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
316 "``cc 10``" - GHC convention
317 This calling convention has been implemented specifically for use by
318 the `Glasgow Haskell Compiler (GHC) <http://www.haskell.org/ghc>`_.
319 It passes everything in registers, going to extremes to achieve this
320 by disabling callee save registers. This calling convention should
321 not be used lightly but only for specific situations such as an
322 alternative to the *register pinning* performance technique often
323 used when implementing functional programming languages. At the
324 moment only X86 supports this convention and it has the following
327 - On *X86-32* only supports up to 4 bit type parameters. No
328 floating point types are supported.
329 - On *X86-64* only supports up to 10 bit type parameters and 6
330 floating point parameters.
332 This calling convention supports `tail call
333 optimization <CodeGenerator.html#id80>`_ but requires both the
334 caller and callee are using it.
335 "``cc 11``" - The HiPE calling convention
336 This calling convention has been implemented specifically for use by
337 the `High-Performance Erlang
338 (HiPE) <http://www.it.uu.se/research/group/hipe/>`_ compiler, *the*
339 native code compiler of the `Ericsson's Open Source Erlang/OTP
340 system <http://www.erlang.org/download.shtml>`_. It uses more
341 registers for argument passing than the ordinary C calling
342 convention and defines no callee-saved registers. The calling
343 convention properly supports `tail call
344 optimization <CodeGenerator.html#id80>`_ but requires that both the
345 caller and the callee use it. It uses a *register pinning*
346 mechanism, similar to GHC's convention, for keeping frequently
347 accessed runtime components pinned to specific hardware registers.
348 At the moment only X86 supports this convention (both 32 and 64
350 "``webkit_jscc``" - WebKit's JavaScript calling convention
351 This calling convention has been implemented for `WebKit FTL JIT
352 <https://trac.webkit.org/wiki/FTLJIT>`_. It passes arguments on the
353 stack right to left (as cdecl does), and returns a value in the
354 platform's customary return register.
355 "``anyregcc``" - Dynamic calling convention for code patching
356 This is a special convention that supports patching an arbitrary code
357 sequence in place of a call site. This convention forces the call
358 arguments into registers but allows them to be dynamcially
359 allocated. This can currently only be used with calls to
360 llvm.experimental.patchpoint because only this intrinsic records
361 the location of its arguments in a side table. See :doc:`StackMaps`.
362 "``preserve_mostcc``" - The `PreserveMost` calling convention
363 This calling convention attempts to make the code in the caller as little
364 intrusive as possible. This calling convention behaves identical to the `C`
365 calling convention on how arguments and return values are passed, but it
366 uses a different set of caller/callee-saved registers. This alleviates the
367 burden of saving and recovering a large register set before and after the
370 - On X86-64 the callee preserves all general purpose registers, except for
371 R11. R11 can be used as a scratch register. Floating-point registers
372 (XMMs/YMMs) are not preserved and need to be saved by the caller.
374 The idea behind this convention is to support calls to runtime functions
375 that have a hot path and a cold path. The hot path is usually a small piece
376 of code that doesn't many registers. The cold path might need to call out to
377 another function and therefore only needs to preserve the caller-saved
378 registers, which haven't already been saved by the caller.
380 This calling convention will be used by a future version of the ObjectiveC
381 runtime and should therefore still be considered experimental at this time.
382 Although this convention was created to optimize certain runtime calls to
383 the ObjectiveC runtime, it is not limited to this runtime and might be used
384 by other runtimes in the future too. The current implementation only
385 supports X86-64, but the intention is to support more architectures in the
387 "``preserve_allcc``" - The `PreserveAll` calling convention
388 This calling convention attempts to make the code in the caller even less
389 intrusive than the `PreserveMost` calling convention. This calling
390 convention also behaves identical to the `C` calling convention on how
391 arguments and return values are passed, but it uses a different set of
392 caller/callee-saved registers. This removes the burden of saving and
393 recovering a large register set before and after the call in the caller.
395 - On X86-64 the callee preserves all general purpose registers, except for
396 R11. R11 can be used as a scratch register. Furthermore it also preserves
397 all floating-point registers (XMMs/YMMs).
399 The idea behind this convention is to support calls to runtime functions
400 that don't need to call out to any other functions.
402 This calling convention, like the `PreserveMost` calling convention, will be
403 used by a future version of the ObjectiveC runtime and should be considered
404 experimental at this time.
405 "``cc <n>``" - Numbered convention
406 Any calling convention may be specified by number, allowing
407 target-specific calling conventions to be used. Target specific
408 calling conventions start at 64.
410 More calling conventions can be added/defined on an as-needed basis, to
411 support Pascal conventions or any other well-known target-independent
414 .. _visibilitystyles:
419 All Global Variables and Functions have one of the following visibility
422 "``default``" - Default style
423 On targets that use the ELF object file format, default visibility
424 means that the declaration is visible to other modules and, in
425 shared libraries, means that the declared entity may be overridden.
426 On Darwin, default visibility means that the declaration is visible
427 to other modules. Default visibility corresponds to "external
428 linkage" in the language.
429 "``hidden``" - Hidden style
430 Two declarations of an object with hidden visibility refer to the
431 same object if they are in the same shared object. Usually, hidden
432 visibility indicates that the symbol will not be placed into the
433 dynamic symbol table, so no other module (executable or shared
434 library) can reference it directly.
435 "``protected``" - Protected style
436 On ELF, protected visibility indicates that the symbol will be
437 placed in the dynamic symbol table, but that references within the
438 defining module will bind to the local symbol. That is, the symbol
439 cannot be overridden by another module.
446 All Global Variables, Functions and Aliases can have one of the following
450 "``dllimport``" causes the compiler to reference a function or variable via
451 a global pointer to a pointer that is set up by the DLL exporting the
452 symbol. On Microsoft Windows targets, the pointer name is formed by
453 combining ``__imp_`` and the function or variable name.
455 "``dllexport``" causes the compiler to provide a global pointer to a pointer
456 in a DLL, so that it can be referenced with the ``dllimport`` attribute. On
457 Microsoft Windows targets, the pointer name is formed by combining
458 ``__imp_`` and the function or variable name. Since this storage class
459 exists for defining a dll interface, the compiler, assembler and linker know
460 it is externally referenced and must refrain from deleting the symbol.
465 LLVM IR allows you to specify name aliases for certain types. This can
466 make it easier to read the IR and make the IR more condensed
467 (particularly when recursive types are involved). An example of a name
472 %mytype = type { %mytype*, i32 }
474 You may give a name to any :ref:`type <typesystem>` except
475 ":ref:`void <t_void>`". Type name aliases may be used anywhere a type is
476 expected with the syntax "%mytype".
478 Note that type names are aliases for the structural type that they
479 indicate, and that you can therefore specify multiple names for the same
480 type. This often leads to confusing behavior when dumping out a .ll
481 file. Since LLVM IR uses structural typing, the name is not part of the
482 type. When printing out LLVM IR, the printer will pick *one name* to
483 render all types of a particular shape. This means that if you have code
484 where two different source types end up having the same LLVM type, that
485 the dumper will sometimes print the "wrong" or unexpected type. This is
486 an important design point and isn't going to change.
493 Global variables define regions of memory allocated at compilation time
496 Global variables definitions must be initialized, may have an explicit section
497 to be placed in, and may have an optional explicit alignment specified.
499 Global variables in other translation units can also be declared, in which
500 case they don't have an initializer.
502 A variable may be defined as ``thread_local``, which means that it will
503 not be shared by threads (each thread will have a separated copy of the
504 variable). Not all targets support thread-local variables. Optionally, a
505 TLS model may be specified:
508 For variables that are only used within the current shared library.
510 For variables in modules that will not be loaded dynamically.
512 For variables defined in the executable and only used within it.
514 The models correspond to the ELF TLS models; see `ELF Handling For
515 Thread-Local Storage <http://people.redhat.com/drepper/tls.pdf>`_ for
516 more information on under which circumstances the different models may
517 be used. The target may choose a different TLS model if the specified
518 model is not supported, or if a better choice of model can be made.
520 A variable may be defined as a global ``constant``, which indicates that
521 the contents of the variable will **never** be modified (enabling better
522 optimization, allowing the global data to be placed in the read-only
523 section of an executable, etc). Note that variables that need runtime
524 initialization cannot be marked ``constant`` as there is a store to the
527 LLVM explicitly allows *declarations* of global variables to be marked
528 constant, even if the final definition of the global is not. This
529 capability can be used to enable slightly better optimization of the
530 program, but requires the language definition to guarantee that
531 optimizations based on the 'constantness' are valid for the translation
532 units that do not include the definition.
534 As SSA values, global variables define pointer values that are in scope
535 (i.e. they dominate) all basic blocks in the program. Global variables
536 always define a pointer to their "content" type because they describe a
537 region of memory, and all memory objects in LLVM are accessed through
540 Global variables can be marked with ``unnamed_addr`` which indicates
541 that the address is not significant, only the content. Constants marked
542 like this can be merged with other constants if they have the same
543 initializer. Note that a constant with significant address *can* be
544 merged with a ``unnamed_addr`` constant, the result being a constant
545 whose address is significant.
547 A global variable may be declared to reside in a target-specific
548 numbered address space. For targets that support them, address spaces
549 may affect how optimizations are performed and/or what target
550 instructions are used to access the variable. The default address space
551 is zero. The address space qualifier must precede any other attributes.
553 LLVM allows an explicit section to be specified for globals. If the
554 target supports it, it will emit globals to the section specified.
556 By default, global initializers are optimized by assuming that global
557 variables defined within the module are not modified from their
558 initial values before the start of the global initializer. This is
559 true even for variables potentially accessible from outside the
560 module, including those with external linkage or appearing in
561 ``@llvm.used`` or dllexported variables. This assumption may be suppressed
562 by marking the variable with ``externally_initialized``.
564 An explicit alignment may be specified for a global, which must be a
565 power of 2. If not present, or if the alignment is set to zero, the
566 alignment of the global is set by the target to whatever it feels
567 convenient. If an explicit alignment is specified, the global is forced
568 to have exactly that alignment. Targets and optimizers are not allowed
569 to over-align the global if the global has an assigned section. In this
570 case, the extra alignment could be observable: for example, code could
571 assume that the globals are densely packed in their section and try to
572 iterate over them as an array, alignment padding would break this
575 Globals can also have a :ref:`DLL storage class <dllstorageclass>`.
579 [@<GlobalVarName> =] [Linkage] [Visibility] [DLLStorageClass] [ThreadLocal]
580 [AddrSpace] [unnamed_addr] [ExternallyInitialized]
581 <global | constant> <Type>
582 [, section "name"] [, align <Alignment>]
584 For example, the following defines a global in a numbered address space
585 with an initializer, section, and alignment:
589 @G = addrspace(5) constant float 1.0, section "foo", align 4
591 The following example just declares a global variable
595 @G = external global i32
597 The following example defines a thread-local global with the
598 ``initialexec`` TLS model:
602 @G = thread_local(initialexec) global i32 0, align 4
604 .. _functionstructure:
609 LLVM function definitions consist of the "``define``" keyword, an
610 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
611 style <visibility>`, an optional :ref:`DLL storage class <dllstorageclass>`,
612 an optional :ref:`calling convention <callingconv>`,
613 an optional ``unnamed_addr`` attribute, a return type, an optional
614 :ref:`parameter attribute <paramattrs>` for the return type, a function
615 name, a (possibly empty) argument list (each with optional :ref:`parameter
616 attributes <paramattrs>`), optional :ref:`function attributes <fnattrs>`,
617 an optional section, an optional alignment, an optional :ref:`garbage
618 collector name <gc>`, an optional :ref:`prefix <prefixdata>`, an opening
619 curly brace, a list of basic blocks, and a closing curly brace.
621 LLVM function declarations consist of the "``declare``" keyword, an
622 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
623 style <visibility>`, an optional :ref:`DLL storage class <dllstorageclass>`,
624 an optional :ref:`calling convention <callingconv>`,
625 an optional ``unnamed_addr`` attribute, a return type, an optional
626 :ref:`parameter attribute <paramattrs>` for the return type, a function
627 name, a possibly empty list of arguments, an optional alignment, an optional
628 :ref:`garbage collector name <gc>` and an optional :ref:`prefix <prefixdata>`.
630 A function definition contains a list of basic blocks, forming the CFG (Control
631 Flow Graph) for the function. Each basic block may optionally start with a label
632 (giving the basic block a symbol table entry), contains a list of instructions,
633 and ends with a :ref:`terminator <terminators>` instruction (such as a branch or
634 function return). If an explicit label is not provided, a block is assigned an
635 implicit numbered label, using the next value from the same counter as used for
636 unnamed temporaries (:ref:`see above<identifiers>`). For example, if a function
637 entry block does not have an explicit label, it will be assigned label "%0",
638 then the first unnamed temporary in that block will be "%1", etc.
640 The first basic block in a function is special in two ways: it is
641 immediately executed on entrance to the function, and it is not allowed
642 to have predecessor basic blocks (i.e. there can not be any branches to
643 the entry block of a function). Because the block can have no
644 predecessors, it also cannot have any :ref:`PHI nodes <i_phi>`.
646 LLVM allows an explicit section to be specified for functions. If the
647 target supports it, it will emit functions to the section specified.
649 An explicit alignment may be specified for a function. If not present,
650 or if the alignment is set to zero, the alignment of the function is set
651 by the target to whatever it feels convenient. If an explicit alignment
652 is specified, the function is forced to have at least that much
653 alignment. All alignments must be a power of 2.
655 If the ``unnamed_addr`` attribute is given, the address is know to not
656 be significant and two identical functions can be merged.
660 define [linkage] [visibility] [DLLStorageClass]
662 <ResultType> @<FunctionName> ([argument list])
663 [fn Attrs] [section "name"] [align N]
664 [gc] [prefix Constant] { ... }
671 Aliases act as "second name" for the aliasee value (which can be either
672 function, global variable, another alias or bitcast of global value).
673 Aliases may have an optional :ref:`linkage type <linkage>`, an optional
674 :ref:`visibility style <visibility>`, and an optional :ref:`DLL storage class
679 @<Name> = [Visibility] [DLLStorageClass] alias [Linkage] <AliaseeTy> @<Aliasee>
681 The linkage must be one of ``private``, ``linker_private``,
682 ``linker_private_weak``, ``internal``, ``linkonce``, ``weak``,
683 ``linkonce_odr``, ``weak_odr``, ``external``. Note that some system linkers
684 might not correctly handle dropping a weak symbol that is aliased by a non-weak
687 .. _namedmetadatastructure:
692 Named metadata is a collection of metadata. :ref:`Metadata
693 nodes <metadata>` (but not metadata strings) are the only valid
694 operands for a named metadata.
698 ; Some unnamed metadata nodes, which are referenced by the named metadata.
699 !0 = metadata !{metadata !"zero"}
700 !1 = metadata !{metadata !"one"}
701 !2 = metadata !{metadata !"two"}
703 !name = !{!0, !1, !2}
710 The return type and each parameter of a function type may have a set of
711 *parameter attributes* associated with them. Parameter attributes are
712 used to communicate additional information about the result or
713 parameters of a function. Parameter attributes are considered to be part
714 of the function, not of the function type, so functions with different
715 parameter attributes can have the same function type.
717 Parameter attributes are simple keywords that follow the type specified.
718 If multiple parameter attributes are needed, they are space separated.
723 declare i32 @printf(i8* noalias nocapture, ...)
724 declare i32 @atoi(i8 zeroext)
725 declare signext i8 @returns_signed_char()
727 Note that any attributes for the function result (``nounwind``,
728 ``readonly``) come immediately after the argument list.
730 Currently, only the following parameter attributes are defined:
733 This indicates to the code generator that the parameter or return
734 value should be zero-extended to the extent required by the target's
735 ABI (which is usually 32-bits, but is 8-bits for a i1 on x86-64) by
736 the caller (for a parameter) or the callee (for a return value).
738 This indicates to the code generator that the parameter or return
739 value should be sign-extended to the extent required by the target's
740 ABI (which is usually 32-bits) by the caller (for a parameter) or
741 the callee (for a return value).
743 This indicates that this parameter or return value should be treated
744 in a special target-dependent fashion during while emitting code for
745 a function call or return (usually, by putting it in a register as
746 opposed to memory, though some targets use it to distinguish between
747 two different kinds of registers). Use of this attribute is
750 This indicates that the pointer parameter should really be passed by
751 value to the function. The attribute implies that a hidden copy of
752 the pointee is made between the caller and the callee, so the callee
753 is unable to modify the value in the caller. This attribute is only
754 valid on LLVM pointer arguments. It is generally used to pass
755 structs and arrays by value, but is also valid on pointers to
756 scalars. The copy is considered to belong to the caller not the
757 callee (for example, ``readonly`` functions should not write to
758 ``byval`` parameters). This is not a valid attribute for return
761 The byval attribute also supports specifying an alignment with the
762 align attribute. It indicates the alignment of the stack slot to
763 form and the known alignment of the pointer specified to the call
764 site. If the alignment is not specified, then the code generator
765 makes a target-specific assumption.
771 .. Warning:: This feature is unstable and not fully implemented.
773 The ``inalloca`` argument attribute allows the caller to take the
774 address of all stack-allocated arguments to a ``call`` or ``invoke``
775 before it executes. It is similar to ``byval`` in that it is used
776 to pass arguments by value, but it guarantees that the argument will
779 To be :ref:`well formed <wellformed>`, an alloca may be used as an
780 ``inalloca`` argument at most once. The attribute can only be
781 applied to the last parameter, and it guarantees that they are
782 passed in memory. The ``inalloca`` attribute cannot be used in
783 conjunction with other attributes that affect argument storage, like
784 ``inreg``, ``nest``, ``sret``, or ``byval``. The ``inalloca`` stack
785 space is considered to be clobbered by any call that uses it, so any
786 ``inalloca`` parameters cannot be marked ``readonly``.
788 When the call site is reached, the argument allocation must have
789 been the most recent stack allocation that is still live, or the
790 results are undefined. It is possible to allocate additional stack
791 space after an argument allocation and before its call site, but it
792 must be cleared off with :ref:`llvm.stackrestore
795 See :doc:`InAlloca` for more information on how to use this
799 This indicates that the pointer parameter specifies the address of a
800 structure that is the return value of the function in the source
801 program. This pointer must be guaranteed by the caller to be valid:
802 loads and stores to the structure may be assumed by the callee
803 not to trap and to be properly aligned. This may only be applied to
804 the first parameter. This is not a valid attribute for return
807 This indicates that pointer values :ref:`based <pointeraliasing>` on
808 the argument or return value do not alias pointer values which are
809 not *based* on it, ignoring certain "irrelevant" dependencies. For a
810 call to the parent function, dependencies between memory references
811 from before or after the call and from those during the call are
812 "irrelevant" to the ``noalias`` keyword for the arguments and return
813 value used in that call. The caller shares the responsibility with
814 the callee for ensuring that these requirements are met. For further
815 details, please see the discussion of the NoAlias response in `alias
816 analysis <AliasAnalysis.html#MustMayNo>`_.
818 Note that this definition of ``noalias`` is intentionally similar
819 to the definition of ``restrict`` in C99 for function arguments,
820 though it is slightly weaker.
822 For function return values, C99's ``restrict`` is not meaningful,
823 while LLVM's ``noalias`` is.
825 This indicates that the callee does not make any copies of the
826 pointer that outlive the callee itself. This is not a valid
827 attribute for return values.
832 This indicates that the pointer parameter can be excised using the
833 :ref:`trampoline intrinsics <int_trampoline>`. This is not a valid
834 attribute for return values and can only be applied to one parameter.
837 This indicates that the function always returns the argument as its return
838 value. This is an optimization hint to the code generator when generating
839 the caller, allowing tail call optimization and omission of register saves
840 and restores in some cases; it is not checked or enforced when generating
841 the callee. The parameter and the function return type must be valid
842 operands for the :ref:`bitcast instruction <i_bitcast>`. This is not a
843 valid attribute for return values and can only be applied to one parameter.
847 Garbage Collector Names
848 -----------------------
850 Each function may specify a garbage collector name, which is simply a
855 define void @f() gc "name" { ... }
857 The compiler declares the supported values of *name*. Specifying a
858 collector which will cause the compiler to alter its output in order to
859 support the named garbage collection algorithm.
866 Prefix data is data associated with a function which the code generator
867 will emit immediately before the function body. The purpose of this feature
868 is to allow frontends to associate language-specific runtime metadata with
869 specific functions and make it available through the function pointer while
870 still allowing the function pointer to be called. To access the data for a
871 given function, a program may bitcast the function pointer to a pointer to
872 the constant's type. This implies that the IR symbol points to the start
875 To maintain the semantics of ordinary function calls, the prefix data must
876 have a particular format. Specifically, it must begin with a sequence of
877 bytes which decode to a sequence of machine instructions, valid for the
878 module's target, which transfer control to the point immediately succeeding
879 the prefix data, without performing any other visible action. This allows
880 the inliner and other passes to reason about the semantics of the function
881 definition without needing to reason about the prefix data. Obviously this
882 makes the format of the prefix data highly target dependent.
884 Prefix data is laid out as if it were an initializer for a global variable
885 of the prefix data's type. No padding is automatically placed between the
886 prefix data and the function body. If padding is required, it must be part
889 A trivial example of valid prefix data for the x86 architecture is ``i8 144``,
890 which encodes the ``nop`` instruction:
894 define void @f() prefix i8 144 { ... }
896 Generally prefix data can be formed by encoding a relative branch instruction
897 which skips the metadata, as in this example of valid prefix data for the
898 x86_64 architecture, where the first two bytes encode ``jmp .+10``:
902 %0 = type <{ i8, i8, i8* }>
904 define void @f() prefix %0 <{ i8 235, i8 8, i8* @md}> { ... }
906 A function may have prefix data but no body. This has similar semantics
907 to the ``available_externally`` linkage in that the data may be used by the
908 optimizers but will not be emitted in the object file.
915 Attribute groups are groups of attributes that are referenced by objects within
916 the IR. They are important for keeping ``.ll`` files readable, because a lot of
917 functions will use the same set of attributes. In the degenerative case of a
918 ``.ll`` file that corresponds to a single ``.c`` file, the single attribute
919 group will capture the important command line flags used to build that file.
921 An attribute group is a module-level object. To use an attribute group, an
922 object references the attribute group's ID (e.g. ``#37``). An object may refer
923 to more than one attribute group. In that situation, the attributes from the
924 different groups are merged.
926 Here is an example of attribute groups for a function that should always be
927 inlined, has a stack alignment of 4, and which shouldn't use SSE instructions:
931 ; Target-independent attributes:
932 attributes #0 = { alwaysinline alignstack=4 }
934 ; Target-dependent attributes:
935 attributes #1 = { "no-sse" }
937 ; Function @f has attributes: alwaysinline, alignstack=4, and "no-sse".
938 define void @f() #0 #1 { ... }
945 Function attributes are set to communicate additional information about
946 a function. Function attributes are considered to be part of the
947 function, not of the function type, so functions with different function
948 attributes can have the same function type.
950 Function attributes are simple keywords that follow the type specified.
951 If multiple attributes are needed, they are space separated. For
956 define void @f() noinline { ... }
957 define void @f() alwaysinline { ... }
958 define void @f() alwaysinline optsize { ... }
959 define void @f() optsize { ... }
962 This attribute indicates that, when emitting the prologue and
963 epilogue, the backend should forcibly align the stack pointer.
964 Specify the desired alignment, which must be a power of two, in
967 This attribute indicates that the inliner should attempt to inline
968 this function into callers whenever possible, ignoring any active
969 inlining size threshold for this caller.
971 This indicates that the callee function at a call site should be
972 recognized as a built-in function, even though the function's declaration
973 uses the ``nobuiltin`` attribute. This is only valid at call sites for
974 direct calls to functions which are declared with the ``nobuiltin``
977 This attribute indicates that this function is rarely called. When
978 computing edge weights, basic blocks post-dominated by a cold
979 function call are also considered to be cold; and, thus, given low
982 This attribute indicates that the source code contained a hint that
983 inlining this function is desirable (such as the "inline" keyword in
984 C/C++). It is just a hint; it imposes no requirements on the
987 This attribute suggests that optimization passes and code generator
988 passes make choices that keep the code size of this function as small
989 as possible and perform optimizations that may sacrifice runtime
990 performance in order to minimize the size of the generated code.
992 This attribute disables prologue / epilogue emission for the
993 function. This can have very system-specific consequences.
995 This indicates that the callee function at a call site is not recognized as
996 a built-in function. LLVM will retain the original call and not replace it
997 with equivalent code based on the semantics of the built-in function, unless
998 the call site uses the ``builtin`` attribute. This is valid at call sites
999 and on function declarations and definitions.
1001 This attribute indicates that calls to the function cannot be
1002 duplicated. A call to a ``noduplicate`` function may be moved
1003 within its parent function, but may not be duplicated within
1004 its parent function.
1006 A function containing a ``noduplicate`` call may still
1007 be an inlining candidate, provided that the call is not
1008 duplicated by inlining. That implies that the function has
1009 internal linkage and only has one call site, so the original
1010 call is dead after inlining.
1012 This attributes disables implicit floating point instructions.
1014 This attribute indicates that the inliner should never inline this
1015 function in any situation. This attribute may not be used together
1016 with the ``alwaysinline`` attribute.
1018 This attribute suppresses lazy symbol binding for the function. This
1019 may make calls to the function faster, at the cost of extra program
1020 startup time if the function is not called during program startup.
1022 This attribute indicates that the code generator should not use a
1023 red zone, even if the target-specific ABI normally permits it.
1025 This function attribute indicates that the function never returns
1026 normally. This produces undefined behavior at runtime if the
1027 function ever does dynamically return.
1029 This function attribute indicates that the function never returns
1030 with an unwind or exceptional control flow. If the function does
1031 unwind, its runtime behavior is undefined.
1033 This function attribute indicates that the function is not optimized
1034 by any optimization or code generator passes with the
1035 exception of interprocedural optimization passes.
1036 This attribute cannot be used together with the ``alwaysinline``
1037 attribute; this attribute is also incompatible
1038 with the ``minsize`` attribute and the ``optsize`` attribute.
1040 This attribute requires the ``noinline`` attribute to be specified on
1041 the function as well, so the function is never inlined into any caller.
1042 Only functions with the ``alwaysinline`` attribute are valid
1043 candidates for inlining into the body of this function.
1045 This attribute suggests that optimization passes and code generator
1046 passes make choices that keep the code size of this function low,
1047 and otherwise do optimizations specifically to reduce code size as
1048 long as they do not significantly impact runtime performance.
1050 On a function, this attribute indicates that the function computes its
1051 result (or decides to unwind an exception) based strictly on its arguments,
1052 without dereferencing any pointer arguments or otherwise accessing
1053 any mutable state (e.g. memory, control registers, etc) visible to
1054 caller functions. It does not write through any pointer arguments
1055 (including ``byval`` arguments) and never changes any state visible
1056 to callers. This means that it cannot unwind exceptions by calling
1057 the ``C++`` exception throwing methods.
1059 On an argument, this attribute indicates that the function does not
1060 dereference that pointer argument, even though it may read or write the
1061 memory that the pointer points to if accessed through other pointers.
1063 On a function, this attribute indicates that the function does not write
1064 through any pointer arguments (including ``byval`` arguments) or otherwise
1065 modify any state (e.g. memory, control registers, etc) visible to
1066 caller functions. It may dereference pointer arguments and read
1067 state that may be set in the caller. A readonly function always
1068 returns the same value (or unwinds an exception identically) when
1069 called with the same set of arguments and global state. It cannot
1070 unwind an exception by calling the ``C++`` exception throwing
1073 On an argument, this attribute indicates that the function does not write
1074 through this pointer argument, even though it may write to the memory that
1075 the pointer points to.
1077 This attribute indicates that this function can return twice. The C
1078 ``setjmp`` is an example of such a function. The compiler disables
1079 some optimizations (like tail calls) in the caller of these
1081 ``sanitize_address``
1082 This attribute indicates that AddressSanitizer checks
1083 (dynamic address safety analysis) are enabled for this function.
1085 This attribute indicates that MemorySanitizer checks (dynamic detection
1086 of accesses to uninitialized memory) are enabled for this function.
1088 This attribute indicates that ThreadSanitizer checks
1089 (dynamic thread safety analysis) are enabled for this function.
1091 This attribute indicates that the function should emit a stack
1092 smashing protector. It is in the form of a "canary" --- a random value
1093 placed on the stack before the local variables that's checked upon
1094 return from the function to see if it has been overwritten. A
1095 heuristic is used to determine if a function needs stack protectors
1096 or not. The heuristic used will enable protectors for functions with:
1098 - Character arrays larger than ``ssp-buffer-size`` (default 8).
1099 - Aggregates containing character arrays larger than ``ssp-buffer-size``.
1100 - Calls to alloca() with variable sizes or constant sizes greater than
1101 ``ssp-buffer-size``.
1103 If a function that has an ``ssp`` attribute is inlined into a
1104 function that doesn't have an ``ssp`` attribute, then the resulting
1105 function will have an ``ssp`` attribute.
1107 This attribute indicates that the function should *always* emit a
1108 stack smashing protector. This overrides the ``ssp`` function
1111 If a function that has an ``sspreq`` attribute is inlined into a
1112 function that doesn't have an ``sspreq`` attribute or which has an
1113 ``ssp`` or ``sspstrong`` attribute, then the resulting function will have
1114 an ``sspreq`` attribute.
1116 This attribute indicates that the function should emit a stack smashing
1117 protector. This attribute causes a strong heuristic to be used when
1118 determining if a function needs stack protectors. The strong heuristic
1119 will enable protectors for functions with:
1121 - Arrays of any size and type
1122 - Aggregates containing an array of any size and type.
1123 - Calls to alloca().
1124 - Local variables that have had their address taken.
1126 This overrides the ``ssp`` function attribute.
1128 If a function that has an ``sspstrong`` attribute is inlined into a
1129 function that doesn't have an ``sspstrong`` attribute, then the
1130 resulting function will have an ``sspstrong`` attribute.
1132 This attribute indicates that the ABI being targeted requires that
1133 an unwind table entry be produce for this function even if we can
1134 show that no exceptions passes by it. This is normally the case for
1135 the ELF x86-64 abi, but it can be disabled for some compilation
1140 Module-Level Inline Assembly
1141 ----------------------------
1143 Modules may contain "module-level inline asm" blocks, which corresponds
1144 to the GCC "file scope inline asm" blocks. These blocks are internally
1145 concatenated by LLVM and treated as a single unit, but may be separated
1146 in the ``.ll`` file if desired. The syntax is very simple:
1148 .. code-block:: llvm
1150 module asm "inline asm code goes here"
1151 module asm "more can go here"
1153 The strings can contain any character by escaping non-printable
1154 characters. The escape sequence used is simply "\\xx" where "xx" is the
1155 two digit hex code for the number.
1157 The inline asm code is simply printed to the machine code .s file when
1158 assembly code is generated.
1160 .. _langref_datalayout:
1165 A module may specify a target specific data layout string that specifies
1166 how data is to be laid out in memory. The syntax for the data layout is
1169 .. code-block:: llvm
1171 target datalayout = "layout specification"
1173 The *layout specification* consists of a list of specifications
1174 separated by the minus sign character ('-'). Each specification starts
1175 with a letter and may include other information after the letter to
1176 define some aspect of the data layout. The specifications accepted are
1180 Specifies that the target lays out data in big-endian form. That is,
1181 the bits with the most significance have the lowest address
1184 Specifies that the target lays out data in little-endian form. That
1185 is, the bits with the least significance have the lowest address
1188 Specifies the natural alignment of the stack in bits. Alignment
1189 promotion of stack variables is limited to the natural stack
1190 alignment to avoid dynamic stack realignment. The stack alignment
1191 must be a multiple of 8-bits. If omitted, the natural stack
1192 alignment defaults to "unspecified", which does not prevent any
1193 alignment promotions.
