument points, so that the next call
6494 to ``va_arg`` will produce the first variable argument passed to the
6495 function. Unlike the C ``va_start`` macro, this intrinsic does not need
6496 to know the last argument of the function as the compiler can figure
6499 '``llvm.va_end``' Intrinsic
6500 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6507 declare void @llvm.va_end(i8* <arglist>)
6512 The '``llvm.va_end``' intrinsic destroys ``*<arglist>``, which has been
6513 initialized previously with ``llvm.va_start`` or ``llvm.va_copy``.
6518 The argument is a pointer to a ``va_list`` to destroy.
6523 The '``llvm.va_end``' intrinsic works just like the ``va_end`` macro
6524 available in C. In a target-dependent way, it destroys the ``va_list``
6525 element to which the argument points. Calls to
6526 :ref:`llvm.va_start <int_va_start>` and
6527 :ref:`llvm.va_copy <int_va_copy>` must be matched exactly with calls to
6532 '``llvm.va_copy``' Intrinsic
6533 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6540 declare void @llvm.va_copy(i8* <destarglist>, i8* <srcarglist>)
6545 The '``llvm.va_copy``' intrinsic copies the current argument position
6546 from the source argument list to the destination argument list.
6551 The first argument is a pointer to a ``va_list`` element to initialize.
6552 The second argument is a pointer to a ``va_list`` element to copy from.
6557 The '``llvm.va_copy``' intrinsic works just like the ``va_copy`` macro
6558 available in C. In a target-dependent way, it copies the source
6559 ``va_list`` element into the destination ``va_list`` element. This
6560 intrinsic is necessary because the `` llvm.va_start`` intrinsic may be
6561 arbitrarily complex and require, for example, memory allocation.
6563 Accurate Garbage Collection Intrinsics
6564 --------------------------------------
6566 LLVM support for `Accurate Garbage Collection <GarbageCollection.html>`_
6567 (GC) requires the implementation and generation of these intrinsics.
6568 These intrinsics allow identification of :ref:`GC roots on the
6569 stack <int_gcroot>`, as well as garbage collector implementations that
6570 require :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers.
6571 Front-ends for type-safe garbage collected languages should generate
6572 these intrinsics to make use of the LLVM garbage collectors. For more
6573 details, see `Accurate Garbage Collection with
6574 LLVM <GarbageCollection.html>`_.
6576 The garbage collection intrinsics only operate on objects in the generic
6577 address space (address space zero).
6581 '``llvm.gcroot``' Intrinsic
6582 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6589 declare void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
6594 The '``llvm.gcroot``' intrinsic declares the existence of a GC root to
6595 the code generator, and allows some metadata to be associated with it.
6600 The first argument specifies the address of a stack object that contains
6601 the root pointer. The second pointer (which must be either a constant or
6602 a global value address) contains the meta-data to be associated with the
6608 At runtime, a call to this intrinsic stores a null pointer into the
6609 "ptrloc" location. At compile-time, the code generator generates
6610 information to allow the runtime to find the pointer at GC safe points.
6611 The '``llvm.gcroot``' intrinsic may only be used in a function which
6612 :ref:`specifies a GC algorithm <gc>`.
6616 '``llvm.gcread``' Intrinsic
6617 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6624 declare i8* @llvm.gcread(i8* %ObjPtr, i8** %Ptr)
6629 The '``llvm.gcread``' intrinsic identifies reads of references from heap
6630 locations, allowing garbage collector implementations that require read
6636 The second argument is the address to read from, which should be an
6637 address allocated from the garbage collector. The first object is a
6638 pointer to the start of the referenced object, if needed by the language
6639 runtime (otherwise null).
6644 The '``llvm.gcread``' intrinsic has the same semantics as a load
6645 instruction, but may be replaced with substantially more complex code by
6646 the garbage collector runtime, as needed. The '``llvm.gcread``'
6647 intrinsic may only be used in a function which :ref:`specifies a GC
6652 '``llvm.gcwrite``' Intrinsic
6653 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6660 declare void @llvm.gcwrite(i8* %P1, i8* %Obj, i8** %P2)
6665 The '``llvm.gcwrite``' intrinsic identifies writes of references to heap
6666 locations, allowing garbage collector implementations that require write
6667 barriers (such as generational or reference counting collectors).
6672 The first argument is the reference to store, the second is the start of
6673 the object to store it to, and the third is the address of the field of
6674 Obj to store to. If the runtime does not require a pointer to the
6675 object, Obj may be null.
6680 The '``llvm.gcwrite``' intrinsic has the same semantics as a store
6681 instruction, but may be replaced with substantially more complex code by
6682 the garbage collector runtime, as needed. The '``llvm.gcwrite``'
6683 intrinsic may only be used in a function which :ref:`specifies a GC
6686 Code Generator Intrinsics
6687 -------------------------
6689 These intrinsics are provided by LLVM to expose special features that
6690 may only be implemented with code generator support.
6692 '``llvm.returnaddress``' Intrinsic
6693 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6700 declare i8 *@llvm.returnaddress(i32 <level>)
6705 The '``llvm.returnaddress``' intrinsic attempts to compute a
6706 target-specific value indicating the return address of the current
6707 function or one of its callers.
6712 The argument to this intrinsic indicates which function to return the
6713 address for. Zero indicates the calling function, one indicates its
6714 caller, etc. The argument is **required** to be a constant integer
6720 The '``llvm.returnaddress``' intrinsic either returns a pointer
6721 indicating the return address of the specified call frame, or zero if it
6722 cannot be identified. The value returned by this intrinsic is likely to
6723 be incorrect or 0 for arguments other than zero, so it should only be
6724 used for debugging purposes.
6726 Note that calling this intrinsic does not prevent function inlining or
6727 other aggressive transformations, so the value returned may not be that
6728 of the obvious source-language caller.
6730 '``llvm.frameaddress``' Intrinsic
6731 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6738 declare i8* @llvm.frameaddress(i32 <level>)
6743 The '``llvm.frameaddress``' intrinsic attempts to return the
6744 target-specific frame pointer value for the specified stack frame.
6749 The argument to this intrinsic indicates which function to return the
6750 frame pointer for. Zero indicates the calling function, one indicates
6751 its caller, etc. The argument is **required** to be a constant integer
6757 The '``llvm.frameaddress``' intrinsic either returns a pointer
6758 indicating the frame address of the specified call frame, or zero if it
6759 cannot be identified. The value returned by this intrinsic is likely to
6760 be incorrect or 0 for arguments other than zero, so it should only be
6761 used for debugging purposes.
6763 Note that calling this intrinsic does not prevent function inlining or
6764 other aggressive transformations, so the value returned may not be that
6765 of the obvious source-language caller.
6769 '``llvm.stacksave``' Intrinsic
6770 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6777 declare i8* @llvm.stacksave()
6782 The '``llvm.stacksave``' intrinsic is used to remember the current state
6783 of the function stack, for use with
6784 :ref:`llvm.stackrestore <int_stackrestore>`. This is useful for
6785 implementing language features like scoped automatic variable sized
6791 This intrinsic returns a opaque pointer value that can be passed to
6792 :ref:`llvm.stackrestore <int_stackrestore>`. When an
6793 ``llvm.stackrestore`` intrinsic is executed with a value saved from
6794 ``llvm.stacksave``, it effectively restores the state of the stack to
6795 the state it was in when the ``llvm.stacksave`` intrinsic executed. In
6796 practice, this pops any :ref:`alloca <i_alloca>` blocks from the stack that
6797 were allocated after the ``llvm.stacksave`` was executed.
6799 .. _int_stackrestore:
6801 '``llvm.stackrestore``' Intrinsic
6802 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6809 declare void @llvm.stackrestore(i8* %ptr)
6814 The '``llvm.stackrestore``' intrinsic is used to restore the state of
6815 the function stack to the state it was in when the corresponding
6816 :ref:`llvm.stacksave <int_stacksave>` intrinsic executed. This is
6817 useful for implementing language features like scoped automatic variable
6818 sized arrays in C99.
