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 that 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. The ``"\01"`` prefix
83 can be used on global variables to suppress mangling.
84 #. Unnamed values are represented as an unsigned numeric value with
85 their prefix. For example, ``%12``, ``@2``, ``%44``.
86 #. Constants, which are described in the section Constants_ below.
88 LLVM requires that values start with a prefix for two reasons: Compilers
89 don't need to worry about name clashes with reserved words, and the set
90 of reserved words may be expanded in the future without penalty.
91 Additionally, unnamed identifiers allow a compiler to quickly come up
92 with a temporary variable without having to avoid symbol table
95 Reserved words in LLVM are very similar to reserved words in other
96 languages. There are keywords for different opcodes ('``add``',
97 '``bitcast``', '``ret``', etc...), for primitive type names ('``void``',
98 '``i32``', etc...), and others. These reserved words cannot conflict
99 with variable names, because none of them start with a prefix character
100 (``'%'`` or ``'@'``).
102 Here is an example of LLVM code to multiply the integer variable
109 %result = mul i32 %X, 8
111 After strength reduction:
115 %result = shl i32 %X, 3
121 %0 = add i32 %X, %X ; yields i32:%0
122 %1 = add i32 %0, %0 ; yields i32:%1
123 %result = add i32 %1, %1
125 This last way of multiplying ``%X`` by 8 illustrates several important
126 lexical features of LLVM:
128 #. Comments are delimited with a '``;``' and go until the end of line.
129 #. Unnamed temporaries are created when the result of a computation is
130 not assigned to a named value.
131 #. Unnamed temporaries are numbered sequentially (using a per-function
132 incrementing counter, starting with 0). Note that basic blocks are
133 included in this numbering. For example, if the entry basic block is not
134 given a label name, then it will get number 0.
136 It also shows a convention that we follow in this document. When
137 demonstrating instructions, we will follow an instruction with a comment
138 that defines the type and name of value produced.
146 LLVM programs are composed of ``Module``'s, each of which is a
147 translation unit of the input programs. Each module consists of
148 functions, global variables, and symbol table entries. Modules may be
149 combined together with the LLVM linker, which merges function (and
150 global variable) definitions, resolves forward declarations, and merges
151 symbol table entries. Here is an example of the "hello world" module:
155 ; Declare the string constant as a global constant.
156 @.str = private unnamed_addr constant [13 x i8] c"hello world\0A\00"
158 ; External declaration of the puts function
159 declare i32 @puts(i8* nocapture) nounwind
161 ; Definition of main function
162 define i32 @main() { ; i32()*
163 ; Convert [13 x i8]* to i8 *...
164 %cast210 = getelementptr [13 x i8]* @.str, i64 0, i64 0
166 ; Call puts function to write out the string to stdout.
167 call i32 @puts(i8* %cast210)
172 !0 = metadata !{i32 42, null, metadata !"string"}
175 This example is made up of a :ref:`global variable <globalvars>` named
176 "``.str``", an external declaration of the "``puts``" function, a
177 :ref:`function definition <functionstructure>` for "``main``" and
178 :ref:`named metadata <namedmetadatastructure>` "``foo``".
180 In general, a module is made up of a list of global values (where both
181 functions and global variables are global values). Global values are
182 represented by a pointer to a memory location (in this case, a pointer
183 to an array of char, and a pointer to a function), and have one of the
184 following :ref:`linkage types <linkage>`.
191 All Global Variables and Functions have one of the following types of
195 Global values with "``private``" linkage are only directly
196 accessible by objects in the current module. In particular, linking
197 code into a module with an private global value may cause the
198 private to be renamed as necessary to avoid collisions. Because the
199 symbol is private to the module, all references can be updated. This
200 doesn't show up in any symbol table in the object file.
202 Similar to private, but the value shows as a local symbol
203 (``STB_LOCAL`` in the case of ELF) in the object file. This
204 corresponds to the notion of the '``static``' keyword in C.
205 ``available_externally``
206 Globals with "``available_externally``" linkage are never emitted
207 into the object file corresponding to the LLVM module. They exist to
208 allow inlining and other optimizations to take place given knowledge
209 of the definition of the global, which is known to be somewhere
210 outside the module. Globals with ``available_externally`` linkage
211 are allowed to be discarded at will, and are otherwise the same as
212 ``linkonce_odr``. This linkage type is only allowed on definitions,
215 Globals with "``linkonce``" linkage are merged with other globals of
216 the same name when linkage occurs. This can be used to implement
217 some forms of inline functions, templates, or other code which must
218 be generated in each translation unit that uses it, but where the
219 body may be overridden with a more definitive definition later.
220 Unreferenced ``linkonce`` globals are allowed to be discarded. Note
221 that ``linkonce`` linkage does not actually allow the optimizer to
222 inline the body of this function into callers because it doesn't
223 know if this definition of the function is the definitive definition
224 within the program or whether it will be overridden by a stronger
225 definition. To enable inlining and other optimizations, use
226 "``linkonce_odr``" linkage.
228 "``weak``" linkage has the same merging semantics as ``linkonce``
229 linkage, except that unreferenced globals with ``weak`` linkage may
230 not be discarded. This is used for globals that are declared "weak"
233 "``common``" linkage is most similar to "``weak``" linkage, but they
234 are used for tentative definitions in C, such as "``int X;``" at
235 global scope. Symbols with "``common``" linkage are merged in the
236 same way as ``weak symbols``, and they may not be deleted if
237 unreferenced. ``common`` symbols may not have an explicit section,
238 must have a zero initializer, and may not be marked
239 ':ref:`constant <globalvars>`'. Functions and aliases may not have
242 .. _linkage_appending:
245 "``appending``" linkage may only be applied to global variables of
246 pointer to array type. When two global variables with appending
247 linkage are linked together, the two global arrays are appended
248 together. This is the LLVM, typesafe, equivalent of having the
249 system linker append together "sections" with identical names when
252 The semantics of this linkage follow the ELF object file model: the
253 symbol is weak until linked, if not linked, the symbol becomes null
254 instead of being an undefined reference.
255 ``linkonce_odr``, ``weak_odr``
256 Some languages allow differing globals to be merged, such as two
257 functions with different semantics. Other languages, such as
258 ``C++``, ensure that only equivalent globals are ever merged (the
259 "one definition rule" --- "ODR"). Such languages can use the
260 ``linkonce_odr`` and ``weak_odr`` linkage types to indicate that the
261 global will only be merged with equivalent globals. These linkage
262 types are otherwise the same as their non-``odr`` versions.
264 If none of the above identifiers are used, the global is externally
265 visible, meaning that it participates in linkage and can be used to
266 resolve external symbol references.
268 It is illegal for a function *declaration* to have any linkage type
269 other than ``external`` or ``extern_weak``.
276 LLVM :ref:`functions <functionstructure>`, :ref:`calls <i_call>` and
277 :ref:`invokes <i_invoke>` can all have an optional calling convention
278 specified for the call. The calling convention of any pair of dynamic
279 caller/callee must match, or the behavior of the program is undefined.
280 The following calling conventions are supported by LLVM, and more may be
283 "``ccc``" - The C calling convention
284 This calling convention (the default if no other calling convention
285 is specified) matches the target C calling conventions. This calling
286 convention supports varargs function calls and tolerates some
287 mismatch in the declared prototype and implemented declaration of
288 the function (as does normal C).
289 "``fastcc``" - The fast calling convention
290 This calling convention attempts to make calls as fast as possible
291 (e.g. by passing things in registers). This calling convention
292 allows the target to use whatever tricks it wants to produce fast
293 code for the target, without having to conform to an externally
294 specified ABI (Application Binary Interface). `Tail calls can only
295 be optimized when this, the GHC or the HiPE convention is
296 used. <CodeGenerator.html#id80>`_ This calling convention does not
297 support varargs and requires the prototype of all callees to exactly
298 match the prototype of the function definition.
299 "``coldcc``" - The cold calling convention
300 This calling convention attempts to make code in the caller as
301 efficient as possible under the assumption that the call is not
302 commonly executed. As such, these calls often preserve all registers
303 so that the call does not break any live ranges in the caller side.
304 This calling convention does not support varargs and requires the
305 prototype of all callees to exactly match the prototype of the
306 function definition. Furthermore the inliner doesn't consider such function
308 "``cc 10``" - GHC convention
309 This calling convention has been implemented specifically for use by
310 the `Glasgow Haskell Compiler (GHC) <http://www.haskell.org/ghc>`_.
311 It passes everything in registers, going to extremes to achieve this
312 by disabling callee save registers. This calling convention should
313 not be used lightly but only for specific situations such as an
314 alternative to the *register pinning* performance technique often
315 used when implementing functional programming languages. At the
316 moment only X86 supports this convention and it has the following
319 - On *X86-32* only supports up to 4 bit type parameters. No
320 floating point types are supported.
321 - On *X86-64* only supports up to 10 bit type parameters and 6
322 floating point parameters.
324 This calling convention supports `tail call
325 optimization <CodeGenerator.html#id80>`_ but requires both the
326 caller and callee are using it.