1194 ``p[n]:<size>:<abi>:<pref>``
1195 This specifies the *size* of a pointer and its ``<abi>`` and
1196 ``<pref>``\erred alignments for address space ``n``. All sizes are in
1197 bits. The address space, ``n`` is optional, and if not specified,
1198 denotes the default address space 0. The value of ``n`` must be
1199 in the range [1,2^23).
1200 ``i<size>:<abi>:<pref>``
1201 This specifies the alignment for an integer type of a given bit
1202 ``<size>``. The value of ``<size>`` must be in the range [1,2^23).
1203 ``v<size>:<abi>:<pref>``
1204 This specifies the alignment for a vector type of a given bit
1206 ``f<size>:<abi>:<pref>``
1207 This specifies the alignment for a floating point type of a given bit
1208 ``<size>``. Only values of ``<size>`` that are supported by the target
1209 will work. 32 (float) and 64 (double) are supported on all targets; 80
1210 or 128 (different flavors of long double) are also supported on some
1213 This specifies the alignment for an object of aggregate type.
1215 If present, specifies that llvm names are mangled in the output. The
1218 * ``e``: ELF mangling: Private symbols get a ``.L`` prefix.
1219 * ``m``: Mips mangling: Private symbols get a ``$`` prefix.
1220 * ``o``: Mach-O mangling: Private symbols get ``L`` prefix. Other
1221 symbols get a ``_`` prefix.
1222 * ``w``: Windows COFF prefix: Similar to Mach-O, but stdcall and fastcall
1223 functions also get a suffix based on the frame size.
1224 ``n<size1>:<size2>:<size3>...``
1225 This specifies a set of native integer widths for the target CPU in
1226 bits. For example, it might contain ``n32`` for 32-bit PowerPC,
1227 ``n32:64`` for PowerPC 64, or ``n8:16:32:64`` for X86-64. Elements of
1228 this set are considered to support most general arithmetic operations
1231 On every specification that takes a ``<abi>:<pref>``, specifying the
1232 ``<pref>`` alignment is optional. If omitted, the preceding ``:``
1233 should be omitted too and ``<pref>`` will be equal to ``<abi>``.
1235 When constructing the data layout for a given target, LLVM starts with a
1236 default set of specifications which are then (possibly) overridden by
1237 the specifications in the ``datalayout`` keyword. The default
1238 specifications are given in this list:
1240 - ``E`` - big endian
1241 - ``p:64:64:64`` - 64-bit pointers with 64-bit alignment.
1242 - ``p[n]:64:64:64`` - Other address spaces are assumed to be the
1243 same as the default address space.
1244 - ``S0`` - natural stack alignment is unspecified
1245 - ``i1:8:8`` - i1 is 8-bit (byte) aligned
1246 - ``i8:8:8`` - i8 is 8-bit (byte) aligned
1247 - ``i16:16:16`` - i16 is 16-bit aligned
1248 - ``i32:32:32`` - i32 is 32-bit aligned
1249 - ``i64:32:64`` - i64 has ABI alignment of 32-bits but preferred
1250 alignment of 64-bits
1251 - ``f16:16:16`` - half is 16-bit aligned
1252 - ``f32:32:32`` - float is 32-bit aligned
1253 - ``f64:64:64`` - double is 64-bit aligned
1254 - ``f128:128:128`` - quad is 128-bit aligned
1255 - ``v64:64:64`` - 64-bit vector is 64-bit aligned
1256 - ``v128:128:128`` - 128-bit vector is 128-bit aligned
1257 - ``a:0:64`` - aggregates are 64-bit aligned
1259 When LLVM is determining the alignment for a given type, it uses the
1262 #. If the type sought is an exact match for one of the specifications,
1263 that specification is used.
1264 #. If no match is found, and the type sought is an integer type, then
1265 the smallest integer type that is larger than the bitwidth of the
1266 sought type is used. If none of the specifications are larger than
1267 the bitwidth then the largest integer type is used. For example,
1268 given the default specifications above, the i7 type will use the
1269 alignment of i8 (next largest) while both i65 and i256 will use the
1270 alignment of i64 (largest specified).
1271 #. If no match is found, and the type sought is a vector type, then the
1272 largest vector type that is smaller than the sought vector type will
1273 be used as a fall back. This happens because <128 x double> can be
1274 implemented in terms of 64 <2 x double>, for example.
1276 The function of the data layout string may not be what you expect.
1277 Notably, this is not a specification from the frontend of what alignment
1278 the code generator should use.
1280 Instead, if specified, the target data layout is required to match what
1281 the ultimate *code generator* expects. This string is used by the
1282 mid-level optimizers to improve code, and this only works if it matches
1283 what the ultimate code generator uses. If you would like to generate IR
1284 that does not embed this target-specific detail into the IR, then you
1285 don't have to specify the string. This will disable some optimizations
1286 that require precise layout information, but this also prevents those
1287 optimizations from introducing target specificity into the IR.
1294 A module may specify a target triple string that describes the target
1295 host. The syntax for the target triple is simply:
1297 .. code-block:: llvm
1299 target triple = "x86_64-apple-macosx10.7.0"
1301 The *target triple* string consists of a series of identifiers delimited
1302 by the minus sign character ('-'). The canonical forms are:
1306 ARCHITECTURE-VENDOR-OPERATING_SYSTEM
1307 ARCHITECTURE-VENDOR-OPERATING_SYSTEM-ENVIRONMENT
1309 This information is passed along to the backend so that it generates
1310 code for the proper architecture. It's possible to override this on the
1311 command line with the ``-mtriple`` command line option.
1313 .. _pointeraliasing:
1315 Pointer Aliasing Rules
1316 ----------------------
1318 Any memory access must be done through a pointer value associated with
1319 an address range of the memory access, otherwise the behavior is
1320 undefined. Pointer values are associated with address ranges according
1321 to the following rules:
1323 - A pointer value is associated with the addresses associated with any
1324 value it is *based* on.
1325 - An address of a global variable is associated with the address range
1326 of the variable's storage.
1327 - The result value of an allocation instruction is associated with the
1328 address range of the allocated storage.
1329 - A null pointer in the default address-space is associated with no
1331 - An integer constant other than zero or a pointer value returned from
1332 a function not defined within LLVM may be associated with address
1333 ranges allocated through mechanisms other than those provided by
1334 LLVM. Such ranges shall not overlap with any ranges of addresses
1335 allocated by mechanisms provided by LLVM.
1337 A pointer value is *based* on another pointer value according to the
1340 - A pointer value formed from a ``getelementptr`` operation is *based*
1341 on the first operand of the ``getelementptr``.
1342 - The result value of a ``bitcast`` is *based* on the operand of the
1344 - A pointer value formed by an ``inttoptr`` is *based* on all pointer
1345 values that contribute (directly or indirectly) to the computation of
1346 the pointer's value.
1347 - The "*based* on" relationship is transitive.
1349 Note that this definition of *"based"* is intentionally similar to the
1350 definition of *"based"* in C99, though it is slightly weaker.
1352 LLVM IR does not associate types with memory. The result type of a
1353 ``load`` merely indicates the size and alignment of the memory from
1354 which to load, as well as the interpretation of the value. The first
1355 operand type of a ``store`` similarly only indicates the size and
1356 alignment of the store.
1358 Consequently, type-based alias analysis, aka TBAA, aka
1359 ``-fstrict-aliasing``, is not applicable to general unadorned LLVM IR.
1360 :ref:`Metadata <metadata>` may be used to encode additional information
1361 which specialized optimization passes may use to implement type-based
1366 Volatile Memory Accesses
1367 ------------------------
1369 Certain memory accesses, such as :ref:`load <i_load>`'s,
1370 :ref:`store <i_store>`'s, and :ref:`llvm.memcpy <int_memcpy>`'s may be
1371 marked ``volatile``. The optimizers must not change the number of
1372 volatile operations or change their order of execution relative to other
1373 volatile operations. The optimizers *may* change the order of volatile
1374 operations relative to non-volatile operations. This is not Java's
1375 "volatile" and has no cross-thread synchronization behavior.
1377 IR-level volatile loads and stores cannot safely be optimized into
1378 llvm.memcpy or llvm.memmove intrinsics even when those intrinsics are
1379 flagged volatile. Likewise, the backend should never split or merge
1380 target-legal volatile load/store instructions.
1382 .. admonition:: Rationale
1384 Platforms may rely on volatile loads and stores of natively supported
1385 data width to be executed as single instruction. For example, in C
1386 this holds for an l-value of volatile primitive type with native
1387 hardware support, but not necessarily for aggregate types. The
1388 frontend upholds these expectations, which are intentionally
1389 unspecified in the IR. The rules above ensure that IR transformation
1390 do not violate the frontend's contract with the language.
1394 Memory Model for Concurrent Operations
1395 --------------------------------------
1397 The LLVM IR does not define any way to start parallel threads of
1398 execution or to register signal handlers. Nonetheless, there are
1399 platform-specific ways to create them, and we define LLVM IR's behavior
1400 in their presence. This model is inspired by the C++0x memory model.
1402 For a more informal introduction to this model, see the :doc:`Atomics`.
1404 We define a *happens-before* partial order as the least partial order
1407 - Is a superset of single-thread program order, and
1408 - When a *synchronizes-with* ``b``, includes an edge from ``a`` to
1409 ``b``. *Synchronizes-with* pairs are introduced by platform-specific
1410 techniques, like pthread locks, thread creation, thread joining,
1411 etc., and by atomic instructions. (See also :ref:`Atomic Memory Ordering
1412 Constraints <ordering>`).
1414 Note that program order does not introduce *happens-before* edges
1415 between a thread and signals executing inside that thread.
1417 Every (defined) read operation (load instructions, memcpy, atomic
1418 loads/read-modify-writes, etc.) R reads a series of bytes written by
1419 (defined) write operations (store instructions, atomic
1420 stores/read-modify-writes, memcpy, etc.). For the purposes of this
1421 section, initialized globals are considered to have a write of the
1422 initializer which is atomic and happens before any other read or write
1423 of the memory in question. For each byte of a read R, R\ :sub:`byte`
1424 may see any write to the same byte, except:
1426 - If write\ :sub:`1` happens before write\ :sub:`2`, and
1427 write\ :sub:`2` happens before R\ :sub:`byte`, then
1428 R\ :sub:`byte` does not see write\ :sub:`1`.
1429 - If R\ :sub:`byte` happens before write\ :sub:`3`, then
1430 R\ :sub:`byte` does not see write\ :sub:`3`.
1432 Given that definition, R\ :sub:`byte` is defined as follows:
1434 - If R is volatile, the result is target-dependent. (Volatile is
1435 supposed to give guarantees which can support ``sig_atomic_t`` in
1436 C/C++, and may be used for accesses to addresses which do not behave
1437 like normal memory. It does not generally provide cross-thread
1439 - Otherwise, if there is no write to the same byte that happens before
1440 R\ :sub:`byte`, R\ :sub:`byte` returns ``undef`` for that byte.
1441 - Otherwise, if R\ :sub:`byte` may see exactly one write,
1442 R\ :sub:`byte` returns the value written by that write.
1443 - Otherwise, if R is atomic, and all the writes R\ :sub:`byte` may
1444 see are atomic, it chooses one of the values written. See the :ref:`Atomic
1445 Memory Ordering Constraints <ordering>` section for additional
1446 constraints on how the choice is made.
1447 - Otherwise R\ :sub:`byte` returns ``undef``.
1449 R returns the value composed of the series of bytes it read. This
1450 implies that some bytes within the value may be ``undef`` **without**
1451 the entire value being ``undef``. Note that this only defines the
1452 semantics of the operation; it doesn't mean that targets will emit more
1453 than one instruction to read the series of bytes.
1455 Note that in cases where none of the atomic intrinsics are used, this
1456 model places only one restriction on IR transformations on top of what
1457 is required for single-threaded execution: introducing a store to a byte
1458 which might not otherwise be stored is not allowed in general.
1459 (Specifically, in the case where another thread might write to and read
1460 from an address, introducing a store can change a load that may see
1461 exactly one write into a load that may see multiple writes.)
1465 Atomic Memory Ordering Constraints
1466 ----------------------------------
1468 Atomic instructions (:ref:`cmpxchg <i_cmpxchg>`,
1469 :ref:`atomicrmw <i_atomicrmw>`, :ref:`fence <i_fence>`,
1470 :ref:`atomic load <i_load>`, and :ref:`atomic store <i_store>`) take
1471 an ordering parameter that determines which other atomic instructions on
1472 the same address they *synchronize with*. These semantics are borrowed
1473 from Java and C++0x, but are somewhat more colloquial. If these
1474 descriptions aren't precise enough, check those specs (see spec
1475 references in the :doc:`atomics guide <Atomics>`).
1476 :ref:`fence <i_fence>` instructions treat these orderings somewhat
1477 differently since they don't take an address. See that instruction's
1478 documentation for details.
1480 For a simpler introduction to the ordering constraints, see the
1484 The set of values that can be read is governed by the happens-before
1485 partial order. A value cannot be read unless some operation wrote
1486 it. This is intended to provide a guarantee strong enough to model
1487 Java's non-volatile shared variables. This ordering cannot be
1488 specified for read-modify-write operations; it is not strong enough
1489 to make them atomic in any interesting way.
1491 In addition to the guarantees of ``unordered``, there is a single
1492 total order for modifications by ``monotonic`` operations on each
1493 address. All modification orders must be compatible with the
1494 happens-before order. There is no guarantee that the modification
1495 orders can be combined to a global total order for the whole program
1496 (and this often will not be possible). The read in an atomic
1497 read-modify-write operation (:ref:`cmpxchg <i_cmpxchg>` and
1498 :ref:`atomicrmw <i_atomicrmw>`) reads the value in the modification
1499 order immediately before the value it writes. If one atomic read
1500 happens before another atomic read of the same address, the later
1501 read must see the same value or a later value in the address's
1502 modification order. This disallows reordering of ``monotonic`` (or
1503 stronger) operations on the same address. If an address is written
1504 ``monotonic``-ally by one thread, and other threads ``monotonic``-ally
1505 read that address repeatedly, the other threads must eventually see
1506 the write. This corresponds to the C++0x/C1x
1507 ``memory_order_relaxed``.
1509 In addition to the guarantees of ``monotonic``, a
1510 *synchronizes-with* edge may be formed with a ``release`` operation.
1511 This is intended to model C++'s ``memory_order_acquire``.
1513 In addition to the guarantees of ``monotonic``, if this operation
1514 writes a value which is subsequently read by an ``acquire``
1515 operation, it *synchronizes-with* that operation. (This isn't a
1516 complete description; see the C++0x definition of a release
1517 sequence.) This corresponds to the C++0x/C1x
1518 ``memory_order_release``.
1519 ``acq_rel`` (acquire+release)
1520 Acts as both an ``acquire`` and ``release`` operation on its
1521 address. This corresponds to the C++0x/C1x ``memory_order_acq_rel``.
1522 ``seq_cst`` (sequentially consistent)
1523 In addition to the guarantees of ``acq_rel`` (``acquire`` for an
1524 operation which only reads, ``release`` for an operation which only
1525 writes), there is a global total order on all
1526 sequentially-consistent operations on all addresses, which is
1527 consistent with the *happens-before* partial order and with the
1528 modification orders of all the affected addresses. Each
1529 sequentially-consistent read sees the last preceding write to the
1530 same address in this global order. This corresponds to the C++0x/C1x
1531 ``memory_order_seq_cst`` and Java volatile.
1535 If an atomic operation is marked ``singlethread``, it only *synchronizes
1536 with* or participates in modification and seq\_cst total orderings with
1537 other operations running in the same thread (for example, in signal
1545 LLVM IR floating-point binary ops (:ref:`fadd <i_fadd>`,
1546 :ref:`fsub <i_fsub>`, :ref:`fmul <i_fmul>`, :ref:`fdiv <i_fdiv>`,
1547 :ref:`frem <i_frem>`) have the following flags that can set to enable
1548 otherwise unsafe floating point operations
1551 No NaNs - Allow optimizations to assume the arguments and result are not
1552 NaN. Such optimizations are required to retain defined behavior over
1553 NaNs, but the value of the result is undefined.
1556 No Infs - Allow optimizations to assume the arguments and result are not
1557 +/-Inf. Such optimizations are required to retain defined behavior over
1558 +/-Inf, but the value of the result is undefined.
1561 No Signed Zeros - Allow optimizations to treat the sign of a zero
1562 argument or result as insignificant.
1565 Allow Reciprocal - Allow optimizations to use the reciprocal of an
1566 argument rather than perform division.
1569 Fast - Allow algebraically equivalent transformations that may
1570 dramatically change results in floating point (e.g. reassociate). This
1571 flag implies all the others.
1578 The LLVM type system is one of the most important features of the
1579 intermediate representation. Being typed enables a number of
1580 optimizations to be performed on the intermediate representation
1581 directly, without having to do extra analyses on the side before the
1582 transformation. A strong type system makes it easier to read the
1583 generated code and enables novel analyses and transformations that are
1584 not feasible to perform on normal three address code representations.
1594 The void type does not represent any value and has no size.
1612 The function type can be thought of as a function signature. It consists of a
1613 return type and a list of formal parameter types. The return type of a function
1614 type is a void type or first class type --- except for :ref:`label <t_label>`
1615 and :ref:`metadata <t_metadata>` types.
1621 <returntype> (<parameter list>)
1623 ...where '``<parameter list>``' is a comma-separated list of type
1624 specifiers. Optionally, the parameter list may include a type ``...``, which
1625 indicates that the function takes a variable number of arguments. Variable
1626 argument functions can access their arguments with the :ref:`variable argument
1627 handling intrinsic <int_varargs>` functions. '``<returntype>``' is any type
1628 except :ref:`label <t_label>` and :ref:`metadata <t_metadata>`.
1632 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1633 | ``i32 (i32)`` | function taking an ``i32``, returning an ``i32`` |
1634 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1635 | ``float (i16, i32 *) *`` | :ref:`Pointer <t_pointer>` to a function that takes an ``i16`` and a :ref:`pointer <t_pointer>` to ``i32``, returning ``float``. |
1636 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1637 | ``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. |
1638 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1639 | ``{i32, i32} (i32)`` | A function taking an ``i32``, returning a :ref:`structure <t_struct>` containing two ``i32`` values |
1640 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1647 The :ref:`first class <t_firstclass>` types are perhaps the most important.
1648 Values of these types are the only ones which can be produced by
1656 These are the types that are valid in registers from CodeGen's perspective.
1665 The integer type is a very simple type that simply specifies an
1666 arbitrary bit width for the integer type desired. Any bit width from 1
1667 bit to 2\ :sup:`23`\ -1 (about 8 million) can be specified.
1675 The number of bits the integer will occupy is specified by the ``N``
1681 +----------------+------------------------------------------------+
1682 | ``i1`` | a single-bit integer. |
1683 +----------------+------------------------------------------------+
1684 | ``i32`` | a 32-bit integer. |
1685 +----------------+------------------------------------------------+
1686 | ``i1942652`` | a really big integer of over 1 million bits. |
1687 +----------------+------------------------------------------------+
1691 Floating Point Types
1692 """"""""""""""""""""
1701 - 16-bit floating point value
1704 - 32-bit floating point value
1707 - 64-bit floating point value
1710 - 128-bit floating point value (112-bit mantissa)
1713 - 80-bit floating point value (X87)
1716 - 128-bit floating point value (two 64-bits)
1725 The x86mmx type represents a value held in an MMX register on an x86
1726 machine. The operations allowed on it are quite limited: parameters and
1727 return values, load and store, and bitcast. User-specified MMX
1728 instructions are represented as intrinsic or asm calls with arguments
1729 and/or results of this type. There are no arrays, vectors or constants
1746 The pointer type is used to specify memory locations. Pointers are
1747 commonly used to reference objects in memory.
1749 Pointer types may have an optional address space attribute defining the
1750 numbered address space where the pointed-to object resides. The default
1751 address space is number zero. The semantics of non-zero address spaces
1752 are target-specific.
1754 Note that LLVM does not permit pointers to void (``void*``) nor does it
1755 permit pointers to labels (``label*``). Use ``i8*`` instead.
1765 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1766 | ``[4 x i32]*`` | A :ref:`pointer <t_pointer>` to :ref:`array <t_array>` of four ``i32`` values. |
1767 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1768 | ``i32 (i32*) *`` | A :ref:`pointer <t_pointer>` to a :ref:`function <t_function>` that takes an ``i32*``, returning an ``i32``. |
1769 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1770 | ``i32 addrspace(5)*`` | A :ref:`pointer <t_pointer>` to an ``i32`` value that resides in address space #5. |
1771 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1780 A vector type is a simple derived type that represents a vector of
1781 elements. Vector types are used when multiple primitive data are
1782 operated in parallel using a single instruction (SIMD). A vector type
1783 requires a size (number of elements) and an underlying primitive data
1784 type. Vector types are considered :ref:`first class <t_firstclass>`.
1790 < <# elements> x <elementtype> >
1792 The number of elements is a constant integer value larger than 0;
1793 elementtype may be any integer or floating point type, or a pointer to
1794 these types. Vectors of size zero are not allowed.
1798 +-------------------+--------------------------------------------------+
1799 | ``<4 x i32>`` | Vector of 4 32-bit integer values. |
1800 +-------------------+--------------------------------------------------+
1801 | ``<8 x float>`` | Vector of 8 32-bit floating-point values. |
1802 +-------------------+--------------------------------------------------+
1803 | ``<2 x i64>`` | Vector of 2 64-bit integer values. |
1804 +-------------------+--------------------------------------------------+
1805 | ``<4 x i64*>`` | Vector of 4 pointers to 64-bit integer values. |
1806 +-------------------+--------------------------------------------------+
1815 The label type represents code labels.
1830 The metadata type represents embedded metadata. No derived types may be
1831 created from metadata except for :ref:`function <t_function>` arguments.
1844 Aggregate Types are a subset of derived types that can contain multiple
1845 member types. :ref:`Arrays <t_array>` and :ref:`structs <t_struct>` are
1846 aggregate types. :ref:`Vectors <t_vector>` are not considered to be
1856 The array type is a very simple derived type that arranges elements
1857 sequentially in memory. The array type requires a size (number of
1858 elements) and an underlying data type.
1864 [<# elements> x <elementtype>]
1866 The number of elements is a constant integer value; ``elementtype`` may
1867 be any type with a size.
1871 +------------------+--------------------------------------+
1872 | ``[40 x i32]`` | Array of 40 32-bit integer values. |
1873 +------------------+--------------------------------------+
1874 | ``[41 x i32]`` | Array of 41 32-bit integer values. |
1875 +------------------+--------------------------------------+
1876 | ``[4 x i8]`` | Array of 4 8-bit integer values. |
1877 +------------------+--------------------------------------+
1879 Here are some examples of multidimensional arrays:
1881 +-----------------------------+----------------------------------------------------------+
1882 | ``[3 x [4 x i32]]`` | 3x4 array of 32-bit integer values. |
1883 +-----------------------------+----------------------------------------------------------+
1884 | ``[12 x [10 x float]]`` | 12x10 array of single precision floating point values. |
1885 +-----------------------------+----------------------------------------------------------+
1886 | ``[2 x [3 x [4 x i16]]]`` | 2x3x4 array of 16-bit integer values. |
1887 +-----------------------------+----------------------------------------------------------+
1889 There is no restriction on indexing beyond the end of the array implied
1890 by a static type (though there are restrictions on indexing beyond the
1891 bounds of an allocated object in some cases). This means that
1892 single-dimension 'variable sized array' addressing can be implemented in
1893 LLVM with a zero length array type. An implementation of 'pascal style
1894 arrays' in LLVM could use the type "``{ i32, [0 x float]}``", for
1904 The structure type is used to represent a collection of data members
1905 together in memory. The elements of a structure may be any type that has
1908 Structures in memory are accessed using '``load``' and '``store``' by
1909 getting a pointer to a field with the '``getelementptr``' instruction.
1910 Structures in registers are accessed using the '``extractvalue``' and
1911 '``insertvalue``' instructions.
1913 Structures may optionally be "packed" structures, which indicate that
1914 the alignment of the struct is one byte, and that there is no padding
1915 between the elements. In non-packed structs, padding between field types
1916 is inserted as defined by the DataLayout string in the module, which is
1917 required to match what the underlying code generator expects.
1919 Structures can either be "literal" or "identified". A literal structure
1920 is defined inline with other types (e.g. ``{i32, i32}*``) whereas
1921 identified types are always defined at the top level with a name.
1922 Literal types are uniqued by their contents and can never be recursive
1923 or opaque since there is no way to write one. Identified types can be
1924 recursive, can be opaqued, and are never uniqued.
1930 %T1 = type { <type list> } ; Identified normal struct type
1931 %T2 = type <{ <type list> }> ; Identified packed struct type
1935 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1936 | ``{ i32, i32, i32 }`` | A triple of three ``i32`` values |
1937 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1938 | ``{ 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``. |
1939 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1940 | ``<{ i8, i32 }>`` | A packed struct known to be 5 bytes in size. |
1941 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1945 Opaque Structure Types
1946 """"""""""""""""""""""
1950 Opaque structure types are used to represent named structure types that
1951 do not have a body specified. This corresponds (for example) to the C
1952 notion of a forward declared structure.
1963 +--------------+-------------------+
1964 | ``opaque`` | An opaque type. |
1965 +--------------+-------------------+
1970 LLVM has several different basic types of constants. This section
1971 describes them all and their syntax.
1976 **Boolean constants**
1977 The two strings '``true``' and '``false``' are both valid constants
1979 **Integer constants**
1980 Standard integers (such as '4') are constants of the
1981 :ref:`integer <t_integer>` type. Negative numbers may be used with
1983 **Floating point constants**
1984 Floating point constants use standard decimal notation (e.g.
1985 123.421), exponential notation (e.g. 1.23421e+2), or a more precise
1986 hexadecimal notation (see below). The assembler requires the exact
1987 decimal value of a floating-point constant. For example, the
1988 assembler accepts 1.25 but rejects 1.3 because 1.3 is a repeating
1989 decimal in binary. Floating point constants must have a :ref:`floating
1990 point <t_floating>` type.
1991 **Null pointer constants**
1992 The identifier '``null``' is recognized as a null pointer constant
1993 and must be of :ref:`pointer type <t_pointer>`.
1995 The one non-intuitive notation for constants is the hexadecimal form of
1996 floating point constants. For example, the form
1997 '``double 0x432ff973cafa8000``' is equivalent to (but harder to read
1998 than) '``double 4.5e+15``'. The only time hexadecimal floating point
1999 constants are required (and the only time that they are generated by the
2000 disassembler) is when a floating point constant must be emitted but it
2001 cannot be represented as a decimal floating point number in a reasonable
2002 number of digits. For example, NaN's, infinities, and other special
2003 values are represented in their IEEE hexadecimal format so that assembly
2004 and disassembly do not cause any bits to change in the constants.
2006 When using the hexadecimal form, constants of types half, float, and
2007 double are represented using the 16-digit form shown above (which
2008 matches the IEEE754 representation for double); half and float values
2009 must, however, be exactly representable as IEEE 754 half and single
2010 precision, respectively. Hexadecimal format is always used for long
2011 double, and there are three forms of long double. The 80-bit format used
2012 by x86 is represented as ``0xK`` followed by 20 hexadecimal digits. The
2013 128-bit format used by PowerPC (two adjacent doubles) is represented by
2014 ``0xM`` followed by 32 hexadecimal digits. The IEEE 128-bit format is
2015 represented by ``0xL`` followed by 32 hexadecimal digits. Long doubles
2016 will only work if they match the long double format on your target.
2017 The IEEE 16-bit format (half precision) is represented by ``0xH``
2018 followed by 4 hexadecimal digits. All hexadecimal formats are big-endian
2019 (sign bit at the left).
2021 There are no constants of type x86mmx.
2023 .. _complexconstants:
2028 Complex constants are a (potentially recursive) combination of simple
2029 constants and smaller complex constants.
2031 **Structure constants**
2032 Structure constants are represented with notation similar to
2033 structure type definitions (a comma separated list of elements,
2034 surrounded by braces (``{}``)). For example:
2035 "``{ i32 4, float 17.0, i32* @G }``", where "``@G``" is declared as
2036 "``@G = external global i32``". Structure constants must have
2037 :ref:`structure type <t_struct>`, and the number and types of elements
2038 must match those specified by the type.
2040 Array constants are represented with notation similar to array type
2041 definitions (a comma separated list of elements, surrounded by
2042 square brackets (``[]``)). For example:
2043 "``[ i32 42, i32 11, i32 74 ]``". Array constants must have
2044 :ref:`array type <t_array>`, and the number and types of elements must
2045 match those specified by the type.
2046 **Vector constants**
2047 Vector constants are represented with notation similar to vector
2048 type definitions (a comma separated list of elements, surrounded by
2049 less-than/greater-than's (``<>``)). For example:
2050 "``< i32 42, i32 11, i32 74, i32 100 >``". Vector constants
2051 must have :ref:`vector type <t_vector>`, and the number and types of
2052 elements must match those specified by the type.
2053 **Zero initialization**
2054 The string '``zeroinitializer``' can be used to zero initialize a
2055 value to zero of *any* type, including scalar and
2056 :ref:`aggregate <t_aggregate>` types. This is often used to avoid
2057 having to print large zero initializers (e.g. for large arrays) and
2058 is always exactly equivalent to using explicit zero initializers.
2060 A metadata node is a structure-like constant with :ref:`metadata
2061 type <t_metadata>`. For example:
2062 "``metadata !{ i32 0, metadata !"test" }``". Unlike other
2063 constants that are meant to be interpreted as part of the
2064 instruction stream, metadata is a place to attach additional
2065 information such as debug info.
2067 Global Variable and Function Addresses
2068 --------------------------------------
2070 The addresses of :ref:`global variables <globalvars>` and
2071 :ref:`functions <functionstructure>` are always implicitly valid
2072 (link-time) constants. These constants are explicitly referenced when
2073 the :ref:`identifier for the global <identifiers>` is used and always have
2074 :ref:`pointer <t_pointer>` type. For example, the following is a legal LLVM
2077 .. code-block:: llvm
2081 @Z = global [2 x i32*] [ i32* @X, i32* @Y ]
2088 The string '``undef``' can be used anywhere a constant is expected, and
2089 indicates that the user of the value may receive an unspecified
2090 bit-pattern. Undefined values may be of any type (other than '``label``'
2091 or '``void``') and be used anywhere a constant is permitted.
2093 Undefined values are useful because they indicate to the compiler that
2094 the program is well defined no matter what value is used. This gives the
2095 compiler more freedom to optimize. Here are some examples of
2096 (potentially surprising) transformations that are valid (in pseudo IR):
2098 .. code-block:: llvm
2108 This is safe because all of the output bits are affected by the undef
2109 bits. Any output bit can have a zero or one depending on the input bits.
2111 .. code-block:: llvm
2122 These logical operations have bits that are not always affected by the
2123 input. For example, if ``%X`` has a zero bit, then the output of the
2124 '``and``' operation will always be a zero for that bit, no matter what
2125 the corresponding bit from the '``undef``' is. As such, it is unsafe to
2126 optimize or assume that the result of the '``and``' is '``undef``'.
2127 However, it is safe to assume that all bits of the '``undef``' could be
2128 0, and optimize the '``and``' to 0. Likewise, it is safe to assume that
2129 all the bits of the '``undef``' operand to the '``or``' could be set,
2130 allowing the '``or``' to be folded to -1.
2132 .. code-block:: llvm
2134 %A = select undef, %X, %Y
2135 %B = select undef, 42, %Y
2136 %C = select %X, %Y, undef
2146 This set of examples shows that undefined '``select``' (and conditional
2147 branch) conditions can go *either way*, but they have to come from one
2148 of the two operands. In the ``%A`` example, if ``%X`` and ``%Y`` were
2149 both known to have a clear low bit, then ``%A`` would have to have a
2150 cleared low bit. However, in the ``%C`` example, the optimizer is
2151 allowed to assume that the '``undef``' operand could be the same as
2152 ``%Y``, allowing the whole '``select``' to be eliminated.