6823 See the description for :ref:`llvm.stacksave <int_stacksave>`.
6825 '``llvm.prefetch``' Intrinsic
6826 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6833 declare void @llvm.prefetch(i8* <address>, i32 <rw>, i32 <locality>, i32 <cache type>)
6838 The '``llvm.prefetch``' intrinsic is a hint to the code generator to
6839 insert a prefetch instruction if supported; otherwise, it is a noop.
6840 Prefetches have no effect on the behavior of the program but can change
6841 its performance characteristics.
6846 ``address`` is the address to be prefetched, ``rw`` is the specifier
6847 determining if the fetch should be for a read (0) or write (1), and
6848 ``locality`` is a temporal locality specifier ranging from (0) - no
6849 locality, to (3) - extremely local keep in cache. The ``cache type``
6850 specifies whether the prefetch is performed on the data (1) or
6851 instruction (0) cache. The ``rw``, ``locality`` and ``cache type``
6852 arguments must be constant integers.
6857 This intrinsic does not modify the behavior of the program. In
6858 particular, prefetches cannot trap and do not produce a value. On
6859 targets that support this intrinsic, the prefetch can provide hints to
6860 the processor cache for better performance.
6862 '``llvm.pcmarker``' Intrinsic
6863 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6870 declare void @llvm.pcmarker(i32 <id>)
6875 The '``llvm.pcmarker``' intrinsic is a method to export a Program
6876 Counter (PC) in a region of code to simulators and other tools. The
6877 method is target specific, but it is expected that the marker will use
6878 exported symbols to transmit the PC of the marker. The marker makes no
6879 guarantees that it will remain with any specific instruction after
6880 optimizations. It is possible that the presence of a marker will inhibit
6881 optimizations. The intended use is to be inserted after optimizations to
6882 allow correlations of simulation runs.
6887 ``id`` is a numerical id identifying the marker.
6892 This intrinsic does not modify the behavior of the program. Backends
6893 that do not support this intrinsic may ignore it.
6895 '``llvm.readcyclecounter``' Intrinsic
6896 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6903 declare i64 @llvm.readcyclecounter()
6908 The '``llvm.readcyclecounter``' intrinsic provides access to the cycle
6909 counter register (or similar low latency, high accuracy clocks) on those
6910 targets that support it. On X86, it should map to RDTSC. On Alpha, it
6911 should map to RPCC. As the backing counters overflow quickly (on the
6912 order of 9 seconds on alpha), this should only be used for small
6918 When directly supported, reading the cycle counter should not modify any
6919 memory. Implementations are allowed to either return a application
6920 specific value or a system wide value. On backends without support, this
6921 is lowered to a constant 0.
6923 Note that runtime support may be conditional on the privilege-level code is
6924 running at and the host platform.
6926 Standard C Library Intrinsics
6927 -----------------------------
6929 LLVM provides intrinsics for a few important standard C library
6930 functions. These intrinsics allow source-language front-ends to pass
6931 information about the alignment of the pointer arguments to the code
6932 generator, providing opportunity for more efficient code generation.
6936 '``llvm.memcpy``' Intrinsic
6937 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6942 This is an overloaded intrinsic. You can use ``llvm.memcpy`` on any
6943 integer bit width and for different address spaces. Not all targets
6944 support all bit widths however.
6948 declare void @llvm.memcpy.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
6949 i32 <len>, i32 <align>, i1 <isvolatile>)
6950 declare void @llvm.memcpy.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
6951 i64 <len>, i32 <align>, i1 <isvolatile>)
6956 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
6957 source location to the destination location.
6959 Note that, unlike the standard libc function, the ``llvm.memcpy.*``
6960 intrinsics do not return a value, takes extra alignment/isvolatile
6961 arguments and the pointers can be in specified address spaces.
6966 The first argument is a pointer to the destination, the second is a
6967 pointer to the source. The third argument is an integer argument
6968 specifying the number of bytes to copy, the fourth argument is the
6969 alignment of the source and destination locations, and the fifth is a
6970 boolean indicating a volatile access.
6972 If the call to this intrinsic has an alignment value that is not 0 or 1,
6973 then the caller guarantees that both the source and destination pointers
6974 are aligned to that boundary.
6976 If the ``isvolatile`` parameter is ``true``, the ``llvm.memcpy`` call is
6977 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
6978 very cleanly specified and it is unwise to depend on it.
6983 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
6984 source location to the destination location, which are not allowed to
6985 overlap. It copies "len" bytes of memory over. If the argument is known
6986 to be aligned to some boundary, this can be specified as the fourth
6987 argument, otherwise it should be set to 0 or 1 (both meaning no alignment).
6989 '``llvm.memmove``' Intrinsic
6990 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6995 This is an overloaded intrinsic. You can use llvm.memmove on any integer
6996 bit width and for different address space. Not all targets support all
7001 declare void @llvm.memmove.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
7002 i32 <len>, i32 <align>, i1 <isvolatile>)
7003 declare void @llvm.memmove.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
7004 i64 <len>, i32 <align>, i1 <isvolatile>)
7009 The '``llvm.memmove.*``' intrinsics move a block of memory from the
7010 source location to the destination location. It is similar to the
7011 '``llvm.memcpy``' intrinsic but allows the two memory locations to
7014 Note that, unlike the standard libc function, the ``llvm.memmove.*``
7015 intrinsics do not return a value, takes extra alignment/isvolatile
7016 arguments and the pointers can be in specified address spaces.
7021 The first argument is a pointer to the destination, the second is a
7022 pointer to the source. The third argument is an integer argument
7023 specifying the number of bytes to copy, the fourth argument is the
7024 alignment of the source and destination locations, and the fifth is a
7025 boolean indicating a volatile access.
7027 If the call to this intrinsic has an alignment value that is not 0 or 1,
7028 then the caller guarantees that the source and destination pointers are
7029 aligned to that boundary.
7031 If the ``isvolatile`` parameter is ``true``, the ``llvm.memmove`` call
7032 is a :ref:`volatile operation <volatile>`. The detailed access behavior is
7033 not very cleanly specified and it is unwise to depend on it.
7038 The '``llvm.memmove.*``' intrinsics copy a block of memory from the
7039 source location to the destination location, which may overlap. It
7040 copies "len" bytes of memory over. If the argument is known to be
7041 aligned to some boundary, this can be specified as the fourth argument,
7042 otherwise it should be set to 0 or 1 (both meaning no alignment).
7044 '``llvm.memset.*``' Intrinsics
7045 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7050 This is an overloaded intrinsic. You can use llvm.memset on any integer
7051 bit width and for different address spaces. However, not all targets
7052 support all bit widths.
7056 declare void @llvm.memset.p0i8.i32(i8* <dest>, i8 <val>,
7057 i32 <len>, i32 <align>, i1 <isvolatile>)
7058 declare void @llvm.memset.p0i8.i64(i8* <dest>, i8 <val>,
7059 i64 <len>, i32 <align>, i1 <isvolatile>)
7064 The '``llvm.memset.*``' intrinsics fill a block of memory with a
7065 particular byte value.
7067 Note that, unlike the standard libc function, the ``llvm.memset``
7068 intrinsic does not return a value and takes extra alignment/volatile
7069 arguments. Also, the destination can be in an arbitrary address space.
7074 The first argument is a pointer to the destination to fill, the second
7075 is the byte value with which to fill it, the third argument is an
7076 integer argument specifying the number of bytes to fill, and the fourth
7077 argument is the known alignment of the destination location.
7079 If the call to this intrinsic has an alignment value that is not 0 or 1,
7080 then the caller guarantees that the destination pointer is aligned to
7083 If the ``isvolatile`` parameter is ``true``, the ``llvm.memset`` call is
7084 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
7085 very cleanly specified and it is unwise to depend on it.
7090 The '``llvm.memset.*``' intrinsics fill "len" bytes of memory starting
7091 at the destination location. If the argument is known to be aligned to
7092 some boundary, this can be specified as the fourth argument, otherwise
7093 it should be set to 0 or 1 (both meaning no alignment).