327 "``cc 11``" - The HiPE calling convention
328 This calling convention has been implemented specifically for use by
329 the `High-Performance Erlang
330 (HiPE) <http://www.it.uu.se/research/group/hipe/>`_ compiler, *the*
331 native code compiler of the `Ericsson's Open Source Erlang/OTP
332 system <http://www.erlang.org/download.shtml>`_. It uses more
333 registers for argument passing than the ordinary C calling
334 convention and defines no callee-saved registers. The calling
335 convention properly supports `tail call
336 optimization <CodeGenerator.html#id80>`_ but requires that both the
337 caller and the callee use it. It uses a *register pinning*
338 mechanism, similar to GHC's convention, for keeping frequently
339 accessed runtime components pinned to specific hardware registers.
340 At the moment only X86 supports this convention (both 32 and 64
342 "``webkit_jscc``" - WebKit's JavaScript calling convention
343 This calling convention has been implemented for `WebKit FTL JIT
344 <https://trac.webkit.org/wiki/FTLJIT>`_. It passes arguments on the
345 stack right to left (as cdecl does), and returns a value in the
346 platform's customary return register.
347 "``anyregcc``" - Dynamic calling convention for code patching
348 This is a special convention that supports patching an arbitrary code
349 sequence in place of a call site. This convention forces the call
350 arguments into registers but allows them to be dynamcially
351 allocated. This can currently only be used with calls to
352 llvm.experimental.patchpoint because only this intrinsic records
353 the location of its arguments in a side table. See :doc:`StackMaps`.
354 "``preserve_mostcc``" - The `PreserveMost` calling convention
355 This calling convention attempts to make the code in the caller as little
356 intrusive as possible. This calling convention behaves identical to the `C`
357 calling convention on how arguments and return values are passed, but it
358 uses a different set of caller/callee-saved registers. This alleviates the
359 burden of saving and recovering a large register set before and after the
360 call in the caller. If the arguments are passed in callee-saved registers,
361 then they will be preserved by the callee across the call. This doesn't
362 apply for values returned in callee-saved registers.
364 - On X86-64 the callee preserves all general purpose registers, except for
365 R11. R11 can be used as a scratch register. Floating-point registers
366 (XMMs/YMMs) are not preserved and need to be saved by the caller.
368 The idea behind this convention is to support calls to runtime functions
369 that have a hot path and a cold path. The hot path is usually a small piece
370 of code that doesn't many registers. The cold path might need to call out to
371 another function and therefore only needs to preserve the caller-saved
372 registers, which haven't already been saved by the caller. The
373 `PreserveMost` calling convention is very similar to the `cold` calling
374 convention in terms of caller/callee-saved registers, but they are used for
375 different types of function calls. `coldcc` is for function calls that are
376 rarely executed, whereas `preserve_mostcc` function calls are intended to be
377 on the hot path and definitely executed a lot. Furthermore `preserve_mostcc`
378 doesn't prevent the inliner from inlining the function call.
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. If
394 the arguments are passed in callee-saved registers, then they will be
395 preserved by the callee across the call. This doesn't apply for values
396 returned in callee-saved registers.
398 - On X86-64 the callee preserves all general purpose registers, except for
399 R11. R11 can be used as a scratch register. Furthermore it also preserves
400 all floating-point registers (XMMs/YMMs).
402 The idea behind this convention is to support calls to runtime functions
403 that don't need to call out to any other functions.
405 This calling convention, like the `PreserveMost` calling convention, will be
406 used by a future version of the ObjectiveC runtime and should be considered
407 experimental at this time.
408 "``cc <n>``" - Numbered convention
409 Any calling convention may be specified by number, allowing
410 target-specific calling conventions to be used. Target specific
411 calling conventions start at 64.
413 More calling conventions can be added/defined on an as-needed basis, to
414 support Pascal conventions or any other well-known target-independent
417 .. _visibilitystyles:
422 All Global Variables and Functions have one of the following visibility
425 "``default``" - Default style
426 On targets that use the ELF object file format, default visibility
427 means that the declaration is visible to other modules and, in
428 shared libraries, means that the declared entity may be overridden.
429 On Darwin, default visibility means that the declaration is visible
430 to other modules. Default visibility corresponds to "external
431 linkage" in the language.
432 "``hidden``" - Hidden style
433 Two declarations of an object with hidden visibility refer to the
434 same object if they are in the same shared object. Usually, hidden
435 visibility indicates that the symbol will not be placed into the
436 dynamic symbol table, so no other module (executable or shared
437 library) can reference it directly.
438 "``protected``" - Protected style
439 On ELF, protected visibility indicates that the symbol will be
440 placed in the dynamic symbol table, but that references within the
441 defining module will bind to the local symbol. That is, the symbol
442 cannot be overridden by another module.
444 A symbol with ``internal`` or ``private`` linkage must have ``default``
452 All Global Variables, Functions and Aliases can have one of the following
456 "``dllimport``" causes the compiler to reference a function or variable via
457 a global pointer to a pointer that is set up by the DLL exporting the
458 symbol. On Microsoft Windows targets, the pointer name is formed by
459 combining ``__imp_`` and the function or variable name.
461 "``dllexport``" causes the compiler to provide a global pointer to a pointer
462 in a DLL, so that it can be referenced with the ``dllimport`` attribute. On
463 Microsoft Windows targets, the pointer name is formed by combining
464 ``__imp_`` and the function or variable name. Since this storage class
465 exists for defining a dll interface, the compiler, assembler and linker know
466 it is externally referenced and must refrain from deleting the symbol.
470 Thread Local Storage Models
471 ---------------------------
473 A variable may be defined as ``thread_local``, which means that it will
474 not be shared by threads (each thread will have a separated copy of the
475 variable). Not all targets support thread-local variables. Optionally, a
476 TLS model may be specified:
479 For variables that are only used within the current shared library.
481 For variables in modules that will not be loaded dynamically.
483 For variables defined in the executable and only used within it.
485 If no explicit model is given, the "general dynamic" model is used.
487 The models correspond to the ELF TLS models; see `ELF Handling For
488 Thread-Local Storage <http://people.redhat.com/drepper/tls.pdf>`_ for
489 more information on under which circumstances the different models may
490 be used. The target may choose a different TLS model if the specified
491 model is not supported, or if a better choice of model can be made.
493 A model can also be specified in a alias, but then it only governs how
494 the alias is accessed. It will not have any effect in the aliasee.
501 LLVM IR allows you to specify both "identified" and "literal" :ref:`structure
502 types <t_struct>`. Literal types are uniqued structurally, but identified types
503 are never uniqued. An :ref:`opaque structural type <t_opaque>` can also be used
504 to forward declare a type that is not yet available.
506 An example of a identified structure specification is:
510 %mytype = type { %mytype*, i32 }
512 Prior to the LLVM 3.0 release, identified types were structurally uniqued. Only
513 literal types are uniqued in recent versions of LLVM.
520 Global variables define regions of memory allocated at compilation time
523 Global variables definitions must be initialized.
525 Global variables in other translation units can also be declared, in which
526 case they don't have an initializer.
528 Either global variable definitions or declarations may have an explicit section
529 to be placed in and may have an optional explicit alignment specified.
531 A variable may be defined as a global ``constant``, which indicates that
532 the contents of the variable will **never** be modified (enabling better
533 optimization, allowing the global data to be placed in the read-only
534 section of an executable, etc). Note that variables that need runtime
535 initialization cannot be marked ``constant`` as there is a store to the
538 LLVM explicitly allows *declarations* of global variables to be marked
539 constant, even if the final definition of the global is not. This
540 capability can be used to enable slightly better optimization of the
541 program, but requires the language definition to guarantee that
542 optimizations based on the 'constantness' are valid for the translation
543 units that do not include the definition.
545 As SSA values, global variables define pointer values that are in scope
546 (i.e. they dominate) all basic blocks in the program. Global variables
547 always define a pointer to their "content" type because they describe a
548 region of memory, and all memory objects in LLVM are accessed through
551 Global variables can be marked with ``unnamed_addr`` which indicates
552 that the address is not significant, only the content. Constants marked
553 like this can be merged with other constants if they have the same
554 initializer. Note that a constant with significant address *can* be
555 merged with a ``unnamed_addr`` constant, the result being a constant
556 whose address is significant.
558 A global variable may be declared to reside in a target-specific
559 numbered address space. For targets that support them, address spaces
560 may affect how optimizations are performed and/or what target
561 instructions are used to access the variable. The default address space
562 is zero. The address space qualifier must precede any other attributes.
564 LLVM allows an explicit section to be specified for globals. If the
565 target supports it, it will emit globals to the section specified.
566 Additionally, the global can placed in a comdat if the target has the necessary
569 By default, global initializers are optimized by assuming that global
570 variables defined within the module are not modified from their
571 initial values before the start of the global initializer. This is
572 true even for variables potentially accessible from outside the
573 module, including those with external linkage or appearing in
574 ``@llvm.used`` or dllexported variables. This assumption may be suppressed
575 by marking the variable with ``externally_initialized``.