2154 .. code-block:: llvm
2156 %A = xor undef, undef
2173 This example points out that two '``undef``' operands are not
2174 necessarily the same. This can be surprising to people (and also matches
2175 C semantics) where they assume that "``X^X``" is always zero, even if
2176 ``X`` is undefined. This isn't true for a number of reasons, but the
2177 short answer is that an '``undef``' "variable" can arbitrarily change
2178 its value over its "live range". This is true because the variable
2179 doesn't actually *have a live range*. Instead, the value is logically
2180 read from arbitrary registers that happen to be around when needed, so
2181 the value is not necessarily consistent over time. In fact, ``%A`` and
2182 ``%C`` need to have the same semantics or the core LLVM "replace all
2183 uses with" concept would not hold.
2185 .. code-block:: llvm
2193 These examples show the crucial difference between an *undefined value*
2194 and *undefined behavior*. An undefined value (like '``undef``') is
2195 allowed to have an arbitrary bit-pattern. This means that the ``%A``
2196 operation can be constant folded to '``undef``', because the '``undef``'
2197 could be an SNaN, and ``fdiv`` is not (currently) defined on SNaN's.
2198 However, in the second example, we can make a more aggressive
2199 assumption: because the ``undef`` is allowed to be an arbitrary value,
2200 we are allowed to assume that it could be zero. Since a divide by zero
2201 has *undefined behavior*, we are allowed to assume that the operation
2202 does not execute at all. This allows us to delete the divide and all
2203 code after it. Because the undefined operation "can't happen", the
2204 optimizer can assume that it occurs in dead code.
2206 .. code-block:: llvm
2208 a: store undef -> %X
2209 b: store %X -> undef
2214 These examples reiterate the ``fdiv`` example: a store *of* an undefined
2215 value can be assumed to not have any effect; we can assume that the
2216 value is overwritten with bits that happen to match what was already
2217 there. However, a store *to* an undefined location could clobber
2218 arbitrary memory, therefore, it has undefined behavior.
2225 Poison values are similar to :ref:`undef values <undefvalues>`, however
2226 they also represent the fact that an instruction or constant expression
2227 which cannot evoke side effects has nevertheless detected a condition
2228 which results in undefined behavior.
2230 There is currently no way of representing a poison value in the IR; they
2231 only exist when produced by operations such as :ref:`add <i_add>` with
2234 Poison value behavior is defined in terms of value *dependence*:
2236 - Values other than :ref:`phi <i_phi>` nodes depend on their operands.
2237 - :ref:`Phi <i_phi>` nodes depend on the operand corresponding to
2238 their dynamic predecessor basic block.
2239 - Function arguments depend on the corresponding actual argument values
2240 in the dynamic callers of their functions.
2241 - :ref:`Call <i_call>` instructions depend on the :ref:`ret <i_ret>`
2242 instructions that dynamically transfer control back to them.
2243 - :ref:`Invoke <i_invoke>` instructions depend on the
2244 :ref:`ret <i_ret>`, :ref:`resume <i_resume>`, or exception-throwing
2245 call instructions that dynamically transfer control back to them.
2246 - Non-volatile loads and stores depend on the most recent stores to all
2247 of the referenced memory addresses, following the order in the IR
2248 (including loads and stores implied by intrinsics such as
2249 :ref:`@llvm.memcpy <int_memcpy>`.)
2250 - An instruction with externally visible side effects depends on the
2251 most recent preceding instruction with externally visible side
2252 effects, following the order in the IR. (This includes :ref:`volatile
2253 operations <volatile>`.)
2254 - An instruction *control-depends* on a :ref:`terminator
2255 instruction <terminators>` if the terminator instruction has
2256 multiple successors and the instruction is always executed when
2257 control transfers to one of the successors, and may not be executed
2258 when control is transferred to another.
2259 - Additionally, an instruction also *control-depends* on a terminator
2260 instruction if the set of instructions it otherwise depends on would
2261 be different if the terminator had transferred control to a different
2263 - Dependence is transitive.
2265 Poison Values have the same behavior as :ref:`undef values <undefvalues>`,
2266 with the additional affect that any instruction which has a *dependence*
2267 on a poison value has undefined behavior.
2269 Here are some examples:
2271 .. code-block:: llvm
2274 %poison = sub nuw i32 0, 1 ; Results in a poison value.
2275 %still_poison = and i32 %poison, 0 ; 0, but also poison.
2276 %poison_yet_again = getelementptr i32* @h, i32 %still_poison
2277 store i32 0, i32* %poison_yet_again ; memory at @h[0] is poisoned
2279 store i32 %poison, i32* @g ; Poison value stored to memory.
2280 %poison2 = load i32* @g ; Poison value loaded back from memory.
2282 store volatile i32 %poison, i32* @g ; External observation; undefined behavior.
2284 %narrowaddr = bitcast i32* @g to i16*
2285 %wideaddr = bitcast i32* @g to i64*
2286 %poison3 = load i16* %narrowaddr ; Returns a poison value.
2287 %poison4 = load i64* %wideaddr ; Returns a poison value.
2289 %cmp = icmp slt i32 %poison, 0 ; Returns a poison value.
2290 br i1 %cmp, label %true, label %end ; Branch to either destination.
2293 store volatile i32 0, i32* @g ; This is control-dependent on %cmp, so
2294 ; it has undefined behavior.
2298 %p = phi i32 [ 0, %entry ], [ 1, %true ]
2299 ; Both edges into this PHI are
2300 ; control-dependent on %cmp, so this
2301 ; always results in a poison value.
2303 store volatile i32 0, i32* @g ; This would depend on the store in %true
2304 ; if %cmp is true, or the store in %entry
2305 ; otherwise, so this is undefined behavior.
2307 br i1 %cmp, label %second_true, label %second_end
2308 ; The same branch again, but this time the
2309 ; true block doesn't have side effects.
2316 store volatile i32 0, i32* @g ; This time, the instruction always depends
2317 ; on the store in %end. Also, it is
2318 ; control-equivalent to %end, so this is
2319 ; well-defined (ignoring earlier undefined
2320 ; behavior in this example).
2324 Addresses of Basic Blocks
2325 -------------------------
2327 ``blockaddress(@function, %block)``
2329 The '``blockaddress``' constant computes the address of the specified
2330 basic block in the specified function, and always has an ``i8*`` type.
2331 Taking the address of the entry block is illegal.
2333 This value only has defined behavior when used as an operand to the
2334 ':ref:`indirectbr <i_indirectbr>`' instruction, or for comparisons
2335 against null. Pointer equality tests between labels addresses results in
2336 undefined behavior --- though, again, comparison against null is ok, and
2337 no label is equal to the null pointer. This may be passed around as an
2338 opaque pointer sized value as long as the bits are not inspected. This
2339 allows ``ptrtoint`` and arithmetic to be performed on these values so
2340 long as the original value is reconstituted before the ``indirectbr``
2343 Finally, some targets may provide defined semantics when using the value
2344 as the operand to an inline assembly, but that is target specific.
2348 Constant Expressions
2349 --------------------
2351 Constant expressions are used to allow expressions involving other
2352 constants to be used as constants. Constant expressions may be of any
2353 :ref:`first class <t_firstclass>` type and may involve any LLVM operation
2354 that does not have side effects (e.g. load and call are not supported).
2355 The following is the syntax for constant expressions:
2357 ``trunc (CST to TYPE)``
2358 Truncate a constant to another type. The bit size of CST must be
2359 larger than the bit size of TYPE. Both types must be integers.
2360 ``zext (CST to TYPE)``
2361 Zero extend a constant to another type. The bit size of CST must be
2362 smaller than the bit size of TYPE. Both types must be integers.
2363 ``sext (CST to TYPE)``
2364 Sign extend a constant to another type. The bit size of CST must be
2365 smaller than the bit size of TYPE. Both types must be integers.
2366 ``fptrunc (CST to TYPE)``
2367 Truncate a floating point constant to another floating point type.
2368 The size of CST must be larger than the size of TYPE. Both types
2369 must be floating point.
2370 ``fpext (CST to TYPE)``
2371 Floating point extend a constant to another type. The size of CST
2372 must be smaller or equal to the size of TYPE. Both types must be
2374 ``fptoui (CST to TYPE)``
2375 Convert a floating point constant to the corresponding unsigned
2376 integer constant. TYPE must be a scalar or vector integer type. CST
2377 must be of scalar or vector floating point type. Both CST and TYPE
2378 must be scalars, or vectors of the same number of elements. If the
2379 value won't fit in the integer type, the results are undefined.
2380 ``fptosi (CST to TYPE)``
2381 Convert a floating point constant to the corresponding signed
2382 integer constant. TYPE must be a scalar or vector integer type. CST
2383 must be of scalar or vector floating point type. Both CST and TYPE
2384 must be scalars, or vectors of the same number of elements. If the
2385 value won't fit in the integer type, the results are undefined.
2386 ``uitofp (CST to TYPE)``
2387 Convert an unsigned integer constant to the corresponding floating
2388 point constant. TYPE must be a scalar or vector floating point type.
2389 CST must be of scalar or vector integer type. Both CST and TYPE must
2390 be scalars, or vectors of the same number of elements. If the value
2391 won't fit in the floating point type, the results are undefined.
2392 ``sitofp (CST to TYPE)``
2393 Convert a signed integer constant to the corresponding floating
2394 point constant. TYPE must be a scalar or vector floating point type.
2395 CST must be of scalar or vector integer type. Both CST and TYPE must
2396 be scalars, or vectors of the same number of elements. If the value
2397 won't fit in the floating point type, the results are undefined.
2398 ``ptrtoint (CST to TYPE)``
2399 Convert a pointer typed constant to the corresponding integer
2400 constant. ``TYPE`` must be an integer type. ``CST`` must be of
2401 pointer type. The ``CST`` value is zero extended, truncated, or
2402 unchanged to make it fit in ``TYPE``.
2403 ``inttoptr (CST to TYPE)``
2404 Convert an integer constant to a pointer constant. TYPE must be a
2405 pointer type. CST must be of integer type. The CST value is zero
2406 extended, truncated, or unchanged to make it fit in a pointer size.
2407 This one is *really* dangerous!
2408 ``bitcast (CST to TYPE)``
2409 Convert a constant, CST, to another TYPE. The constraints of the
2410 operands are the same as those for the :ref:`bitcast
2411 instruction <i_bitcast>`.
2412 ``addrspacecast (CST to TYPE)``
2413 Convert a constant pointer or constant vector of pointer, CST, to another
2414 TYPE in a different address space. The constraints of the operands are the
2415 same as those for the :ref:`addrspacecast instruction <i_addrspacecast>`.
2416 ``getelementptr (CSTPTR, IDX0, IDX1, ...)``, ``getelementptr inbounds (CSTPTR, IDX0, IDX1, ...)``
2417 Perform the :ref:`getelementptr operation <i_getelementptr>` on
2418 constants. As with the :ref:`getelementptr <i_getelementptr>`
2419 instruction, the index list may have zero or more indexes, which are
2420 required to make sense for the type of "CSTPTR".
2421 ``select (COND, VAL1, VAL2)``
2422 Perform the :ref:`select operation <i_select>` on constants.
2423 ``icmp COND (VAL1, VAL2)``
2424 Performs the :ref:`icmp operation <i_icmp>` on constants.
2425 ``fcmp COND (VAL1, VAL2)``
2426 Performs the :ref:`fcmp operation <i_fcmp>` on constants.
2427 ``extractelement (VAL, IDX)``
2428 Perform the :ref:`extractelement operation <i_extractelement>` on
2430 ``insertelement (VAL, ELT, IDX)``
2431 Perform the :ref:`insertelement operation <i_insertelement>` on
2433 ``shufflevector (VEC1, VEC2, IDXMASK)``
2434 Perform the :ref:`shufflevector operation <i_shufflevector>` on
2436 ``extractvalue (VAL, IDX0, IDX1, ...)``
2437 Perform the :ref:`extractvalue operation <i_extractvalue>` on
2438 constants. The index list is interpreted in a similar manner as
2439 indices in a ':ref:`getelementptr <i_getelementptr>`' operation. At
2440 least one index value must be specified.
2441 ``insertvalue (VAL, ELT, IDX0, IDX1, ...)``
2442 Perform the :ref:`insertvalue operation <i_insertvalue>` on constants.
2443 The index list is interpreted in a similar manner as indices in a
2444 ':ref:`getelementptr <i_getelementptr>`' operation. At least one index
2445 value must be specified.
2446 ``OPCODE (LHS, RHS)``
2447 Perform the specified operation of the LHS and RHS constants. OPCODE
2448 may be any of the :ref:`binary <binaryops>` or :ref:`bitwise
2449 binary <bitwiseops>` operations. The constraints on operands are
2450 the same as those for the corresponding instruction (e.g. no bitwise
2451 operations on floating point values are allowed).
2458 Inline Assembler Expressions
2459 ----------------------------
2461 LLVM supports inline assembler expressions (as opposed to :ref:`Module-Level
2462 Inline Assembly <moduleasm>`) through the use of a special value. This
2463 value represents the inline assembler as a string (containing the
2464 instructions to emit), a list of operand constraints (stored as a
2465 string), a flag that indicates whether or not the inline asm expression
2466 has side effects, and a flag indicating whether the function containing
2467 the asm needs to align its stack conservatively. An example inline
2468 assembler expression is:
2470 .. code-block:: llvm
2472 i32 (i32) asm "bswap $0", "=r,r"
2474 Inline assembler expressions may **only** be used as the callee operand
2475 of a :ref:`call <i_call>` or an :ref:`invoke <i_invoke>` instruction.
2476 Thus, typically we have:
2478 .. code-block:: llvm
2480 %X = call i32 asm "bswap $0", "=r,r"(i32 %Y)
2482 Inline asms with side effects not visible in the constraint list must be
2483 marked as having side effects. This is done through the use of the
2484 '``sideeffect``' keyword, like so:
2486 .. code-block:: llvm
2488 call void asm sideeffect "eieio", ""()
2490 In some cases inline asms will contain code that will not work unless
2491 the stack is aligned in some way, such as calls or SSE instructions on
2492 x86, yet will not contain code that does that alignment within the asm.
2493 The compiler should make conservative assumptions about what the asm
2494 might contain and should generate its usual stack alignment code in the
2495 prologue if the '``alignstack``' keyword is present:
2497 .. code-block:: llvm
2499 call void asm alignstack "eieio", ""()
2501 Inline asms also support using non-standard assembly dialects. The
2502 assumed dialect is ATT. When the '``inteldialect``' keyword is present,
2503 the inline asm is using the Intel dialect. Currently, ATT and Intel are
2504 the only supported dialects. An example is:
2506 .. code-block:: llvm
2508 call void asm inteldialect "eieio", ""()
2510 If multiple keywords appear the '``sideeffect``' keyword must come
2511 first, the '``alignstack``' keyword second and the '``inteldialect``'
2517 The call instructions that wrap inline asm nodes may have a
2518 "``!srcloc``" MDNode attached to it that contains a list of constant
2519 integers. If present, the code generator will use the integer as the
2520 location cookie value when report errors through the ``LLVMContext``
2521 error reporting mechanisms. This allows a front-end to correlate backend
2522 errors that occur with inline asm back to the source code that produced
2525 .. code-block:: llvm
2527 call void asm sideeffect "something bad", ""(), !srcloc !42
2529 !42 = !{ i32 1234567 }
2531 It is up to the front-end to make sense of the magic numbers it places
2532 in the IR. If the MDNode contains multiple constants, the code generator
2533 will use the one that corresponds to the line of the asm that the error
2538 Metadata Nodes and Metadata Strings
2539 -----------------------------------
2541 LLVM IR allows metadata to be attached to instructions in the program
2542 that can convey extra information about the code to the optimizers and
2543 code generator. One example application of metadata is source-level
2544 debug information. There are two metadata primitives: strings and nodes.
2545 All metadata has the ``metadata`` type and is identified in syntax by a
2546 preceding exclamation point ('``!``').
2548 A metadata string is a string surrounded by double quotes. It can
2549 contain any character by escaping non-printable characters with
2550 "``\xx``" where "``xx``" is the two digit hex code. For example:
2553 Metadata nodes are represented with notation similar to structure
2554 constants (a comma separated list of elements, surrounded by braces and
2555 preceded by an exclamation point). Metadata nodes can have any values as
2556 their operand. For example:
2558 .. code-block:: llvm
2560 !{ metadata !"test\00", i32 10}
2562 A :ref:`named metadata <namedmetadatastructure>` is a collection of
2563 metadata nodes, which can be looked up in the module symbol table. For
2566 .. code-block:: llvm
2568 !foo = metadata !{!4, !3}
2570 Metadata can be used as function arguments. Here ``llvm.dbg.value``
2571 function is using two metadata arguments:
2573 .. code-block:: llvm
2575 call void @llvm.dbg.value(metadata !24, i64 0, metadata !25)
2577 Metadata can be attached with an instruction. Here metadata ``!21`` is
2578 attached to the ``add`` instruction using the ``!dbg`` identifier:
2580 .. code-block:: llvm
2582 %indvar.next = add i64 %indvar, 1, !dbg !21
2584 More information about specific metadata nodes recognized by the
2585 optimizers and code generator is found below.
2590 In LLVM IR, memory does not have types, so LLVM's own type system is not
2591 suitable for doing TBAA. Instead, metadata is added to the IR to
2592 describe a type system of a higher level language. This can be used to
2593 implement typical C/C++ TBAA, but it can also be used to implement
2594 custom alias analysis behavior for other languages.
2596 The current metadata format is very simple. TBAA metadata nodes have up
2597 to three fields, e.g.:
2599 .. code-block:: llvm
2601 !0 = metadata !{ metadata !"an example type tree" }
2602 !1 = metadata !{ metadata !"int", metadata !0 }
2603 !2 = metadata !{ metadata !"float", metadata !0 }
2604 !3 = metadata !{ metadata !"const float", metadata !2, i64 1 }
2606 The first field is an identity field. It can be any value, usually a
2607 metadata string, which uniquely identifies the type. The most important
2608 name in the tree is the name of the root node. Two trees with different
2609 root node names are entirely disjoint, even if they have leaves with
2612 The second field identifies the type's parent node in the tree, or is
2613 null or omitted for a root node. A type is considered to alias all of
2614 its descendants and all of its ancestors in the tree. Also, a type is
2615 considered to alias all types in other trees, so that bitcode produced
2616 from multiple front-ends is handled conservatively.
2618 If the third field is present, it's an integer which if equal to 1
2619 indicates that the type is "constant" (meaning
2620 ``pointsToConstantMemory`` should return true; see `other useful
2621 AliasAnalysis methods <AliasAnalysis.html#OtherItfs>`_).
2623 '``tbaa.struct``' Metadata
2624 ^^^^^^^^^^^^^^^^^^^^^^^^^^
2626 The :ref:`llvm.memcpy <int_memcpy>` is often used to implement
2627 aggregate assignment operations in C and similar languages, however it
2628 is defined to copy a contiguous region of memory, which is more than
2629 strictly necessary for aggregate types which contain holes due to
2630 padding. Also, it doesn't contain any TBAA information about the fields
2633 ``!tbaa.struct`` metadata can describe which memory subregions in a
2634 memcpy are padding and what the TBAA tags of the struct are.
2636 The current metadata format is very simple. ``!tbaa.struct`` metadata
2637 nodes are a list of operands which are in conceptual groups of three.
2638 For each group of three, the first operand gives the byte offset of a
2639 field in bytes, the second gives its size in bytes, and the third gives
2642 .. code-block:: llvm
2644 !4 = metadata !{ i64 0, i64 4, metadata !1, i64 8, i64 4, metadata !2 }
2646 This describes a struct with two fields. The first is at offset 0 bytes
2647 with size 4 bytes, and has tbaa tag !1. The second is at offset 8 bytes
2648 and has size 4 bytes and has tbaa tag !2.
2650 Note that the fields need not be contiguous. In this example, there is a
2651 4 byte gap between the two fields. This gap represents padding which
2652 does not carry useful data and need not be preserved.
2654 '``fpmath``' Metadata
2655 ^^^^^^^^^^^^^^^^^^^^^
2657 ``fpmath`` metadata may be attached to any instruction of floating point
2658 type. It can be used to express the maximum acceptable error in the
2659 result of that instruction, in ULPs, thus potentially allowing the
2660 compiler to use a more efficient but less accurate method of computing
2661 it. ULP is defined as follows:
2663 If ``x`` is a real number that lies between two finite consecutive
2664 floating-point numbers ``a`` and ``b``, without being equal to one
2665 of them, then ``ulp(x) = |b - a|``, otherwise ``ulp(x)`` is the
2666 distance between the two non-equal finite floating-point numbers
2667 nearest ``x``. Moreover, ``ulp(NaN)`` is ``NaN``.
2669 The metadata node shall consist of a single positive floating point
2670 number representing the maximum relative error, for example:
2672 .. code-block:: llvm
2674 !0 = metadata !{ float 2.5 } ; maximum acceptable inaccuracy is 2.5 ULPs
2676 '``range``' Metadata
2677 ^^^^^^^^^^^^^^^^^^^^
2679 ``range`` metadata may be attached only to loads of integer types. It
2680 expresses the possible ranges the loaded value is in. The ranges are
2681 represented with a flattened list of integers. The loaded value is known
2682 to be in the union of the ranges defined by each consecutive pair. Each
2683 pair has the following properties:
2685 - The type must match the type loaded by the instruction.
2686 - The pair ``a,b`` represents the range ``[a,b)``.
2687 - Both ``a`` and ``b`` are constants.
2688 - The range is allowed to wrap.
2689 - The range should not represent the full or empty set. That is,
2692 In addition, the pairs must be in signed order of the lower bound and
2693 they must be non-contiguous.
2697 .. code-block:: llvm
2699 %a = load i8* %x, align 1, !range !0 ; Can only be 0 or 1
2700 %b = load i8* %y, align 1, !range !1 ; Can only be 255 (-1), 0 or 1
2701 %c = load i8* %z, align 1, !range !2 ; Can only be 0, 1, 3, 4 or 5
2702 %d = load i8* %z, align 1, !range !3 ; Can only be -2, -1, 3, 4 or 5
2704 !0 = metadata !{ i8 0, i8 2 }
2705 !1 = metadata !{ i8 255, i8 2 }
2706 !2 = metadata !{ i8 0, i8 2, i8 3, i8 6 }
2707 !3 = metadata !{ i8 -2, i8 0, i8 3, i8 6 }
2712 It is sometimes useful to attach information to loop constructs. Currently,
2713 loop metadata is implemented as metadata attached to the branch instruction
2714 in the loop latch block. This type of metadata refer to a metadata node that is
2715 guaranteed to be separate for each loop. The loop identifier metadata is
2716 specified with the name ``llvm.loop``.
2718 The loop identifier metadata is implemented using a metadata that refers to
2719 itself to avoid merging it with any other identifier metadata, e.g.,
2720 during module linkage or function inlining. That is, each loop should refer
2721 to their own identification metadata even if they reside in separate functions.
2722 The following example contains loop identifier metadata for two separate loop
2725 .. code-block:: llvm
2727 !0 = metadata !{ metadata !0 }
2728 !1 = metadata !{ metadata !1 }
2730 The loop identifier metadata can be used to specify additional per-loop
2731 metadata. Any operands after the first operand can be treated as user-defined
2732 metadata. For example the ``llvm.vectorizer.unroll`` metadata is understood
2733 by the loop vectorizer to indicate how many times to unroll the loop:
2735 .. code-block:: llvm
2737 br i1 %exitcond, label %._crit_edge, label %.lr.ph, !llvm.loop !0
2739 !0 = metadata !{ metadata !0, metadata !1 }
2740 !1 = metadata !{ metadata !"llvm.vectorizer.unroll", i32 2 }
2745 Metadata types used to annotate memory accesses with information helpful
2746 for optimizations are prefixed with ``llvm.mem``.
2748 '``llvm.mem.parallel_loop_access``' Metadata
2749 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2751 For a loop to be parallel, in addition to using
2752 the ``llvm.loop`` metadata to mark the loop latch branch instruction,
2753 also all of the memory accessing instructions in the loop body need to be
2754 marked with the ``llvm.mem.parallel_loop_access`` metadata. If there
2755 is at least one memory accessing instruction not marked with the metadata,
2756 the loop must be considered a sequential loop. This causes parallel loops to be
2757 converted to sequential loops due to optimization passes that are unaware of
2758 the parallel semantics and that insert new memory instructions to the loop
2761 Example of a loop that is considered parallel due to its correct use of
2762 both ``llvm.loop`` and ``llvm.mem.parallel_loop_access``
2763 metadata types that refer to the same loop identifier metadata.
2765 .. code-block:: llvm
2769 %0 = load i32* %arrayidx, align 4, !llvm.mem.parallel_loop_access !0
2771 store i32 %0, i32* %arrayidx4, align 4, !llvm.mem.parallel_loop_access !0
2773 br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !0
2777 !0 = metadata !{ metadata !0 }
2779 It is also possible to have nested parallel loops. In that case the
2780 memory accesses refer to a list of loop identifier metadata nodes instead of
2781 the loop identifier metadata node directly:
2783 .. code-block:: llvm
2790 %0 = load i32* %arrayidx, align 4, !llvm.mem.parallel_loop_access !0
2792 store i32 %0, i32* %arrayidx4, align 4, !llvm.mem.parallel_loop_access !0
2794 br i1 %exitcond, label %inner.for.end, label %inner.for.body, !llvm.loop !1
2798 %0 = load i32* %arrayidx, align 4, !llvm.mem.parallel_loop_access !0
2800 store i32 %0, i32* %arrayidx4, align 4, !llvm.mem.parallel_loop_access !0
2802 br i1 %exitcond, label %outer.for.end, label %outer.for.body, !llvm.loop !2
2804 outer.for.end: ; preds = %for.body
2806 !0 = metadata !{ metadata !1, metadata !2 } ; a list of loop identifiers
2807 !1 = metadata !{ metadata !1 } ; an identifier for the inner loop
2808 !2 = metadata !{ metadata !2 } ; an identifier for the outer loop
2810 '``llvm.vectorizer``'
2811 ^^^^^^^^^^^^^^^^^^^^^
2813 Metadata prefixed with ``llvm.vectorizer`` is used to control per-loop
2814 vectorization parameters such as vectorization factor and unroll factor.
2816 ``llvm.vectorizer`` metadata should be used in conjunction with ``llvm.loop``
2817 loop identification metadata.
2819 '``llvm.vectorizer.unroll``' Metadata
2820 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2822 This metadata instructs the loop vectorizer to unroll the specified
2823 loop exactly ``N`` times.
2825 The first operand is the string ``llvm.vectorizer.unroll`` and the second
2826 operand is an integer specifying the unroll factor. For example:
2828 .. code-block:: llvm
2830 !0 = metadata !{ metadata !"llvm.vectorizer.unroll", i32 4 }
2832 Note that setting ``llvm.vectorizer.unroll`` to 1 disables unrolling of the
2835 If ``llvm.vectorizer.unroll`` is set to 0 then the amount of unrolling will be
2836 determined automatically.
2838 '``llvm.vectorizer.width``' Metadata
2839 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2841 This metadata sets the target width of the vectorizer to ``N``. Without
2842 this metadata, the vectorizer will choose a width automatically.
2843 Regardless of this metadata, the vectorizer will only vectorize loops if
2844 it believes it is valid to do so.
2846 The first operand is the string ``llvm.vectorizer.width`` and the second
2847 operand is an integer specifying the width. For example:
2849 .. code-block:: llvm
2851 !0 = metadata !{ metadata !"llvm.vectorizer.width", i32 4 }
2853 Note that setting ``llvm.vectorizer.width`` to 1 disables vectorization of the
2856 If ``llvm.vectorizer.width`` is set to 0 then the width will be determined
2859 Module Flags Metadata
2860 =====================
2862 Information about the module as a whole is difficult to convey to LLVM's
2863 subsystems. The LLVM IR isn't sufficient to transmit this information.
2864 The ``llvm.module.flags`` named metadata exists in order to facilitate
2865 this. These flags are in the form of key / value pairs --- much like a
2866 dictionary --- making it easy for any subsystem who cares about a flag to
2869 The ``llvm.module.flags`` metadata contains a list of metadata triplets.
2870 Each triplet has the following form:
2872 - The first element is a *behavior* flag, which specifies the behavior
2873 when two (or more) modules are merged together, and it encounters two
2874 (or more) metadata with the same ID. The supported behaviors are
2876 - The second element is a metadata string that is a unique ID for the
2877 metadata. Each module may only have one flag entry for each unique ID (not
2878 including entries with the **Require** behavior).
2879 - The third element is the value of the flag.
2881 When two (or more) modules are merged together, the resulting
2882 ``llvm.module.flags`` metadata is the union of the modules' flags. That is, for
2883 each unique metadata ID string, there will be exactly one entry in the merged
2884 modules ``llvm.module.flags`` metadata table, and the value for that entry will
2885 be determined by the merge behavior flag, as described below. The only exception
2886 is that entries with the *Require* behavior are always preserved.
2888 The following behaviors are supported:
2899 Emits an error if two values disagree, otherwise the resulting value
2900 is that of the operands.
2904 Emits a warning if two values disagree. The result value will be the
2905 operand for the flag from the first module being linked.
2909 Adds a requirement that another module flag be present and have a
2910 specified value after linking is performed. The value must be a
2911 metadata pair, where the first element of the pair is the ID of the
2912 module flag to be restricted, and the second element of the pair is
2913 the value the module flag should be restricted to. This behavior can
2914 be used to restrict the allowable results (via triggering of an
2915 error) of linking IDs with the **Override** behavior.
2919 Uses the specified value, regardless of the behavior or value of the
2920 other module. If both modules specify **Override**, but the values
2921 differ, an error will be emitted.
2925 Appends the two values, which are required to be metadata nodes.
2929 Appends the two values, which are required to be metadata
2930 nodes. However, duplicate entries in the second list are dropped
2931 during the append operation.
2933 It is an error for a particular unique flag ID to have multiple behaviors,
2934 except in the case of **Require** (which adds restrictions on another metadata
2935 value) or **Override**.
2937 An example of module flags:
2939 .. code-block:: llvm
2941 !0 = metadata !{ i32 1, metadata !"foo", i32 1 }
2942 !1 = metadata !{ i32 4, metadata !"bar", i32 37 }
2943 !2 = metadata !{ i32 2, metadata !"qux", i32 42 }
2944 !3 = metadata !{ i32 3, metadata !"qux",
2946 metadata !"foo", i32 1
2949 !llvm.module.flags = !{ !0, !1, !2, !3 }
2951 - Metadata ``!0`` has the ID ``!"foo"`` and the value '1'. The behavior
2952 if two or more ``!"foo"`` flags are seen is to emit an error if their
2953 values are not equal.
2955 - Metadata ``!1`` has the ID ``!"bar"`` and the value '37'. The
2956 behavior if two or more ``!"bar"`` flags are seen is to use the value
2959 - Metadata ``!2`` has the ID ``!"qux"`` and the value '42'. The
2960 behavior if two or more ``!"qux"`` flags are seen is to emit a
2961 warning if their values are not equal.
2963 - Metadata ``!3`` has the ID ``!"qux"`` and the value:
2967 metadata !{ metadata !"foo", i32 1 }
2969 The behavior is to emit an error if the ``llvm.module.flags`` does not
2970 contain a flag with the ID ``!"foo"`` that has the value '1' after linking is
2973 Objective-C Garbage Collection Module Flags Metadata
2974 ----------------------------------------------------
2976 On the Mach-O platform, Objective-C stores metadata about garbage
2977 collection in a special section called "image info". The metadata
2978 consists of a version number and a bitmask specifying what types of
2979 garbage collection are supported (if any) by the file. If two or more
2980 modules are linked together their garbage collection metadata needs to
2981 be merged rather than appended together.