7095 '``llvm.sqrt.*``' Intrinsic
7096 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7101 This is an overloaded intrinsic. You can use ``llvm.sqrt`` on any
7102 floating point or vector of floating point type. Not all targets support
7107 declare float @llvm.sqrt.f32(float %Val)
7108 declare double @llvm.sqrt.f64(double %Val)
7109 declare x86_fp80 @llvm.sqrt.f80(x86_fp80 %Val)
7110 declare fp128 @llvm.sqrt.f128(fp128 %Val)
7111 declare ppc_fp128 @llvm.sqrt.ppcf128(ppc_fp128 %Val)
7116 The '``llvm.sqrt``' intrinsics return the sqrt of the specified operand,
7117 returning the same value as the libm '``sqrt``' functions would. Unlike
7118 ``sqrt`` in libm, however, ``llvm.sqrt`` has undefined behavior for
7119 negative numbers other than -0.0 (which allows for better optimization,
7120 because there is no need to worry about errno being set).
7121 ``llvm.sqrt(-0.0)`` is defined to return -0.0 like IEEE sqrt.
7126 The argument and return value are floating point numbers of the same
7132 This function returns the sqrt of the specified operand if it is a
7133 nonnegative floating point number.
7135 '``llvm.powi.*``' Intrinsic
7136 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7141 This is an overloaded intrinsic. You can use ``llvm.powi`` on any
7142 floating point or vector of floating point type. Not all targets support
7147 declare float @llvm.powi.f32(float %Val, i32 %power)
7148 declare double @llvm.powi.f64(double %Val, i32 %power)
7149 declare x86_fp80 @llvm.powi.f80(x86_fp80 %Val, i32 %power)
7150 declare fp128 @llvm.powi.f128(fp128 %Val, i32 %power)
7151 declare ppc_fp128 @llvm.powi.ppcf128(ppc_fp128 %Val, i32 %power)
7156 The '``llvm.powi.*``' intrinsics return the first operand raised to the
7157 specified (positive or negative) power. The order of evaluation of
7158 multiplications is not defined. When a vector of floating point type is
7159 used, the second argument remains a scalar integer value.
7164 The second argument is an integer power, and the first is a value to
7165 raise to that power.
7170 This function returns the first value raised to the second power with an
7171 unspecified sequence of rounding operations.
7173 '``llvm.sin.*``' Intrinsic
7174 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7179 This is an overloaded intrinsic. You can use ``llvm.sin`` on any
7180 floating point or vector of floating point type. Not all targets support
7185 declare float @llvm.sin.f32(float %Val)
7186 declare double @llvm.sin.f64(double %Val)
7187 declare x86_fp80 @llvm.sin.f80(x86_fp80 %Val)
7188 declare fp128 @llvm.sin.f128(fp128 %Val)
7189 declare ppc_fp128 @llvm.sin.ppcf128(ppc_fp128 %Val)
7194 The '``llvm.sin.*``' intrinsics return the sine of the operand.
7199 The argument and return value are floating point numbers of the same
7205 This function returns the sine of the specified operand, returning the
7206 same values as the libm ``sin`` functions would, and handles error
7207 conditions in the same way.
7209 '``llvm.cos.*``' Intrinsic
7210 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7215 This is an overloaded intrinsic. You can use ``llvm.cos`` on any
7216 floating point or vector of floating point type. Not all targets support
7221 declare float @llvm.cos.f32(float %Val)
7222 declare double @llvm.cos.f64(double %Val)
7223 declare x86_fp80 @llvm.cos.f80(x86_fp80 %Val)
7224 declare fp128 @llvm.cos.f128(fp128 %Val)
7225 declare ppc_fp128 @llvm.cos.ppcf128(ppc_fp128 %Val)
7230 The '``llvm.cos.*``' intrinsics return the cosine of the operand.
7235 The argument and return value are floating point numbers of the same
7241 This function returns the cosine of the specified operand, returning the
7242 same values as the libm ``cos`` functions would, and handles error
7243 conditions in the same way.
7245 '``llvm.pow.*``' Intrinsic
7246 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7251 This is an overloaded intrinsic. You can use ``llvm.pow`` on any
7252 floating point or vector of floating point type. Not all targets support
7257 declare float @llvm.pow.f32(float %Val, float %Power)
7258 declare double @llvm.pow.f64(double %Val, double %Power)
7259 declare x86_fp80 @llvm.pow.f80(x86_fp80 %Val, x86_fp80 %Power)
7260 declare fp128 @llvm.pow.f128(fp128 %Val, fp128 %Power)
7261 declare ppc_fp128 @llvm.pow.ppcf128(ppc_fp128 %Val, ppc_fp128 Power)
7266 The '``llvm.pow.*``' intrinsics return the first operand raised to the
7267 specified (positive or negative) power.
7272 The second argument is a floating point power, and the first is a value
7273 to raise to that power.
7278 This function returns the first value raised to the second power,
7279 returning the same values as the libm ``pow`` functions would, and
7280 handles error conditions in the same way.
7282 '``llvm.exp.*``' Intrinsic
7283 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7288 This is an overloaded intrinsic. You can use ``llvm.exp`` on any
7289 floating point or vector of floating point type. Not all targets support
7294 declare float @llvm.exp.f32(float %Val)
7295 declare double @llvm.exp.f64(double %Val)
7296 declare x86_fp80 @llvm.exp.f80(x86_fp80 %Val)
7297 declare fp128 @llvm.exp.f128(fp128 %Val)
7298 declare ppc_fp128 @llvm.exp.ppcf128(ppc_fp128 %Val)
7303 The '``llvm.exp.*``' intrinsics perform the exp function.
7308 The argument and return value are floating point numbers of the same
7314 This function returns the same values as the libm ``exp`` functions
7315 would, and handles error conditions in the same way.
7317 '``llvm.exp2.*``' Intrinsic
7318 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7323 This is an overloaded intrinsic. You can use ``llvm.exp2`` on any
7324 floating point or vector of floating point type. Not all targets support
7329 declare float @llvm.exp2.f32(float %Val)
7330 declare double @llvm.exp2.f64(double %Val)
7331 declare x86_fp80 @llvm.exp2.f80(x86_fp80 %Val)
7332 declare fp128 @llvm.exp2.f128(fp128 %Val)
7333 declare ppc_fp128 @llvm.exp2.ppcf128(ppc_fp128 %Val)
7338 The '``llvm.exp2.*``' intrinsics perform the exp2 function.
7343 The argument and return value are floating point numbers of the same
7349 This function returns the same values as the libm ``exp2`` functions
7350 would, and handles error conditions in the same way.
7352 '``llvm.log.*``' Intrinsic
7353 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7358 This is an overloaded intrinsic. You can use ``llvm.log`` on any
7359 floating point or vector of floating point type. Not all targets support
7364 declare float @llvm.log.f32(float %Val)
7365 declare double @llvm.log.f64(double %Val)
7366 declare x86_fp80 @llvm.log.f80(x86_fp80 %Val)
7367 declare fp128 @llvm.log.f128(fp128 %Val)
7368 declare ppc_fp128 @llvm.log.ppcf128(ppc_fp128 %Val)
7373 The '``llvm.log.*``' intrinsics perform the log function.
7378 The argument and return value are floating point numbers of the same
7384 This function returns the same values as the libm ``log`` functions
7385 would, and handles error conditions in the same way.
7387 '``llvm.log10.*``' Intrinsic
7388 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7393 This is an overloaded intrinsic. You can use ``llvm.log10`` on any
7394 floating point or vector of floating point type. Not all targets support
7399 declare float @llvm.log10.f32(float %Val)
7400 declare double @llvm.log10.f64(double %Val)
7401 declare x86_fp80 @llvm.log10.f80(x86_fp80 %Val)
7402 declare fp128 @llvm.log10.f128(fp128 %Val)
7403 declare ppc_fp128 @llvm.log10.ppcf128(ppc_fp128 %Val)
7408 The '``llvm.log10.*``' intrinsics perform the log10 function.
7413 The argument and return value are floating point numbers of the same
7419 This function returns the same values as the libm ``log10`` functions
7420 would, and handles error conditions in the same way.