577 An explicit alignment may be specified for a global, which must be a
578 power of 2. If not present, or if the alignment is set to zero, the
579 alignment of the global is set by the target to whatever it feels
580 convenient. If an explicit alignment is specified, the global is forced
581 to have exactly that alignment. Targets and optimizers are not allowed
582 to over-align the global if the global has an assigned section. In this
583 case, the extra alignment could be observable: for example, code could
584 assume that the globals are densely packed in their section and try to
585 iterate over them as an array, alignment padding would break this
586 iteration. The maximum alignment is ``1 << 29``.
588 Globals can also have a :ref:`DLL storage class <dllstorageclass>`.
590 Variables and aliasaes can have a
591 :ref:`Thread Local Storage Model <tls_model>`.
595 [@<GlobalVarName> =] [Linkage] [Visibility] [DLLStorageClass] [ThreadLocal]
596 [unnamed_addr] [AddrSpace] [ExternallyInitialized]
597 <global | constant> <Type> [<InitializerConstant>]
598 [, section "name"] [, align <Alignment>]
600 For example, the following defines a global in a numbered address space
601 with an initializer, section, and alignment:
605 @G = addrspace(5) constant float 1.0, section "foo", align 4
607 The following example just declares a global variable
611 @G = external global i32
613 The following example defines a thread-local global with the
614 ``initialexec`` TLS model:
618 @G = thread_local(initialexec) global i32 0, align 4
620 .. _functionstructure:
625 LLVM function definitions consist of the "``define``" keyword, an
626 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
627 style <visibility>`, an optional :ref:`DLL storage class <dllstorageclass>`,
628 an optional :ref:`calling convention <callingconv>`,
629 an optional ``unnamed_addr`` attribute, a return type, an optional
630 :ref:`parameter attribute <paramattrs>` for the return type, a function
631 name, a (possibly empty) argument list (each with optional :ref:`parameter
632 attributes <paramattrs>`), optional :ref:`function attributes <fnattrs>`,
633 an optional section, an optional alignment,
634 an optional :ref:`comdat <langref_comdats>`,
635 an optional :ref:`garbage collector name <gc>`, an optional :ref:`prefix <prefixdata>`, an opening
636 curly brace, a list of basic blocks, and a closing curly brace.
638 LLVM function declarations consist of the "``declare``" keyword, an
639 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
640 style <visibility>`, an optional :ref:`DLL storage class <dllstorageclass>`,
641 an optional :ref:`calling convention <callingconv>`,
642 an optional ``unnamed_addr`` attribute, a return type, an optional
643 :ref:`parameter attribute <paramattrs>` for the return type, a function
644 name, a possibly empty list of arguments, an optional alignment, an optional
645 :ref:`garbage collector name <gc>` and an optional :ref:`prefix <prefixdata>`.
647 A function definition contains a list of basic blocks, forming the CFG (Control
648 Flow Graph) for the function. Each basic block may optionally start with a label
649 (giving the basic block a symbol table entry), contains a list of instructions,
650 and ends with a :ref:`terminator <terminators>` instruction (such as a branch or
651 function return). If an explicit label is not provided, a block is assigned an
652 implicit numbered label, using the next value from the same counter as used for
653 unnamed temporaries (:ref:`see above<identifiers>`). For example, if a function
654 entry block does not have an explicit label, it will be assigned label "%0",
655 then the first unnamed temporary in that block will be "%1", etc.
657 The first basic block in a function is special in two ways: it is
658 immediately executed on entrance to the function, and it is not allowed
659 to have predecessor basic blocks (i.e. there can not be any branches to
660 the entry block of a function). Because the block can have no
661 predecessors, it also cannot have any :ref:`PHI nodes <i_phi>`.
663 LLVM allows an explicit section to be specified for functions. If the
664 target supports it, it will emit functions to the section specified.
665 Additionally, the function can placed in a COMDAT.
667 An explicit alignment may be specified for a function. If not present,
668 or if the alignment is set to zero, the alignment of the function is set
669 by the target to whatever it feels convenient. If an explicit alignment
670 is specified, the function is forced to have at least that much
671 alignment. All alignments must be a power of 2.
673 If the ``unnamed_addr`` attribute is given, the address is know to not
674 be significant and two identical functions can be merged.
678 define [linkage] [visibility] [DLLStorageClass]
680 <ResultType> @<FunctionName> ([argument list])
681 [unnamed_addr] [fn Attrs] [section "name"] [comdat $<ComdatName>]
682 [align N] [gc] [prefix Constant] { ... }
689 Aliases, unlike function or variables, don't create any new data. They
690 are just a new symbol and metadata for an existing position.
692 Aliases have a name and an aliasee that is either a global value or a
695 Aliases may have an optional :ref:`linkage type <linkage>`, an optional
696 :ref:`visibility style <visibility>`, an optional :ref:`DLL storage class
697 <dllstorageclass>` and an optional :ref:`tls model <tls_model>`.
701 @<Name> = [Linkage] [Visibility] [DLLStorageClass] [ThreadLocal] [unnamed_addr] alias <AliaseeTy> @<Aliasee>
703 The linkage must be one of ``private``, ``internal``, ``linkonce``, ``weak``,
704 ``linkonce_odr``, ``weak_odr``, ``external``. Note that some system linkers
705 might not correctly handle dropping a weak symbol that is aliased.
707 Alias that are not ``unnamed_addr`` are guaranteed to have the same address as
708 the aliasee expression. ``unnamed_addr`` ones are only guaranteed to point
711 Since aliases are only a second name, some restrictions apply, of which
712 some can only be checked when producing an object file:
714 * The expression defining the aliasee must be computable at assembly
715 time. Since it is just a name, no relocations can be used.
717 * No alias in the expression can be weak as the possibility of the
718 intermediate alias being overridden cannot be represented in an
721 * No global value in the expression can be a declaration, since that
722 would require a relocation, which is not possible.
729 Comdat IR provides access to COFF and ELF object file COMDAT functionality.
731 Comdats have a name which represents the COMDAT key. All global objects that
732 specify this key will only end up in the final object file if the linker chooses
733 that key over some other key. Aliases are placed in the same COMDAT that their
734 aliasee computes to, if any.
736 Comdats have a selection kind to provide input on how the linker should
737 choose between keys in two different object files.
741 $<Name> = comdat SelectionKind
743 The selection kind must be one of the following:
746 The linker may choose any COMDAT key, the choice is arbitrary.
748 The linker may choose any COMDAT key but the sections must contain the
751 The linker will choose the section containing the largest COMDAT key.
753 The linker requires that only section with this COMDAT key exist.
755 The linker may choose any COMDAT key but the sections must contain the
758 Note that the Mach-O platform doesn't support COMDATs and ELF only supports
759 ``any`` as a selection kind.
761 Here is an example of a COMDAT group where a function will only be selected if
762 the COMDAT key's section is the largest:
766 $foo = comdat largest
767 @foo = global i32 2, comdat $foo
769 define void @bar() comdat $foo {
773 In a COFF object file, this will create a COMDAT section with selection kind
774 ``IMAGE_COMDAT_SELECT_LARGEST`` containing the contents of the ``@foo`` symbol
775 and another COMDAT section with selection kind
776 ``IMAGE_COMDAT_SELECT_ASSOCIATIVE`` which is associated with the first COMDAT
777 section and contains the contents of the ``@baz`` symbol.
779 There are some restrictions on the properties of the global object.
780 It, or an alias to it, must have the same name as the COMDAT group when
782 The contents and size of this object may be used during link-time to determine
783 which COMDAT groups get selected depending on the selection kind.
784 Because the name of the object must match the name of the COMDAT group, the
785 linkage of the global object must not be local; local symbols can get renamed
786 if a collision occurs in the symbol table.
788 The combined use of COMDATS and section attributes may yield surprising results.
795 @g1 = global i32 42, section "sec", comdat $foo
796 @g2 = global i32 42, section "sec", comdat $bar
798 From the object file perspective, this requires the creation of two sections
799 with the same name. This is necessary because both globals belong to different
800 COMDAT groups and COMDATs, at the object file level, are represented by
803 Note that certain IR constructs like global variables and functions may create
804 COMDATs in the object file in addition to any which are specified using COMDAT
805 IR. This arises, for example, when a global variable has linkonce_odr linkage.
807 .. _namedmetadatastructure:
812 Named metadata is a collection of metadata. :ref:`Metadata
813 nodes <metadata>` (but not metadata strings) are the only valid
814 operands for a named metadata.
818 ; Some unnamed metadata nodes, which are referenced by the named metadata.
819 !0 = metadata !{metadata !"zero"}
820 !1 = metadata !{metadata !"one"}
821 !2 = metadata !{metadata !"two"}
823 !name = !{!0, !1, !2}
830 The return type and each parameter of a function type may have a set of
831 *parameter attributes* associated with them. Parameter attributes are
832 used to communicate additional information about the result or
833 parameters of a function. Parameter attributes are considered to be part
834 of the function, not of the function type, so functions with different
835 parameter attributes can have the same function type.
837 Parameter attributes are simple keywords that follow the type specified.
838 If multiple parameter attributes are needed, they are space separated.
843 declare i32 @printf(i8* noalias nocapture, ...)
844 declare i32 @atoi(i8 zeroext)
845 declare signext i8 @returns_signed_char()
847 Note that any attributes for the function result (``nounwind``,
848 ``readonly``) come immediately after the argument list.