2983 The Objective-C garbage collection module flags metadata consists of the
2984 following key-value pairs:
2993 * - ``Objective-C Version``
2994 - **[Required]** --- The Objective-C ABI version. Valid values are 1 and 2.
2996 * - ``Objective-C Image Info Version``
2997 - **[Required]** --- The version of the image info section. Currently
3000 * - ``Objective-C Image Info Section``
3001 - **[Required]** --- The section to place the metadata. Valid values are
3002 ``"__OBJC, __image_info, regular"`` for Objective-C ABI version 1, and
3003 ``"__DATA,__objc_imageinfo, regular, no_dead_strip"`` for
3004 Objective-C ABI version 2.
3006 * - ``Objective-C Garbage Collection``
3007 - **[Required]** --- Specifies whether garbage collection is supported or
3008 not. Valid values are 0, for no garbage collection, and 2, for garbage
3009 collection supported.
3011 * - ``Objective-C GC Only``
3012 - **[Optional]** --- Specifies that only garbage collection is supported.
3013 If present, its value must be 6. This flag requires that the
3014 ``Objective-C Garbage Collection`` flag have the value 2.
3016 Some important flag interactions:
3018 - If a module with ``Objective-C Garbage Collection`` set to 0 is
3019 merged with a module with ``Objective-C Garbage Collection`` set to
3020 2, then the resulting module has the
3021 ``Objective-C Garbage Collection`` flag set to 0.
3022 - A module with ``Objective-C Garbage Collection`` set to 0 cannot be
3023 merged with a module with ``Objective-C GC Only`` set to 6.
3025 Automatic Linker Flags Module Flags Metadata
3026 --------------------------------------------
3028 Some targets support embedding flags to the linker inside individual object
3029 files. Typically this is used in conjunction with language extensions which
3030 allow source files to explicitly declare the libraries they depend on, and have
3031 these automatically be transmitted to the linker via object files.
3033 These flags are encoded in the IR using metadata in the module flags section,
3034 using the ``Linker Options`` key. The merge behavior for this flag is required
3035 to be ``AppendUnique``, and the value for the key is expected to be a metadata
3036 node which should be a list of other metadata nodes, each of which should be a
3037 list of metadata strings defining linker options.
3039 For example, the following metadata section specifies two separate sets of
3040 linker options, presumably to link against ``libz`` and the ``Cocoa``
3043 !0 = metadata !{ i32 6, metadata !"Linker Options",
3045 metadata !{ metadata !"-lz" },
3046 metadata !{ metadata !"-framework", metadata !"Cocoa" } } }
3047 !llvm.module.flags = !{ !0 }
3049 The metadata encoding as lists of lists of options, as opposed to a collapsed
3050 list of options, is chosen so that the IR encoding can use multiple option
3051 strings to specify e.g., a single library, while still having that specifier be
3052 preserved as an atomic element that can be recognized by a target specific
3053 assembly writer or object file emitter.
3055 Each individual option is required to be either a valid option for the target's
3056 linker, or an option that is reserved by the target specific assembly writer or
3057 object file emitter. No other aspect of these options is defined by the IR.
3059 .. _intrinsicglobalvariables:
3061 Intrinsic Global Variables
3062 ==========================
3064 LLVM has a number of "magic" global variables that contain data that
3065 affect code generation or other IR semantics. These are documented here.
3066 All globals of this sort should have a section specified as
3067 "``llvm.metadata``". This section and all globals that start with
3068 "``llvm.``" are reserved for use by LLVM.
3072 The '``llvm.used``' Global Variable
3073 -----------------------------------
3075 The ``@llvm.used`` global is an array which has
3076 :ref:`appending linkage <linkage_appending>`. This array contains a list of
3077 pointers to named global variables, functions and aliases which may optionally
3078 have a pointer cast formed of bitcast or getelementptr. For example, a legal
3081 .. code-block:: llvm
3086 @llvm.used = appending global [2 x i8*] [
3088 i8* bitcast (i32* @Y to i8*)
3089 ], section "llvm.metadata"
3091 If a symbol appears in the ``@llvm.used`` list, then the compiler, assembler,
3092 and linker are required to treat the symbol as if there is a reference to the
3093 symbol that it cannot see (which is why they have to be named). For example, if
3094 a variable has internal linkage and no references other than that from the
3095 ``@llvm.used`` list, it cannot be deleted. This is commonly used to represent
3096 references from inline asms and other things the compiler cannot "see", and
3097 corresponds to "``attribute((used))``" in GNU C.
3099 On some targets, the code generator must emit a directive to the
3100 assembler or object file to prevent the assembler and linker from
3101 molesting the symbol.
3103 .. _gv_llvmcompilerused:
3105 The '``llvm.compiler.used``' Global Variable
3106 --------------------------------------------
3108 The ``@llvm.compiler.used`` directive is the same as the ``@llvm.used``
3109 directive, except that it only prevents the compiler from touching the
3110 symbol. On targets that support it, this allows an intelligent linker to
3111 optimize references to the symbol without being impeded as it would be
3114 This is a rare construct that should only be used in rare circumstances,
3115 and should not be exposed to source languages.
3117 .. _gv_llvmglobalctors:
3119 The '``llvm.global_ctors``' Global Variable
3120 -------------------------------------------
3122 .. code-block:: llvm
3124 %0 = type { i32, void ()* }
3125 @llvm.global_ctors = appending global [1 x %0] [%0 { i32 65535, void ()* @ctor }]
3127 The ``@llvm.global_ctors`` array contains a list of constructor
3128 functions and associated priorities. The functions referenced by this
3129 array will be called in ascending order of priority (i.e. lowest first)
3130 when the module is loaded. The order of functions with the same priority
3133 .. _llvmglobaldtors:
3135 The '``llvm.global_dtors``' Global Variable
3136 -------------------------------------------
3138 .. code-block:: llvm
3140 %0 = type { i32, void ()* }
3141 @llvm.global_dtors = appending global [1 x %0] [%0 { i32 65535, void ()* @dtor }]
3143 The ``@llvm.global_dtors`` array contains a list of destructor functions
3144 and associated priorities. The functions referenced by this array will
3145 be called in descending order of priority (i.e. highest first) when the
3146 module is loaded. The order of functions with the same priority is not
3149 Instruction Reference
3150 =====================
3152 The LLVM instruction set consists of several different classifications
3153 of instructions: :ref:`terminator instructions <terminators>`, :ref:`binary
3154 instructions <binaryops>`, :ref:`bitwise binary
3155 instructions <bitwiseops>`, :ref:`memory instructions <memoryops>`, and
3156 :ref:`other instructions <otherops>`.
3160 Terminator Instructions
3161 -----------------------
3163 As mentioned :ref:`previously <functionstructure>`, every basic block in a
3164 program ends with a "Terminator" instruction, which indicates which
3165 block should be executed after the current block is finished. These
3166 terminator instructions typically yield a '``void``' value: they produce
3167 control flow, not values (the one exception being the
3168 ':ref:`invoke <i_invoke>`' instruction).
3170 The terminator instructions are: ':ref:`ret <i_ret>`',
3171 ':ref:`br <i_br>`', ':ref:`switch <i_switch>`',
3172 ':ref:`indirectbr <i_indirectbr>`', ':ref:`invoke <i_invoke>`',
3173 ':ref:`resume <i_resume>`', and ':ref:`unreachable <i_unreachable>`'.
3177 '``ret``' Instruction
3178 ^^^^^^^^^^^^^^^^^^^^^
3185 ret <type> <value> ; Return a value from a non-void function
3186 ret void ; Return from void function
3191 The '``ret``' instruction is used to return control flow (and optionally
3192 a value) from a function back to the caller.
3194 There are two forms of the '``ret``' instruction: one that returns a
3195 value and then causes control flow, and one that just causes control
3201 The '``ret``' instruction optionally accepts a single argument, the
3202 return value. The type of the return value must be a ':ref:`first
3203 class <t_firstclass>`' type.
3205 A function is not :ref:`well formed <wellformed>` if it it has a non-void
3206 return type and contains a '``ret``' instruction with no return value or
3207 a return value with a type that does not match its type, or if it has a
3208 void return type and contains a '``ret``' instruction with a return
3214 When the '``ret``' instruction is executed, control flow returns back to
3215 the calling function's context. If the caller is a
3216 ":ref:`call <i_call>`" instruction, execution continues at the
3217 instruction after the call. If the caller was an
3218 ":ref:`invoke <i_invoke>`" instruction, execution continues at the
3219 beginning of the "normal" destination block. If the instruction returns
3220 a value, that value shall set the call or invoke instruction's return
3226 .. code-block:: llvm
3228 ret i32 5 ; Return an integer value of 5
3229 ret void ; Return from a void function
3230 ret { i32, i8 } { i32 4, i8 2 } ; Return a struct of values 4 and 2
3234 '``br``' Instruction
3235 ^^^^^^^^^^^^^^^^^^^^
3242 br i1 <cond>, label <iftrue>, label <iffalse>
3243 br label <dest> ; Unconditional branch
3248 The '``br``' instruction is used to cause control flow to transfer to a
3249 different basic block in the current function. There are two forms of
3250 this instruction, corresponding to a conditional branch and an
3251 unconditional branch.
3256 The conditional branch form of the '``br``' instruction takes a single
3257 '``i1``' value and two '``label``' values. The unconditional form of the
3258 '``br``' instruction takes a single '``label``' value as a target.
3263 Upon execution of a conditional '``br``' instruction, the '``i1``'
3264 argument is evaluated. If the value is ``true``, control flows to the
3265 '``iftrue``' ``label`` argument. If "cond" is ``false``, control flows
3266 to the '``iffalse``' ``label`` argument.
3271 .. code-block:: llvm
3274 %cond = icmp eq i32 %a, %b
3275 br i1 %cond, label %IfEqual, label %IfUnequal
3283 '``switch``' Instruction
3284 ^^^^^^^^^^^^^^^^^^^^^^^^
3291 switch <intty> <value>, label <defaultdest> [ <intty> <val>, label <dest> ... ]
3296 The '``switch``' instruction is used to transfer control flow to one of
3297 several different places. It is a generalization of the '``br``'
3298 instruction, allowing a branch to occur to one of many possible
3304 The '``switch``' instruction uses three parameters: an integer
3305 comparison value '``value``', a default '``label``' destination, and an
3306 array of pairs of comparison value constants and '``label``'s. The table
3307 is not allowed to contain duplicate constant entries.
3312 The ``switch`` instruction specifies a table of values and destinations.
3313 When the '``switch``' instruction is executed, this table is searched
3314 for the given value. If the value is found, control flow is transferred
3315 to the corresponding destination; otherwise, control flow is transferred
3316 to the default destination.
3321 Depending on properties of the target machine and the particular
3322 ``switch`` instruction, this instruction may be code generated in
3323 different ways. For example, it could be generated as a series of
3324 chained conditional branches or with a lookup table.
3329 .. code-block:: llvm
3331 ; Emulate a conditional br instruction
3332 %Val = zext i1 %value to i32
3333 switch i32 %Val, label %truedest [ i32 0, label %falsedest ]
3335 ; Emulate an unconditional br instruction
3336 switch i32 0, label %dest [ ]
3338 ; Implement a jump table:
3339 switch i32 %val, label %otherwise [ i32 0, label %onzero
3341 i32 2, label %ontwo ]
3345 '``indirectbr``' Instruction
3346 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3353 indirectbr <somety>* <address>, [ label <dest1>, label <dest2>, ... ]
3358 The '``indirectbr``' instruction implements an indirect branch to a
3359 label within the current function, whose address is specified by
3360 "``address``". Address must be derived from a
3361 :ref:`blockaddress <blockaddress>` constant.
3366 The '``address``' argument is the address of the label to jump to. The
3367 rest of the arguments indicate the full set of possible destinations
3368 that the address may point to. Blocks are allowed to occur multiple
3369 times in the destination list, though this isn't particularly useful.
3371 This destination list is required so that dataflow analysis has an
3372 accurate understanding of the CFG.
3377 Control transfers to the block specified in the address argument. All
3378 possible destination blocks must be listed in the label list, otherwise
3379 this instruction has undefined behavior. This implies that jumps to
3380 labels defined in other functions have undefined behavior as well.
3385 This is typically implemented with a jump through a register.
3390 .. code-block:: llvm
3392 indirectbr i8* %Addr, [ label %bb1, label %bb2, label %bb3 ]
3396 '``invoke``' Instruction
3397 ^^^^^^^^^^^^^^^^^^^^^^^^
3404 <result> = invoke [cconv] [ret attrs] <ptr to function ty> <function ptr val>(<function args>) [fn attrs]
3405 to label <normal label> unwind label <exception label>
3410 The '``invoke``' instruction causes control to transfer to a specified
3411 function, with the possibility of control flow transfer to either the
3412 '``normal``' label or the '``exception``' label. If the callee function
3413 returns with the "``ret``" instruction, control flow will return to the
3414 "normal" label. If the callee (or any indirect callees) returns via the
3415 ":ref:`resume <i_resume>`" instruction or other exception handling
3416 mechanism, control is interrupted and continued at the dynamically
3417 nearest "exception" label.
3419 The '``exception``' label is a `landing
3420 pad <ExceptionHandling.html#overview>`_ for the exception. As such,
3421 '``exception``' label is required to have the
3422 ":ref:`landingpad <i_landingpad>`" instruction, which contains the
3423 information about the behavior of the program after unwinding happens,
3424 as its first non-PHI instruction. The restrictions on the
3425 "``landingpad``" instruction's tightly couples it to the "``invoke``"
3426 instruction, so that the important information contained within the
3427 "``landingpad``" instruction can't be lost through normal code motion.
3432 This instruction requires several arguments:
3434 #. The optional "cconv" marker indicates which :ref:`calling
3435 convention <callingconv>` the call should use. If none is
3436 specified, the call defaults to using C calling conventions.
3437 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
3438 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
3440 #. '``ptr to function ty``': shall be the signature of the pointer to
3441 function value being invoked. In most cases, this is a direct
3442 function invocation, but indirect ``invoke``'s are just as possible,
3443 branching off an arbitrary pointer to function value.
3444 #. '``function ptr val``': An LLVM value containing a pointer to a
3445 function to be invoked.
3446 #. '``function args``': argument list whose types match the function
3447 signature argument types and parameter attributes. All arguments must
3448 be of :ref:`first class <t_firstclass>` type. If the function signature
3449 indicates the function accepts a variable number of arguments, the
3450 extra arguments can be specified.
3451 #. '``normal label``': the label reached when the called function
3452 executes a '``ret``' instruction.
3453 #. '``exception label``': the label reached when a callee returns via
3454 the :ref:`resume <i_resume>` instruction or other exception handling
3456 #. The optional :ref:`function attributes <fnattrs>` list. Only
3457 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
3458 attributes are valid here.
3463 This instruction is designed to operate as a standard '``call``'
3464 instruction in most regards. The primary difference is that it
3465 establishes an association with a label, which is used by the runtime
3466 library to unwind the stack.
3468 This instruction is used in languages with destructors to ensure that
3469 proper cleanup is performed in the case of either a ``longjmp`` or a
3470 thrown exception. Additionally, this is important for implementation of
3471 '``catch``' clauses in high-level languages that support them.
3473 For the purposes of the SSA form, the definition of the value returned
3474 by the '``invoke``' instruction is deemed to occur on the edge from the
3475 current block to the "normal" label. If the callee unwinds then no
3476 return value is available.
3481 .. code-block:: llvm
3483 %retval = invoke i32 @Test(i32 15) to label %Continue
3484 unwind label %TestCleanup ; {i32}:retval set
3485 %retval = invoke coldcc i32 %Testfnptr(i32 15) to label %Continue
3486 unwind label %TestCleanup ; {i32}:retval set
3490 '``resume``' Instruction
3491 ^^^^^^^^^^^^^^^^^^^^^^^^
3498 resume <type> <value>
3503 The '``resume``' instruction is a terminator instruction that has no
3509 The '``resume``' instruction requires one argument, which must have the
3510 same type as the result of any '``landingpad``' instruction in the same
3516 The '``resume``' instruction resumes propagation of an existing
3517 (in-flight) exception whose unwinding was interrupted with a
3518 :ref:`landingpad <i_landingpad>` instruction.
3523 .. code-block:: llvm
3525 resume { i8*, i32 } %exn
3529 '``unreachable``' Instruction
3530 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3542 The '``unreachable``' instruction has no defined semantics. This
3543 instruction is used to inform the optimizer that a particular portion of
3544 the code is not reachable. This can be used to indicate that the code
3545 after a no-return function cannot be reached, and other facts.
3550 The '``unreachable``' instruction has no defined semantics.
3557 Binary operators are used to do most of the computation in a program.
3558 They require two operands of the same type, execute an operation on
3559 them, and produce a single value. The operands might represent multiple
3560 data, as is the case with the :ref:`vector <t_vector>` data type. The
3561 result value has the same type as its operands.
3563 There are several different binary operators:
3567 '``add``' Instruction
3568 ^^^^^^^^^^^^^^^^^^^^^
3575 <result> = add <ty> <op1>, <op2> ; yields {ty}:result
3576 <result> = add nuw <ty> <op1>, <op2> ; yields {ty}:result
3577 <result> = add nsw <ty> <op1>, <op2> ; yields {ty}:result
3578 <result> = add nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3583 The '``add``' instruction returns the sum of its two operands.
3588 The two arguments to the '``add``' instruction must be
3589 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3590 arguments must have identical types.
3595 The value produced is the integer sum of the two operands.
3597 If the sum has unsigned overflow, the result returned is the
3598 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
3601 Because LLVM integers use a two's complement representation, this
3602 instruction is appropriate for both signed and unsigned integers.
3604 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3605 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3606 result value of the ``add`` is a :ref:`poison value <poisonvalues>` if
3607 unsigned and/or signed overflow, respectively, occurs.
3612 .. code-block:: llvm
3614 <result> = add i32 4, %var ; yields {i32}:result = 4 + %var
3618 '``fadd``' Instruction
3619 ^^^^^^^^^^^^^^^^^^^^^^
3626 <result> = fadd [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3631 The '``fadd``' instruction returns the sum of its two operands.
3636 The two arguments to the '``fadd``' instruction must be :ref:`floating
3637 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3638 Both arguments must have identical types.
3643 The value produced is the floating point sum of the two operands. This
3644 instruction can also take any number of :ref:`fast-math flags <fastmath>`,
3645 which are optimization hints to enable otherwise unsafe floating point
3651 .. code-block:: llvm
3653 <result> = fadd float 4.0, %var ; yields {float}:result = 4.0 + %var
3655 '``sub``' Instruction
3656 ^^^^^^^^^^^^^^^^^^^^^
3663 <result> = sub <ty> <op1>, <op2> ; yields {ty}:result
3664 <result> = sub nuw <ty> <op1>, <op2> ; yields {ty}:result
3665 <result> = sub nsw <ty> <op1>, <op2> ; yields {ty}:result
3666 <result> = sub nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3671 The '``sub``' instruction returns the difference of its two operands.
3673 Note that the '``sub``' instruction is used to represent the '``neg``'
3674 instruction present in most other intermediate representations.
3679 The two arguments to the '``sub``' instruction must be
3680 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3681 arguments must have identical types.
3686 The value produced is the integer difference of the two operands.
3688 If the difference has unsigned overflow, the result returned is the
3689 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
3692 Because LLVM integers use a two's complement representation, this
3693 instruction is appropriate for both signed and unsigned integers.
3695 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3696 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3697 result value of the ``sub`` is a :ref:`poison value <poisonvalues>` if
3698 unsigned and/or signed overflow, respectively, occurs.
3703 .. code-block:: llvm
3705 <result> = sub i32 4, %var ; yields {i32}:result = 4 - %var
3706 <result> = sub i32 0, %val ; yields {i32}:result = -%var
3710 '``fsub``' Instruction
3711 ^^^^^^^^^^^^^^^^^^^^^^
3718 <result> = fsub [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3723 The '``fsub``' instruction returns the difference of its two operands.
3725 Note that the '``fsub``' instruction is used to represent the '``fneg``'
3726 instruction present in most other intermediate representations.
3731 The two arguments to the '``fsub``' instruction must be :ref:`floating
3732 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3733 Both arguments must have identical types.
3738 The value produced is the floating point difference of the two operands.
3739 This instruction can also take any number of :ref:`fast-math
3740 flags <fastmath>`, which are optimization hints to enable otherwise
3741 unsafe floating point optimizations:
3746 .. code-block:: llvm
3748 <result> = fsub float 4.0, %var ; yields {float}:result = 4.0 - %var
3749 <result> = fsub float -0.0, %val ; yields {float}:result = -%var
3751 '``mul``' Instruction
3752 ^^^^^^^^^^^^^^^^^^^^^
3759 <result> = mul <ty> <op1>, <op2> ; yields {ty}:result
3760 <result> = mul nuw <ty> <op1>, <op2> ; yields {ty}:result
3761 <result> = mul nsw <ty> <op1>, <op2> ; yields {ty}:result
3762 <result> = mul nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3767 The '``mul``' instruction returns the product of its two operands.
3772 The two arguments to the '``mul``' instruction must be
3773 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3774 arguments must have identical types.
3779 The value produced is the integer product of the two operands.
3781 If the result of the multiplication has unsigned overflow, the result
3782 returned is the mathematical result modulo 2\ :sup:`n`\ , where n is the
3783 bit width of the result.
3785 Because LLVM integers use a two's complement representation, and the
3786 result is the same width as the operands, this instruction returns the
3787 correct result for both signed and unsigned integers. If a full product
3788 (e.g. ``i32`` * ``i32`` -> ``i64``) is needed, the operands should be
3789 sign-extended or zero-extended as appropriate to the width of the full
3792 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3793 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3794 result value of the ``mul`` is a :ref:`poison value <poisonvalues>` if
3795 unsigned and/or signed overflow, respectively, occurs.
3800 .. code-block:: llvm
3802 <result> = mul i32 4, %var ; yields {i32}:result = 4 * %var
3806 '``fmul``' Instruction
3807 ^^^^^^^^^^^^^^^^^^^^^^
3814 <result> = fmul [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3819 The '``fmul``' instruction returns the product of its two operands.
3824 The two arguments to the '``fmul``' instruction must be :ref:`floating
3825 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3826 Both arguments must have identical types.
3831 The value produced is the floating point product of the two operands.
3832 This instruction can also take any number of :ref:`fast-math
3833 flags <fastmath>`, which are optimization hints to enable otherwise
3834 unsafe floating point optimizations:
3839 .. code-block:: llvm
3841 <result> = fmul float 4.0, %var ; yields {float}:result = 4.0 * %var
3843 '``udiv``' Instruction
3844 ^^^^^^^^^^^^^^^^^^^^^^
3851 <result> = udiv <ty> <op1>, <op2> ; yields {ty}:result
3852 <result> = udiv exact <ty> <op1>, <op2> ; yields {ty}:result
3857 The '``udiv``' instruction returns the quotient of its two operands.
3862 The two arguments to the '``udiv``' instruction must be
3863 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3864 arguments must have identical types.
3869 The value produced is the unsigned integer quotient of the two operands.
3871 Note that unsigned integer division and signed integer division are
3872 distinct operations; for signed integer division, use '``sdiv``'.
3874 Division by zero leads to undefined behavior.
3876 If the ``exact`` keyword is present, the result value of the ``udiv`` is
3877 a :ref:`poison value <poisonvalues>` if %op1 is not a multiple of %op2 (as
3878 such, "((a udiv exact b) mul b) == a").
3883 .. code-block:: llvm
3885 <result> = udiv i32 4, %var ; yields {i32}:result = 4 / %var
3887 '``sdiv``' Instruction
3888 ^^^^^^^^^^^^^^^^^^^^^^
3895 <result> = sdiv <ty> <op1>, <op2> ; yields {ty}:result
3896 <result> = sdiv exact <ty> <op1>, <op2> ; yields {ty}:result
3901 The '``sdiv``' instruction returns the quotient of its two operands.
3906 The two arguments to the '``sdiv``' instruction must be
3907 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3908 arguments must have identical types.
3913 The value produced is the signed integer quotient of the two operands
3914 rounded towards zero.
3916 Note that signed integer division and unsigned integer division are
3917 distinct operations; for unsigned integer division, use '``udiv``'.
3919 Division by zero leads to undefined behavior. Overflow also leads to
3920 undefined behavior; this is a rare case, but can occur, for example, by
3921 doing a 32-bit division of -2147483648 by -1.
3923 If the ``exact`` keyword is present, the result value of the ``sdiv`` is
3924 a :ref:`poison value <poisonvalues>` if the result would be rounded.
3929 .. code-block:: llvm
3931 <result> = sdiv i32 4, %var ; yields {i32}:result = 4 / %var
3935 '``fdiv``' Instruction
3936 ^^^^^^^^^^^^^^^^^^^^^^
3943 <result> = fdiv [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3948 The '``fdiv``' instruction returns the quotient of its two operands.
3953 The two arguments to the '``fdiv``' instruction must be :ref:`floating
3954 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3955 Both arguments must have identical types.
3960 The value produced is the floating point quotient of the two operands.
3961 This instruction can also take any number of :ref:`fast-math
3962 flags <fastmath>`, which are optimization hints to enable otherwise
3963 unsafe floating point optimizations:
3968 .. code-block:: llvm
3970 <result> = fdiv float 4.0, %var ; yields {float}:result = 4.0 / %var
3972 '``urem``' Instruction
3973 ^^^^^^^^^^^^^^^^^^^^^^
3980 <result> = urem <ty> <op1>, <op2> ; yields {ty}:result
3985 The '``urem``' instruction returns the remainder from the unsigned
3986 division of its two arguments.
3991 The two arguments to the '``urem``' instruction must be
3992 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3993 arguments must have identical types.
3998 This instruction returns the unsigned integer *remainder* of a division.
3999 This instruction always performs an unsigned division to get the
4002 Note that unsigned integer remainder and signed integer remainder are
4003 distinct operations; for signed integer remainder, use '``srem``'.
4005 Taking the remainder of a division by zero leads to undefined behavior.
4010 .. code-block:: llvm
4012 <result> = urem i32 4, %var ; yields {i32}:result = 4 % %var
4014 '``srem``' Instruction
4015 ^^^^^^^^^^^^^^^^^^^^^^
4022 <result> = srem <ty> <op1>, <op2> ; yields {ty}:result
4027 The '``srem``' instruction returns the remainder from the signed
4028 division of its two operands. This instruction can also take
4029 :ref:`vector <t_vector>` versions of the values in which case the elements
4035 The two arguments to the '``srem``' instruction must be
4036 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4037 arguments must have identical types.
4042 This instruction returns the *remainder* of a division (where the result
4043 is either zero or has the same sign as the dividend, ``op1``), not the
4044 *modulo* operator (where the result is either zero or has the same sign
4045 as the divisor, ``op2``) of a value. For more information about the
4046 difference, see `The Math
4047 Forum <http://mathforum.org/dr.math/problems/anne.4.28.99.html>`_. For a
4048 table of how this is implemented in various languages, please see
4050 operation <http://en.wikipedia.org/wiki/Modulo_operation>`_.
4052 Note that signed integer remainder and unsigned integer remainder are
4053 distinct operations; for unsigned integer remainder, use '``urem``'.
4055 Taking the remainder of a division by zero leads to undefined behavior.
4056 Overflow also leads to undefined behavior; this is a rare case, but can
4057 occur, for example, by taking the remainder of a 32-bit division of
4058 -2147483648 by -1. (The remainder doesn't actually overflow, but this
4059 rule lets srem be implemented using instructions that return both the
4060 result of the division and the remainder.)
4065 .. code-block:: llvm
4067 <result> = srem i32 4, %var ; yields {i32}:result = 4 % %var
4071 '``frem``' Instruction
4072 ^^^^^^^^^^^^^^^^^^^^^^
4079 <result> = frem [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
4084 The '``frem``' instruction returns the remainder from the division of
4090 The two arguments to the '``frem``' instruction must be :ref:`floating
4091 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4092 Both arguments must have identical types.
4097 This instruction returns the *remainder* of a division. The remainder
4098 has the same sign as the dividend. This instruction can also take any
4099 number of :ref:`fast-math flags <fastmath>`, which are optimization hints
4100 to enable otherwise unsafe floating point optimizations:
4105 .. code-block:: llvm
4107 <result> = frem float 4.0, %var ; yields {float}:result = 4.0 % %var
4111 Bitwise Binary Operations
4112 -------------------------
4114 Bitwise binary operators are used to do various forms of bit-twiddling
4115 in a program. They are generally very efficient instructions and can
4116 commonly be strength reduced from other instructions. They require two
4117 operands of the same type, execute an operation on them, and produce a
4118 single value. The resulting value is the same type as its operands.
4120 '``shl``' Instruction
4121 ^^^^^^^^^^^^^^^^^^^^^
4128 <result> = shl <ty> <op1>, <op2> ; yields {ty}:result
4129 <result> = shl nuw <ty> <op1>, <op2> ; yields {ty}:result
4130 <result> = shl nsw <ty> <op1>, <op2> ; yields {ty}:result
4131 <result> = shl nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
4136 The '``shl``' instruction returns the first operand shifted to the left
4137 a specified number of bits.
4142 Both arguments to the '``shl``' instruction must be the same
4143 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4144 '``op2``' is treated as an unsigned value.
4149 The value produced is ``op1`` \* 2\ :sup:`op2` mod 2\ :sup:`n`,
4150 where ``n`` is the width of the result. If ``op2`` is (statically or
4151 dynamically) negative or equal to or larger than the number of bits in
4152 ``op1``, the result is undefined. If the arguments are vectors, each
4153 vector element of ``op1`` is shifted by the corresponding shift amount
4156 If the ``nuw`` keyword is present, then the shift produces a :ref:`poison
4157 value <poisonvalues>` if it shifts out any non-zero bits. If the
4158 ``nsw`` keyword is present, then the shift produces a :ref:`poison
4159 value <poisonvalues>` if it shifts out any bits that disagree with the
4160 resultant sign bit. As such, NUW/NSW have the same semantics as they
4161 would if the shift were expressed as a mul instruction with the same
4162 nsw/nuw bits in (mul %op1, (shl 1, %op2)).
4167 .. code-block:: llvm
4169 <result> = shl i32 4, %var ; yields {i32}: 4 << %var
4170 <result> = shl i32 4, 2 ; yields {i32}: 16
4171 <result> = shl i32 1, 10 ; yields {i32}: 1024
4172 <result> = shl i32 1, 32 ; undefined
4173 <result> = shl <2 x i32> < i32 1, i32 1>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 2, i32 4>
4175 '``lshr``' Instruction
4176 ^^^^^^^^^^^^^^^^^^^^^^
4183 <result> = lshr <ty> <op1>, <op2> ; yields {ty}:result
4184 <result> = lshr exact <ty> <op1>, <op2> ; yields {ty}:result
4189 The '``lshr``' instruction (logical shift right) returns the first
4190 operand shifted to the right a specified number of bits with zero fill.
4195 Both arguments to the '``lshr``' instruction must be the same
4196 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4197 '``op2``' is treated as an unsigned value.
4202 This instruction always performs a logical shift right operation. The
4203 most significant bits of the result will be filled with zero bits after
4204 the shift. If ``op2`` is (statically or dynamically) equal to or larger
4205 than the number of bits in ``op1``, the result is undefined. If the
4206 arguments are vectors, each vector element of ``op1`` is shifted by the
4207 corresponding shift amount in ``op2``.