7422 '``llvm.log2.*``' Intrinsic
7423 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7428 This is an overloaded intrinsic. You can use ``llvm.log2`` on any
7429 floating point or vector of floating point type. Not all targets support
7434 declare float @llvm.log2.f32(float %Val)
7435 declare double @llvm.log2.f64(double %Val)
7436 declare x86_fp80 @llvm.log2.f80(x86_fp80 %Val)
7437 declare fp128 @llvm.log2.f128(fp128 %Val)
7438 declare ppc_fp128 @llvm.log2.ppcf128(ppc_fp128 %Val)
7443 The '``llvm.log2.*``' intrinsics perform the log2 function.
7448 The argument and return value are floating point numbers of the same
7454 This function returns the same values as the libm ``log2`` functions
7455 would, and handles error conditions in the same way.
7457 '``llvm.fma.*``' Intrinsic
7458 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7463 This is an overloaded intrinsic. You can use ``llvm.fma`` on any
7464 floating point or vector of floating point type. Not all targets support
7469 declare float @llvm.fma.f32(float %a, float %b, float %c)
7470 declare double @llvm.fma.f64(double %a, double %b, double %c)
7471 declare x86_fp80 @llvm.fma.f80(x86_fp80 %a, x86_fp80 %b, x86_fp80 %c)
7472 declare fp128 @llvm.fma.f128(fp128 %a, fp128 %b, fp128 %c)
7473 declare ppc_fp128 @llvm.fma.ppcf128(ppc_fp128 %a, ppc_fp128 %b, ppc_fp128 %c)
7478 The '``llvm.fma.*``' intrinsics perform the fused multiply-add
7484 The argument and return value are floating point numbers of the same
7490 This function returns the same values as the libm ``fma`` functions
7493 '``llvm.fabs.*``' Intrinsic
7494 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7499 This is an overloaded intrinsic. You can use ``llvm.fabs`` on any
7500 floating point or vector of floating point type. Not all targets support
7505 declare float @llvm.fabs.f32(float %Val)
7506 declare double @llvm.fabs.f64(double %Val)
7507 declare x86_fp80 @llvm.fabs.f80(x86_fp80 %Val)
7508 declare fp128 @llvm.fabs.f128(fp128 %Val)
7509 declare ppc_fp128 @llvm.fabs.ppcf128(ppc_fp128 %Val)
7514 The '``llvm.fabs.*``' intrinsics return the absolute value of the
7520 The argument and return value are floating point numbers of the same
7526 This function returns the same values as the libm ``fabs`` functions
7527 would, and handles error conditions in the same way.
7529 '``llvm.copysign.*``' Intrinsic
7530 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7535 This is an overloaded intrinsic. You can use ``llvm.copysign`` on any
7536 floating point or vector of floating point type. Not all targets support
7541 declare float @llvm.copysign.f32(float %Mag, float %Sgn)
7542 declare double @llvm.copysign.f64(double %Mag, double %Sgn)
7543 declare x86_fp80 @llvm.copysign.f80(x86_fp80 %Mag, x86_fp80 %Sgn)
7544 declare fp128 @llvm.copysign.f128(fp128 %Mag, fp128 %Sgn)
7545 declare ppc_fp128 @llvm.copysign.ppcf128(ppc_fp128 %Mag, ppc_fp128 %Sgn)
7550 The '``llvm.copysign.*``' intrinsics return a value with the magnitude of the
7551 first operand and the sign of the second operand.
7556 The arguments and return value are floating point numbers of the same
7562 This function returns the same values as the libm ``copysign``
7563 functions would, and handles error conditions in the same way.
7565 '``llvm.floor.*``' Intrinsic
7566 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7571 This is an overloaded intrinsic. You can use ``llvm.floor`` on any
7572 floating point or vector of floating point type. Not all targets support
7577 declare float @llvm.floor.f32(float %Val)
7578 declare double @llvm.floor.f64(double %Val)
7579 declare x86_fp80 @llvm.floor.f80(x86_fp80 %Val)
7580 declare fp128 @llvm.floor.f128(fp128 %Val)
7581 declare ppc_fp128 @llvm.floor.ppcf128(ppc_fp128 %Val)
7586 The '``llvm.floor.*``' intrinsics return the floor of the operand.
7591 The argument and return value are floating point numbers of the same
7597 This function returns the same values as the libm ``floor`` functions
7598 would, and handles error conditions in the same way.
7600 '``llvm.ceil.*``' Intrinsic
7601 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7606 This is an overloaded intrinsic. You can use ``llvm.ceil`` on any
7607 floating point or vector of floating point type. Not all targets support
7612 declare float @llvm.ceil.f32(float %Val)
7613 declare double @llvm.ceil.f64(double %Val)
7614 declare x86_fp80 @llvm.ceil.f80(x86_fp80 %Val)
7615 declare fp128 @llvm.ceil.f128(fp128 %Val)
7616 declare ppc_fp128 @llvm.ceil.ppcf128(ppc_fp128 %Val)
7621 The '``llvm.ceil.*``' intrinsics return the ceiling of the operand.
7626 The argument and return value are floating point numbers of the same
7632 This function returns the same values as the libm ``ceil`` functions
7633 would, and handles error conditions in the same way.
7635 '``llvm.trunc.*``' Intrinsic
7636 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7641 This is an overloaded intrinsic. You can use ``llvm.trunc`` on any
7642 floating point or vector of floating point type. Not all targets support
7647 declare float @llvm.trunc.f32(float %Val)
7648 declare double @llvm.trunc.f64(double %Val)
7649 declare x86_fp80 @llvm.trunc.f80(x86_fp80 %Val)
7650 declare fp128 @llvm.trunc.f128(fp128 %Val)
7651 declare ppc_fp128 @llvm.trunc.ppcf128(ppc_fp128 %Val)
7656 The '``llvm.trunc.*``' intrinsics returns the operand rounded to the
7657 nearest integer not larger in magnitude than the operand.
7662 The argument and return value are floating point numbers of the same
7668 This function returns the same values as the libm ``trunc`` functions
7669 would, and handles error conditions in the same way.
7671 '``llvm.rint.*``' Intrinsic
7672 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7677 This is an overloaded intrinsic. You can use ``llvm.rint`` on any
7678 floating point or vector of floating point type. Not all targets support
7683 declare float @llvm.rint.f32(float %Val)
7684 declare double @llvm.rint.f64(double %Val)
7685 declare x86_fp80 @llvm.rint.f80(x86_fp80 %Val)
7686 declare fp128 @llvm.rint.f128(fp128 %Val)
7687 declare ppc_fp128 @llvm.rint.ppcf128(ppc_fp128 %Val)
7692 The '``llvm.rint.*``' intrinsics returns the operand rounded to the
7693 nearest integer. It may raise an inexact floating-point exception if the
7694 operand isn't an integer.
7699 The argument and return value are floating point numbers of the same
7705 This function returns the same values as the libm ``rint`` functions
7706 would, and handles error conditions in the same way.
7708 '``llvm.nearbyint.*``' Intrinsic
7709 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7714 This is an overloaded intrinsic. You can use ``llvm.nearbyint`` on any
7715 floating point or vector of floating point type. Not all targets support
7720 declare float @llvm.nearbyint.f32(float %Val)
7721 declare double @llvm.nearbyint.f64(double %Val)
7722 declare x86_fp80 @llvm.nearbyint.f80(x86_fp80 %Val)
7723 declare fp128 @llvm.nearbyint.f128(fp128 %Val)
7724 declare ppc_fp128 @llvm.nearbyint.ppcf128(ppc_fp128 %Val)
7729 The '``llvm.nearbyint.*``' intrinsics returns the operand rounded to the
7735 The argument and return value are floating point numbers of the same
7741 This function returns the same values as the libm ``nearbyint``
7742 functions would, and handles error conditions in the same way.