850 Currently, only the following parameter attributes are defined:
853 This indicates to the code generator that the parameter or return
854 value should be zero-extended to the extent required by the target's
855 ABI (which is usually 32-bits, but is 8-bits for a i1 on x86-64) by
856 the caller (for a parameter) or the callee (for a return value).
858 This indicates to the code generator that the parameter or return
859 value should be sign-extended to the extent required by the target's
860 ABI (which is usually 32-bits) by the caller (for a parameter) or
861 the callee (for a return value).
863 This indicates that this parameter or return value should be treated
864 in a special target-dependent fashion during while emitting code for
865 a function call or return (usually, by putting it in a register as
866 opposed to memory, though some targets use it to distinguish between
867 two different kinds of registers). Use of this attribute is
870 This indicates that the pointer parameter should really be passed by
871 value to the function. The attribute implies that a hidden copy of
872 the pointee is made between the caller and the callee, so the callee
873 is unable to modify the value in the caller. This attribute is only
874 valid on LLVM pointer arguments. It is generally used to pass
875 structs and arrays by value, but is also valid on pointers to
876 scalars. The copy is considered to belong to the caller not the
877 callee (for example, ``readonly`` functions should not write to
878 ``byval`` parameters). This is not a valid attribute for return
881 The byval attribute also supports specifying an alignment with the
882 align attribute. It indicates the alignment of the stack slot to
883 form and the known alignment of the pointer specified to the call
884 site. If the alignment is not specified, then the code generator
885 makes a target-specific assumption.
891 The ``inalloca`` argument attribute allows the caller to take the
892 address of outgoing stack arguments. An ``inalloca`` argument must
893 be a pointer to stack memory produced by an ``alloca`` instruction.
894 The alloca, or argument allocation, must also be tagged with the
895 inalloca keyword. Only the last argument may have the ``inalloca``
896 attribute, and that argument is guaranteed to be passed in memory.
898 An argument allocation may be used by a call at most once because
899 the call may deallocate it. The ``inalloca`` attribute cannot be
900 used in conjunction with other attributes that affect argument
901 storage, like ``inreg``, ``nest``, ``sret``, or ``byval``. The
902 ``inalloca`` attribute also disables LLVM's implicit lowering of
903 large aggregate return values, which means that frontend authors
904 must lower them with ``sret`` pointers.
906 When the call site is reached, the argument allocation must have
907 been the most recent stack allocation that is still live, or the
908 results are undefined. It is possible to allocate additional stack
909 space after an argument allocation and before its call site, but it
910 must be cleared off with :ref:`llvm.stackrestore
913 See :doc:`InAlloca` for more information on how to use this
917 This indicates that the pointer parameter specifies the address of a
918 structure that is the return value of the function in the source
919 program. This pointer must be guaranteed by the caller to be valid:
920 loads and stores to the structure may be assumed by the callee
921 not to trap and to be properly aligned. This may only be applied to
922 the first parameter. This is not a valid attribute for return
926 This indicates that the pointer value may be assumed by the optimizer to
927 have the specified alignment.
929 Note that this attribute has additional semantics when combined with the
935 This indicates that pointer values :ref:`based <pointeraliasing>` on
936 the argument or return value do not alias pointer values that are
937 not *based* on it, ignoring certain "irrelevant" dependencies. For a
938 call to the parent function, dependencies between memory references
939 from before or after the call and from those during the call are
940 "irrelevant" to the ``noalias`` keyword for the arguments and return
941 value used in that call. The caller shares the responsibility with
942 the callee for ensuring that these requirements are met. For further
943 details, please see the discussion of the NoAlias response in :ref:`alias
944 analysis <Must, May, or No>`.
946 Note that this definition of ``noalias`` is intentionally similar
947 to the definition of ``restrict`` in C99 for function arguments,
948 though it is slightly weaker.
950 For function return values, C99's ``restrict`` is not meaningful,
951 while LLVM's ``noalias`` is.
953 This indicates that the callee does not make any copies of the
954 pointer that outlive the callee itself. This is not a valid
955 attribute for return values.
960 This indicates that the pointer parameter can be excised using the
961 :ref:`trampoline intrinsics <int_trampoline>`. This is not a valid
962 attribute for return values and can only be applied to one parameter.
965 This indicates that the function always returns the argument as its return
966 value. This is an optimization hint to the code generator when generating
967 the caller, allowing tail call optimization and omission of register saves
968 and restores in some cases; it is not checked or enforced when generating
969 the callee. The parameter and the function return type must be valid
970 operands for the :ref:`bitcast instruction <i_bitcast>`. This is not a
971 valid attribute for return values and can only be applied to one parameter.
974 This indicates that the parameter or return pointer is not null. This
975 attribute may only be applied to pointer typed parameters. This is not
976 checked or enforced by LLVM, the caller must ensure that the pointer
977 passed in is non-null, or the callee must ensure that the returned pointer
980 ``dereferenceable(<n>)``
981 This indicates that the parameter or return pointer is dereferenceable. This
982 attribute may only be applied to pointer typed parameters. A pointer that
983 is dereferenceable can be loaded from speculatively without a risk of
984 trapping. The number of bytes known to be dereferenceable must be provided
985 in parentheses. It is legal for the number of bytes to be less than the
986 size of the pointee type. The ``nonnull`` attribute does not imply
987 dereferenceability (consider a pointer to one element past the end of an
988 array), however ``dereferenceable(<n>)`` does imply ``nonnull`` in
989 ``addrspace(0)`` (which is the default address space).
993 Garbage Collector Names
994 -----------------------
996 Each function may specify a garbage collector name, which is simply a
1001 define void @f() gc "name" { ... }
1003 The compiler declares the supported values of *name*. Specifying a
1004 collector will cause the compiler to alter its output in order to
1005 support the named garbage collection algorithm.
1012 Prefix data is data associated with a function which the code generator
1013 will emit immediately before the function body. The purpose of this feature
1014 is to allow frontends to associate language-specific runtime metadata with
1015 specific functions and make it available through the function pointer while
1016 still allowing the function pointer to be called. To access the data for a
1017 given function, a program may bitcast the function pointer to a pointer to
1018 the constant's type. This implies that the IR symbol points to the start
1021 To maintain the semantics of ordinary function calls, the prefix data must
1022 have a particular format. Specifically, it must begin with a sequence of
1023 bytes which decode to a sequence of machine instructions, valid for the
1024 module's target, which transfer control to the point immediately succeeding
1025 the prefix data, without performing any other visible action. This allows
1026 the inliner and other passes to reason about the semantics of the function
1027 definition without needing to reason about the prefix data. Obviously this
1028 makes the format of the prefix data highly target dependent.
1030 Prefix data is laid out as if it were an initializer for a global variable
1031 of the prefix data's type. No padding is automatically placed between the
1032 prefix data and the function body. If padding is required, it must be part
1035 A trivial example of valid prefix data for the x86 architecture is ``i8 144``,
1036 which encodes the ``nop`` instruction:
1038 .. code-block:: llvm
1040 define void @f() prefix i8 144 { ... }
1042 Generally prefix data can be formed by encoding a relative branch instruction
1043 which skips the metadata, as in this example of valid prefix data for the
1044 x86_64 architecture, where the first two bytes encode ``jmp .+10``:
1046 .. code-block:: llvm
1048 %0 = type <{ i8, i8, i8* }>
1050 define void @f() prefix %0 <{ i8 235, i8 8, i8* @md}> { ... }
1052 A function may have prefix data but no body. This has similar semantics
1053 to the ``available_externally`` linkage in that the data may be used by the
1054 optimizers but will not be emitted in the object file.
1061 Attribute groups are groups of attributes that are referenced by objects within
1062 the IR. They are important for keeping ``.ll`` files readable, because a lot of
1063 functions will use the same set of attributes. In the degenerative case of a
1064 ``.ll`` file that corresponds to a single ``.c`` file, the single attribute
1065 group will capture the important command line flags used to build that file.
1067 An attribute group is a module-level object. To use an attribute group, an
1068 object references the attribute group's ID (e.g. ``#37``). An object may refer
1069 to more than one attribute group. In that situation, the attributes from the
1070 different groups are merged.
1072 Here is an example of attribute groups for a function that should always be
1073 inlined, has a stack alignment of 4, and which shouldn't use SSE instructions:
1075 .. code-block:: llvm
1077 ; Target-independent attributes:
1078 attributes #0 = { alwaysinline alignstack=4 }
1080 ; Target-dependent attributes:
1081 attributes #1 = { "no-sse" }
1083 ; Function @f has attributes: alwaysinline, alignstack=4, and "no-sse".
1084 define void @f() #0 #1 { ... }
1091 Function attributes are set to communicate additional information about
1092 a function. Function attributes are considered to be part of the
1093 function, not of the function type, so functions with different function
1094 attributes can have the same function type.
1096 Function attributes are simple keywords that follow the type specified.
1097 If multiple attributes are needed, they are space separated. For
1100 .. code-block:: llvm
1102 define void @f() noinline { ... }
1103 define void @f() alwaysinline { ... }
1104 define void @f() alwaysinline optsize { ... }
1105 define void @f() optsize { ... }
1108 This attribute indicates that, when emitting the prologue and
1109 epilogue, the backend should forcibly align the stack pointer.