4209 If the ``exact`` keyword is present, the result value of the ``lshr`` is
4210 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
4216 .. code-block:: llvm
4218 <result> = lshr i32 4, 1 ; yields {i32}:result = 2
4219 <result> = lshr i32 4, 2 ; yields {i32}:result = 1
4220 <result> = lshr i8 4, 3 ; yields {i8}:result = 0
4221 <result> = lshr i8 -2, 1 ; yields {i8}:result = 0x7F
4222 <result> = lshr i32 1, 32 ; undefined
4223 <result> = lshr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 0x7FFFFFFF, i32 1>
4225 '``ashr``' Instruction
4226 ^^^^^^^^^^^^^^^^^^^^^^
4233 <result> = ashr <ty> <op1>, <op2> ; yields {ty}:result
4234 <result> = ashr exact <ty> <op1>, <op2> ; yields {ty}:result
4239 The '``ashr``' instruction (arithmetic shift right) returns the first
4240 operand shifted to the right a specified number of bits with sign
4246 Both arguments to the '``ashr``' instruction must be the same
4247 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4248 '``op2``' is treated as an unsigned value.
4253 This instruction always performs an arithmetic shift right operation,
4254 The most significant bits of the result will be filled with the sign bit
4255 of ``op1``. If ``op2`` is (statically or dynamically) equal to or larger
4256 than the number of bits in ``op1``, the result is undefined. If the
4257 arguments are vectors, each vector element of ``op1`` is shifted by the
4258 corresponding shift amount in ``op2``.
4260 If the ``exact`` keyword is present, the result value of the ``ashr`` is
4261 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
4267 .. code-block:: llvm
4269 <result> = ashr i32 4, 1 ; yields {i32}:result = 2
4270 <result> = ashr i32 4, 2 ; yields {i32}:result = 1
4271 <result> = ashr i8 4, 3 ; yields {i8}:result = 0
4272 <result> = ashr i8 -2, 1 ; yields {i8}:result = -1
4273 <result> = ashr i32 1, 32 ; undefined
4274 <result> = ashr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 3> ; yields: result=<2 x i32> < i32 -1, i32 0>
4276 '``and``' Instruction
4277 ^^^^^^^^^^^^^^^^^^^^^
4284 <result> = and <ty> <op1>, <op2> ; yields {ty}:result
4289 The '``and``' instruction returns the bitwise logical and of its two
4295 The two arguments to the '``and``' instruction must be
4296 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4297 arguments must have identical types.
4302 The truth table used for the '``and``' instruction is:
4319 .. code-block:: llvm
4321 <result> = and i32 4, %var ; yields {i32}:result = 4 & %var
4322 <result> = and i32 15, 40 ; yields {i32}:result = 8
4323 <result> = and i32 4, 8 ; yields {i32}:result = 0
4325 '``or``' Instruction
4326 ^^^^^^^^^^^^^^^^^^^^
4333 <result> = or <ty> <op1>, <op2> ; yields {ty}:result
4338 The '``or``' instruction returns the bitwise logical inclusive or of its
4344 The two arguments to the '``or``' instruction must be
4345 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4346 arguments must have identical types.
4351 The truth table used for the '``or``' instruction is:
4370 <result> = or i32 4, %var ; yields {i32}:result = 4 | %var
4371 <result> = or i32 15, 40 ; yields {i32}:result = 47
4372 <result> = or i32 4, 8 ; yields {i32}:result = 12
4374 '``xor``' Instruction
4375 ^^^^^^^^^^^^^^^^^^^^^
4382 <result> = xor <ty> <op1>, <op2> ; yields {ty}:result
4387 The '``xor``' instruction returns the bitwise logical exclusive or of
4388 its two operands. The ``xor`` is used to implement the "one's
4389 complement" operation, which is the "~" operator in C.
4394 The two arguments to the '``xor``' instruction must be
4395 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4396 arguments must have identical types.
4401 The truth table used for the '``xor``' instruction is:
4418 .. code-block:: llvm
4420 <result> = xor i32 4, %var ; yields {i32}:result = 4 ^ %var
4421 <result> = xor i32 15, 40 ; yields {i32}:result = 39
4422 <result> = xor i32 4, 8 ; yields {i32}:result = 12
4423 <result> = xor i32 %V, -1 ; yields {i32}:result = ~%V
4428 LLVM supports several instructions to represent vector operations in a
4429 target-independent manner. These instructions cover the element-access
4430 and vector-specific operations needed to process vectors effectively.
4431 While LLVM does directly support these vector operations, many
4432 sophisticated algorithms will want to use target-specific intrinsics to
4433 take full advantage of a specific target.
4435 .. _i_extractelement:
4437 '``extractelement``' Instruction
4438 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4445 <result> = extractelement <n x <ty>> <val>, i32 <idx> ; yields <ty>
4450 The '``extractelement``' instruction extracts a single scalar element
4451 from a vector at a specified index.
4456 The first operand of an '``extractelement``' instruction is a value of
4457 :ref:`vector <t_vector>` type. The second operand is an index indicating
4458 the position from which to extract the element. The index may be a
4464 The result is a scalar of the same type as the element type of ``val``.
4465 Its value is the value at position ``idx`` of ``val``. If ``idx``
4466 exceeds the length of ``val``, the results are undefined.
4471 .. code-block:: llvm
4473 <result> = extractelement <4 x i32> %vec, i32 0 ; yields i32
4475 .. _i_insertelement:
4477 '``insertelement``' Instruction
4478 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4485 <result> = insertelement <n x <ty>> <val>, <ty> <elt>, i32 <idx> ; yields <n x <ty>>
4490 The '``insertelement``' instruction inserts a scalar element into a
4491 vector at a specified index.
4496 The first operand of an '``insertelement``' instruction is a value of
4497 :ref:`vector <t_vector>` type. The second operand is a scalar value whose
4498 type must equal the element type of the first operand. The third operand
4499 is an index indicating the position at which to insert the value. The
4500 index may be a variable.
4505 The result is a vector of the same type as ``val``. Its element values
4506 are those of ``val`` except at position ``idx``, where it gets the value
4507 ``elt``. If ``idx`` exceeds the length of ``val``, the results are
4513 .. code-block:: llvm
4515 <result> = insertelement <4 x i32> %vec, i32 1, i32 0 ; yields <4 x i32>
4517 .. _i_shufflevector:
4519 '``shufflevector``' Instruction
4520 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4527 <result> = shufflevector <n x <ty>> <v1>, <n x <ty>> <v2>, <m x i32> <mask> ; yields <m x <ty>>
4532 The '``shufflevector``' instruction constructs a permutation of elements
4533 from two input vectors, returning a vector with the same element type as
4534 the input and length that is the same as the shuffle mask.
4539 The first two operands of a '``shufflevector``' instruction are vectors
4540 with the same type. The third argument is a shuffle mask whose element
4541 type is always 'i32'. The result of the instruction is a vector whose
4542 length is the same as the shuffle mask and whose element type is the
4543 same as the element type of the first two operands.
4545 The shuffle mask operand is required to be a constant vector with either
4546 constant integer or undef values.
4551 The elements of the two input vectors are numbered from left to right
4552 across both of the vectors. The shuffle mask operand specifies, for each
4553 element of the result vector, which element of the two input vectors the
4554 result element gets. The element selector may be undef (meaning "don't
4555 care") and the second operand may be undef if performing a shuffle from
4561 .. code-block:: llvm
4563 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4564 <4 x i32> <i32 0, i32 4, i32 1, i32 5> ; yields <4 x i32>
4565 <result> = shufflevector <4 x i32> %v1, <4 x i32> undef,
4566 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32> - Identity shuffle.
4567 <result> = shufflevector <8 x i32> %v1, <8 x i32> undef,
4568 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32>
4569 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4570 <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7 > ; yields <8 x i32>
4572 Aggregate Operations
4573 --------------------
4575 LLVM supports several instructions for working with
4576 :ref:`aggregate <t_aggregate>` values.
4580 '``extractvalue``' Instruction
4581 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4588 <result> = extractvalue <aggregate type> <val>, <idx>{, <idx>}*
4593 The '``extractvalue``' instruction extracts the value of a member field
4594 from an :ref:`aggregate <t_aggregate>` value.
4599 The first operand of an '``extractvalue``' instruction is a value of
4600 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The operands are
4601 constant indices to specify which value to extract in a similar manner
4602 as indices in a '``getelementptr``' instruction.
4604 The major differences to ``getelementptr`` indexing are:
4606 - Since the value being indexed is not a pointer, the first index is
4607 omitted and assumed to be zero.
4608 - At least one index must be specified.
4609 - Not only struct indices but also array indices must be in bounds.
4614 The result is the value at the position in the aggregate specified by
4620 .. code-block:: llvm
4622 <result> = extractvalue {i32, float} %agg, 0 ; yields i32
4626 '``insertvalue``' Instruction
4627 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4634 <result> = insertvalue <aggregate type> <val>, <ty> <elt>, <idx>{, <idx>}* ; yields <aggregate type>
4639 The '``insertvalue``' instruction inserts a value into a member field in
4640 an :ref:`aggregate <t_aggregate>` value.
4645 The first operand of an '``insertvalue``' instruction is a value of
4646 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The second operand is
4647 a first-class value to insert. The following operands are constant
4648 indices indicating the position at which to insert the value in a
4649 similar manner as indices in a '``extractvalue``' instruction. The value
4650 to insert must have the same type as the value identified by the
4656 The result is an aggregate of the same type as ``val``. Its value is
4657 that of ``val`` except that the value at the position specified by the
4658 indices is that of ``elt``.
4663 .. code-block:: llvm
4665 %agg1 = insertvalue {i32, float} undef, i32 1, 0 ; yields {i32 1, float undef}
4666 %agg2 = insertvalue {i32, float} %agg1, float %val, 1 ; yields {i32 1, float %val}
4667 %agg3 = insertvalue {i32, {float}} %agg1, float %val, 1, 0 ; yields {i32 1, float %val}
4671 Memory Access and Addressing Operations
4672 ---------------------------------------
4674 A key design point of an SSA-based representation is how it represents
4675 memory. In LLVM, no memory locations are in SSA form, which makes things
4676 very simple. This section describes how to read, write, and allocate
4681 '``alloca``' Instruction
4682 ^^^^^^^^^^^^^^^^^^^^^^^^
4689 <result> = alloca <type>[, <ty> <NumElements>][, align <alignment>] ; yields {type*}:result
4694 The '``alloca``' instruction allocates memory on the stack frame of the
4695 currently executing function, to be automatically released when this
4696 function returns to its caller. The object is always allocated in the
4697 generic address space (address space zero).
4702 The '``alloca``' instruction allocates ``sizeof(<type>)*NumElements``
4703 bytes of memory on the runtime stack, returning a pointer of the
4704 appropriate type to the program. If "NumElements" is specified, it is
4705 the number of elements allocated, otherwise "NumElements" is defaulted
4706 to be one. If a constant alignment is specified, the value result of the
4707 allocation is guaranteed to be aligned to at least that boundary. If not
4708 specified, or if zero, the target can choose to align the allocation on
4709 any convenient boundary compatible with the type.
4711 '``type``' may be any sized type.
4716 Memory is allocated; a pointer is returned. The operation is undefined
4717 if there is insufficient stack space for the allocation. '``alloca``'d
4718 memory is automatically released when the function returns. The
4719 '``alloca``' instruction is commonly used to represent automatic
4720 variables that must have an address available. When the function returns
4721 (either with the ``ret`` or ``resume`` instructions), the memory is
4722 reclaimed. Allocating zero bytes is legal, but the result is undefined.
4723 The order in which memory is allocated (ie., which way the stack grows)
4729 .. code-block:: llvm
4731 %ptr = alloca i32 ; yields {i32*}:ptr
4732 %ptr = alloca i32, i32 4 ; yields {i32*}:ptr
4733 %ptr = alloca i32, i32 4, align 1024 ; yields {i32*}:ptr
4734 %ptr = alloca i32, align 1024 ; yields {i32*}:ptr
4738 '``load``' Instruction
4739 ^^^^^^^^^^^^^^^^^^^^^^
4746 <result> = load [volatile] <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>][, !invariant.load !<index>]
4747 <result> = load atomic [volatile] <ty>* <pointer> [singlethread] <ordering>, align <alignment>
4748 !<index> = !{ i32 1 }
4753 The '``load``' instruction is used to read from memory.
4758 The argument to the ``load`` instruction specifies the memory address
4759 from which to load. The pointer must point to a :ref:`first
4760 class <t_firstclass>` type. If the ``load`` is marked as ``volatile``,
4761 then the optimizer is not allowed to modify the number or order of
4762 execution of this ``load`` with other :ref:`volatile
4763 operations <volatile>`.
4765 If the ``load`` is marked as ``atomic``, it takes an extra
4766 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
4767 ``release`` and ``acq_rel`` orderings are not valid on ``load``
4768 instructions. Atomic loads produce :ref:`defined <memmodel>` results
4769 when they may see multiple atomic stores. The type of the pointee must
4770 be an integer type whose bit width is a power of two greater than or
4771 equal to eight and less than or equal to a target-specific size limit.
4772 ``align`` must be explicitly specified on atomic loads, and the load has
4773 undefined behavior if the alignment is not set to a value which is at
4774 least the size in bytes of the pointee. ``!nontemporal`` does not have
4775 any defined semantics for atomic loads.
4777 The optional constant ``align`` argument specifies the alignment of the
4778 operation (that is, the alignment of the memory address). A value of 0
4779 or an omitted ``align`` argument means that the operation has the ABI
4780 alignment for the target. It is the responsibility of the code emitter
4781 to ensure that the alignment information is correct. Overestimating the
4782 alignment results in undefined behavior. Underestimating the alignment
4783 may produce less efficient code. An alignment of 1 is always safe.
4785 The optional ``!nontemporal`` metadata must reference a single
4786 metadata name ``<index>`` corresponding to a metadata node with one
4787 ``i32`` entry of value 1. The existence of the ``!nontemporal``
4788 metadata on the instruction tells the optimizer and code generator
4789 that this load is not expected to be reused in the cache. The code
4790 generator may select special instructions to save cache bandwidth, such
4791 as the ``MOVNT`` instruction on x86.
4793 The optional ``!invariant.load`` metadata must reference a single
4794 metadata name ``<index>`` corresponding to a metadata node with no
4795 entries. The existence of the ``!invariant.load`` metadata on the
4796 instruction tells the optimizer and code generator that this load
4797 address points to memory which does not change value during program
4798 execution. The optimizer may then move this load around, for example, by
4799 hoisting it out of loops using loop invariant code motion.
4804 The location of memory pointed to is loaded. If the value being loaded
4805 is of scalar type then the number of bytes read does not exceed the
4806 minimum number of bytes needed to hold all bits of the type. For
4807 example, loading an ``i24`` reads at most three bytes. When loading a
4808 value of a type like ``i20`` with a size that is not an integral number
4809 of bytes, the result is undefined if the value was not originally
4810 written using a store of the same type.
4815 .. code-block:: llvm
4817 %ptr = alloca i32 ; yields {i32*}:ptr
4818 store i32 3, i32* %ptr ; yields {void}
4819 %val = load i32* %ptr ; yields {i32}:val = i32 3
4823 '``store``' Instruction
4824 ^^^^^^^^^^^^^^^^^^^^^^^
4831 store [volatile] <ty> <value>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>] ; yields {void}
4832 store atomic [volatile] <ty> <value>, <ty>* <pointer> [singlethread] <ordering>, align <alignment> ; yields {void}
4837 The '``store``' instruction is used to write to memory.
4842 There are two arguments to the ``store`` instruction: a value to store
4843 and an address at which to store it. The type of the ``<pointer>``
4844 operand must be a pointer to the :ref:`first class <t_firstclass>` type of
4845 the ``<value>`` operand. If the ``store`` is marked as ``volatile``,
4846 then the optimizer is not allowed to modify the number or order of
4847 execution of this ``store`` with other :ref:`volatile
4848 operations <volatile>`.
4850 If the ``store`` is marked as ``atomic``, it takes an extra
4851 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
4852 ``acquire`` and ``acq_rel`` orderings aren't valid on ``store``
4853 instructions. Atomic loads produce :ref:`defined <memmodel>` results
4854 when they may see multiple atomic stores. The type of the pointee must
4855 be an integer type whose bit width is a power of two greater than or
4856 equal to eight and less than or equal to a target-specific size limit.
4857 ``align`` must be explicitly specified on atomic stores, and the store
4858 has undefined behavior if the alignment is not set to a value which is
4859 at least the size in bytes of the pointee. ``!nontemporal`` does not
4860 have any defined semantics for atomic stores.
4862 The optional constant ``align`` argument specifies the alignment of the
4863 operation (that is, the alignment of the memory address). A value of 0
4864 or an omitted ``align`` argument means that the operation has the ABI
4865 alignment for the target. It is the responsibility of the code emitter
4866 to ensure that the alignment information is correct. Overestimating the
4867 alignment results in undefined behavior. Underestimating the
4868 alignment may produce less efficient code. An alignment of 1 is always
4871 The optional ``!nontemporal`` metadata must reference a single metadata
4872 name ``<index>`` corresponding to a metadata node with one ``i32`` entry of
4873 value 1. The existence of the ``!nontemporal`` metadata on the instruction
4874 tells the optimizer and code generator that this load is not expected to
4875 be reused in the cache. The code generator may select special
4876 instructions to save cache bandwidth, such as the MOVNT instruction on
4882 The contents of memory are updated to contain ``<value>`` at the
4883 location specified by the ``<pointer>`` operand. If ``<value>`` is
4884 of scalar type then the number of bytes written does not exceed the
4885 minimum number of bytes needed to hold all bits of the type. For
4886 example, storing an ``i24`` writes at most three bytes. When writing a
4887 value of a type like ``i20`` with a size that is not an integral number
4888 of bytes, it is unspecified what happens to the extra bits that do not
4889 belong to the type, but they will typically be overwritten.
4894 .. code-block:: llvm
4896 %ptr = alloca i32 ; yields {i32*}:ptr
4897 store i32 3, i32* %ptr ; yields {void}
4898 %val = load i32* %ptr ; yields {i32}:val = i32 3
4902 '``fence``' Instruction
4903 ^^^^^^^^^^^^^^^^^^^^^^^
4910 fence [singlethread] <ordering> ; yields {void}
4915 The '``fence``' instruction is used to introduce happens-before edges
4921 '``fence``' instructions take an :ref:`ordering <ordering>` argument which
4922 defines what *synchronizes-with* edges they add. They can only be given
4923 ``acquire``, ``release``, ``acq_rel``, and ``seq_cst`` orderings.
4928 A fence A which has (at least) ``release`` ordering semantics
4929 *synchronizes with* a fence B with (at least) ``acquire`` ordering
4930 semantics if and only if there exist atomic operations X and Y, both
4931 operating on some atomic object M, such that A is sequenced before X, X
4932 modifies M (either directly or through some side effect of a sequence
4933 headed by X), Y is sequenced before B, and Y observes M. This provides a
4934 *happens-before* dependency between A and B. Rather than an explicit
4935 ``fence``, one (but not both) of the atomic operations X or Y might
4936 provide a ``release`` or ``acquire`` (resp.) ordering constraint and
4937 still *synchronize-with* the explicit ``fence`` and establish the
4938 *happens-before* edge.
4940 A ``fence`` which has ``seq_cst`` ordering, in addition to having both
4941 ``acquire`` and ``release`` semantics specified above, participates in
4942 the global program order of other ``seq_cst`` operations and/or fences.
4944 The optional ":ref:`singlethread <singlethread>`" argument specifies
4945 that the fence only synchronizes with other fences in the same thread.
4946 (This is useful for interacting with signal handlers.)
4951 .. code-block:: llvm
4953 fence acquire ; yields {void}
4954 fence singlethread seq_cst ; yields {void}
4958 '``cmpxchg``' Instruction
4959 ^^^^^^^^^^^^^^^^^^^^^^^^^
4966 cmpxchg [volatile] <ty>* <pointer>, <ty> <cmp>, <ty> <new> [singlethread] <ordering> ; yields {ty}
4971 The '``cmpxchg``' instruction is used to atomically modify memory. It
4972 loads a value in memory and compares it to a given value. If they are
4973 equal, it stores a new value into the memory.
4978 There are three arguments to the '``cmpxchg``' instruction: an address
4979 to operate on, a value to compare to the value currently be at that
4980 address, and a new value to place at that address if the compared values
4981 are equal. The type of '<cmp>' must be an integer type whose bit width
4982 is a power of two greater than or equal to eight and less than or equal
4983 to a target-specific size limit. '<cmp>' and '<new>' must have the same
4984 type, and the type of '<pointer>' must be a pointer to that type. If the
4985 ``cmpxchg`` is marked as ``volatile``, then the optimizer is not allowed
4986 to modify the number or order of execution of this ``cmpxchg`` with
4987 other :ref:`volatile operations <volatile>`.
4989 The :ref:`ordering <ordering>` argument specifies how this ``cmpxchg``
4990 synchronizes with other atomic operations.
4992 The optional "``singlethread``" argument declares that the ``cmpxchg``
4993 is only atomic with respect to code (usually signal handlers) running in
4994 the same thread as the ``cmpxchg``. Otherwise the cmpxchg is atomic with
4995 respect to all other code in the system.
4997 The pointer passed into cmpxchg must have alignment greater than or
4998 equal to the size in memory of the operand.
5003 The contents of memory at the location specified by the '``<pointer>``'
5004 operand is read and compared to '``<cmp>``'; if the read value is the
5005 equal, '``<new>``' is written. The original value at the location is
5008 A successful ``cmpxchg`` is a read-modify-write instruction for the purpose
5009 of identifying release sequences. A failed ``cmpxchg`` is equivalent to an
5010 atomic load with an ordering parameter determined by dropping any
5011 ``release`` part of the ``cmpxchg``'s ordering.
5016 .. code-block:: llvm
5019 %orig = atomic load i32* %ptr unordered ; yields {i32}
5023 %cmp = phi i32 [ %orig, %entry ], [%old, %loop]
5024 %squared = mul i32 %cmp, %cmp
5025 %old = cmpxchg i32* %ptr, i32 %cmp, i32 %squared ; yields {i32}
5026 %success = icmp eq i32 %cmp, %old
5027 br i1 %success, label %done, label %loop
5034 '``atomicrmw``' Instruction
5035 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
5042 atomicrmw [volatile] <operation> <ty>* <pointer>, <ty> <value> [singlethread] <ordering> ; yields {ty}
5047 The '``atomicrmw``' instruction is used to atomically modify memory.
5052 There are three arguments to the '``atomicrmw``' instruction: an
5053 operation to apply, an address whose value to modify, an argument to the
5054 operation. The operation must be one of the following keywords:
5068 The type of '<value>' must be an integer type whose bit width is a power
5069 of two greater than or equal to eight and less than or equal to a
5070 target-specific size limit. The type of the '``<pointer>``' operand must
5071 be a pointer to that type. If the ``atomicrmw`` is marked as
5072 ``volatile``, then the optimizer is not allowed to modify the number or
5073 order of execution of this ``atomicrmw`` with other :ref:`volatile
5074 operations <volatile>`.
5079 The contents of memory at the location specified by the '``<pointer>``'
5080 operand are atomically read, modified, and written back. The original
5081 value at the location is returned. The modification is specified by the
5084 - xchg: ``*ptr = val``
5085 - add: ``*ptr = *ptr + val``
5086 - sub: ``*ptr = *ptr - val``
5087 - and: ``*ptr = *ptr & val``
5088 - nand: ``*ptr = ~(*ptr & val)``
5089 - or: ``*ptr = *ptr | val``
5090 - xor: ``*ptr = *ptr ^ val``
5091 - max: ``*ptr = *ptr > val ? *ptr : val`` (using a signed comparison)
5092 - min: ``*ptr = *ptr < val ? *ptr : val`` (using a signed comparison)
5093 - umax: ``*ptr = *ptr > val ? *ptr : val`` (using an unsigned
5095 - umin: ``*ptr = *ptr < val ? *ptr : val`` (using an unsigned
5101 .. code-block:: llvm
5103 %old = atomicrmw add i32* %ptr, i32 1 acquire ; yields {i32}
5105 .. _i_getelementptr:
5107 '``getelementptr``' Instruction
5108 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5115 <result> = getelementptr <pty>* <ptrval>{, <ty> <idx>}*
5116 <result> = getelementptr inbounds <pty>* <ptrval>{, <ty> <idx>}*
5117 <result> = getelementptr <ptr vector> ptrval, <vector index type> idx
5122 The '``getelementptr``' instruction is used to get the address of a
5123 subelement of an :ref:`aggregate <t_aggregate>` data structure. It performs
5124 address calculation only and does not access memory.
5129 The first argument is always a pointer or a vector of pointers, and
5130 forms the basis of the calculation. The remaining arguments are indices
5131 that indicate which of the elements of the aggregate object are indexed.
5132 The interpretation of each index is dependent on the type being indexed
5133 into. The first index always indexes the pointer value given as the
5134 first argument, the second index indexes a value of the type pointed to
5135 (not necessarily the value directly pointed to, since the first index
5136 can be non-zero), etc. The first type indexed into must be a pointer
5137 value, subsequent types can be arrays, vectors, and structs. Note that
5138 subsequent types being indexed into can never be pointers, since that
5139 would require loading the pointer before continuing calculation.
5141 The type of each index argument depends on the type it is indexing into.
5142 When indexing into a (optionally packed) structure, only ``i32`` integer
5143 **constants** are allowed (when using a vector of indices they must all
5144 be the **same** ``i32`` integer constant). When indexing into an array,
5145 pointer or vector, integers of any width are allowed, and they are not
5146 required to be constant. These integers are treated as signed values
5149 For example, let's consider a C code fragment and how it gets compiled
5165 int *foo(struct ST *s) {
5166 return &s[1].Z.B[5][13];
5169 The LLVM code generated by Clang is:
5171 .. code-block:: llvm
5173 %struct.RT = type { i8, [10 x [20 x i32]], i8 }
5174 %struct.ST = type { i32, double, %struct.RT }
5176 define i32* @foo(%struct.ST* %s) nounwind uwtable readnone optsize ssp {
5178 %arrayidx = getelementptr inbounds %struct.ST* %s, i64 1, i32 2, i32 1, i64 5, i64 13
5185 In the example above, the first index is indexing into the
5186 '``%struct.ST*``' type, which is a pointer, yielding a '``%struct.ST``'
5187 = '``{ i32, double, %struct.RT }``' type, a structure. The second index
5188 indexes into the third element of the structure, yielding a
5189 '``%struct.RT``' = '``{ i8 , [10 x [20 x i32]], i8 }``' type, another
5190 structure. The third index indexes into the second element of the
5191 structure, yielding a '``[10 x [20 x i32]]``' type, an array. The two
5192 dimensions of the array are subscripted into, yielding an '``i32``'
5193 type. The '``getelementptr``' instruction returns a pointer to this
5194 element, thus computing a value of '``i32*``' type.
5196 Note that it is perfectly legal to index partially through a structure,
5197 returning a pointer to an inner element. Because of this, the LLVM code
5198 for the given testcase is equivalent to:
5200 .. code-block:: llvm
5202 define i32* @foo(%struct.ST* %s) {
5203 %t1 = getelementptr %struct.ST* %s, i32 1 ; yields %struct.ST*:%t1
5204 %t2 = getelementptr %struct.ST* %t1, i32 0, i32 2 ; yields %struct.RT*:%t2
5205 %t3 = getelementptr %struct.RT* %t2, i32 0, i32 1 ; yields [10 x [20 x i32]]*:%t3
5206 %t4 = getelementptr [10 x [20 x i32]]* %t3, i32 0, i32 5 ; yields [20 x i32]*:%t4
5207 %t5 = getelementptr [20 x i32]* %t4, i32 0, i32 13 ; yields i32*:%t5
5211 If the ``inbounds`` keyword is present, the result value of the
5212 ``getelementptr`` is a :ref:`poison value <poisonvalues>` if the base
5213 pointer is not an *in bounds* address of an allocated object, or if any
5214 of the addresses that would be formed by successive addition of the
5215 offsets implied by the indices to the base address with infinitely
5216 precise signed arithmetic are not an *in bounds* address of that
5217 allocated object. The *in bounds* addresses for an allocated object are
5218 all the addresses that point into the object, plus the address one byte
5219 past the end. In cases where the base is a vector of pointers the
5220 ``inbounds`` keyword applies to each of the computations element-wise.
5222 If the ``inbounds`` keyword is not present, the offsets are added to the
5223 base address with silently-wrapping two's complement arithmetic. If the
5224 offsets have a different width from the pointer, they are sign-extended
5225 or truncated to the width of the pointer. The result value of the
5226 ``getelementptr`` may be outside the object pointed to by the base
5227 pointer. The result value may not necessarily be used to access memory
5228 though, even if it happens to point into allocated storage. See the
5229 :ref:`Pointer Aliasing Rules <pointeraliasing>` section for more
5232 The getelementptr instruction is often confusing. For some more insight
5233 into how it works, see :doc:`the getelementptr FAQ <GetElementPtr>`.
5238 .. code-block:: llvm
5240 ; yields [12 x i8]*:aptr
5241 %aptr = getelementptr {i32, [12 x i8]}* %saptr, i64 0, i32 1
5243 %vptr = getelementptr {i32, <2 x i8>}* %svptr, i64 0, i32 1, i32 1
5245 %eptr = getelementptr [12 x i8]* %aptr, i64 0, i32 1
5247 %iptr = getelementptr [10 x i32]* @arr, i16 0, i16 0
5249 In cases where the pointer argument is a vector of pointers, each index
5250 must be a vector with the same number of elements. For example:
5252 .. code-block:: llvm
5254 %A = getelementptr <4 x i8*> %ptrs, <4 x i64> %offsets,
5256 Conversion Operations
5257 ---------------------
5259 The instructions in this category are the conversion instructions
5260 (casting) which all take a single operand and a type. They perform
5261 various bit conversions on the operand.
5263 '``trunc .. to``' Instruction
5264 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5271 <result> = trunc <ty> <value> to <ty2> ; yields ty2
5276 The '``trunc``' instruction truncates its operand to the type ``ty2``.
5281 The '``trunc``' instruction takes a value to trunc, and a type to trunc
5282 it to. Both types must be of :ref:`integer <t_integer>` types, or vectors
5283 of the same number of integers. The bit size of the ``value`` must be
5284 larger than the bit size of the destination type, ``ty2``. Equal sized
5285 types are not allowed.
5290 The '``trunc``' instruction truncates the high order bits in ``value``
5291 and converts the remaining bits to ``ty2``. Since the source size must
5292 be larger than the destination size, ``trunc`` cannot be a *no-op cast*.
5293 It will always truncate bits.
5298 .. code-block:: llvm
5300 %X = trunc i32 257 to i8 ; yields i8:1
5301 %Y = trunc i32 123 to i1 ; yields i1:true
5302 %Z = trunc i32 122 to i1 ; yields i1:false
5303 %W = trunc <2 x i16> <i16 8, i16 7> to <2 x i8> ; yields <i8 8, i8 7>
5305 '``zext .. to``' Instruction
5306 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5313 <result> = zext <ty> <value> to <ty2> ; yields ty2
5318 The '``zext``' instruction zero extends its operand to type ``ty2``.
5323 The '``zext``' instruction takes a value to cast, and a type to cast it
5324 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
5325 the same number of integers. The bit size of the ``value`` must be
5326 smaller than the bit size of the destination type, ``ty2``.
5331 The ``zext`` fills the high order bits of the ``value`` with zero bits
5332 until it reaches the size of the destination type, ``ty2``.
5334 When zero extending from i1, the result will always be either 0 or 1.
5339 .. code-block:: llvm
5341 %X = zext i32 257 to i64 ; yields i64:257
5342 %Y = zext i1 true to i32 ; yields i32:1
5343 %Z = zext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5345 '``sext .. to``' Instruction
5346 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5353 <result> = sext <ty> <value> to <ty2> ; yields ty2
5358 The '``sext``' sign extends ``value`` to the type ``ty2``.
5363 The '``sext``' instruction takes a value to cast, and a type to cast it
5364 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
5365 the same number of integers. The bit size of the ``value`` must be
5366 smaller than the bit size of the destination type, ``ty2``.
5371 The '``sext``' instruction performs a sign extension by copying the sign
5372 bit (highest order bit) of the ``value`` until it reaches the bit size
5373 of the type ``ty2``.