7744 '``llvm.round.*``' Intrinsic
7745 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7750 This is an overloaded intrinsic. You can use ``llvm.round`` on any
7751 floating point or vector of floating point type. Not all targets support
7756 declare float @llvm.round.f32(float %Val)
7757 declare double @llvm.round.f64(double %Val)
7758 declare x86_fp80 @llvm.round.f80(x86_fp80 %Val)
7759 declare fp128 @llvm.round.f128(fp128 %Val)
7760 declare ppc_fp128 @llvm.round.ppcf128(ppc_fp128 %Val)
7765 The '``llvm.round.*``' intrinsics returns the operand rounded to the
7771 The argument and return value are floating point numbers of the same
7777 This function returns the same values as the libm ``round``
7778 functions would, and handles error conditions in the same way.
7780 Bit Manipulation Intrinsics
7781 ---------------------------
7783 LLVM provides intrinsics for a few important bit manipulation
7784 operations. These allow efficient code generation for some algorithms.
7786 '``llvm.bswap.*``' Intrinsics
7787 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7792 This is an overloaded intrinsic function. You can use bswap on any
7793 integer type that is an even number of bytes (i.e. BitWidth % 16 == 0).
7797 declare i16 @llvm.bswap.i16(i16 <id>)
7798 declare i32 @llvm.bswap.i32(i32 <id>)
7799 declare i64 @llvm.bswap.i64(i64 <id>)
7804 The '``llvm.bswap``' family of intrinsics is used to byte swap integer
7805 values with an even number of bytes (positive multiple of 16 bits).
7806 These are useful for performing operations on data that is not in the
7807 target's native byte order.
7812 The ``llvm.bswap.i16`` intrinsic returns an i16 value that has the high
7813 and low byte of the input i16 swapped. Similarly, the ``llvm.bswap.i32``
7814 intrinsic returns an i32 value that has the four bytes of the input i32
7815 swapped, so that if the input bytes are numbered 0, 1, 2, 3 then the
7816 returned i32 will have its bytes in 3, 2, 1, 0 order. The
7817 ``llvm.bswap.i48``, ``llvm.bswap.i64`` and other intrinsics extend this
7818 concept to additional even-byte lengths (6 bytes, 8 bytes and more,
7821 '``llvm.ctpop.*``' Intrinsic
7822 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7827 This is an overloaded intrinsic. You can use llvm.ctpop on any integer
7828 bit width, or on any vector with integer elements. Not all targets
7829 support all bit widths or vector types, however.
7833 declare i8 @llvm.ctpop.i8(i8 <src>)
7834 declare i16 @llvm.ctpop.i16(i16 <src>)
7835 declare i32 @llvm.ctpop.i32(i32 <src>)
7836 declare i64 @llvm.ctpop.i64(i64 <src>)
7837 declare i256 @llvm.ctpop.i256(i256 <src>)
7838 declare <2 x i32> @llvm.ctpop.v2i32(<2 x i32> <src>)
7843 The '``llvm.ctpop``' family of intrinsics counts the number of bits set
7849 The only argument is the value to be counted. The argument may be of any
7850 integer type, or a vector with integer elements. The return type must
7851 match the argument type.
7856 The '``llvm.ctpop``' intrinsic counts the 1's in a variable, or within
7857 each element of a vector.
7859 '``llvm.ctlz.*``' Intrinsic
7860 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7865 This is an overloaded intrinsic. You can use ``llvm.ctlz`` on any
7866 integer bit width, or any vector whose elements are integers. Not all
7867 targets support all bit widths or vector types, however.
7871 declare i8 @llvm.ctlz.i8 (i8 <src>, i1 <is_zero_undef>)
7872 declare i16 @llvm.ctlz.i16 (i16 <src>, i1 <is_zero_undef>)
7873 declare i32 @llvm.ctlz.i32 (i32 <src>, i1 <is_zero_undef>)
7874 declare i64 @llvm.ctlz.i64 (i64 <src>, i1 <is_zero_undef>)
7875 declare i256 @llvm.ctlz.i256(i256 <src>, i1 <is_zero_undef>)
7876 declase <2 x i32> @llvm.ctlz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
7881 The '``llvm.ctlz``' family of intrinsic functions counts the number of
7882 leading zeros in a variable.
7887 The first argument is the value to be counted. This argument may be of
7888 any integer type, or a vectory with integer element type. The return
7889 type must match the first argument type.
7891 The second argument must be a constant and is a flag to indicate whether
7892 the intrinsic should ensure that a zero as the first argument produces a
7893 defined result. Historically some architectures did not provide a
7894 defined result for zero values as efficiently, and many algorithms are
7895 now predicated on avoiding zero-value inputs.
7900 The '``llvm.ctlz``' intrinsic counts the leading (most significant)
7901 zeros in a variable, or within each element of the vector. If
7902 ``src == 0`` then the result is the size in bits of the type of ``src``
7903 if ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
7904 ``llvm.ctlz(i32 2) = 30``.
7906 '``llvm.cttz.*``' Intrinsic
7907 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7912 This is an overloaded intrinsic. You can use ``llvm.cttz`` on any
7913 integer bit width, or any vector of integer elements. Not all targets
7914 support all bit widths or vector types, however.
7918 declare i8 @llvm.cttz.i8 (i8 <src>, i1 <is_zero_undef>)
7919 declare i16 @llvm.cttz.i16 (i16 <src>, i1 <is_zero_undef>)
7920 declare i32 @llvm.cttz.i32 (i32 <src>, i1 <is_zero_undef>)
7921 declare i64 @llvm.cttz.i64 (i64 <src>, i1 <is_zero_undef>)
7922 declare i256 @llvm.cttz.i256(i256 <src>, i1 <is_zero_undef>)
7923 declase <2 x i32> @llvm.cttz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
7928 The '``llvm.cttz``' family of intrinsic functions counts the number of
7934 The first argument is the value to be counted. This argument may be of
7935 any integer type, or a vectory with integer element type. The return
7936 type must match the first argument type.
7938 The second argument must be a constant and is a flag to indicate whether
7939 the intrinsic should ensure that a zero as the first argument produces a
7940 defined result. Historically some architectures did not provide a
7941 defined result for zero values as efficiently, and many algorithms are
7942 now predicated on avoiding zero-value inputs.
7947 The '``llvm.cttz``' intrinsic counts the trailing (least significant)
7948 zeros in a variable, or within each element of a vector. If ``src == 0``
7949 then the result is the size in bits of the type of ``src`` if
7950 ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
7951 ``llvm.cttz(2) = 1``.
7953 Arithmetic with Overflow Intrinsics
7954 -----------------------------------
7956 LLVM provides intrinsics for some arithmetic with overflow operations.
7958 '``llvm.sadd.with.overflow.*``' Intrinsics
7959 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7964 This is an overloaded intrinsic. You can use ``llvm.sadd.with.overflow``
7965 on any integer bit width.
7969 declare {i16, i1} @llvm.sadd.with.overflow.i16(i16 %a, i16 %b)
7970 declare {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
7971 declare {i64, i1} @llvm.sadd.with.overflow.i64(i64 %a, i64 %b)
7976 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
7977 a signed addition of the two arguments, and indicate whether an overflow
7978 occurred during the signed summation.
7983 The arguments (%a and %b) and the first element of the result structure
7984 may be of integer types of any bit width, but they must have the same
7985 bit width. The second element of the result structure must be of type
7986 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
7992 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
7993 a signed addition of the two variables. They return a structure --- the
7994 first element of which is the signed summation, and the second element
7995 of which is a bit specifying if the signed summation resulted in an
8001 .. code-block:: llvm
8003 %res = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
8004 %sum = extractvalue {i32, i1} %res, 0
8005 %obit = extractvalue {i32, i1} %res, 1
8006 br i1 %obit, label %overflow, label %normal
8008 '``llvm.uadd.with.overflow.*``' Intrinsics
8009 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8014 This is an overloaded intrinsic. You can use ``llvm.uadd.with.overflow``
8015 on any integer bit width.