1110 Specify the desired alignment, which must be a power of two, in
1113 This attribute indicates that the inliner should attempt to inline
1114 this function into callers whenever possible, ignoring any active
1115 inlining size threshold for this caller.
1117 This indicates that the callee function at a call site should be
1118 recognized as a built-in function, even though the function's declaration
1119 uses the ``nobuiltin`` attribute. This is only valid at call sites for
1120 direct calls to functions that are declared with the ``nobuiltin``
1123 This attribute indicates that this function is rarely called. When
1124 computing edge weights, basic blocks post-dominated by a cold
1125 function call are also considered to be cold; and, thus, given low
1128 This attribute indicates that the source code contained a hint that
1129 inlining this function is desirable (such as the "inline" keyword in
1130 C/C++). It is just a hint; it imposes no requirements on the
1133 This attribute indicates that the function should be added to a
1134 jump-instruction table at code-generation time, and that all address-taken
1135 references to this function should be replaced with a reference to the
1136 appropriate jump-instruction-table function pointer. Note that this creates
1137 a new pointer for the original function, which means that code that depends
1138 on function-pointer identity can break. So, any function annotated with
1139 ``jumptable`` must also be ``unnamed_addr``.
1141 This attribute suggests that optimization passes and code generator
1142 passes make choices that keep the code size of this function as small
1143 as possible and perform optimizations that may sacrifice runtime
1144 performance in order to minimize the size of the generated code.
1146 This attribute disables prologue / epilogue emission for the
1147 function. This can have very system-specific consequences.
1149 This indicates that the callee function at a call site is not recognized as
1150 a built-in function. LLVM will retain the original call and not replace it
1151 with equivalent code based on the semantics of the built-in function, unless
1152 the call site uses the ``builtin`` attribute. This is valid at call sites
1153 and on function declarations and definitions.
1155 This attribute indicates that calls to the function cannot be
1156 duplicated. A call to a ``noduplicate`` function may be moved
1157 within its parent function, but may not be duplicated within
1158 its parent function.
1160 A function containing a ``noduplicate`` call may still
1161 be an inlining candidate, provided that the call is not
1162 duplicated by inlining. That implies that the function has
1163 internal linkage and only has one call site, so the original
1164 call is dead after inlining.
1166 This attributes disables implicit floating point instructions.
1168 This attribute indicates that the inliner should never inline this
1169 function in any situation. This attribute may not be used together
1170 with the ``alwaysinline`` attribute.
1172 This attribute suppresses lazy symbol binding for the function. This
1173 may make calls to the function faster, at the cost of extra program
1174 startup time if the function is not called during program startup.
1176 This attribute indicates that the code generator should not use a
1177 red zone, even if the target-specific ABI normally permits it.
1179 This function attribute indicates that the function never returns
1180 normally. This produces undefined behavior at runtime if the
1181 function ever does dynamically return.
1183 This function attribute indicates that the function never returns
1184 with an unwind or exceptional control flow. If the function does
1185 unwind, its runtime behavior is undefined.
1187 This function attribute indicates that the function is not optimized
1188 by any optimization or code generator passes with the
1189 exception of interprocedural optimization passes.
1190 This attribute cannot be used together with the ``alwaysinline``
1191 attribute; this attribute is also incompatible
1192 with the ``minsize`` attribute and the ``optsize`` attribute.
1194 This attribute requires the ``noinline`` attribute to be specified on
1195 the function as well, so the function is never inlined into any caller.
1196 Only functions with the ``alwaysinline`` attribute are valid
1197 candidates for inlining into the body of this function.
1199 This attribute suggests that optimization passes and code generator
1200 passes make choices that keep the code size of this function low,
1201 and otherwise do optimizations specifically to reduce code size as
1202 long as they do not significantly impact runtime performance.
1204 On a function, this attribute indicates that the function computes its
1205 result (or decides to unwind an exception) based strictly on its arguments,
1206 without dereferencing any pointer arguments or otherwise accessing
1207 any mutable state (e.g. memory, control registers, etc) visible to
1208 caller functions. It does not write through any pointer arguments
1209 (including ``byval`` arguments) and never changes any state visible
1210 to callers. This means that it cannot unwind exceptions by calling
1211 the ``C++`` exception throwing methods.
1213 On an argument, this attribute indicates that the function does not
1214 dereference that pointer argument, even though it may read or write the
1215 memory that the pointer points to if accessed through other pointers.
1217 On a function, this attribute indicates that the function does not write
1218 through any pointer arguments (including ``byval`` arguments) or otherwise
1219 modify any state (e.g. memory, control registers, etc) visible to
1220 caller functions. It may dereference pointer arguments and read
1221 state that may be set in the caller. A readonly function always
1222 returns the same value (or unwinds an exception identically) when
1223 called with the same set of arguments and global state. It cannot
1224 unwind an exception by calling the ``C++`` exception throwing
1227 On an argument, this attribute indicates that the function does not write
1228 through this pointer argument, even though it may write to the memory that
1229 the pointer points to.
1231 This attribute indicates that this function can return twice. The C
1232 ``setjmp`` is an example of such a function. The compiler disables
1233 some optimizations (like tail calls) in the caller of these
1235 ``sanitize_address``
1236 This attribute indicates that AddressSanitizer checks
1237 (dynamic address safety analysis) are enabled for this function.
1239 This attribute indicates that MemorySanitizer checks (dynamic detection
1240 of accesses to uninitialized memory) are enabled for this function.
1242 This attribute indicates that ThreadSanitizer checks
1243 (dynamic thread safety analysis) are enabled for this function.
1245 This attribute indicates that the function should emit a stack
1246 smashing protector. It is in the form of a "canary" --- a random value
1247 placed on the stack before the local variables that's checked upon
1248 return from the function to see if it has been overwritten. A
1249 heuristic is used to determine if a function needs stack protectors
1250 or not. The heuristic used will enable protectors for functions with:
1252 - Character arrays larger than ``ssp-buffer-size`` (default 8).
1253 - Aggregates containing character arrays larger than ``ssp-buffer-size``.
1254 - Calls to alloca() with variable sizes or constant sizes greater than
1255 ``ssp-buffer-size``.
1257 Variables that are identified as requiring a protector will be arranged
1258 on the stack such that they are adjacent to the stack protector guard.
1260 If a function that has an ``ssp`` attribute is inlined into a
1261 function that doesn't have an ``ssp`` attribute, then the resulting
1262 function will have an ``ssp`` attribute.
1264 This attribute indicates that the function should *always* emit a
1265 stack smashing protector. This overrides the ``ssp`` function
1268 Variables that are identified as requiring a protector will be arranged
1269 on the stack such that they are adjacent to the stack protector guard.
1270 The specific layout rules are:
1272 #. Large arrays and structures containing large arrays
1273 (``>= ssp-buffer-size``) are closest to the stack protector.
1274 #. Small arrays and structures containing small arrays
1275 (``< ssp-buffer-size``) are 2nd closest to the protector.
1276 #. Variables that have had their address taken are 3rd closest to the
1279 If a function that has an ``sspreq`` attribute is inlined into a
1280 function that doesn't have an ``sspreq`` attribute or which has an
1281 ``ssp`` or ``sspstrong`` attribute, then the resulting function will have
1282 an ``sspreq`` attribute.
1284 This attribute indicates that the function should emit a stack smashing
1285 protector. This attribute causes a strong heuristic to be used when
1286 determining if a function needs stack protectors. The strong heuristic
1287 will enable protectors for functions with:
1289 - Arrays of any size and type
1290 - Aggregates containing an array of any size and type.
1291 - Calls to alloca().
1292 - Local variables that have had their address taken.
1294 Variables that are identified as requiring a protector will be arranged
1295 on the stack such that they are adjacent to the stack protector guard.
1296 The specific layout rules are:
1298 #. Large arrays and structures containing large arrays
1299 (``>= ssp-buffer-size``) are closest to the stack protector.
1300 #. Small arrays and structures containing small arrays
1301 (``< ssp-buffer-size``) are 2nd closest to the protector.
1302 #. Variables that have had their address taken are 3rd closest to the
1305 This overrides the ``ssp`` function attribute.
1307 If a function that has an ``sspstrong`` attribute is inlined into a
1308 function that doesn't have an ``sspstrong`` attribute, then the
1309 resulting function will have an ``sspstrong`` attribute.
1311 This attribute indicates that the ABI being targeted requires that
1312 an unwind table entry be produce for this function even if we can
1313 show that no exceptions passes by it. This is normally the case for
1314 the ELF x86-64 abi, but it can be disabled for some compilation
1319 Module-Level Inline Assembly
1320 ----------------------------
1322 Modules may contain "module-level inline asm" blocks, which corresponds
1323 to the GCC "file scope inline asm" blocks. These blocks are internally
1324 concatenated by LLVM and treated as a single unit, but may be separated
1325 in the ``.ll`` file if desired. The syntax is very simple:
1327 .. code-block:: llvm
1329 module asm "inline asm code goes here"
1330 module asm "more can go here"
1332 The strings can contain any character by escaping non-printable
1333 characters. The escape sequence used is simply "\\xx" where "xx" is the
1334 two digit hex code for the number.