5375 When sign extending from i1, the extension always results in -1 or 0.
5380 .. code-block:: llvm
5382 %X = sext i8 -1 to i16 ; yields i16 :65535
5383 %Y = sext i1 true to i32 ; yields i32:-1
5384 %Z = sext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5386 '``fptrunc .. to``' Instruction
5387 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5394 <result> = fptrunc <ty> <value> to <ty2> ; yields ty2
5399 The '``fptrunc``' instruction truncates ``value`` to type ``ty2``.
5404 The '``fptrunc``' instruction takes a :ref:`floating point <t_floating>`
5405 value to cast and a :ref:`floating point <t_floating>` type to cast it to.
5406 The size of ``value`` must be larger than the size of ``ty2``. This
5407 implies that ``fptrunc`` cannot be used to make a *no-op cast*.
5412 The '``fptrunc``' instruction truncates a ``value`` from a larger
5413 :ref:`floating point <t_floating>` type to a smaller :ref:`floating
5414 point <t_floating>` type. If the value cannot fit within the
5415 destination type, ``ty2``, then the results are undefined.
5420 .. code-block:: llvm
5422 %X = fptrunc double 123.0 to float ; yields float:123.0
5423 %Y = fptrunc double 1.0E+300 to float ; yields undefined
5425 '``fpext .. to``' Instruction
5426 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5433 <result> = fpext <ty> <value> to <ty2> ; yields ty2
5438 The '``fpext``' extends a floating point ``value`` to a larger floating
5444 The '``fpext``' instruction takes a :ref:`floating point <t_floating>`
5445 ``value`` to cast, and a :ref:`floating point <t_floating>` type to cast it
5446 to. The source type must be smaller than the destination type.
5451 The '``fpext``' instruction extends the ``value`` from a smaller
5452 :ref:`floating point <t_floating>` type to a larger :ref:`floating
5453 point <t_floating>` type. The ``fpext`` cannot be used to make a
5454 *no-op cast* because it always changes bits. Use ``bitcast`` to make a
5455 *no-op cast* for a floating point cast.
5460 .. code-block:: llvm
5462 %X = fpext float 3.125 to double ; yields double:3.125000e+00
5463 %Y = fpext double %X to fp128 ; yields fp128:0xL00000000000000004000900000000000
5465 '``fptoui .. to``' Instruction
5466 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5473 <result> = fptoui <ty> <value> to <ty2> ; yields ty2
5478 The '``fptoui``' converts a floating point ``value`` to its unsigned
5479 integer equivalent of type ``ty2``.
5484 The '``fptoui``' instruction takes a value to cast, which must be a
5485 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5486 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5487 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5488 type with the same number of elements as ``ty``
5493 The '``fptoui``' instruction converts its :ref:`floating
5494 point <t_floating>` operand into the nearest (rounding towards zero)
5495 unsigned integer value. If the value cannot fit in ``ty2``, the results
5501 .. code-block:: llvm
5503 %X = fptoui double 123.0 to i32 ; yields i32:123
5504 %Y = fptoui float 1.0E+300 to i1 ; yields undefined:1
5505 %Z = fptoui float 1.04E+17 to i8 ; yields undefined:1
5507 '``fptosi .. to``' Instruction
5508 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5515 <result> = fptosi <ty> <value> to <ty2> ; yields ty2
5520 The '``fptosi``' instruction converts :ref:`floating point <t_floating>`
5521 ``value`` to type ``ty2``.
5526 The '``fptosi``' instruction takes a value to cast, which must be a
5527 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5528 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5529 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5530 type with the same number of elements as ``ty``
5535 The '``fptosi``' instruction converts its :ref:`floating
5536 point <t_floating>` operand into the nearest (rounding towards zero)
5537 signed integer value. If the value cannot fit in ``ty2``, the results
5543 .. code-block:: llvm
5545 %X = fptosi double -123.0 to i32 ; yields i32:-123
5546 %Y = fptosi float 1.0E-247 to i1 ; yields undefined:1
5547 %Z = fptosi float 1.04E+17 to i8 ; yields undefined:1
5549 '``uitofp .. to``' Instruction
5550 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5557 <result> = uitofp <ty> <value> to <ty2> ; yields ty2
5562 The '``uitofp``' instruction regards ``value`` as an unsigned integer
5563 and converts that value to the ``ty2`` type.
5568 The '``uitofp``' instruction takes a value to cast, which must be a
5569 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
5570 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
5571 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
5572 type with the same number of elements as ``ty``
5577 The '``uitofp``' instruction interprets its operand as an unsigned
5578 integer quantity and converts it to the corresponding floating point
5579 value. If the value cannot fit in the floating point value, the results
5585 .. code-block:: llvm
5587 %X = uitofp i32 257 to float ; yields float:257.0
5588 %Y = uitofp i8 -1 to double ; yields double:255.0
5590 '``sitofp .. to``' Instruction
5591 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5598 <result> = sitofp <ty> <value> to <ty2> ; yields ty2
5603 The '``sitofp``' instruction regards ``value`` as a signed integer and
5604 converts that value to the ``ty2`` type.
5609 The '``sitofp``' instruction takes a value to cast, which must be a
5610 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
5611 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
5612 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
5613 type with the same number of elements as ``ty``
5618 The '``sitofp``' instruction interprets its operand as a signed integer
5619 quantity and converts it to the corresponding floating point value. If
5620 the value cannot fit in the floating point value, the results are
5626 .. code-block:: llvm
5628 %X = sitofp i32 257 to float ; yields float:257.0
5629 %Y = sitofp i8 -1 to double ; yields double:-1.0
5633 '``ptrtoint .. to``' Instruction
5634 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5641 <result> = ptrtoint <ty> <value> to <ty2> ; yields ty2
5646 The '``ptrtoint``' instruction converts the pointer or a vector of
5647 pointers ``value`` to the integer (or vector of integers) type ``ty2``.
5652 The '``ptrtoint``' instruction takes a ``value`` to cast, which must be
5653 a a value of type :ref:`pointer <t_pointer>` or a vector of pointers, and a
5654 type to cast it to ``ty2``, which must be an :ref:`integer <t_integer>` or
5655 a vector of integers type.
5660 The '``ptrtoint``' instruction converts ``value`` to integer type
5661 ``ty2`` by interpreting the pointer value as an integer and either
5662 truncating or zero extending that value to the size of the integer type.
5663 If ``value`` is smaller than ``ty2`` then a zero extension is done. If
5664 ``value`` is larger than ``ty2`` then a truncation is done. If they are
5665 the same size, then nothing is done (*no-op cast*) other than a type
5671 .. code-block:: llvm
5673 %X = ptrtoint i32* %P to i8 ; yields truncation on 32-bit architecture
5674 %Y = ptrtoint i32* %P to i64 ; yields zero extension on 32-bit architecture
5675 %Z = ptrtoint <4 x i32*> %P to <4 x i64>; yields vector zero extension for a vector of addresses on 32-bit architecture
5679 '``inttoptr .. to``' Instruction
5680 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5687 <result> = inttoptr <ty> <value> to <ty2> ; yields ty2
5692 The '``inttoptr``' instruction converts an integer ``value`` to a
5693 pointer type, ``ty2``.
5698 The '``inttoptr``' instruction takes an :ref:`integer <t_integer>` value to
5699 cast, and a type to cast it to, which must be a :ref:`pointer <t_pointer>`
5705 The '``inttoptr``' instruction converts ``value`` to type ``ty2`` by
5706 applying either a zero extension or a truncation depending on the size
5707 of the integer ``value``. If ``value`` is larger than the size of a
5708 pointer then a truncation is done. If ``value`` is smaller than the size
5709 of a pointer then a zero extension is done. If they are the same size,
5710 nothing is done (*no-op cast*).
5715 .. code-block:: llvm
5717 %X = inttoptr i32 255 to i32* ; yields zero extension on 64-bit architecture
5718 %Y = inttoptr i32 255 to i32* ; yields no-op on 32-bit architecture
5719 %Z = inttoptr i64 0 to i32* ; yields truncation on 32-bit architecture
5720 %Z = inttoptr <4 x i32> %G to <4 x i8*>; yields truncation of vector G to four pointers
5724 '``bitcast .. to``' Instruction
5725 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5732 <result> = bitcast <ty> <value> to <ty2> ; yields ty2
5737 The '``bitcast``' instruction converts ``value`` to type ``ty2`` without
5743 The '``bitcast``' instruction takes a value to cast, which must be a
5744 non-aggregate first class value, and a type to cast it to, which must
5745 also be a non-aggregate :ref:`first class <t_firstclass>` type. The
5746 bit sizes of ``value`` and the destination type, ``ty2``, must be
5747 identical. If the source type is a pointer, the destination type must
5748 also be a pointer of the same size. This instruction supports bitwise
5749 conversion of vectors to integers and to vectors of other types (as
5750 long as they have the same size).
5755 The '``bitcast``' instruction converts ``value`` to type ``ty2``. It
5756 is always a *no-op cast* because no bits change with this
5757 conversion. The conversion is done as if the ``value`` had been stored
5758 to memory and read back as type ``ty2``. Pointer (or vector of
5759 pointers) types may only be converted to other pointer (or vector of
5760 pointers) types with the same address space through this instruction.
5761 To convert pointers to other types, use the :ref:`inttoptr <i_inttoptr>`
5762 or :ref:`ptrtoint <i_ptrtoint>` instructions first.
5767 .. code-block:: llvm
5769 %X = bitcast i8 255 to i8 ; yields i8 :-1
5770 %Y = bitcast i32* %x to sint* ; yields sint*:%x
5771 %Z = bitcast <2 x int> %V to i64; ; yields i64: %V
5772 %Z = bitcast <2 x i32*> %V to <2 x i64*> ; yields <2 x i64*>
5774 .. _i_addrspacecast:
5776 '``addrspacecast .. to``' Instruction
5777 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5784 <result> = addrspacecast <pty> <ptrval> to <pty2> ; yields pty2
5789 The '``addrspacecast``' instruction converts ``ptrval`` from ``pty`` in
5790 address space ``n`` to type ``pty2`` in address space ``m``.
5795 The '``addrspacecast``' instruction takes a pointer or vector of pointer value
5796 to cast and a pointer type to cast it to, which must have a different
5802 The '``addrspacecast``' instruction converts the pointer value
5803 ``ptrval`` to type ``pty2``. It can be a *no-op cast* or a complex
5804 value modification, depending on the target and the address space
5805 pair. Pointer conversions within the same address space must be
5806 performed with the ``bitcast`` instruction. Note that if the address space
5807 conversion is legal then both result and operand refer to the same memory
5813 .. code-block:: llvm
5815 %X = addrspacecast i32* %x to i32 addrspace(1)* ; yields i32 addrspace(1)*:%x
5816 %Y = addrspacecast i32 addrspace(1)* %y to i64 addrspace(2)* ; yields i64 addrspace(2)*:%y
5817 %Z = addrspacecast <4 x i32*> %z to <4 x float addrspace(3)*> ; yields <4 x float addrspace(3)*>:%z
5824 The instructions in this category are the "miscellaneous" instructions,
5825 which defy better classification.
5829 '``icmp``' Instruction
5830 ^^^^^^^^^^^^^^^^^^^^^^
5837 <result> = icmp <cond> <ty> <op1>, <op2> ; yields {i1} or {<N x i1>}:result
5842 The '``icmp``' instruction returns a boolean value or a vector of
5843 boolean values based on comparison of its two integer, integer vector,
5844 pointer, or pointer vector operands.
5849 The '``icmp``' instruction takes three operands. The first operand is
5850 the condition code indicating the kind of comparison to perform. It is
5851 not a value, just a keyword. The possible condition code are:
5854 #. ``ne``: not equal
5855 #. ``ugt``: unsigned greater than
5856 #. ``uge``: unsigned greater or equal
5857 #. ``ult``: unsigned less than
5858 #. ``ule``: unsigned less or equal
5859 #. ``sgt``: signed greater than
5860 #. ``sge``: signed greater or equal
5861 #. ``slt``: signed less than
5862 #. ``sle``: signed less or equal
5864 The remaining two arguments must be :ref:`integer <t_integer>` or
5865 :ref:`pointer <t_pointer>` or integer :ref:`vector <t_vector>` typed. They
5866 must also be identical types.
5871 The '``icmp``' compares ``op1`` and ``op2`` according to the condition
5872 code given as ``cond``. The comparison performed always yields either an
5873 :ref:`i1 <t_integer>` or vector of ``i1`` result, as follows:
5875 #. ``eq``: yields ``true`` if the operands are equal, ``false``
5876 otherwise. No sign interpretation is necessary or performed.
5877 #. ``ne``: yields ``true`` if the operands are unequal, ``false``
5878 otherwise. No sign interpretation is necessary or performed.
5879 #. ``ugt``: interprets the operands as unsigned values and yields
5880 ``true`` if ``op1`` is greater than ``op2``.
5881 #. ``uge``: interprets the operands as unsigned values and yields
5882 ``true`` if ``op1`` is greater than or equal to ``op2``.
5883 #. ``ult``: interprets the operands as unsigned values and yields
5884 ``true`` if ``op1`` is less than ``op2``.
5885 #. ``ule``: interprets the operands as unsigned values and yields
5886 ``true`` if ``op1`` is less than or equal to ``op2``.
5887 #. ``sgt``: interprets the operands as signed values and yields ``true``
5888 if ``op1`` is greater than ``op2``.
5889 #. ``sge``: interprets the operands as signed values and yields ``true``
5890 if ``op1`` is greater than or equal to ``op2``.
5891 #. ``slt``: interprets the operands as signed values and yields ``true``
5892 if ``op1`` is less than ``op2``.
5893 #. ``sle``: interprets the operands as signed values and yields ``true``
5894 if ``op1`` is less than or equal to ``op2``.
5896 If the operands are :ref:`pointer <t_pointer>` typed, the pointer values
5897 are compared as if they were integers.
5899 If the operands are integer vectors, then they are compared element by
5900 element. The result is an ``i1`` vector with the same number of elements
5901 as the values being compared. Otherwise, the result is an ``i1``.
5906 .. code-block:: llvm
5908 <result> = icmp eq i32 4, 5 ; yields: result=false
5909 <result> = icmp ne float* %X, %X ; yields: result=false
5910 <result> = icmp ult i16 4, 5 ; yields: result=true
5911 <result> = icmp sgt i16 4, 5 ; yields: result=false
5912 <result> = icmp ule i16 -4, 5 ; yields: result=false
5913 <result> = icmp sge i16 4, 5 ; yields: result=false
5915 Note that the code generator does not yet support vector types with the
5916 ``icmp`` instruction.
5920 '``fcmp``' Instruction
5921 ^^^^^^^^^^^^^^^^^^^^^^
5928 <result> = fcmp <cond> <ty> <op1>, <op2> ; yields {i1} or {<N x i1>}:result
5933 The '``fcmp``' instruction returns a boolean value or vector of boolean
5934 values based on comparison of its operands.
5936 If the operands are floating point scalars, then the result type is a
5937 boolean (:ref:`i1 <t_integer>`).
5939 If the operands are floating point vectors, then the result type is a
5940 vector of boolean with the same number of elements as the operands being
5946 The '``fcmp``' instruction takes three operands. The first operand is
5947 the condition code indicating the kind of comparison to perform. It is
5948 not a value, just a keyword. The possible condition code are:
5950 #. ``false``: no comparison, always returns false
5951 #. ``oeq``: ordered and equal
5952 #. ``ogt``: ordered and greater than
5953 #. ``oge``: ordered and greater than or equal
5954 #. ``olt``: ordered and less than
5955 #. ``ole``: ordered and less than or equal
5956 #. ``one``: ordered and not equal
5957 #. ``ord``: ordered (no nans)
5958 #. ``ueq``: unordered or equal
5959 #. ``ugt``: unordered or greater than
5960 #. ``uge``: unordered or greater than or equal
5961 #. ``ult``: unordered or less than
5962 #. ``ule``: unordered or less than or equal
5963 #. ``une``: unordered or not equal
5964 #. ``uno``: unordered (either nans)
5965 #. ``true``: no comparison, always returns true
5967 *Ordered* means that neither operand is a QNAN while *unordered* means
5968 that either operand may be a QNAN.
5970 Each of ``val1`` and ``val2`` arguments must be either a :ref:`floating
5971 point <t_floating>` type or a :ref:`vector <t_vector>` of floating point
5972 type. They must have identical types.
5977 The '``fcmp``' instruction compares ``op1`` and ``op2`` according to the
5978 condition code given as ``cond``. If the operands are vectors, then the
5979 vectors are compared element by element. Each comparison performed
5980 always yields an :ref:`i1 <t_integer>` result, as follows:
5982 #. ``false``: always yields ``false``, regardless of operands.
5983 #. ``oeq``: yields ``true`` if both operands are not a QNAN and ``op1``
5984 is equal to ``op2``.
5985 #. ``ogt``: yields ``true`` if both operands are not a QNAN and ``op1``
5986 is greater than ``op2``.
5987 #. ``oge``: yields ``true`` if both operands are not a QNAN and ``op1``
5988 is greater than or equal to ``op2``.
5989 #. ``olt``: yields ``true`` if both operands are not a QNAN and ``op1``
5990 is less than ``op2``.
5991 #. ``ole``: yields ``true`` if both operands are not a QNAN and ``op1``
5992 is less than or equal to ``op2``.
5993 #. ``one``: yields ``true`` if both operands are not a QNAN and ``op1``
5994 is not equal to ``op2``.
5995 #. ``ord``: yields ``true`` if both operands are not a QNAN.
5996 #. ``ueq``: yields ``true`` if either operand is a QNAN or ``op1`` is
5998 #. ``ugt``: yields ``true`` if either operand is a QNAN or ``op1`` is
5999 greater than ``op2``.
6000 #. ``uge``: yields ``true`` if either operand is a QNAN or ``op1`` is
6001 greater than or equal to ``op2``.
6002 #. ``ult``: yields ``true`` if either operand is a QNAN or ``op1`` is
6004 #. ``ule``: yields ``true`` if either operand is a QNAN or ``op1`` is
6005 less than or equal to ``op2``.
6006 #. ``une``: yields ``true`` if either operand is a QNAN or ``op1`` is
6007 not equal to ``op2``.
6008 #. ``uno``: yields ``true`` if either operand is a QNAN.
6009 #. ``true``: always yields ``true``, regardless of operands.
6014 .. code-block:: llvm
6016 <result> = fcmp oeq float 4.0, 5.0 ; yields: result=false
6017 <result> = fcmp one float 4.0, 5.0 ; yields: result=true
6018 <result> = fcmp olt float 4.0, 5.0 ; yields: result=true
6019 <result> = fcmp ueq double 1.0, 2.0 ; yields: result=false
6021 Note that the code generator does not yet support vector types with the
6022 ``fcmp`` instruction.
6026 '``phi``' Instruction
6027 ^^^^^^^^^^^^^^^^^^^^^
6034 <result> = phi <ty> [ <val0>, <label0>], ...
6039 The '``phi``' instruction is used to implement the φ node in the SSA
6040 graph representing the function.
6045 The type of the incoming values is specified with the first type field.
6046 After this, the '``phi``' instruction takes a list of pairs as
6047 arguments, with one pair for each predecessor basic block of the current
6048 block. Only values of :ref:`first class <t_firstclass>` type may be used as
6049 the value arguments to the PHI node. Only labels may be used as the
6052 There must be no non-phi instructions between the start of a basic block
6053 and the PHI instructions: i.e. PHI instructions must be first in a basic
6056 For the purposes of the SSA form, the use of each incoming value is
6057 deemed to occur on the edge from the corresponding predecessor block to
6058 the current block (but after any definition of an '``invoke``'
6059 instruction's return value on the same edge).
6064 At runtime, the '``phi``' instruction logically takes on the value
6065 specified by the pair corresponding to the predecessor basic block that
6066 executed just prior to the current block.
6071 .. code-block:: llvm
6073 Loop: ; Infinite loop that counts from 0 on up...
6074 %indvar = phi i32 [ 0, %LoopHeader ], [ %nextindvar, %Loop ]
6075 %nextindvar = add i32 %indvar, 1
6080 '``select``' Instruction
6081 ^^^^^^^^^^^^^^^^^^^^^^^^
6088 <result> = select selty <cond>, <ty> <val1>, <ty> <val2> ; yields ty
6090 selty is either i1 or {<N x i1>}
6095 The '``select``' instruction is used to choose one value based on a
6096 condition, without branching.
6101 The '``select``' instruction requires an 'i1' value or a vector of 'i1'
6102 values indicating the condition, and two values of the same :ref:`first
6103 class <t_firstclass>` type. If the val1/val2 are vectors and the
6104 condition is a scalar, then entire vectors are selected, not individual
6110 If the condition is an i1 and it evaluates to 1, the instruction returns
6111 the first value argument; otherwise, it returns the second value
6114 If the condition is a vector of i1, then the value arguments must be
6115 vectors of the same size, and the selection is done element by element.
6120 .. code-block:: llvm
6122 %X = select i1 true, i8 17, i8 42 ; yields i8:17
6126 '``call``' Instruction
6127 ^^^^^^^^^^^^^^^^^^^^^^
6134 <result> = [tail] call [cconv] [ret attrs] <ty> [<fnty>*] <fnptrval>(<function args>) [fn attrs]
6139 The '``call``' instruction represents a simple function call.
6144 This instruction requires several arguments:
6146 #. The optional "tail" marker indicates that the callee function does
6147 not access any allocas or varargs in the caller. Note that calls may
6148 be marked "tail" even if they do not occur before a
6149 :ref:`ret <i_ret>` instruction. If the "tail" marker is present, the
6150 function call is eligible for tail call optimization, but `might not
6151 in fact be optimized into a jump <CodeGenerator.html#tailcallopt>`_.
6152 The code generator may optimize calls marked "tail" with either 1)
6153 automatic `sibling call
6154 optimization <CodeGenerator.html#sibcallopt>`_ when the caller and
6155 callee have matching signatures, or 2) forced tail call optimization
6156 when the following extra requirements are met:
6158 - Caller and callee both have the calling convention ``fastcc``.
6159 - The call is in tail position (ret immediately follows call and ret
6160 uses value of call or is void).
6161 - Option ``-tailcallopt`` is enabled, or
6162 ``llvm::GuaranteedTailCallOpt`` is ``true``.
6163 - `Platform specific constraints are
6164 met. <CodeGenerator.html#tailcallopt>`_
6166 #. The optional "cconv" marker indicates which :ref:`calling
6167 convention <callingconv>` the call should use. If none is
6168 specified, the call defaults to using C calling conventions. The
6169 calling convention of the call must match the calling convention of
6170 the target function, or else the behavior is undefined.
6171 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
6172 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
6174 #. '``ty``': the type of the call instruction itself which is also the
6175 type of the return value. Functions that return no value are marked
6177 #. '``fnty``': shall be the signature of the pointer to function value
6178 being invoked. The argument types must match the types implied by
6179 this signature. This type can be omitted if the function is not
6180 varargs and if the function type does not return a pointer to a
6182 #. '``fnptrval``': An LLVM value containing a pointer to a function to
6183 be invoked. In most cases, this is a direct function invocation, but
6184 indirect ``call``'s are just as possible, calling an arbitrary pointer
6186 #. '``function args``': argument list whose types match the function
6187 signature argument types and parameter attributes. All arguments must
6188 be of :ref:`first class <t_firstclass>` type. If the function signature
6189 indicates the function accepts a variable number of arguments, the
6190 extra arguments can be specified.
6191 #. The optional :ref:`function attributes <fnattrs>` list. Only
6192 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
6193 attributes are valid here.
6198 The '``call``' instruction is used to cause control flow to transfer to
6199 a specified function, with its incoming arguments bound to the specified
6200 values. Upon a '``ret``' instruction in the called function, control
6201 flow continues with the instruction after the function call, and the
6202 return value of the function is bound to the result argument.
6207 .. code-block:: llvm
6209 %retval = call i32 @test(i32 %argc)
6210 call i32 (i8*, ...)* @printf(i8* %msg, i32 12, i8 42) ; yields i32
6211 %X = tail call i32 @foo() ; yields i32
6212 %Y = tail call fastcc i32 @foo() ; yields i32
6213 call void %foo(i8 97 signext)
6215 %struct.A = type { i32, i8 }
6216 %r = call %struct.A @foo() ; yields { 32, i8 }
6217 %gr = extractvalue %struct.A %r, 0 ; yields i32
6218 %gr1 = extractvalue %struct.A %r, 1 ; yields i8
6219 %Z = call void @foo() noreturn ; indicates that %foo never returns normally
6220 %ZZ = call zeroext i32 @bar() ; Return value is %zero extended
6222 llvm treats calls to some functions with names and arguments that match
6223 the standard C99 library as being the C99 library functions, and may
6224 perform optimizations or generate code for them under that assumption.
6225 This is something we'd like to change in the future to provide better
6226 support for freestanding environments and non-C-based languages.
6230 '``va_arg``' Instruction
6231 ^^^^^^^^^^^^^^^^^^^^^^^^
6238 <resultval> = va_arg <va_list*> <arglist>, <argty>
6243 The '``va_arg``' instruction is used to access arguments passed through
6244 the "variable argument" area of a function call. It is used to implement
6245 the ``va_arg`` macro in C.
6250 This instruction takes a ``va_list*`` value and the type of the
6251 argument. It returns a value of the specified argument type and
6252 increments the ``va_list`` to point to the next argument. The actual
6253 type of ``va_list`` is target specific.
6258 The '``va_arg``' instruction loads an argument of the specified type
6259 from the specified ``va_list`` and causes the ``va_list`` to point to
6260 the next argument. For more information, see the variable argument
6261 handling :ref:`Intrinsic Functions <int_varargs>`.
6263 It is legal for this instruction to be called in a function which does
6264 not take a variable number of arguments, for example, the ``vfprintf``
6267 ``va_arg`` is an LLVM instruction instead of an :ref:`intrinsic
6268 function <intrinsics>` because it takes a type as an argument.
6273 See the :ref:`variable argument processing <int_varargs>` section.
6275 Note that the code generator does not yet fully support va\_arg on many
6276 targets. Also, it does not currently support va\_arg with aggregate
6277 types on any target.
6281 '``landingpad``' Instruction
6282 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6289 <resultval> = landingpad <resultty> personality <type> <pers_fn> <clause>+
6290 <resultval> = landingpad <resultty> personality <type> <pers_fn> cleanup <clause>*
6292 <clause> := catch <type> <value>
6293 <clause> := filter <array constant type> <array constant>
6298 The '``landingpad``' instruction is used by `LLVM's exception handling
6299 system <ExceptionHandling.html#overview>`_ to specify that a basic block
6300 is a landing pad --- one where the exception lands, and corresponds to the
6301 code found in the ``catch`` portion of a ``try``/``catch`` sequence. It
6302 defines values supplied by the personality function (``pers_fn``) upon
6303 re-entry to the function. The ``resultval`` has the type ``resultty``.
6308 This instruction takes a ``pers_fn`` value. This is the personality
6309 function associated with the unwinding mechanism. The optional
6310 ``cleanup`` flag indicates that the landing pad block is a cleanup.
6312 A ``clause`` begins with the clause type --- ``catch`` or ``filter`` --- and
6313 contains the global variable representing the "type" that may be caught
6314 or filtered respectively. Unlike the ``catch`` clause, the ``filter``
6315 clause takes an array constant as its argument. Use
6316 "``[0 x i8**] undef``" for a filter which cannot throw. The
6317 '``landingpad``' instruction must contain *at least* one ``clause`` or
6318 the ``cleanup`` flag.
6323 The '``landingpad``' instruction defines the values which are set by the
6324 personality function (``pers_fn``) upon re-entry to the function, and
6325 therefore the "result type" of the ``landingpad`` instruction. As with
6326 calling conventions, how the personality function results are
6327 represented in LLVM IR is target specific.
6329 The clauses are applied in order from top to bottom. If two
6330 ``landingpad`` instructions are merged together through inlining, the
6331 clauses from the calling function are appended to the list of clauses.
6332 When the call stack is being unwound due to an exception being thrown,
6333 the exception is compared against each ``clause`` in turn. If it doesn't
6334 match any of the clauses, and the ``cleanup`` flag is not set, then
6335 unwinding continues further up the call stack.
6337 The ``landingpad`` instruction has several restrictions:
6339 - A landing pad block is a basic block which is the unwind destination
6340 of an '``invoke``' instruction.
6341 - A landing pad block must have a '``landingpad``' instruction as its
6342 first non-PHI instruction.
6343 - There can be only one '``landingpad``' instruction within the landing
6345 - A basic block that is not a landing pad block may not include a
6346 '``landingpad``' instruction.
6347 - All '``landingpad``' instructions in a function must have the same
6348 personality function.
6353 .. code-block:: llvm
6355 ;; A landing pad which can catch an integer.
6356 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6358 ;; A landing pad that is a cleanup.
6359 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6361 ;; A landing pad which can catch an integer and can only throw a double.
6362 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6364 filter [1 x i8**] [@_ZTId]
6371 LLVM supports the notion of an "intrinsic function". These functions
6372 have well known names and semantics and are required to follow certain
6373 restrictions. Overall, these intrinsics represent an extension mechanism
6374 for the LLVM language that does not require changing all of the
6375 transformations in LLVM when adding to the language (or the bitcode
6376 reader/writer, the parser, etc...).
6378 Intrinsic function names must all start with an "``llvm.``" prefix. This
6379 prefix is reserved in LLVM for intrinsic names; thus, function names may
6380 not begin with this prefix. Intrinsic functions must always be external
6381 functions: you cannot define the body of intrinsic functions. Intrinsic
6382 functions may only be used in call or invoke instructions: it is illegal
6383 to take the address of an intrinsic function. Additionally, because
6384 intrinsic functions are part of the LLVM language, it is required if any
6385 are added that they be documented here.
6387 Some intrinsic functions can be overloaded, i.e., the intrinsic
6388 represents a family of functions that perform the same operation but on
6389 different data types. Because LLVM can represent over 8 million
6390 different integer types, overloading is used commonly to allow an
6391 intrinsic function to operate on any integer type. One or more of the
6392 argument types or the result type can be overloaded to accept any
6393 integer type. Argument types may also be defined as exactly matching a
6394 previous argument's type or the result type. This allows an intrinsic
6395 function which accepts multiple arguments, but needs all of them to be
6396 of the same type, to only be overloaded with respect to a single
6397 argument or the result.
6399 Overloaded intrinsics will have the names of its overloaded argument
6400 types encoded into its function name, each preceded by a period. Only
6401 those types which are overloaded result in a name suffix. Arguments
6402 whose type is matched against another type do not. For example, the
6403 ``llvm.ctpop`` function can take an integer of any width and returns an
6404 integer of exactly the same integer width. This leads to a family of
6405 functions such as ``i8 @llvm.ctpop.i8(i8 %val)`` and
6406 ``i29 @llvm.ctpop.i29(i29 %val)``. Only one type, the return type, is
6407 overloaded, and only one type suffix is required. Because the argument's
6408 type is matched against the return type, it does not require its own
6411 To learn how to add an intrinsic function, please see the `Extending
6412 LLVM Guide <ExtendingLLVM.html>`_.
6416 Variable Argument Handling Intrinsics
6417 -------------------------------------
6419 Variable argument support is defined in LLVM with the
6420 :ref:`va_arg <i_va_arg>` instruction and these three intrinsic
6421 functions. These functions are related to the similarly named macros
6422 defined in the ``<stdarg.h>`` header file.
6424 All of these functions operate on arguments that use a target-specific
6425 value type "``va_list``". The LLVM assembly language reference manual
6426 does not define what this type is, so all transformations should be
6427 prepared to handle these functions regardless of the type used.
6429 This example shows how the :ref:`va_arg <i_va_arg>` instruction and the
6430 variable argument handling intrinsic functions are used.