8019 declare {i16, i1} @llvm.uadd.with.overflow.i16(i16 %a, i16 %b)
8020 declare {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
8021 declare {i64, i1} @llvm.uadd.with.overflow.i64(i64 %a, i64 %b)
8026 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
8027 an unsigned addition of the two arguments, and indicate whether a carry
8028 occurred during the unsigned summation.
8033 The arguments (%a and %b) and the first element of the result structure
8034 may be of integer types of any bit width, but they must have the same
8035 bit width. The second element of the result structure must be of type
8036 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8042 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
8043 an unsigned addition of the two arguments. They return a structure --- the
8044 first element of which is the sum, and the second element of which is a
8045 bit specifying if the unsigned summation resulted in a carry.
8050 .. code-block:: llvm
8052 %res = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
8053 %sum = extractvalue {i32, i1} %res, 0
8054 %obit = extractvalue {i32, i1} %res, 1
8055 br i1 %obit, label %carry, label %normal
8057 '``llvm.ssub.with.overflow.*``' Intrinsics
8058 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8063 This is an overloaded intrinsic. You can use ``llvm.ssub.with.overflow``
8064 on any integer bit width.
8068 declare {i16, i1} @llvm.ssub.with.overflow.i16(i16 %a, i16 %b)
8069 declare {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
8070 declare {i64, i1} @llvm.ssub.with.overflow.i64(i64 %a, i64 %b)
8075 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
8076 a signed subtraction of the two arguments, and indicate whether an
8077 overflow occurred during the signed subtraction.
8082 The arguments (%a and %b) and the first element of the result structure
8083 may be of integer types of any bit width, but they must have the same
8084 bit width. The second element of the result structure must be of type
8085 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
8091 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
8092 a signed subtraction of the two arguments. They return a structure --- the
8093 first element of which is the subtraction, and the second element of
8094 which is a bit specifying if the signed subtraction resulted in an
8100 .. code-block:: llvm
8102 %res = call {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
8103 %sum = extractvalue {i32, i1} %res, 0
8104 %obit = extractvalue {i32, i1} %res, 1
8105 br i1 %obit, label %overflow, label %normal
8107 '``llvm.usub.with.overflow.*``' Intrinsics
8108 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8113 This is an overloaded intrinsic. You can use ``llvm.usub.with.overflow``
8114 on any integer bit width.
8118 declare {i16, i1} @llvm.usub.with.overflow.i16(i16 %a, i16 %b)
8119 declare {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
8120 declare {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
8125 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
8126 an unsigned subtraction of the two arguments, and indicate whether an
8127 overflow occurred during the unsigned subtraction.
8132 The arguments (%a and %b) and the first element of the result structure
8133 may be of integer types of any bit width, but they must have the same
8134 bit width. The second element of the result structure must be of type
8135 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8141 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
8142 an unsigned subtraction of the two arguments. They return a structure ---
8143 the first element of which is the subtraction, and the second element of
8144 which is a bit specifying if the unsigned subtraction resulted in an
8150 .. code-block:: llvm
8152 %res = call {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
8153 %sum = extractvalue {i32, i1} %res, 0
8154 %obit = extractvalue {i32, i1} %res, 1
8155 br i1 %obit, label %overflow, label %normal
8157 '``llvm.smul.with.overflow.*``' Intrinsics
8158 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8163 This is an overloaded intrinsic. You can use ``llvm.smul.with.overflow``
8164 on any integer bit width.
8168 declare {i16, i1} @llvm.smul.with.overflow.i16(i16 %a, i16 %b)
8169 declare {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
8170 declare {i64, i1} @llvm.smul.with.overflow.i64(i64 %a, i64 %b)
8175 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
8176 a signed multiplication of the two arguments, and indicate whether an
8177 overflow occurred during the signed multiplication.
8182 The arguments (%a and %b) and the first element of the result structure
8183 may be of integer types of any bit width, but they must have the same
8184 bit width. The second element of the result structure must be of type
8185 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
8191 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
8192 a signed multiplication of the two arguments. They return a structure ---
8193 the first element of which is the multiplication, and the second element
8194 of which is a bit specifying if the signed multiplication resulted in an
8200 .. code-block:: llvm
8202 %res = call {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
8203 %sum = extractvalue {i32, i1} %res, 0
8204 %obit = extractvalue {i32, i1} %res, 1
8205 br i1 %obit, label %overflow, label %normal
8207 '``llvm.umul.with.overflow.*``' Intrinsics
8208 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8213 This is an overloaded intrinsic. You can use ``llvm.umul.with.overflow``
8214 on any integer bit width.
8218 declare {i16, i1} @llvm.umul.with.overflow.i16(i16 %a, i16 %b)
8219 declare {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
8220 declare {i64, i1} @llvm.umul.with.overflow.i64(i64 %a, i64 %b)
8225 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
8226 a unsigned multiplication of the two arguments, and indicate whether an
8227 overflow occurred during the unsigned multiplication.
8232 The arguments (%a and %b) and the first element of the result structure
8233 may be of integer types of any bit width, but they must have the same
8234 bit width. The second element of the result structure must be of type
8235 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8241 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
8242 an unsigned multiplication of the two arguments. They return a structure ---
8243 the first element of which is the multiplication, and the second
8244 element of which is a bit specifying if the unsigned multiplication
8245 resulted in an overflow.
8250 .. code-block:: llvm
8252 %res = call {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
8253 %sum = extractvalue {i32, i1} %res, 0
8254 %obit = extractvalue {i32, i1} %res, 1
8255 br i1 %obit, label %overflow, label %normal
8257 Specialised Arithmetic Intrinsics
8258 ---------------------------------
8260 '``llvm.fmuladd.*``' Intrinsic
8261 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8268 declare float @llvm.fmuladd.f32(float %a, float %b, float %c)
8269 declare double @llvm.fmuladd.f64(double %a, double %b, double %c)
8274 The '``llvm.fmuladd.*``' intrinsic functions represent multiply-add
8275 expressions that can be fused if the code generator determines that (a) the
8276 target instruction set has support for a fused operation, and (b) that the
8277 fused operation is more efficient than the equivalent, separate pair of mul
8278 and add instructions.
8283 The '``llvm.fmuladd.*``' intrinsics each take three arguments: two
8284 multiplicands, a and b, and an addend c.
8293 %0 = call float @llvm.fmuladd.f32(%a, %b, %c)
8295 is equivalent to the expression a \* b + c, except that rounding will
8296 not be performed between the multiplication and addition steps if the
8297 code generator fuses the operations. Fusion is not guaranteed, even if
8298 the target platform supports it. If a fused multiply-add is required the
8299 corresponding llvm.fma.\* intrinsic function should be used instead.
8304 .. code-block:: llvm
8306 %r2 = call float @llvm.fmuladd.f32(float %a, float %b, float %c) ; yields {float}:r2 = (a * b) + c
8308 Half Precision Floating Point Intrinsics
8309 ----------------------------------------
8311 For most target platforms, half precision floating point is a
8312 storage-only format. This means that it is a dense encoding (in memory)
8313 but does not support computation in the format.
8315 This means that code must first load the half-precision floating point
8316 value as an i16, then convert it to float with
8317 :ref:`llvm.convert.from.fp16 <int_convert_from_fp16>`. Computation can
8318 then be performed on the float value (including extending to double
8319 etc). To store the value back to memory, it is first converted to float
8320 if needed, then converted to i16 with
8321 :ref:`llvm.convert.to.fp16 <int_convert_to_fp16>`, then storing as an
8324 .. _int_convert_to_fp16:
8326 '``llvm.convert.to.fp16``' Intrinsic
8327 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8334 declare i16 @llvm.convert.to.fp16(f32 %a)
8339 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion
8340 from single precision floating point format to half precision floating
8346 The intrinsic function contains single argument - the value to be
8352 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion
8353 from single precision floating point format to half precision floating
8354 point format. The return value is an ``i16`` which contains the
8360 .. code-block:: llvm
8362 %res = call i16 @llvm.convert.to.fp16(f32 %a)
8363 store i16 %res, i16* @x, align 2
8365 .. _int_convert_from_fp16:
8367 '``llvm.convert.from.fp16``' Intrinsic
8368 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8375 declare f32 @llvm.convert.from.fp16(i16 %a)
8380 The '``llvm.convert.from.fp16``' intrinsic function performs a
8381 conversion from half precision floating point format to single precision
8382 floating point format.