1336 The inline asm code is simply printed to the machine code .s file when
1337 assembly code is generated.
1339 .. _langref_datalayout:
1344 A module may specify a target specific data layout string that specifies
1345 how data is to be laid out in memory. The syntax for the data layout is
1348 .. code-block:: llvm
1350 target datalayout = "layout specification"
1352 The *layout specification* consists of a list of specifications
1353 separated by the minus sign character ('-'). Each specification starts
1354 with a letter and may include other information after the letter to
1355 define some aspect of the data layout. The specifications accepted are
1359 Specifies that the target lays out data in big-endian form. That is,
1360 the bits with the most significance have the lowest address
1363 Specifies that the target lays out data in little-endian form. That
1364 is, the bits with the least significance have the lowest address
1367 Specifies the natural alignment of the stack in bits. Alignment
1368 promotion of stack variables is limited to the natural stack
1369 alignment to avoid dynamic stack realignment. The stack alignment
1370 must be a multiple of 8-bits. If omitted, the natural stack
1371 alignment defaults to "unspecified", which does not prevent any
1372 alignment promotions.
1373 ``p[n]:<size>:<abi>:<pref>``
1374 This specifies the *size* of a pointer and its ``<abi>`` and
1375 ``<pref>``\erred alignments for address space ``n``. All sizes are in
1376 bits. The address space, ``n`` is optional, and if not specified,
1377 denotes the default address space 0. The value of ``n`` must be
1378 in the range [1,2^23).
1379 ``i<size>:<abi>:<pref>``
1380 This specifies the alignment for an integer type of a given bit
1381 ``<size>``. The value of ``<size>`` must be in the range [1,2^23).
1382 ``v<size>:<abi>:<pref>``
1383 This specifies the alignment for a vector type of a given bit
1385 ``f<size>:<abi>:<pref>``
1386 This specifies the alignment for a floating point type of a given bit
1387 ``<size>``. Only values of ``<size>`` that are supported by the target
1388 will work. 32 (float) and 64 (double) are supported on all targets; 80
1389 or 128 (different flavors of long double) are also supported on some
1392 This specifies the alignment for an object of aggregate type.
1394 If present, specifies that llvm names are mangled in the output. The
1397 * ``e``: ELF mangling: Private symbols get a ``.L`` prefix.
1398 * ``m``: Mips mangling: Private symbols get a ``$`` prefix.
1399 * ``o``: Mach-O mangling: Private symbols get ``L`` prefix. Other
1400 symbols get a ``_`` prefix.
1401 * ``w``: Windows COFF prefix: Similar to Mach-O, but stdcall and fastcall
1402 functions also get a suffix based on the frame size.
1403 ``n<size1>:<size2>:<size3>...``
1404 This specifies a set of native integer widths for the target CPU in
1405 bits. For example, it might contain ``n32`` for 32-bit PowerPC,
1406 ``n32:64`` for PowerPC 64, or ``n8:16:32:64`` for X86-64. Elements of
1407 this set are considered to support most general arithmetic operations
1410 On every specification that takes a ``<abi>:<pref>``, specifying the
1411 ``<pref>`` alignment is optional. If omitted, the preceding ``:``
1412 should be omitted too and ``<pref>`` will be equal to ``<abi>``.
1414 When constructing the data layout for a given target, LLVM starts with a
1415 default set of specifications which are then (possibly) overridden by
1416 the specifications in the ``datalayout`` keyword. The default
1417 specifications are given in this list:
1419 - ``E`` - big endian
1420 - ``p:64:64:64`` - 64-bit pointers with 64-bit alignment.
1421 - ``p[n]:64:64:64`` - Other address spaces are assumed to be the
1422 same as the default address space.
1423 - ``S0`` - natural stack alignment is unspecified
1424 - ``i1:8:8`` - i1 is 8-bit (byte) aligned
1425 - ``i8:8:8`` - i8 is 8-bit (byte) aligned
1426 - ``i16:16:16`` - i16 is 16-bit aligned
1427 - ``i32:32:32`` - i32 is 32-bit aligned
1428 - ``i64:32:64`` - i64 has ABI alignment of 32-bits but preferred
1429 alignment of 64-bits
1430 - ``f16:16:16`` - half is 16-bit aligned
1431 - ``f32:32:32`` - float is 32-bit aligned
1432 - ``f64:64:64`` - double is 64-bit aligned
1433 - ``f128:128:128`` - quad is 128-bit aligned
1434 - ``v64:64:64`` - 64-bit vector is 64-bit aligned
1435 - ``v128:128:128`` - 128-bit vector is 128-bit aligned
1436 - ``a:0:64`` - aggregates are 64-bit aligned
1438 When LLVM is determining the alignment for a given type, it uses the
1441 #. If the type sought is an exact match for one of the specifications,
1442 that specification is used.
1443 #. If no match is found, and the type sought is an integer type, then
1444 the smallest integer type that is larger than the bitwidth of the
1445 sought type is used. If none of the specifications are larger than
1446 the bitwidth then the largest integer type is used. For example,
1447 given the default specifications above, the i7 type will use the
1448 alignment of i8 (next largest) while both i65 and i256 will use the
1449 alignment of i64 (largest specified).
1450 #. If no match is found, and the type sought is a vector type, then the
1451 largest vector type that is smaller than the sought vector type will
1452 be used as a fall back. This happens because <128 x double> can be
1453 implemented in terms of 64 <2 x double>, for example.
1455 The function of the data layout string may not be what you expect.
1456 Notably, this is not a specification from the frontend of what alignment
1457 the code generator should use.
1459 Instead, if specified, the target data layout is required to match what
1460 the ultimate *code generator* expects. This string is used by the
1461 mid-level optimizers to improve code, and this only works if it matches
1462 what the ultimate code generator uses. If you would like to generate IR
1463 that does not embed this target-specific detail into the IR, then you
1464 don't have to specify the string. This will disable some optimizations
1465 that require precise layout information, but this also prevents those
1466 optimizations from introducing target specificity into the IR.
1473 A module may specify a target triple string that describes the target
1474 host. The syntax for the target triple is simply:
1476 .. code-block:: llvm
1478 target triple = "x86_64-apple-macosx10.7.0"
1480 The *target triple* string consists of a series of identifiers delimited
1481 by the minus sign character ('-'). The canonical forms are:
1485 ARCHITECTURE-VENDOR-OPERATING_SYSTEM
1486 ARCHITECTURE-VENDOR-OPERATING_SYSTEM-ENVIRONMENT
1488 This information is passed along to the backend so that it generates
1489 code for the proper architecture. It's possible to override this on the
1490 command line with the ``-mtriple`` command line option.
1492 .. _pointeraliasing:
1494 Pointer Aliasing Rules
1495 ----------------------
1497 Any memory access must be done through a pointer value associated with
1498 an address range of the memory access, otherwise the behavior is
1499 undefined. Pointer values are associated with address ranges according
1500 to the following rules:
1502 - A pointer value is associated with the addresses associated with any
1503 value it is *based* on.
1504 - An address of a global variable is associated with the address range
1505 of the variable's storage.
1506 - The result value of an allocation instruction is associated with the
1507 address range of the allocated storage.
1508 - A null pointer in the default address-space is associated with no
1510 - An integer constant other than zero or a pointer value returned from
1511 a function not defined within LLVM may be associated with address
1512 ranges allocated through mechanisms other than those provided by
1513 LLVM. Such ranges shall not overlap with any ranges of addresses
1514 allocated by mechanisms provided by LLVM.
1516 A pointer value is *based* on another pointer value according to the
1519 - A pointer value formed from a ``getelementptr`` operation is *based*
1520 on the first operand of the ``getelementptr``.
1521 - The result value of a ``bitcast`` is *based* on the operand of the
1523 - A pointer value formed by an ``inttoptr`` is *based* on all pointer
1524 values that contribute (directly or indirectly) to the computation of
1525 the pointer's value.
1526 - The "*based* on" relationship is transitive.
1528 Note that this definition of *"based"* is intentionally similar to the
1529 definition of *"based"* in C99, though it is slightly weaker.
1531 LLVM IR does not associate types with memory. The result type of a
1532 ``load`` merely indicates the size and alignment of the memory from
1533 which to load, as well as the interpretation of the value. The first
1534 operand type of a ``store`` similarly only indicates the size and
1535 alignment of the store.
1537 Consequently, type-based alias analysis, aka TBAA, aka
1538 ``-fstrict-aliasing``, is not applicable to general unadorned LLVM IR.
1539 :ref:`Metadata <metadata>` may be used to encode additional information
1540 which specialized optimization passes may use to implement type-based
1545 Volatile Memory Accesses
1546 ------------------------
1548 Certain memory accesses, such as :ref:`load <i_load>`'s,
1549 :ref:`store <i_store>`'s, and :ref:`llvm.memcpy <int_memcpy>`'s may be
1550 marked ``volatile``. The optimizers must not change the number of
1551 volatile operations or change their order of execution relative to other
1552 volatile operations. The optimizers *may* change the order of volatile
1553 operations relative to non-volatile operations. This is not Java's
1554 "volatile" and has no cross-thread synchronization behavior.