6432 .. code-block:: llvm
6434 define i32 @test(i32 %X, ...) {
6435 ; Initialize variable argument processing
6437 %ap2 = bitcast i8** %ap to i8*
6438 call void @llvm.va_start(i8* %ap2)
6440 ; Read a single integer argument
6441 %tmp = va_arg i8** %ap, i32
6443 ; Demonstrate usage of llvm.va_copy and llvm.va_end
6445 %aq2 = bitcast i8** %aq to i8*
6446 call void @llvm.va_copy(i8* %aq2, i8* %ap2)
6447 call void @llvm.va_end(i8* %aq2)
6449 ; Stop processing of arguments.
6450 call void @llvm.va_end(i8* %ap2)
6454 declare void @llvm.va_start(i8*)
6455 declare void @llvm.va_copy(i8*, i8*)
6456 declare void @llvm.va_end(i8*)
6460 '``llvm.va_start``' Intrinsic
6461 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6468 declare void @llvm.va_start(i8* <arglist>)
6473 The '``llvm.va_start``' intrinsic initializes ``*<arglist>`` for
6474 subsequent use by ``va_arg``.
6479 The argument is a pointer to a ``va_list`` element to initialize.
6484 The '``llvm.va_start``' intrinsic works just like the ``va_start`` macro
6485 available in C. In a target-dependent way, it initializes the
6486 ``va_list`` element to which the argument points, so that the next call
6487 to ``va_arg`` will produce the first variable argument passed to the
6488 function. Unlike the C ``va_start`` macro, this intrinsic does not need
6489 to know the last argument of the function as the compiler can figure
6492 '``llvm.va_end``' Intrinsic
6493 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6500 declare void @llvm.va_end(i8* <arglist>)
6505 The '``llvm.va_end``' intrinsic destroys ``*<arglist>``, which has been
6506 initialized previously with ``llvm.va_start`` or ``llvm.va_copy``.
6511 The argument is a pointer to a ``va_list`` to destroy.
6516 The '``llvm.va_end``' intrinsic works just like the ``va_end`` macro
6517 available in C. In a target-dependent way, it destroys the ``va_list``
6518 element to which the argument points. Calls to
6519 :ref:`llvm.va_start <int_va_start>` and
6520 :ref:`llvm.va_copy <int_va_copy>` must be matched exactly with calls to
6525 '``llvm.va_copy``' Intrinsic
6526 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6533 declare void @llvm.va_copy(i8* <destarglist>, i8* <srcarglist>)
6538 The '``llvm.va_copy``' intrinsic copies the current argument position
6539 from the source argument list to the destination argument list.
6544 The first argument is a pointer to a ``va_list`` element to initialize.
6545 The second argument is a pointer to a ``va_list`` element to copy from.
6550 The '``llvm.va_copy``' intrinsic works just like the ``va_copy`` macro
6551 available in C. In a target-dependent way, it copies the source
6552 ``va_list`` element into the destination ``va_list`` element. This
6553 intrinsic is necessary because the `` llvm.va_start`` intrinsic may be
6554 arbitrarily complex and require, for example, memory allocation.
6556 Accurate Garbage Collection Intrinsics
6557 --------------------------------------
6559 LLVM support for `Accurate Garbage Collection <GarbageCollection.html>`_
6560 (GC) requires the implementation and generation of these intrinsics.
6561 These intrinsics allow identification of :ref:`GC roots on the
6562 stack <int_gcroot>`, as well as garbage collector implementations that
6563 require :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers.
6564 Front-ends for type-safe garbage collected languages should generate
6565 these intrinsics to make use of the LLVM garbage collectors. For more
6566 details, see `Accurate Garbage Collection with
6567 LLVM <GarbageCollection.html>`_.
6569 The garbage collection intrinsics only operate on objects in the generic
6570 address space (address space zero).
6574 '``llvm.gcroot``' Intrinsic
6575 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6582 declare void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
6587 The '``llvm.gcroot``' intrinsic declares the existence of a GC root to
6588 the code generator, and allows some metadata to be associated with it.
6593 The first argument specifies the address of a stack object that contains
6594 the root pointer. The second pointer (which must be either a constant or
6595 a global value address) contains the meta-data to be associated with the
6601 At runtime, a call to this intrinsic stores a null pointer into the
6602 "ptrloc" location. At compile-time, the code generator generates
6603 information to allow the runtime to find the pointer at GC safe points.
6604 The '``llvm.gcroot``' intrinsic may only be used in a function which
6605 :ref:`specifies a GC algorithm <gc>`.
6609 '``llvm.gcread``' Intrinsic
6610 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6617 declare i8* @llvm.gcread(i8* %ObjPtr, i8** %Ptr)
6622 The '``llvm.gcread``' intrinsic identifies reads of references from heap
6623 locations, allowing garbage collector implementations that require read
6629 The second argument is the address to read from, which should be an
6630 address allocated from the garbage collector. The first object is a
6631 pointer to the start of the referenced object, if needed by the language
6632 runtime (otherwise null).
6637 The '``llvm.gcread``' intrinsic has the same semantics as a load
6638 instruction, but may be replaced with substantially more complex code by
6639 the garbage collector runtime, as needed. The '``llvm.gcread``'
6640 intrinsic may only be used in a function which :ref:`specifies a GC
6645 '``llvm.gcwrite``' Intrinsic
6646 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6653 declare void @llvm.gcwrite(i8* %P1, i8* %Obj, i8** %P2)
6658 The '``llvm.gcwrite``' intrinsic identifies writes of references to heap
6659 locations, allowing garbage collector implementations that require write
6660 barriers (such as generational or reference counting collectors).
6665 The first argument is the reference to store, the second is the start of
6666 the object to store it to, and the third is the address of the field of
6667 Obj to store to. If the runtime does not require a pointer to the
6668 object, Obj may be null.
6673 The '``llvm.gcwrite``' intrinsic has the same semantics as a store
6674 instruction, but may be replaced with substantially more complex code by
6675 the garbage collector runtime, as needed. The '``llvm.gcwrite``'
6676 intrinsic may only be used in a function which :ref:`specifies a GC
6679 Code Generator Intrinsics
6680 -------------------------
6682 These intrinsics are provided by LLVM to expose special features that
6683 may only be implemented with code generator support.
6685 '``llvm.returnaddress``' Intrinsic
6686 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6693 declare i8 *@llvm.returnaddress(i32 <level>)
6698 The '``llvm.returnaddress``' intrinsic attempts to compute a
6699 target-specific value indicating the return address of the current
6700 function or one of its callers.
6705 The argument to this intrinsic indicates which function to return the
6706 address for. Zero indicates the calling function, one indicates its
6707 caller, etc. The argument is **required** to be a constant integer
6713 The '``llvm.returnaddress``' intrinsic either returns a pointer
6714 indicating the return address of the specified call frame, or zero if it
6715 cannot be identified. The value returned by this intrinsic is likely to
6716 be incorrect or 0 for arguments other than zero, so it should only be
6717 used for debugging purposes.
6719 Note that calling this intrinsic does not prevent function inlining or
6720 other aggressive transformations, so the value returned may not be that
6721 of the obvious source-language caller.
6723 '``llvm.frameaddress``' Intrinsic
6724 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6731 declare i8* @llvm.frameaddress(i32 <level>)
6736 The '``llvm.frameaddress``' intrinsic attempts to return the
6737 target-specific frame pointer value for the specified stack frame.
6742 The argument to this intrinsic indicates which function to return the
6743 frame pointer for. Zero indicates the calling function, one indicates
6744 its caller, etc. The argument is **required** to be a constant integer
6750 The '``llvm.frameaddress``' intrinsic either returns a pointer
6751 indicating the frame address of the specified call frame, or zero if it
6752 cannot be identified. The value returned by this intrinsic is likely to
6753 be incorrect or 0 for arguments other than zero, so it should only be
6754 used for debugging purposes.
6756 Note that calling this intrinsic does not prevent function inlining or
6757 other aggressive transformations, so the value returned may not be that
6758 of the obvious source-language caller.
6762 '``llvm.stacksave``' Intrinsic
6763 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6770 declare i8* @llvm.stacksave()
6775 The '``llvm.stacksave``' intrinsic is used to remember the current state
6776 of the function stack, for use with
6777 :ref:`llvm.stackrestore <int_stackrestore>`. This is useful for
6778 implementing language features like scoped automatic variable sized
6784 This intrinsic returns a opaque pointer value that can be passed to
6785 :ref:`llvm.stackrestore <int_stackrestore>`. When an
6786 ``llvm.stackrestore`` intrinsic is executed with a value saved from
6787 ``llvm.stacksave``, it effectively restores the state of the stack to
6788 the state it was in when the ``llvm.stacksave`` intrinsic executed. In
6789 practice, this pops any :ref:`alloca <i_alloca>` blocks from the stack that
6790 were allocated after the ``llvm.stacksave`` was executed.
6792 .. _int_stackrestore:
6794 '``llvm.stackrestore``' Intrinsic
6795 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6802 declare void @llvm.stackrestore(i8* %ptr)
6807 The '``llvm.stackrestore``' intrinsic is used to restore the state of
6808 the function stack to the state it was in when the corresponding
6809 :ref:`llvm.stacksave <int_stacksave>` intrinsic executed. This is
6810 useful for implementing language features like scoped automatic variable
6811 sized arrays in C99.
6816 See the description for :ref:`llvm.stacksave <int_stacksave>`.
6818 '``llvm.prefetch``' Intrinsic
6819 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6826 declare void @llvm.prefetch(i8* <address>, i32 <rw>, i32 <locality>, i32 <cache type>)
6831 The '``llvm.prefetch``' intrinsic is a hint to the code generator to
6832 insert a prefetch instruction if supported; otherwise, it is a noop.
6833 Prefetches have no effect on the behavior of the program but can change
6834 its performance characteristics.
6839 ``address`` is the address to be prefetched, ``rw`` is the specifier
6840 determining if the fetch should be for a read (0) or write (1), and
6841 ``locality`` is a temporal locality specifier ranging from (0) - no
6842 locality, to (3) - extremely local keep in cache. The ``cache type``
6843 specifies whether the prefetch is performed on the data (1) or
6844 instruction (0) cache. The ``rw``, ``locality`` and ``cache type``
6845 arguments must be constant integers.
6850 This intrinsic does not modify the behavior of the program. In
6851 particular, prefetches cannot trap and do not produce a value. On
6852 targets that support this intrinsic, the prefetch can provide hints to
6853 the processor cache for better performance.
6855 '``llvm.pcmarker``' Intrinsic
6856 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6863 declare void @llvm.pcmarker(i32 <id>)
6868 The '``llvm.pcmarker``' intrinsic is a method to export a Program
6869 Counter (PC) in a region of code to simulators and other tools. The
6870 method is target specific, but it is expected that the marker will use
6871 exported symbols to transmit the PC of the marker. The marker makes no
6872 guarantees that it will remain with any specific instruction after
6873 optimizations. It is possible that the presence of a marker will inhibit
6874 optimizations. The intended use is to be inserted after optimizations to
6875 allow correlations of simulation runs.
6880 ``id`` is a numerical id identifying the marker.
6885 This intrinsic does not modify the behavior of the program. Backends
6886 that do not support this intrinsic may ignore it.
6888 '``llvm.readcyclecounter``' Intrinsic
6889 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6896 declare i64 @llvm.readcyclecounter()
6901 The '``llvm.readcyclecounter``' intrinsic provides access to the cycle
6902 counter register (or similar low latency, high accuracy clocks) on those
6903 targets that support it. On X86, it should map to RDTSC. On Alpha, it
6904 should map to RPCC. As the backing counters overflow quickly (on the
6905 order of 9 seconds on alpha), this should only be used for small
6911 When directly supported, reading the cycle counter should not modify any
6912 memory. Implementations are allowed to either return a application
6913 specific value or a system wide value. On backends without support, this
6914 is lowered to a constant 0.
6916 Note that runtime support may be conditional on the privilege-level code is
6917 running at and the host platform.
6919 Standard C Library Intrinsics
6920 -----------------------------
6922 LLVM provides intrinsics for a few important standard C library
6923 functions. These intrinsics allow source-language front-ends to pass
6924 information about the alignment of the pointer arguments to the code
6925 generator, providing opportunity for more efficient code generation.
6929 '``llvm.memcpy``' Intrinsic
6930 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6935 This is an overloaded intrinsic. You can use ``llvm.memcpy`` on any
6936 integer bit width and for different address spaces. Not all targets
6937 support all bit widths however.
6941 declare void @llvm.memcpy.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
6942 i32 <len>, i32 <align>, i1 <isvolatile>)
6943 declare void @llvm.memcpy.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
6944 i64 <len>, i32 <align>, i1 <isvolatile>)
6949 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
6950 source location to the destination location.
6952 Note that, unlike the standard libc function, the ``llvm.memcpy.*``
6953 intrinsics do not return a value, takes extra alignment/isvolatile
6954 arguments and the pointers can be in specified address spaces.
6959 The first argument is a pointer to the destination, the second is a
6960 pointer to the source. The third argument is an integer argument
6961 specifying the number of bytes to copy, the fourth argument is the
6962 alignment of the source and destination locations, and the fifth is a
6963 boolean indicating a volatile access.
6965 If the call to this intrinsic has an alignment value that is not 0 or 1,
6966 then the caller guarantees that both the source and destination pointers
6967 are aligned to that boundary.
6969 If the ``isvolatile`` parameter is ``true``, the ``llvm.memcpy`` call is
6970 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
6971 very cleanly specified and it is unwise to depend on it.
6976 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
6977 source location to the destination location, which are not allowed to
6978 overlap. It copies "len" bytes of memory over. If the argument is known
6979 to be aligned to some boundary, this can be specified as the fourth
6980 argument, otherwise it should be set to 0 or 1 (both meaning no alignment).
6982 '``llvm.memmove``' Intrinsic
6983 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6988 This is an overloaded intrinsic. You can use llvm.memmove on any integer
6989 bit width and for different address space. Not all targets support all
6994 declare void @llvm.memmove.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
6995 i32 <len>, i32 <align>, i1 <isvolatile>)
6996 declare void @llvm.memmove.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
6997 i64 <len>, i32 <align>, i1 <isvolatile>)
7002 The '``llvm.memmove.*``' intrinsics move a block of memory from the
7003 source location to the destination location. It is similar to the
7004 '``llvm.memcpy``' intrinsic but allows the two memory locations to
7007 Note that, unlike the standard libc function, the ``llvm.memmove.*``
7008 intrinsics do not return a value, takes extra alignment/isvolatile
7009 arguments and the pointers can be in specified address spaces.
7014 The first argument is a pointer to the destination, the second is a
7015 pointer to the source. The third argument is an integer argument
7016 specifying the number of bytes to copy, the fourth argument is the
7017 alignment of the source and destination locations, and the fifth is a
7018 boolean indicating a volatile access.
7020 If the call to this intrinsic has an alignment value that is not 0 or 1,
7021 then the caller guarantees that the source and destination pointers are
7022 aligned to that boundary.
7024 If the ``isvolatile`` parameter is ``true``, the ``llvm.memmove`` call
7025 is a :ref:`volatile operation <volatile>`. The detailed access behavior is
7026 not very cleanly specified and it is unwise to depend on it.
7031 The '``llvm.memmove.*``' intrinsics copy a block of memory from the
7032 source location to the destination location, which may overlap. It
7033 copies "len" bytes of memory over. If the argument is known to be
7034 aligned to some boundary, this can be specified as the fourth argument,
7035 otherwise it should be set to 0 or 1 (both meaning no alignment).
7037 '``llvm.memset.*``' Intrinsics
7038 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7043 This is an overloaded intrinsic. You can use llvm.memset on any integer
7044 bit width and for different address spaces. However, not all targets
7045 support all bit widths.
7049 declare void @llvm.memset.p0i8.i32(i8* <dest>, i8 <val>,
7050 i32 <len>, i32 <align>, i1 <isvolatile>)
7051 declare void @llvm.memset.p0i8.i64(i8* <dest>, i8 <val>,
7052 i64 <len>, i32 <align>, i1 <isvolatile>)
7057 The '``llvm.memset.*``' intrinsics fill a block of memory with a
7058 particular byte value.
7060 Note that, unlike the standard libc function, the ``llvm.memset``
7061 intrinsic does not return a value and takes extra alignment/volatile
7062 arguments. Also, the destination can be in an arbitrary address space.
7067 The first argument is a pointer to the destination to fill, the second
7068 is the byte value with which to fill it, the third argument is an
7069 integer argument specifying the number of bytes to fill, and the fourth
7070 argument is the known alignment of the destination location.
7072 If the call to this intrinsic has an alignment value that is not 0 or 1,
7073 then the caller guarantees that the destination pointer is aligned to
7076 If the ``isvolatile`` parameter is ``true``, the ``llvm.memset`` call is
7077 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
7078 very cleanly specified and it is unwise to depend on it.
7083 The '``llvm.memset.*``' intrinsics fill "len" bytes of memory starting
7084 at the destination location. If the argument is known to be aligned to
7085 some boundary, this can be specified as the fourth argument, otherwise
7086 it should be set to 0 or 1 (both meaning no alignment).
7088 '``llvm.sqrt.*``' Intrinsic
7089 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7094 This is an overloaded intrinsic. You can use ``llvm.sqrt`` on any
7095 floating point or vector of floating point type. Not all targets support
7100 declare float @llvm.sqrt.f32(float %Val)
7101 declare double @llvm.sqrt.f64(double %Val)
7102 declare x86_fp80 @llvm.sqrt.f80(x86_fp80 %Val)
7103 declare fp128 @llvm.sqrt.f128(fp128 %Val)
7104 declare ppc_fp128 @llvm.sqrt.ppcf128(ppc_fp128 %Val)
7109 The '``llvm.sqrt``' intrinsics return the sqrt of the specified operand,
7110 returning the same value as the libm '``sqrt``' functions would. Unlike
7111 ``sqrt`` in libm, however, ``llvm.sqrt`` has undefined behavior for
7112 negative numbers other than -0.0 (which allows for better optimization,
7113 because there is no need to worry about errno being set).
7114 ``llvm.sqrt(-0.0)`` is defined to return -0.0 like IEEE sqrt.
7119 The argument and return value are floating point numbers of the same
7125 This function returns the sqrt of the specified operand if it is a
7126 nonnegative floating point number.
7128 '``llvm.powi.*``' Intrinsic
7129 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7134 This is an overloaded intrinsic. You can use ``llvm.powi`` on any
7135 floating point or vector of floating point type. Not all targets support
7140 declare float @llvm.powi.f32(float %Val, i32 %power)
7141 declare double @llvm.powi.f64(double %Val, i32 %power)
7142 declare x86_fp80 @llvm.powi.f80(x86_fp80 %Val, i32 %power)
7143 declare fp128 @llvm.powi.f128(fp128 %Val, i32 %power)
7144 declare ppc_fp128 @llvm.powi.ppcf128(ppc_fp128 %Val, i32 %power)
7149 The '``llvm.powi.*``' intrinsics return the first operand raised to the
7150 specified (positive or negative) power. The order of evaluation of
7151 multiplications is not defined. When a vector of floating point type is
7152 used, the second argument remains a scalar integer value.
7157 The second argument is an integer power, and the first is a value to
7158 raise to that power.
7163 This function returns the first value raised to the second power with an
7164 unspecified sequence of rounding operations.
7166 '``llvm.sin.*``' Intrinsic
7167 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7172 This is an overloaded intrinsic. You can use ``llvm.sin`` on any
7173 floating point or vector of floating point type. Not all targets support
7178 declare float @llvm.sin.f32(float %Val)
7179 declare double @llvm.sin.f64(double %Val)
7180 declare x86_fp80 @llvm.sin.f80(x86_fp80 %Val)
7181 declare fp128 @llvm.sin.f128(fp128 %Val)
7182 declare ppc_fp128 @llvm.sin.ppcf128(ppc_fp128 %Val)
7187 The '``llvm.sin.*``' intrinsics return the sine of the operand.
7192 The argument and return value are floating point numbers of the same
7198 This function returns the sine of the specified operand, returning the
7199 same values as the libm ``sin`` functions would, and handles error
7200 conditions in the same way.
7202 '``llvm.cos.*``' Intrinsic
7203 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7208 This is an overloaded intrinsic. You can use ``llvm.cos`` on any
7209 floating point or vector of floating point type. Not all targets support
7214 declare float @llvm.cos.f32(float %Val)
7215 declare double @llvm.cos.f64(double %Val)
7216 declare x86_fp80 @llvm.cos.f80(x86_fp80 %Val)
7217 declare fp128 @llvm.cos.f128(fp128 %Val)
7218 declare ppc_fp128 @llvm.cos.ppcf128(ppc_fp128 %Val)
7223 The '``llvm.cos.*``' intrinsics return the cosine of the operand.
7228 The argument and return value are floating point numbers of the same
7234 This function returns the cosine of the specified operand, returning the
7235 same values as the libm ``cos`` functions would, and handles error
7236 conditions in the same way.
7238 '``llvm.pow.*``' Intrinsic
7239 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7244 This is an overloaded intrinsic. You can use ``llvm.pow`` on any
7245 floating point or vector of floating point type. Not all targets support
7250 declare float @llvm.pow.f32(float %Val, float %Power)
7251 declare double @llvm.pow.f64(double %Val, double %Power)
7252 declare x86_fp80 @llvm.pow.f80(x86_fp80 %Val, x86_fp80 %Power)
7253 declare fp128 @llvm.pow.f128(fp128 %Val, fp128 %Power)
7254 declare ppc_fp128 @llvm.pow.ppcf128(ppc_fp128 %Val, ppc_fp128 Power)
7259 The '``llvm.pow.*``' intrinsics return the first operand raised to the
7260 specified (positive or negative) power.
7265 The second argument is a floating point power, and the first is a value
7266 to raise to that power.
7271 This function returns the first value raised to the second power,
7272 returning the same values as the libm ``pow`` functions would, and
7273 handles error conditions in the same way.
7275 '``llvm.exp.*``' Intrinsic
7276 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7281 This is an overloaded intrinsic. You can use ``llvm.exp`` on any
7282 floating point or vector of floating point type. Not all targets support
7287 declare float @llvm.exp.f32(float %Val)
7288 declare double @llvm.exp.f64(double %Val)
7289 declare x86_fp80 @llvm.exp.f80(x86_fp80 %Val)
7290 declare fp128 @llvm.exp.f128(fp128 %Val)
7291 declare ppc_fp128 @llvm.exp.ppcf128(ppc_fp128 %Val)
7296 The '``llvm.exp.*``' intrinsics perform the exp function.
7301 The argument and return value are floating point numbers of the same
7307 This function returns the same values as the libm ``exp`` functions
7308 would, and handles error conditions in the same way.
7310 '``llvm.exp2.*``' Intrinsic
7311 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7316 This is an overloaded intrinsic. You can use ``llvm.exp2`` on any
7317 floating point or vector of floating point type. Not all targets support
7322 declare float @llvm.exp2.f32(float %Val)
7323 declare double @llvm.exp2.f64(double %Val)
7324 declare x86_fp80 @llvm.exp2.f80(x86_fp80 %Val)
7325 declare fp128 @llvm.exp2.f128(fp128 %Val)
7326 declare ppc_fp128 @llvm.exp2.ppcf128(ppc_fp128 %Val)
7331 The '``llvm.exp2.*``' intrinsics perform the exp2 function.
7336 The argument and return value are floating point numbers of the same
7342 This function returns the same values as the libm ``exp2`` functions
7343 would, and handles error conditions in the same way.
7345 '``llvm.log.*``' Intrinsic
7346 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7351 This is an overloaded intrinsic. You can use ``llvm.log`` on any
7352 floating point or vector of floating point type. Not all targets support
7357 declare float @llvm.log.f32(float %Val)
7358 declare double @llvm.log.f64(double %Val)
7359 declare x86_fp80 @llvm.log.f80(x86_fp80 %Val)
7360 declare fp128 @llvm.log.f128(fp128 %Val)
7361 declare ppc_fp128 @llvm.log.ppcf128(ppc_fp128 %Val)
7366 The '``llvm.log.*``' intrinsics perform the log function.
7371 The argument and return value are floating point numbers of the same
7377 This function returns the same values as the libm ``log`` functions
7378 would, and handles error conditions in the same way.
7380 '``llvm.log10.*``' Intrinsic
7381 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7386 This is an overloaded intrinsic. You can use ``llvm.log10`` on any
7387 floating point or vector of floating point type. Not all targets support
7392 declare float @llvm.log10.f32(float %Val)
7393 declare double @llvm.log10.f64(double %Val)
7394 declare x86_fp80 @llvm.log10.f80(x86_fp80 %Val)
7395 declare fp128 @llvm.log10.f128(fp128 %Val)
7396 declare ppc_fp128 @llvm.log10.ppcf128(ppc_fp128 %Val)
7401 The '``llvm.log10.*``' intrinsics perform the log10 function.
7406 The argument and return value are floating point numbers of the same
7412 This function returns the same values as the libm ``log10`` functions
7413 would, and handles error conditions in the same way.
7415 '``llvm.log2.*``' Intrinsic
7416 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7421 This is an overloaded intrinsic. You can use ``llvm.log2`` on any
7422 floating point or vector of floating point type. Not all targets support
7427 declare float @llvm.log2.f32(float %Val)
7428 declare double @llvm.log2.f64(double %Val)
7429 declare x86_fp80 @llvm.log2.f80(x86_fp80 %Val)
7430 declare fp128 @llvm.log2.f128(fp128 %Val)
7431 declare ppc_fp128 @llvm.log2.ppcf128(ppc_fp128 %Val)
7436 The '``llvm.log2.*``' intrinsics perform the log2 function.
7441 The argument and return value are floating point numbers of the same
7447 This function returns the same values as the libm ``log2`` functions
7448 would, and handles error conditions in the same way.
7450 '``llvm.fma.*``' Intrinsic
7451 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7456 This is an overloaded intrinsic. You can use ``llvm.fma`` on any
7457 floating point or vector of floating point type. Not all targets support
7462 declare float @llvm.fma.f32(float %a, float %b, float %c)
7463 declare double @llvm.fma.f64(double %a, double %b, double %c)
7464 declare x86_fp80 @llvm.fma.f80(x86_fp80 %a, x86_fp80 %b, x86_fp80 %c)
7465 declare fp128 @llvm.fma.f128(fp128 %a, fp128 %b, fp128 %c)
7466 declare ppc_fp128 @llvm.fma.ppcf128(ppc_fp128 %a, ppc_fp128 %b, ppc_fp128 %c)
7471 The '``llvm.fma.*``' intrinsics perform the fused multiply-add
7477 The argument and return value are floating point numbers of the same
7483 This function returns the same values as the libm ``fma`` functions
7486 '``llvm.fabs.*``' Intrinsic
7487 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7492 This is an overloaded intrinsic. You can use ``llvm.fabs`` on any
7493 floating point or vector of floating point type. Not all targets support
7498 declare float @llvm.fabs.f32(float %Val)
7499 declare double @llvm.fabs.f64(double %Val)
7500 declare x86_fp80 @llvm.fabs.f80(x86_fp80 %Val)
7501 declare fp128 @llvm.fabs.f128(fp128 %Val)
7502 declare ppc_fp128 @llvm.fabs.ppcf128(ppc_fp128 %Val)
7507 The '``llvm.fabs.*``' intrinsics return the absolute value of the
7513 The argument and return value are floating point numbers of the same
7519 This function returns the same values as the libm ``fabs`` functions
7520 would, and handles error conditions in the same way.
7522 '``llvm.copysign.*``' Intrinsic
7523 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7528 This is an overloaded intrinsic. You can use ``llvm.copysign`` on any
7529 floating point or vector of floating point type. Not all targets support
7534 declare float @llvm.copysign.f32(float %Mag, float %Sgn)
7535 declare double @llvm.copysign.f64(double %Mag, double %Sgn)
7536 declare x86_fp80 @llvm.copysign.f80(x86_fp80 %Mag, x86_fp80 %Sgn)
7537 declare fp128 @llvm.copysign.f128(fp128 %Mag, fp128 %Sgn)
7538 declare ppc_fp128 @llvm.copysign.ppcf128(ppc_fp128 %Mag, ppc_fp128 %Sgn)
7543 The '``llvm.copysign.*``' intrinsics return a value with the magnitude of the
7544 first operand and the sign of the second operand.
7549 The arguments and return value are floating point numbers of the same
7555 This function returns the same values as the libm ``copysign``
7556 functions would, and handles error conditions in the same way.
7558 '``llvm.floor.*``' Intrinsic
7559 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7564 This is an overloaded intrinsic. You can use ``llvm.floor`` on any
7565 floating point or vector of floating point type. Not all targets support
7570 declare float @llvm.floor.f32(float %Val)
7571 declare double @llvm.floor.f64(double %Val)
7572 declare x86_fp80 @llvm.floor.f80(x86_fp80 %Val)
7573 declare fp128 @llvm.floor.f128(fp128 %Val)
7574 declare ppc_fp128 @llvm.floor.ppcf128(ppc_fp128 %Val)
7579 The '``llvm.floor.*``' intrinsics return the floor of the operand.
7584 The argument and return value are floating point numbers of the same
7590 This function returns the same values as the libm ``floor`` functions
7591 would, and handles error conditions in the same way.
7593 '``llvm.ceil.*``' Intrinsic
7594 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7599 This is an overloaded intrinsic. You can use ``llvm.ceil`` on any
7600 floating point or vector of floating point type. Not all targets support
7605 declare float @llvm.ceil.f32(float %Val)
7606 declare double @llvm.ceil.f64(double %Val)
7607 declare x86_fp80 @llvm.ceil.f80(x86_fp80 %Val)
7608 declare fp128 @llvm.ceil.f128(fp128 %Val)
7609 declare ppc_fp128 @llvm.ceil.ppcf128(ppc_fp128 %Val)
7614 The '``llvm.ceil.*``' intrinsics return the ceiling of the operand.
7619 The argument and return value are floating point numbers of the same
7625 This function returns the same values as the libm ``ceil`` functions
7626 would, and handles error conditions in the same way.
7628 '``llvm.trunc.*``' Intrinsic
7629 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7634 This is an overloaded intrinsic. You can use ``llvm.trunc`` on any
7635 floating point or vector of floating point type. Not all targets support
7640 declare float @llvm.trunc.f32(float %Val)
7641 declare double @llvm.trunc.f64(double %Val)
7642 declare x86_fp80 @llvm.trunc.f80(x86_fp80 %Val)
7643 declare fp128 @llvm.trunc.f128(fp128 %Val)
7644 declare ppc_fp128 @llvm.trunc.ppcf128(ppc_fp128 %Val)
7649 The '``llvm.trunc.*``' intrinsics returns the operand rounded to the
7650 nearest integer not larger in magnitude than the operand.
7655 The argument and return value are floating point numbers of the same
7661 This function returns the same values as the libm ``trunc`` functions
7662 would, and handles error conditions in the same way.
7664 '``llvm.rint.*``' Intrinsic
7665 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7670 This is an overloaded intrinsic. You can use ``llvm.rint`` on any
7671 floating point or vector of floating point type. Not all targets support
7676 declare float @llvm.rint.f32(float %Val)
7677 declare double @llvm.rint.f64(double %Val)
7678 declare x86_fp80 @llvm.rint.f80(x86_fp80 %Val)
7679 declare fp128 @llvm.rint.f128(fp128 %Val)
7680 declare ppc_fp128 @llvm.rint.ppcf128(ppc_fp128 %Val)
7685 The '``llvm.rint.*``' intrinsics returns the operand rounded to the
7686 nearest integer. It may raise an inexact floating-point exception if the
7687 operand isn't an integer.
7692 The argument and return value are floating point numbers of the same
7698 This function returns the same values as the libm ``rint`` functions
7699 would, and handles error conditions in the same way.