8387 The intrinsic function contains single argument - the value to be
8393 The '``llvm.convert.from.fp16``' intrinsic function performs a
8394 conversion from half single precision floating point format to single
8395 precision floating point format. The input half-float value is
8396 represented by an ``i16`` value.
8401 .. code-block:: llvm
8403 %a = load i16* @x, align 2
8404 %res = call f32 @llvm.convert.from.fp16(i16 %a)
8409 The LLVM debugger intrinsics (which all start with ``llvm.dbg.``
8410 prefix), are described in the `LLVM Source Level
8411 Debugging <SourceLevelDebugging.html#format_common_intrinsics>`_
8414 Exception Handling Intrinsics
8415 -----------------------------
8417 The LLVM exception handling intrinsics (which all start with
8418 ``llvm.eh.`` prefix), are described in the `LLVM Exception
8419 Handling <ExceptionHandling.html#format_common_intrinsics>`_ document.
8423 Trampoline Intrinsics
8424 ---------------------
8426 These intrinsics make it possible to excise one parameter, marked with
8427 the :ref:`nest <nest>` attribute, from a function. The result is a
8428 callable function pointer lacking the nest parameter - the caller does
8429 not need to provide a value for it. Instead, the value to use is stored
8430 in advance in a "trampoline", a block of memory usually allocated on the
8431 stack, which also contains code to splice the nest value into the
8432 argument list. This is used to implement the GCC nested function address
8435 For example, if the function is ``i32 f(i8* nest %c, i32 %x, i32 %y)``
8436 then the resulting function pointer has signature ``i32 (i32, i32)*``.
8437 It can be created as follows:
8439 .. code-block:: llvm
8441 %tramp = alloca [10 x i8], align 4 ; size and alignment only correct for X86
8442 %tramp1 = getelementptr [10 x i8]* %tramp, i32 0, i32 0
8443 call i8* @llvm.init.trampoline(i8* %tramp1, i8* bitcast (i32 (i8*, i32, i32)* @f to i8*), i8* %nval)
8444 %p = call i8* @llvm.adjust.trampoline(i8* %tramp1)
8445 %fp = bitcast i8* %p to i32 (i32, i32)*
8447 The call ``%val = call i32 %fp(i32 %x, i32 %y)`` is then equivalent to
8448 ``%val = call i32 %f(i8* %nval, i32 %x, i32 %y)``.
8452 '``llvm.init.trampoline``' Intrinsic
8453 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8460 declare void @llvm.init.trampoline(i8* <tramp>, i8* <func>, i8* <nval>)
8465 This fills the memory pointed to by ``tramp`` with executable code,
8466 turning it into a trampoline.
8471 The ``llvm.init.trampoline`` intrinsic takes three arguments, all
8472 pointers. The ``tramp`` argument must point to a sufficiently large and
8473 sufficiently aligned block of memory; this memory is written to by the
8474 intrinsic. Note that the size and the alignment are target-specific -
8475 LLVM currently provides no portable way of determining them, so a
8476 front-end that generates this intrinsic needs to have some
8477 target-specific knowledge. The ``func`` argument must hold a function
8478 bitcast to an ``i8*``.
8483 The block of memory pointed to by ``tramp`` is filled with target
8484 dependent code, turning it into a function. Then ``tramp`` needs to be
8485 passed to :ref:`llvm.adjust.trampoline <int_at>` to get a pointer which can
8486 be :ref:`bitcast (to a new function) and called <int_trampoline>`. The new
8487 function's signature is the same as that of ``func`` with any arguments
8488 marked with the ``nest`` attribute removed. At most one such ``nest``
8489 argument is allowed, and it must be of pointer type. Calling the new
8490 function is equivalent to calling ``func`` with the same argument list,
8491 but with ``nval`` used for the missing ``nest`` argument. If, after
8492 calling ``llvm.init.trampoline``, the memory pointed to by ``tramp`` is
8493 modified, then the effect of any later call to the returned function
8494 pointer is undefined.
8498 '``llvm.adjust.trampoline``' Intrinsic
8499 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8506 declare i8* @llvm.adjust.trampoline(i8* <tramp>)
8511 This performs any required machine-specific adjustment to the address of
8512 a trampoline (passed as ``tramp``).
8517 ``tramp`` must point to a block of memory which already has trampoline
8518 code filled in by a previous call to
8519 :ref:`llvm.init.trampoline <int_it>`.
8524 On some architectures the address of the code to be executed needs to be
8525 different to the address where the trampoline is actually stored. This
8526 intrinsic returns the executable address corresponding to ``tramp``
8527 after performing the required machine specific adjustments. The pointer
8528 returned can then be :ref:`bitcast and executed <int_trampoline>`.
8533 This class of intrinsics exists to information about the lifetime of
8534 memory objects and ranges where variables are immutable.
8538 '``llvm.lifetime.start``' Intrinsic
8539 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8546 declare void @llvm.lifetime.start(i64 <size>, i8* nocapture <ptr>)
8551 The '``llvm.lifetime.start``' intrinsic specifies the start of a memory
8557 The first argument is a constant integer representing the size of the
8558 object, or -1 if it is variable sized. The second argument is a pointer
8564 This intrinsic indicates that before this point in the code, the value
8565 of the memory pointed to by ``ptr`` is dead. This means that it is known
8566 to never be used and has an undefined value. A load from the pointer
8567 that precedes this intrinsic can be replaced with ``'undef'``.
8571 '``llvm.lifetime.end``' Intrinsic
8572 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8579 declare void @llvm.lifetime.end(i64 <size>, i8* nocapture <ptr>)
8584 The '``llvm.lifetime.end``' intrinsic specifies the end of a memory
8590 The first argument is a constant integer representing the size of the
8591 object, or -1 if it is variable sized. The second argument is a pointer
8597 This intrinsic indicates that after this point in the code, the value of
8598 the memory pointed to by ``ptr`` is dead. This means that it is known to
8599 never be used and has an undefined value. Any stores into the memory
8600 object following this intrinsic may be removed as dead.
8602 '``llvm.invariant.start``' Intrinsic
8603 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8610 declare {}* @llvm.invariant.start(i64 <size>, i8* nocapture <ptr>)
8615 The '``llvm.invariant.start``' intrinsic specifies that the contents of
8616 a memory object will not change.
8621 The first argument is a constant integer representing the size of the
8622 object, or -1 if it is variable sized. The second argument is a pointer
8628 This intrinsic indicates that until an ``llvm.invariant.end`` that uses
8629 the return value, the referenced memory location is constant and
8632 '``llvm.invariant.end``' Intrinsic
8633 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8640 declare void @llvm.invariant.end({}* <start>, i64 <size>, i8* nocapture <ptr>)
8645 The '``llvm.invariant.end``' intrinsic specifies that the contents of a
8646 memory object are mutable.
8651 The first argument is the matching ``llvm.invariant.start`` intrinsic.
8652 The second argument is a constant integer representing the size of the
8653 object, or -1 if it is variable sized and the third argument is a
8654 pointer to the object.
8659 This intrinsic indicates that the memory is mutable again.
8664 This class of intrinsics is designed to be generic and has no specific
8667 '``llvm.var.annotation``' Intrinsic
8668 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8675 declare void @llvm.var.annotation(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
8680 The '``llvm.var.annotation``' intrinsic.
8685 The first argument is a pointer to a value, the second is a pointer to a
8686 global string, the third is a pointer to a global string which is the
8687 source file name, and the last argument is the line number.
8692 This intrinsic allows annotation of local variables with arbitrary
8693 strings. This can be useful for special purpose optimizations that want
8694 to look for these annotations. These have no other defined use; they are
8695 ignored by code generation and optimization.