1556 IR-level volatile loads and stores cannot safely be optimized into
1557 llvm.memcpy or llvm.memmove intrinsics even when those intrinsics are
1558 flagged volatile. Likewise, the backend should never split or merge
1559 target-legal volatile load/store instructions.
1561 .. admonition:: Rationale
1563 Platforms may rely on volatile loads and stores of natively supported
1564 data width to be executed as single instruction. For example, in C
1565 this holds for an l-value of volatile primitive type with native
1566 hardware support, but not necessarily for aggregate types. The
1567 frontend upholds these expectations, which are intentionally
1568 unspecified in the IR. The rules above ensure that IR transformation
1569 do not violate the frontend's contract with the language.
1573 Memory Model for Concurrent Operations
1574 --------------------------------------
1576 The LLVM IR does not define any way to start parallel threads of
1577 execution or to register signal handlers. Nonetheless, there are
1578 platform-specific ways to create them, and we define LLVM IR's behavior
1579 in their presence. This model is inspired by the C++0x memory model.
1581 For a more informal introduction to this model, see the :doc:`Atomics`.
1583 We define a *happens-before* partial order as the least partial order
1586 - Is a superset of single-thread program order, and
1587 - When a *synchronizes-with* ``b``, includes an edge from ``a`` to
1588 ``b``. *Synchronizes-with* pairs are introduced by platform-specific
1589 techniques, like pthread locks, thread creation, thread joining,
1590 etc., and by atomic instructions. (See also :ref:`Atomic Memory Ordering
1591 Constraints <ordering>`).
1593 Note that program order does not introduce *happens-before* edges
1594 between a thread and signals executing inside that thread.
1596 Every (defined) read operation (load instructions, memcpy, atomic
1597 loads/read-modify-writes, etc.) R reads a series of bytes written by
1598 (defined) write operations (store instructions, atomic
1599 stores/read-modify-writes, memcpy, etc.). For the purposes of this
1600 section, initialized globals are considered to have a write of the
1601 initializer which is atomic and happens before any other read or write
1602 of the memory in question. For each byte of a read R, R\ :sub:`byte`
1603 may see any write to the same byte, except:
1605 - If write\ :sub:`1` happens before write\ :sub:`2`, and
1606 write\ :sub:`2` happens before R\ :sub:`byte`, then
1607 R\ :sub:`byte` does not see write\ :sub:`1`.
1608 - If R\ :sub:`byte` happens before write\ :sub:`3`, then
1609 R\ :sub:`byte` does not see write\ :sub:`3`.
1611 Given that definition, R\ :sub:`byte` is defined as follows:
1613 - If R is volatile, the result is target-dependent. (Volatile is
1614 supposed to give guarantees which can support ``sig_atomic_t`` in
1615 C/C++, and may be used for accesses to addresses that do not behave
1616 like normal memory. It does not generally provide cross-thread
1618 - Otherwise, if there is no write to the same byte that happens before
1619 R\ :sub:`byte`, R\ :sub:`byte` returns ``undef`` for that byte.
1620 - Otherwise, if R\ :sub:`byte` may see exactly one write,
1621 R\ :sub:`byte` returns the value written by that write.
1622 - Otherwise, if R is atomic, and all the writes R\ :sub:`byte` may
1623 see are atomic, it chooses one of the values written. See the :ref:`Atomic
1624 Memory Ordering Constraints <ordering>` section for additional
1625 constraints on how the choice is made.
1626 - Otherwise R\ :sub:`byte` returns ``undef``.
1628 R returns the value composed of the series of bytes it read. This
1629 implies that some bytes within the value may be ``undef`` **without**
1630 the entire value being ``undef``. Note that this only defines the
1631 semantics of the operation; it doesn't mean that targets will emit more
1632 than one instruction to read the series of bytes.
1634 Note that in cases where none of the atomic intrinsics are used, this
1635 model places only one restriction on IR transformations on top of what
1636 is required for single-threaded execution: introducing a store to a byte
1637 which might not otherwise be stored is not allowed in general.
1638 (Specifically, in the case where another thread might write to and read
1639 from an address, introducing a store can change a load that may see
1640 exactly one write into a load that may see multiple writes.)
1644 Atomic Memory Ordering Constraints
1645 ----------------------------------
1647 Atomic instructions (:ref:`cmpxchg <i_cmpxchg>`,
1648 :ref:`atomicrmw <i_atomicrmw>`, :ref:`fence <i_fence>`,
1649 :ref:`atomic load <i_load>`, and :ref:`atomic store <i_store>`) take
1650 ordering parameters that determine which other atomic instructions on
1651 the same address they *synchronize with*. These semantics are borrowed
1652 from Java and C++0x, but are somewhat more colloquial. If these
1653 descriptions aren't precise enough, check those specs (see spec
1654 references in the :doc:`atomics guide <Atomics>`).
1655 :ref:`fence <i_fence>` instructions treat these orderings somewhat
1656 differently since they don't take an address. See that instruction's
1657 documentation for details.
1659 For a simpler introduction to the ordering constraints, see the
1663 The set of values that can be read is governed by the happens-before
1664 partial order. A value cannot be read unless some operation wrote
1665 it. This is intended to provide a guarantee strong enough to model
1666 Java's non-volatile shared variables. This ordering cannot be
1667 specified for read-modify-write operations; it is not strong enough
1668 to make them atomic in any interesting way.
1670 In addition to the guarantees of ``unordered``, there is a single
1671 total order for modifications by ``monotonic`` operations on each
1672 address. All modification orders must be compatible with the
1673 happens-before order. There is no guarantee that the modification
1674 orders can be combined to a global total order for the whole program
1675 (and this often will not be possible). The read in an atomic
1676 read-modify-write operation (:ref:`cmpxchg <i_cmpxchg>` and
1677 :ref:`atomicrmw <i_atomicrmw>`) reads the value in the modification
1678 order immediately before the value it writes. If one atomic read
1679 happens before another atomic read of the same address, the later
1680 read must see the same value or a later value in the address's
1681 modification order. This disallows reordering of ``monotonic`` (or
1682 stronger) operations on the same address. If an address is written
1683 ``monotonic``-ally by one thread, and other threads ``monotonic``-ally
1684 read that address repeatedly, the other threads must eventually see
1685 the write. This corresponds to the C++0x/C1x
1686 ``memory_order_relaxed``.
1688 In addition to the guarantees of ``monotonic``, a
1689 *synchronizes-with* edge may be formed with a ``release`` operation.
1690 This is intended to model C++'s ``memory_order_acquire``.
1692 In addition to the guarantees of ``monotonic``, if this operation
1693 writes a value which is subsequently read by an ``acquire``
1694 operation, it *synchronizes-with* that operation. (This isn't a
1695 complete description; see the C++0x definition of a release
1696 sequence.) This corresponds to the C++0x/C1x
1697 ``memory_order_release``.
1698 ``acq_rel`` (acquire+release)
1699 Acts as both an ``acquire`` and ``release`` operation on its
1700 address. This corresponds to the C++0x/C1x ``memory_order_acq_rel``.
1701 ``seq_cst`` (sequentially consistent)
1702 In addition to the guarantees of ``acq_rel`` (``acquire`` for an
1703 operation that only reads, ``release`` for an operation that only
1704 writes), there is a global total order on all
1705 sequentially-consistent operations on all addresses, which is
1706 consistent with the *happens-before* partial order and with the
1707 modification orders of all the affected addresses. Each
1708 sequentially-consistent read sees the last preceding write to the
1709 same address in this global order. This corresponds to the C++0x/C1x
1710 ``memory_order_seq_cst`` and Java volatile.
1714 If an atomic operation is marked ``singlethread``, it only *synchronizes
1715 with* or participates in modification and seq\_cst total orderings with
1716 other operations running in the same thread (for example, in signal
1724 LLVM IR floating-point binary ops (:ref:`fadd <i_fadd>`,
1725 :ref:`fsub <i_fsub>`, :ref:`fmul <i_fmul>`, :ref:`fdiv <i_fdiv>`,
1726 :ref:`frem <i_frem>`) have the following flags that can set to enable
1727 otherwise unsafe floating point operations
1730 No NaNs - Allow optimizations to assume the arguments and result are not
1731 NaN. Such optimizations are required to retain defined behavior over
1732 NaNs, but the value of the result is undefined.
1735 No Infs - Allow optimizations to assume the arguments and result are not
1736 +/-Inf. Such optimizations are required to retain defined behavior over
1737 +/-Inf, but the value of the result is undefined.
1740 No Signed Zeros - Allow optimizations to treat the sign of a zero
1741 argument or result as insignificant.
1744 Allow Reciprocal - Allow optimizations to use the reciprocal of an
1745 argument rather than perform division.
1748 Fast - Allow algebraically equivalent transformations that may
1749 dramatically change results in floating point (e.g. reassociate). This
1750 flag implies all the others.