7701 '``llvm.nearbyint.*``' Intrinsic
7702 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7707 This is an overloaded intrinsic. You can use ``llvm.nearbyint`` on any
7708 floating point or vector of floating point type. Not all targets support
7713 declare float @llvm.nearbyint.f32(float %Val)
7714 declare double @llvm.nearbyint.f64(double %Val)
7715 declare x86_fp80 @llvm.nearbyint.f80(x86_fp80 %Val)
7716 declare fp128 @llvm.nearbyint.f128(fp128 %Val)
7717 declare ppc_fp128 @llvm.nearbyint.ppcf128(ppc_fp128 %Val)
7722 The '``llvm.nearbyint.*``' intrinsics returns the operand rounded to the
7728 The argument and return value are floating point numbers of the same
7734 This function returns the same values as the libm ``nearbyint``
7735 functions would, and handles error conditions in the same way.
7737 '``llvm.round.*``' Intrinsic
7738 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7743 This is an overloaded intrinsic. You can use ``llvm.round`` on any
7744 floating point or vector of floating point type. Not all targets support
7749 declare float @llvm.round.f32(float %Val)
7750 declare double @llvm.round.f64(double %Val)
7751 declare x86_fp80 @llvm.round.f80(x86_fp80 %Val)
7752 declare fp128 @llvm.round.f128(fp128 %Val)
7753 declare ppc_fp128 @llvm.round.ppcf128(ppc_fp128 %Val)
7758 The '``llvm.round.*``' intrinsics returns the operand rounded to the
7764 The argument and return value are floating point numbers of the same
7770 This function returns the same values as the libm ``round``
7771 functions would, and handles error conditions in the same way.
7773 Bit Manipulation Intrinsics
7774 ---------------------------
7776 LLVM provides intrinsics for a few important bit manipulation
7777 operations. These allow efficient code generation for some algorithms.
7779 '``llvm.bswap.*``' Intrinsics
7780 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7785 This is an overloaded intrinsic function. You can use bswap on any
7786 integer type that is an even number of bytes (i.e. BitWidth % 16 == 0).
7790 declare i16 @llvm.bswap.i16(i16 <id>)
7791 declare i32 @llvm.bswap.i32(i32 <id>)
7792 declare i64 @llvm.bswap.i64(i64 <id>)
7797 The '``llvm.bswap``' family of intrinsics is used to byte swap integer
7798 values with an even number of bytes (positive multiple of 16 bits).
7799 These are useful for performing operations on data that is not in the
7800 target's native byte order.
7805 The ``llvm.bswap.i16`` intrinsic returns an i16 value that has the high
7806 and low byte of the input i16 swapped. Similarly, the ``llvm.bswap.i32``
7807 intrinsic returns an i32 value that has the four bytes of the input i32
7808 swapped, so that if the input bytes are numbered 0, 1, 2, 3 then the
7809 returned i32 will have its bytes in 3, 2, 1, 0 order. The
7810 ``llvm.bswap.i48``, ``llvm.bswap.i64`` and other intrinsics extend this
7811 concept to additional even-byte lengths (6 bytes, 8 bytes and more,
7814 '``llvm.ctpop.*``' Intrinsic
7815 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7820 This is an overloaded intrinsic. You can use llvm.ctpop on any integer
7821 bit width, or on any vector with integer elements. Not all targets
7822 support all bit widths or vector types, however.
7826 declare i8 @llvm.ctpop.i8(i8 <src>)
7827 declare i16 @llvm.ctpop.i16(i16 <src>)
7828 declare i32 @llvm.ctpop.i32(i32 <src>)
7829 declare i64 @llvm.ctpop.i64(i64 <src>)
7830 declare i256 @llvm.ctpop.i256(i256 <src>)
7831 declare <2 x i32> @llvm.ctpop.v2i32(<2 x i32> <src>)
7836 The '``llvm.ctpop``' family of intrinsics counts the number of bits set
7842 The only argument is the value to be counted. The argument may be of any
7843 integer type, or a vector with integer elements. The return type must
7844 match the argument type.
7849 The '``llvm.ctpop``' intrinsic counts the 1's in a variable, or within
7850 each element of a vector.
7852 '``llvm.ctlz.*``' Intrinsic
7853 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7858 This is an overloaded intrinsic. You can use ``llvm.ctlz`` on any
7859 integer bit width, or any vector whose elements are integers. Not all
7860 targets support all bit widths or vector types, however.
7864 declare i8 @llvm.ctlz.i8 (i8 <src>, i1 <is_zero_undef>)
7865 declare i16 @llvm.ctlz.i16 (i16 <src>, i1 <is_zero_undef>)
7866 declare i32 @llvm.ctlz.i32 (i32 <src>, i1 <is_zero_undef>)
7867 declare i64 @llvm.ctlz.i64 (i64 <src>, i1 <is_zero_undef>)
7868 declare i256 @llvm.ctlz.i256(i256 <src>, i1 <is_zero_undef>)
7869 declase <2 x i32> @llvm.ctlz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
7874 The '``llvm.ctlz``' family of intrinsic functions counts the number of
7875 leading zeros in a variable.
7880 The first argument is the value to be counted. This argument may be of
7881 any integer type, or a vectory with integer element type. The return
7882 type must match the first argument type.
7884 The second argument must be a constant and is a flag to indicate whether
7885 the intrinsic should ensure that a zero as the first argument produces a
7886 defined result. Historically some architectures did not provide a
7887 defined result for zero values as efficiently, and many algorithms are
7888 now predicated on avoiding zero-value inputs.
7893 The '``llvm.ctlz``' intrinsic counts the leading (most significant)
7894 zeros in a variable, or within each element of the vector. If
7895 ``src == 0`` then the result is the size in bits of the type of ``src``
7896 if ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
7897 ``llvm.ctlz(i32 2) = 30``.
7899 '``llvm.cttz.*``' Intrinsic
7900 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7905 This is an overloaded intrinsic. You can use ``llvm.cttz`` on any
7906 integer bit width, or any vector of integer elements. Not all targets
7907 support all bit widths or vector types, however.
7911 declare i8 @llvm.cttz.i8 (i8 <src>, i1 <is_zero_undef>)
7912 declare i16 @llvm.cttz.i16 (i16 <src>, i1 <is_zero_undef>)
7913 declare i32 @llvm.cttz.i32 (i32 <src>, i1 <is_zero_undef>)
7914 declare i64 @llvm.cttz.i64 (i64 <src>, i1 <is_zero_undef>)
7915 declare i256 @llvm.cttz.i256(i256 <src>, i1 <is_zero_undef>)
7916 declase <2 x i32> @llvm.cttz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
7921 The '``llvm.cttz``' family of intrinsic functions counts the number of
7927 The first argument is the value to be counted. This argument may be of
7928 any integer type, or a vectory with integer element type. The return
7929 type must match the first argument type.
7931 The second argument must be a constant and is a flag to indicate whether
7932 the intrinsic should ensure that a zero as the first argument produces a
7933 defined result. Historically some architectures did not provide a
7934 defined result for zero values as efficiently, and many algorithms are
7935 now predicated on avoiding zero-value inputs.
7940 The '``llvm.cttz``' intrinsic counts the trailing (least significant)
7941 zeros in a variable, or within each element of a vector. If ``src == 0``
7942 then the result is the size in bits of the type of ``src`` if
7943 ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
7944 ``llvm.cttz(2) = 1``.
7946 Arithmetic with Overflow Intrinsics
7947 -----------------------------------
7949 LLVM provides intrinsics for some arithmetic with overflow operations.
7951 '``llvm.sadd.with.overflow.*``' Intrinsics
7952 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7957 This is an overloaded intrinsic. You can use ``llvm.sadd.with.overflow``
7958 on any integer bit width.
7962 declare {i16, i1} @llvm.sadd.with.overflow.i16(i16 %a, i16 %b)
7963 declare {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
7964 declare {i64, i1} @llvm.sadd.with.overflow.i64(i64 %a, i64 %b)
7969 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
7970 a signed addition of the two arguments, and indicate whether an overflow
7971 occurred during the signed summation.
7976 The arguments (%a and %b) and the first element of the result structure
7977 may be of integer types of any bit width, but they must have the same
7978 bit width. The second element of the result structure must be of type
7979 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
7985 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
7986 a signed addition of the two variables. They return a structure --- the
7987 first element of which is the signed summation, and the second element
7988 of which is a bit specifying if the signed summation resulted in an
7994 .. code-block:: llvm
7996 %res = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
7997 %sum = extractvalue {i32, i1} %res, 0
7998 %obit = extractvalue {i32, i1} %res, 1
7999 br i1 %obit, label %overflow, label %normal
8001 '``llvm.uadd.with.overflow.*``' Intrinsics
8002 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8007 This is an overloaded intrinsic. You can use ``llvm.uadd.with.overflow``
8008 on any integer bit width.
8012 declare {i16, i1} @llvm.uadd.with.overflow.i16(i16 %a, i16 %b)
8013 declare {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
8014 declare {i64, i1} @llvm.uadd.with.overflow.i64(i64 %a, i64 %b)
8019 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
8020 an unsigned addition of the two arguments, and indicate whether a carry
8021 occurred during the unsigned summation.
8026 The arguments (%a and %b) and the first element of the result structure
8027 may be of integer types of any bit width, but they must have the same
8028 bit width. The second element of the result structure must be of type
8029 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8035 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
8036 an unsigned addition of the two arguments. They return a structure --- the
8037 first element of which is the sum, and the second element of which is a
8038 bit specifying if the unsigned summation resulted in a carry.
8043 .. code-block:: llvm
8045 %res = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
8046 %sum = extractvalue {i32, i1} %res, 0
8047 %obit = extractvalue {i32, i1} %res, 1
8048 br i1 %obit, label %carry, label %normal
8050 '``llvm.ssub.with.overflow.*``' Intrinsics
8051 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8056 This is an overloaded intrinsic. You can use ``llvm.ssub.with.overflow``
8057 on any integer bit width.
8061 declare {i16, i1} @llvm.ssub.with.overflow.i16(i16 %a, i16 %b)
8062 declare {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
8063 declare {i64, i1} @llvm.ssub.with.overflow.i64(i64 %a, i64 %b)
8068 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
8069 a signed subtraction of the two arguments, and indicate whether an
8070 overflow occurred during the signed subtraction.
8075 The arguments (%a and %b) and the first element of the result structure
8076 may be of integer types of any bit width, but they must have the same
8077 bit width. The second element of the result structure must be of type
8078 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
8084 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
8085 a signed subtraction of the two arguments. They return a structure --- the
8086 first element of which is the subtraction, and the second element of
8087 which is a bit specifying if the signed subtraction resulted in an
8093 .. code-block:: llvm
8095 %res = call {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
8096 %sum = extractvalue {i32, i1} %res, 0
8097 %obit = extractvalue {i32, i1} %res, 1
8098 br i1 %obit, label %overflow, label %normal
8100 '``llvm.usub.with.overflow.*``' Intrinsics
8101 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8106 This is an overloaded intrinsic. You can use ``llvm.usub.with.overflow``
8107 on any integer bit width.
8111 declare {i16, i1} @llvm.usub.with.overflow.i16(i16 %a, i16 %b)
8112 declare {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
8113 declare {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
8118 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
8119 an unsigned subtraction of the two arguments, and indicate whether an
8120 overflow occurred during the unsigned subtraction.
8125 The arguments (%a and %b) and the first element of the result structure
8126 may be of integer types of any bit width, but they must have the same
8127 bit width. The second element of the result structure must be of type
8128 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8134 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
8135 an unsigned subtraction of the two arguments. They return a structure ---
8136 the first element of which is the subtraction, and the second element of
8137 which is a bit specifying if the unsigned subtraction resulted in an
8143 .. code-block:: llvm
8145 %res = call {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
8146 %sum = extractvalue {i32, i1} %res, 0
8147 %obit = extractvalue {i32, i1} %res, 1
8148 br i1 %obit, label %overflow, label %normal
8150 '``llvm.smul.with.overflow.*``' Intrinsics
8151 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8156 This is an overloaded intrinsic. You can use ``llvm.smul.with.overflow``
8157 on any integer bit width.
8161 declare {i16, i1} @llvm.smul.with.overflow.i16(i16 %a, i16 %b)
8162 declare {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
8163 declare {i64, i1} @llvm.smul.with.overflow.i64(i64 %a, i64 %b)
8168 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
8169 a signed multiplication of the two arguments, and indicate whether an
8170 overflow occurred during the signed multiplication.
8175 The arguments (%a and %b) and the first element of the result structure
8176 may be of integer types of any bit width, but they must have the same
8177 bit width. The second element of the result structure must be of type
8178 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
8184 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
8185 a signed multiplication of the two arguments. They return a structure ---
8186 the first element of which is the multiplication, and the second element
8187 of which is a bit specifying if the signed multiplication resulted in an
8193 .. code-block:: llvm
8195 %res = call {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
8196 %sum = extractvalue {i32, i1} %res, 0
8197 %obit = extractvalue {i32, i1} %res, 1
8198 br i1 %obit, label %overflow, label %normal
8200 '``llvm.umul.with.overflow.*``' Intrinsics
8201 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8206 This is an overloaded intrinsic. You can use ``llvm.umul.with.overflow``
8207 on any integer bit width.
8211 declare {i16, i1} @llvm.umul.with.overflow.i16(i16 %a, i16 %b)
8212 declare {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
8213 declare {i64, i1} @llvm.umul.with.overflow.i64(i64 %a, i64 %b)
8218 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
8219 a unsigned multiplication of the two arguments, and indicate whether an
8220 overflow occurred during the unsigned multiplication.
8225 The arguments (%a and %b) and the first element of the result structure
8226 may be of integer types of any bit width, but they must have the same
8227 bit width. The second element of the result structure must be of type
8228 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8234 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
8235 an unsigned multiplication of the two arguments. They return a structure ---
8236 the first element of which is the multiplication, and the second
8237 element of which is a bit specifying if the unsigned multiplication
8238 resulted in an overflow.
8243 .. code-block:: llvm
8245 %res = call {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
8246 %sum = extractvalue {i32, i1} %res, 0
8247 %obit = extractvalue {i32, i1} %res, 1
8248 br i1 %obit, label %overflow, label %normal
8250 Specialised Arithmetic Intrinsics
8251 ---------------------------------
8253 '``llvm.fmuladd.*``' Intrinsic
8254 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8261 declare float @llvm.fmuladd.f32(float %a, float %b, float %c)
8262 declare double @llvm.fmuladd.f64(double %a, double %b, double %c)
8267 The '``llvm.fmuladd.*``' intrinsic functions represent multiply-add
8268 expressions that can be fused if the code generator determines that (a) the
8269 target instruction set has support for a fused operation, and (b) that the
8270 fused operation is more efficient than the equivalent, separate pair of mul
8271 and add instructions.
8276 The '``llvm.fmuladd.*``' intrinsics each take three arguments: two
8277 multiplicands, a and b, and an addend c.
8286 %0 = call float @llvm.fmuladd.f32(%a, %b, %c)
8288 is equivalent to the expression a \* b + c, except that rounding will
8289 not be performed between the multiplication and addition steps if the
8290 code generator fuses the operations. Fusion is not guaranteed, even if
8291 the target platform supports it. If a fused multiply-add is required the
8292 corresponding llvm.fma.\* intrinsic function should be used instead.
8297 .. code-block:: llvm
8299 %r2 = call float @llvm.fmuladd.f32(float %a, float %b, float %c) ; yields {float}:r2 = (a * b) + c
8301 Half Precision Floating Point Intrinsics
8302 ----------------------------------------
8304 For most target platforms, half precision floating point is a
8305 storage-only format. This means that it is a dense encoding (in memory)
8306 but does not support computation in the format.
8308 This means that code must first load the half-precision floating point
8309 value as an i16, then convert it to float with
8310 :ref:`llvm.convert.from.fp16 <int_convert_from_fp16>`. Computation can
8311 then be performed on the float value (including extending to double
8312 etc). To store the value back to memory, it is first converted to float
8313 if needed, then converted to i16 with
8314 :ref:`llvm.convert.to.fp16 <int_convert_to_fp16>`, then storing as an
8317 .. _int_convert_to_fp16:
8319 '``llvm.convert.to.fp16``' Intrinsic
8320 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8327 declare i16 @llvm.convert.to.fp16(f32 %a)
8332 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion
8333 from single precision floating point format to half precision floating
8339 The intrinsic function contains single argument - the value to be
8345 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion
8346 from single precision floating point format to half precision floating
8347 point format. The return value is an ``i16`` which contains the
8353 .. code-block:: llvm
8355 %res = call i16 @llvm.convert.to.fp16(f32 %a)
8356 store i16 %res, i16* @x, align 2
8358 .. _int_convert_from_fp16:
8360 '``llvm.convert.from.fp16``' Intrinsic
8361 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8368 declare f32 @llvm.convert.from.fp16(i16 %a)
8373 The '``llvm.convert.from.fp16``' intrinsic function performs a
8374 conversion from half precision floating point format to single precision
8375 floating point format.
8380 The intrinsic function contains single argument - the value to be
8386 The '``llvm.convert.from.fp16``' intrinsic function performs a
8387 conversion from half single precision floating point format to single
8388 precision floating point format. The input half-float value is
8389 represented by an ``i16`` value.
8394 .. code-block:: llvm
8396 %a = load i16* @x, align 2
8397 %res = call f32 @llvm.convert.from.fp16(i16 %a)
8402 The LLVM debugger intrinsics (which all start with ``llvm.dbg.``
8403 prefix), are described in the `LLVM Source Level
8404 Debugging <SourceLevelDebugging.html#format_common_intrinsics>`_
8407 Exception Handling Intrinsics
8408 -----------------------------
8410 The LLVM exception handling intrinsics (which all start with
8411 ``llvm.eh.`` prefix), are described in the `LLVM Exception
8412 Handling <ExceptionHandling.html#format_common_intrinsics>`_ document.
8416 Trampoline Intrinsics
8417 ---------------------
8419 These intrinsics make it possible to excise one parameter, marked with
8420 the :ref:`nest <nest>` attribute, from a function. The result is a
8421 callable function pointer lacking the nest parameter - the caller does
8422 not need to provide a value for it. Instead, the value to use is stored
8423 in advance in a "trampoline", a block of memory usually allocated on the
8424 stack, which also contains code to splice the nest value into the
8425 argument list. This is used to implement the GCC nested function address
8428 For example, if the function is ``i32 f(i8* nest %c, i32 %x, i32 %y)``
8429 then the resulting function pointer has signature ``i32 (i32, i32)*``.
8430 It can be created as follows:
8432 .. code-block:: llvm
8434 %tramp = alloca [10 x i8], align 4 ; size and alignment only correct for X86
8435 %tramp1 = getelementptr [10 x i8]* %tramp, i32 0, i32 0
8436 call i8* @llvm.init.trampoline(i8* %tramp1, i8* bitcast (i32 (i8*, i32, i32)* @f to i8*), i8* %nval)
8437 %p = call i8* @llvm.adjust.trampoline(i8* %tramp1)
8438 %fp = bitcast i8* %p to i32 (i32, i32)*
8440 The call ``%val = call i32 %fp(i32 %x, i32 %y)`` is then equivalent to
8441 ``%val = call i32 %f(i8* %nval, i32 %x, i32 %y)``.
8445 '``llvm.init.trampoline``' Intrinsic
8446 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8453 declare void @llvm.init.trampoline(i8* <tramp>, i8* <func>, i8* <nval>)
8458 This fills the memory pointed to by ``tramp`` with executable code,
8459 turning it into a trampoline.
8464 The ``llvm.init.trampoline`` intrinsic takes three arguments, all
8465 pointers. The ``tramp`` argument must point to a sufficiently large and
8466 sufficiently aligned block of memory; this memory is written to by the
8467 intrinsic. Note that the size and the alignment are target-specific -
8468 LLVM currently provides no portable way of determining them, so a
8469 front-end that generates this intrinsic needs to have some
8470 target-specific knowledge. The ``func`` argument must hold a function
8471 bitcast to an ``i8*``.
8476 The block of memory pointed to by ``tramp`` is filled with target
8477 dependent code, turning it into a function. Then ``tramp`` needs to be
8478 passed to :ref:`llvm.adjust.trampoline <int_at>` to get a pointer which can
8479 be :ref:`bitcast (to a new function) and called <int_trampoline>`. The new
8480 function's signature is the same as that of ``func`` with any arguments
8481 marked with the ``nest`` attribute removed. At most one such ``nest``
8482 argument is allowed, and it must be of pointer type. Calling the new
8483 function is equivalent to calling ``func`` with the same argument list,
8484 but with ``nval`` used for the missing ``nest`` argument. If, after
8485 calling ``llvm.init.trampoline``, the memory pointed to by ``tramp`` is
8486 modified, then the effect of any later call to the returned function
8487 pointer is undefined.
8491 '``llvm.adjust.trampoline``' Intrinsic
8492 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8499 declare i8* @llvm.adjust.trampoline(i8* <tramp>)
8504 This performs any required machine-specific adjustment to the address of
8505 a trampoline (passed as ``tramp``).
8510 ``tramp`` must point to a block of memory which already has trampoline
8511 code filled in by a previous call to
8512 :ref:`llvm.init.trampoline <int_it>`.
8517 On some architectures the address of the code to be executed needs to be
8518 different to the address where the trampoline is actually stored. This
8519 intrinsic returns the executable address corresponding to ``tramp``
8520 after performing the required machine specific adjustments. The pointer
8521 returned can then be :ref:`bitcast and executed <int_trampoline>`.
8526 This class of intrinsics exists to information about the lifetime of
8527 memory objects and ranges where variables are immutable.
8531 '``llvm.lifetime.start``' Intrinsic
8532 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8539 declare void @llvm.lifetime.start(i64 <size>, i8* nocapture <ptr>)
8544 The '``llvm.lifetime.start``' intrinsic specifies the start of a memory
8550 The first argument is a constant integer representing the size of the
8551 object, or -1 if it is variable sized. The second argument is a pointer
8557 This intrinsic indicates that before this point in the code, the value
8558 of the memory pointed to by ``ptr`` is dead. This means that it is known
8559 to never be used and has an undefined value. A load from the pointer
8560 that precedes this intrinsic can be replaced with ``'undef'``.
8564 '``llvm.lifetime.end``' Intrinsic
8565 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8572 declare void @llvm.lifetime.end(i64 <size>, i8* nocapture <ptr>)
8577 The '``llvm.lifetime.end``' intrinsic specifies the end of a memory
8583 The first argument is a constant integer representing the size of the
8584 object, or -1 if it is variable sized. The second argument is a pointer
8590 This intrinsic indicates that after this point in the code, the value of
8591 the memory pointed to by ``ptr`` is dead. This means that it is known to
8592 never be used and has an undefined value. Any stores into the memory
8593 object following this intrinsic may be removed as dead.
8595 '``llvm.invariant.start``' Intrinsic
8596 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8603 declare {}* @llvm.invariant.start(i64 <size>, i8* nocapture <ptr>)
8608 The '``llvm.invariant.start``' intrinsic specifies that the contents of
8609 a memory object will not change.
8614 The first argument is a constant integer representing the size of the
8615 object, or -1 if it is variable sized. The second argument is a pointer
8621 This intrinsic indicates that until an ``llvm.invariant.end`` that uses
8622 the return value, the referenced memory location is constant and
8625 '``llvm.invariant.end``' Intrinsic
8626 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8633 declare void @llvm.invariant.end({}* <start>, i64 <size>, i8* nocapture <ptr>)
8638 The '``llvm.invariant.end``' intrinsic specifies that the contents of a
8639 memory object are mutable.
8644 The first argument is the matching ``llvm.invariant.start`` intrinsic.
8645 The second argument is a constant integer representing the size of the
8646 object, or -1 if it is variable sized and the third argument is a
8647 pointer to the object.
8652 This intrinsic indicates that the memory is mutable again.
8657 This class of intrinsics is designed to be generic and has no specific
8660 '``llvm.var.annotation``' Intrinsic
8661 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8668 declare void @llvm.var.annotation(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
8673 The '``llvm.var.annotation``' intrinsic.
8678 The first argument is a pointer to a value, the second is a pointer to a
8679 global string, the third is a pointer to a global string which is the
8680 source file name, and the last argument is the line number.
8685 This intrinsic allows annotation of local variables with arbitrary
8686 strings. This can be useful for special purpose optimizations that want
8687 to look for these annotations. These have no other defined use; they are
8688 ignored by code generation and optimization.
8690 '``llvm.ptr.annotation.*``' Intrinsic
8691 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8696 This is an overloaded intrinsic. You can use '``llvm.ptr.annotation``' on a
8697 pointer to an integer of any width. *NOTE* you must specify an address space for
8698 the pointer. The identifier for the default address space is the integer
8703 declare i8* @llvm.ptr.annotation.p<address space>i8(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
8704 declare i16* @llvm.ptr.annotation.p<address space>i16(i16* <val>, i8* <str>, i8* <str>, i32 <int>)
8705 declare i32* @llvm.ptr.annotation.p<address space>i32(i32* <val>, i8* <str>, i8* <str>, i32 <int>)
8706 declare i64* @llvm.ptr.annotation.p<address space>i64(i64* <val>, i8* <str>, i8* <str>, i32 <int>)
8707 declare i256* @llvm.ptr.annotation.p<address space>i256(i256* <val>, i8* <str>, i8* <str>, i32 <int>)
8712 The '``llvm.ptr.annotation``' intrinsic.
8717 The first argument is a pointer to an integer value of arbitrary bitwidth
8718 (result of some expression), the second is a pointer to a global string, the
8719 third is a pointer to a global string which is the source file name, and the
8720 last argument is the line number. It returns the value of the first argument.
8725 This intrinsic allows annotation of a pointer to an integer with arbitrary
8726 strings. This can be useful for special purpose optimizations that want to look
8727 for these annotations. These have no other defined use; they are ignored by code
8728 generation and optimization.
8730 '``llvm.annotation.*``' Intrinsic
8731 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8736 This is an overloaded intrinsic. You can use '``llvm.annotation``' on
8737 any integer bit width.
8741 declare i8 @llvm.annotation.i8(i8 <val>, i8* <str>, i8* <str>, i32 <int>)
8742 declare i16 @llvm.annotation.i16(i16 <val>, i8* <str>, i8* <str>, i32 <int>)
8743 declare i32 @llvm.annotation.i32(i32 <val>, i8* <str>, i8* <str>, i32 <int>)
8744 declare i64 @llvm.annotation.i64(i64 <val>, i8* <str>, i8* <str>, i32 <int>)
8745 declare i256 @llvm.annotation.i256(i256 <val>, i8* <str>, i8* <str>, i32 <int>)
8750 The '``llvm.annotation``' intrinsic.
8755 The first argument is an integer value (result of some expression), the
8756 second is a pointer to a global string, the third is a pointer to a
8757 global string which is the source file name, and the last argument is
8758 the line number. It returns the value of the first argument.
8763 This intrinsic allows annotations to be put on arbitrary expressions
8764 with arbitrary strings. This can be useful for special purpose
8765 optimizations that want to look for these annotations. These have no
8766 other defined use; they are ignored by code generation and optimization.
8768 '``llvm.trap``' Intrinsic
8769 ^^^^^^^^^^^^^^^^^^^^^^^^^
8776 declare void @llvm.trap() noreturn nounwind
8781 The '``llvm.trap``' intrinsic.
8791 This intrinsic is lowered to the target dependent trap instruction. If
8792 the target does not have a trap instruction, this intrinsic will be
8793 lowered to a call of the ``abort()`` function.
8795 '``llvm.debugtrap``' Intrinsic
8796 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8803 declare void @llvm.debugtrap() nounwind
8808 The '``llvm.debugtrap``' intrinsic.
8818 This intrinsic is lowered to code which is intended to cause an
8819 execution trap with the intention of requesting the attention of a
8822 '``llvm.stackprotector``' Intrinsic
8823 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8830 declare void @llvm.stackprotector(i8* <guard>, i8** <slot>)
8835 The ``llvm.stackprotector`` intrinsic takes the ``guard`` and stores it
8836 onto the stack at ``slot``. The stack slot is adjusted to ensure that it
8837 is placed on the stack before local variables.
8842 The ``llvm.stackprotector`` intrinsic requires two pointer arguments.
8843 The first argument is the value loaded from the stack guard
8844 ``@__stack_chk_guard``. The second variable is an ``alloca`` that has
8845 enough space to hold the value of the guard.
8850 This intrinsic causes the prologue/epilogue inserter to force the position of
8851 the ``AllocaInst`` stack slot to be before local variables on the stack. This is
8852 to ensure that if a local variable on the stack is overwritten, it will destroy
8853 the value of the guard. When the function exits, the guard on the stack is
8854 checked against the original guard by ``llvm.stackprotectorcheck``. If they are
8855 different, then ``llvm.stackprotectorcheck`` causes the program to abort by
8856 calling the ``__stack_chk_fail()`` function.
8858 '``llvm.stackprotectorcheck``' Intrinsic
8859 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8866 declare void @llvm.stackprotectorcheck(i8** <guard>)
8871 The ``llvm.stackprotectorcheck`` intrinsic compares ``guard`` against an already
8872 created stack protector and if they are not equal calls the
8873 ``__stack_chk_fail()`` function.
8878 The ``llvm.stackprotectorcheck`` intrinsic requires one pointer argument, the
8879 the variable ``@__stack_chk_guard``.
8884 This intrinsic is provided to perform the stack protector check by comparing
8885 ``guard`` with the stack slot created by ``llvm.stackprotector`` and if the
8886 values do not match call the ``__stack_chk_fail()`` function.
8888 The reason to provide this as an IR level intrinsic instead of implementing it
8889 via other IR operations is that in order to perform this operation at the IR
8890 level without an intrinsic, one would need to create additional basic blocks to
8891 handle the success/failure cases. This makes it difficult to stop the stack
8892 protector check from disrupting sibling tail calls in Codegen. With this
8893 intrinsic, we are able to generate the stack protector basic blocks late in
8894 codegen after the tail call decision has occurred.
8896 '``llvm.objectsize``' Intrinsic
8897 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8904 declare i32 @llvm.objectsize.i32(i8* <object>, i1 <min>)
8905 declare i64 @llvm.objectsize.i64(i8* <object>, i1 <min>)
8910 The ``llvm.objectsize`` intrinsic is designed to provide information to
8911 the optimizers to determine at compile time whether a) an operation
8912 (like memcpy) will overflow a buffer that corresponds to an object, or
8913 b) that a runtime check for overflow isn't necessary. An object in this
8914 context means an allocation of a specific class, structure, array, or
8920 The ``llvm.objectsize`` intrinsic takes two arguments. The first
8921 argument is a pointer to or into the ``object``. The second argument is
8922 a boolean and determines whether ``llvm.objectsize`` returns 0 (if true)
8923 or -1 (if false) when the object size is unknown. The second argument
8924 only accepts constants.
8929 The ``llvm.objectsize`` intrinsic is lowered to a constant representing
8930 the size of the object concerned. If the size cannot be determined at
8931 compile time, ``llvm.objectsize`` returns ``i32/i64 -1 or 0`` (depending
8932 on the ``min`` argument).
8934 '``llvm.expect``' Intrinsic
8935 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8942 declare i32 @llvm.expect.i32(i32 <val>, i32 <expected_val>)
8943 declare i64 @llvm.expect.i64(i64 <val>, i64 <expected_val>)
8948 The ``llvm.expect`` intrinsic provides information about expected (the
8949 most probable) value of ``val``, which can be used by optimizers.
8954 The ``llvm.expect`` intrinsic takes two arguments. The first argument is
8955 a value. The second argument is an expected value, this needs to be a
8956 constant value, variables are not allowed.
8961 This intrinsic is lowered to the ``val``.
8963 '``llvm.donothing``' Intrinsic
8964 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8971 declare void @llvm.donothing() nounwind readnone
8976 The ``llvm.donothing`` intrinsic doesn't perform any operation. It's the
8977 only intrinsic that can be called with an invoke instruction.
8987 This intrinsic does nothing, and it's removed by optimizers and ignored
8990 Stack Map Intrinsics
8991 --------------------
8993 LLVM provides experimental intrinsics to support runtime patching
8994 mechanisms commonly desired in dynamic language JITs. These intrinsics
8995 are described in :doc:`StackMaps`.