8697 '``llvm.ptr.annotation.*``' Intrinsic
8698 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8703 This is an overloaded intrinsic. You can use '``llvm.ptr.annotation``' on a
8704 pointer to an integer of any width. *NOTE* you must specify an address space for
8705 the pointer. The identifier for the default address space is the integer
8710 declare i8* @llvm.ptr.annotation.p<address space>i8(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
8711 declare i16* @llvm.ptr.annotation.p<address space>i16(i16* <val>, i8* <str>, i8* <str>, i32 <int>)
8712 declare i32* @llvm.ptr.annotation.p<address space>i32(i32* <val>, i8* <str>, i8* <str>, i32 <int>)
8713 declare i64* @llvm.ptr.annotation.p<address space>i64(i64* <val>, i8* <str>, i8* <str>, i32 <int>)
8714 declare i256* @llvm.ptr.annotation.p<address space>i256(i256* <val>, i8* <str>, i8* <str>, i32 <int>)
8719 The '``llvm.ptr.annotation``' intrinsic.
8724 The first argument is a pointer to an integer value of arbitrary bitwidth
8725 (result of some expression), the second is a pointer to a global string, the
8726 third is a pointer to a global string which is the source file name, and the
8727 last argument is the line number. It returns the value of the first argument.
8732 This intrinsic allows annotation of a pointer to an integer with arbitrary
8733 strings. This can be useful for special purpose optimizations that want to look
8734 for these annotations. These have no other defined use; they are ignored by code
8735 generation and optimization.
8737 '``llvm.annotation.*``' Intrinsic
8738 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8743 This is an overloaded intrinsic. You can use '``llvm.annotation``' on
8744 any integer bit width.
8748 declare i8 @llvm.annotation.i8(i8 <val>, i8* <str>, i8* <str>, i32 <int>)
8749 declare i16 @llvm.annotation.i16(i16 <val>, i8* <str>, i8* <str>, i32 <int>)
8750 declare i32 @llvm.annotation.i32(i32 <val>, i8* <str>, i8* <str>, i32 <int>)
8751 declare i64 @llvm.annotation.i64(i64 <val>, i8* <str>, i8* <str>, i32 <int>)
8752 declare i256 @llvm.annotation.i256(i256 <val>, i8* <str>, i8* <str>, i32 <int>)
8757 The '``llvm.annotation``' intrinsic.
8762 The first argument is an integer value (result of some expression), the
8763 second is a pointer to a global string, the third is a pointer to a
8764 global string which is the source file name, and the last argument is
8765 the line number. It returns the value of the first argument.
8770 This intrinsic allows annotations to be put on arbitrary expressions
8771 with arbitrary strings. This can be useful for special purpose
8772 optimizations that want to look for these annotations. These have no
8773 other defined use; they are ignored by code generation and optimization.
8775 '``llvm.trap``' Intrinsic
8776 ^^^^^^^^^^^^^^^^^^^^^^^^^
8783 declare void @llvm.trap() noreturn nounwind
8788 The '``llvm.trap``' intrinsic.
8798 This intrinsic is lowered to the target dependent trap instruction. If
8799 the target does not have a trap instruction, this intrinsic will be
8800 lowered to a call of the ``abort()`` function.
8802 '``llvm.debugtrap``' Intrinsic
8803 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8810 declare void @llvm.debugtrap() nounwind
8815 The '``llvm.debugtrap``' intrinsic.
8825 This intrinsic is lowered to code which is intended to cause an
8826 execution trap with the intention of requesting the attention of a
8829 '``llvm.stackprotector``' Intrinsic
8830 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8837 declare void @llvm.stackprotector(i8* <guard>, i8** <slot>)
8842 The ``llvm.stackprotector`` intrinsic takes the ``guard`` and stores it
8843 onto the stack at ``slot``. The stack slot is adjusted to ensure that it
8844 is placed on the stack before local variables.
8849 The ``llvm.stackprotector`` intrinsic requires two pointer arguments.
8850 The first argument is the value loaded from the stack guard
8851 ``@__stack_chk_guard``. The second variable is an ``alloca`` that has
8852 enough space to hold the value of the guard.
8857 This intrinsic causes the prologue/epilogue inserter to force the position of
8858 the ``AllocaInst`` stack slot to be before local variables on the stack. This is
8859 to ensure that if a local variable on the stack is overwritten, it will destroy
8860 the value of the guard. When the function exits, the guard on the stack is
8861 checked against the original guard by ``llvm.stackprotectorcheck``. If they are
8862 different, then ``llvm.stackprotectorcheck`` causes the program to abort by
8863 calling the ``__stack_chk_fail()`` function.
8865 '``llvm.stackprotectorcheck``' Intrinsic
8866 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8873 declare void @llvm.stackprotectorcheck(i8** <guard>)
8878 The ``llvm.stackprotectorcheck`` intrinsic compares ``guard`` against an already
8879 created stack protector and if they are not equal calls the
8880 ``__stack_chk_fail()`` function.
8885 The ``llvm.stackprotectorcheck`` intrinsic requires one pointer argument, the
8886 the variable ``@__stack_chk_guard``.
8891 This intrinsic is provided to perform the stack protector check by comparing
8892 ``guard`` with the stack slot created by ``llvm.stackprotector`` and if the
8893 values do not match call the ``__stack_chk_fail()`` function.
8895 The reason to provide this as an IR level intrinsic instead of implementing it
8896 via other IR operations is that in order to perform this operation at the IR
8897 level without an intrinsic, one would need to create additional basic blocks to
8898 handle the success/failure cases. This makes it difficult to stop the stack
8899 protector check from disrupting sibling tail calls in Codegen. With this
8900 intrinsic, we are able to generate the stack protector basic blocks late in
8901 codegen after the tail call decision has occurred.
8903 '``llvm.objectsize``' Intrinsic
8904 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8911 declare i32 @llvm.objectsize.i32(i8* <object>, i1 <min>)
8912 declare i64 @llvm.objectsize.i64(i8* <object>, i1 <min>)
8917 The ``llvm.objectsize`` intrinsic is designed to provide information to
8918 the optimizers to determine at compile time whether a) an operation
8919 (like memcpy) will overflow a buffer that corresponds to an object, or
8920 b) that a runtime check for overflow isn't necessary. An object in this
8921 context means an allocation of a specific class, structure, array, or
8927 The ``llvm.objectsize`` intrinsic takes two arguments. The first
8928 argument is a pointer to or into the ``object``. The second argument is
8929 a boolean and determines whether ``llvm.objectsize`` returns 0 (if true)
8930 or -1 (if false) when the object size is unknown. The second argument
8931 only accepts constants.
8936 The ``llvm.objectsize`` intrinsic is lowered to a constant representing
8937 the size of the object concerned. If the size cannot be determined at
8938 compile time, ``llvm.objectsize`` returns ``i32/i64 -1 or 0`` (depending
8939 on the ``min`` argument).
8941 '``llvm.expect``' Intrinsic
8942 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8949 declare i32 @llvm.expect.i32(i32 <val>, i32 <expected_val>)
8950 declare i64 @llvm.expect.i64(i64 <val>, i64 <expected_val>)
8955 The ``llvm.expect`` intrinsic provides information about expected (the
8956 most probable) value of ``val``, which can be used by optimizers.
8961 The ``llvm.expect`` intrinsic takes two arguments. The first argument is
8962 a value. The second argument is an expected value, this needs to be a
8963 constant value, variables are not allowed.
8968 This intrinsic is lowered to the ``val``.
8970 '``llvm.donothing``' Intrinsic
8971 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8978 declare void @llvm.donothing() nounwind readnone
8983 The ``llvm.donothing`` intrinsic doesn't perform any operation. It's the
8984 only intrinsic that can be called with an invoke instruction.
8994 This intrinsic does nothing, and it's removed by optimizers and ignored
8997 Stack Map Intrinsics
8998 --------------------
9000 LLVM provides experimental intrinsics to support runtime patching
9001 mechanisms commonly desired in dynamic language JITs. These intrinsics
9002 are described in :doc:`StackMaps`.