1754 Use-list Order Directives
1755 -------------------------
1757 Use-list directives encode the in-memory order of each use-list, allowing the
1758 order to be recreated. ``<order-indexes>`` is a comma-separated list of
1759 indexes that are assigned to the referenced value's uses. The referenced
1760 value's use-list is immediately sorted by these indexes.
1762 Use-list directives may appear at function scope or global scope. They are not
1763 instructions, and have no effect on the semantics of the IR. When they're at
1764 function scope, they must appear after the terminator of the final basic block.
1766 If basic blocks have their address taken via ``blockaddress()`` expressions,
1767 ``uselistorder_bb`` can be used to reorder their use-lists from outside their
1774 uselistorder <ty> <value>, { <order-indexes> }
1775 uselistorder_bb @function, %block { <order-indexes> }
1781 ; At function scope.
1782 uselistorder i32 %arg1, { 1, 0, 2 }
1783 uselistorder label %bb, { 1, 0 }
1786 uselistorder i32* @global, { 1, 2, 0 }
1787 uselistorder i32 7, { 1, 0 }
1788 uselistorder i32 (i32) @bar, { 1, 0 }
1789 uselistorder_bb @foo, %bb, { 5, 1, 3, 2, 0, 4 }
1796 The LLVM type system is one of the most important features of the
1797 intermediate representation. Being typed enables a number of
1798 optimizations to be performed on the intermediate representation
1799 directly, without having to do extra analyses on the side before the
1800 transformation. A strong type system makes it easier to read the
1801 generated code and enables novel analyses and transformations that are
1802 not feasible to perform on normal three address code representations.
1812 The void type does not represent any value and has no size.
1830 The function type can be thought of as a function signature. It consists of a
1831 return type and a list of formal parameter types. The return type of a function
1832 type is a void type or first class type --- except for :ref:`label <t_label>`
1833 and :ref:`metadata <t_metadata>` types.
1839 <returntype> (<parameter list>)
1841 ...where '``<parameter list>``' is a comma-separated list of type
1842 specifiers. Optionally, the parameter list may include a type ``...``, which
1843 indicates that the function takes a variable number of arguments. Variable
1844 argument functions can access their arguments with the :ref:`variable argument
1845 handling intrinsic <int_varargs>` functions. '``<returntype>``' is any type
1846 except :ref:`label <t_label>` and :ref:`metadata <t_metadata>`.
1850 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1851 | ``i32 (i32)`` | function taking an ``i32``, returning an ``i32`` |
1852 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1853 | ``float (i16, i32 *) *`` | :ref:`Pointer <t_pointer>` to a function that takes an ``i16`` and a :ref:`pointer <t_pointer>` to ``i32``, returning ``float``. |
1854 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1855 | ``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. |
1856 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1857 | ``{i32, i32} (i32)`` | A function taking an ``i32``, returning a :ref:`structure <t_struct>` containing two ``i32`` values |
1858 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1865 The :ref:`first class <t_firstclass>` types are perhaps the most important.
1866 Values of these types are the only ones which can be produced by
1874 These are the types that are valid in registers from CodeGen's perspective.
1883 The integer type is a very simple type that simply specifies an
1884 arbitrary bit width for the integer type desired. Any bit width from 1
1885 bit to 2\ :sup:`23`\ -1 (about 8 million) can be specified.
1893 The number of bits the integer will occupy is specified by the ``N``
1899 +----------------+------------------------------------------------+
1900 | ``i1`` | a single-bit integer. |
1901 +----------------+------------------------------------------------+
1902 | ``i32`` | a 32-bit integer. |
1903 +----------------+------------------------------------------------+
1904 | ``i1942652`` | a really big integer of over 1 million bits. |
1905 +----------------+------------------------------------------------+
1909 Floating Point Types
1910 """"""""""""""""""""
1919 - 16-bit floating point value
1922 - 32-bit floating point value
1925 - 64-bit floating point value
1928 - 128-bit floating point value (112-bit mantissa)
1931 - 80-bit floating point value (X87)
1934 - 128-bit floating point value (two 64-bits)
1941 The x86_mmx type represents a value held in an MMX register on an x86
1942 machine. The operations allowed on it are quite limited: parameters and
1943 return values, load and store, and bitcast. User-specified MMX
1944 instructions are represented as intrinsic or asm calls with arguments
1945 and/or results of this type. There are no arrays, vectors or constants
1962 The pointer type is used to specify memory locations. Pointers are
1963 commonly used to reference objects in memory.
1965 Pointer types may have an optional address space attribute defining the
1966 numbered address space where the pointed-to object resides. The default
1967 address space is number zero. The semantics of non-zero address spaces
1968 are target-specific.
1970 Note that LLVM does not permit pointers to void (``void*``) nor does it
1971 permit pointers to labels (``label*``). Use ``i8*`` instead.
1981 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1982 | ``[4 x i32]*`` | A :ref:`pointer <t_pointer>` to :ref:`array <t_array>` of four ``i32`` values. |
1983 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1984 | ``i32 (i32*) *`` | A :ref:`pointer <t_pointer>` to a :ref:`function <t_function>` that takes an ``i32*``, returning an ``i32``. |
1985 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1986 | ``i32 addrspace(5)*`` | A :ref:`pointer <t_pointer>` to an ``i32`` value that resides in address space #5. |
1987 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1996 A vector type is a simple derived type that represents a vector of
1997 elements. Vector types are used when multiple primitive data are
1998 operated in parallel using a single instruction (SIMD). A vector type
1999 requires a size (number of elements) and an underlying primitive data
2000 type. Vector types are considered :ref:`first class <t_firstclass>`.
2006 < <# elements> x <elementtype> >
2008 The number of elements is a constant integer value larger than 0;
2009 elementtype may be any integer, floating point or pointer type. Vectors
2010 of size zero are not allowed.
2014 +-------------------+--------------------------------------------------+
2015 | ``<4 x i32>`` | Vector of 4 32-bit integer values. |
2016 +-------------------+--------------------------------------------------+
2017 | ``<8 x float>`` | Vector of 8 32-bit floating-point values. |
2018 +-------------------+--------------------------------------------------+
2019 | ``<2 x i64>`` | Vector of 2 64-bit integer values. |
2020 +-------------------+--------------------------------------------------+
2021 | ``<4 x i64*>`` | Vector of 4 pointers to 64-bit integer values. |
2022 +-------------------+--------------------------------------------------+
2031 The label type represents code labels.
2046 The metadata type represents embedded metadata. No derived types may be
2047 created from metadata except for :ref:`function <t_function>` arguments.
2060 Aggregate Types are a subset of derived types that can contain multiple
2061 member types. :ref:`Arrays <t_array>` and :ref:`structs <t_struct>` are
2062 aggregate types. :ref:`Vectors <t_vector>` are not considered to be
2072 The array type is a very simple derived type that arranges elements
2073 sequentially in memory. The array type requires a size (number of
2074 elements) and an underlying data type.
2080 [<# elements> x <elementtype>]
2082 The number of elements is a constant integer value; ``elementtype`` may
2083 be any type with a size.
2087 +------------------+--------------------------------------+
2088 | ``[40 x i32]`` | Array of 40 32-bit integer values. |
2089 +------------------+--------------------------------------+
2090 | ``[41 x i32]`` | Array of 41 32-bit integer values. |
2091 +------------------+--------------------------------------+
2092 | ``[4 x i8]`` | Array of 4 8-bit integer values. |
2093 +------------------+--------------------------------------+
2095 Here are some examples of multidimensional arrays:
2097 +-----------------------------+----------------------------------------------------------+
2098 | ``[3 x [4 x i32]]`` | 3x4 array of 32-bit integer values. |
2099 +-----------------------------+----------------------------------------------------------+
2100 | ``[12 x [10 x float]]`` | 12x10 array of single precision floating point values. |
2101 +-----------------------------+----------------------------------------------------------+
2102 | ``[2 x [3 x [4 x i16]]]`` | 2x3x4 array of 16-bit integer values. |
2103 +-----------------------------+----------------------------------------------------------+
2105 There is no restriction on indexing beyond the end of the array implied
2106 by a static type (though there are restrictions on indexing beyond the
2107 bounds of an allocated object in some cases). This means that
2108 single-dimension 'variable sized array' addressing can be implemented in
2109 LLVM with a zero length array type. An implementation of 'pascal style
2110 arrays' in LLVM could use the type "``{ i32, [0 x float]}``", for
2120 The structure type is used to represent a collection of data members
2121 together in memory. The elements of a structure may be any type that has
2124 Structures in memory are accessed using '``load``' and '``store``' by
2125 getting a pointer to a field with the '``getelementptr``' instruction.
2126 Structures in registers are accessed using the '``extractvalue``' and
2127 '``insertvalue``' instructions.
2129 Structures may optionally be "packed" structures, which indicate that
2130 the alignment of the struct is one byte, and that there is no padding
2131 between the elements. In non-packed structs, padding between field types
2132 is inserted as defined by the DataLayout string in the module, which is
2133 required to match what the underlying code generator expects.
2135 Structures can either be "literal" or "identified". A literal structure
2136 is defined inline with other types (e.g. ``{i32, i32}*``) whereas
2137 identified types are always defined at the top level with a name.
2138 Literal types are uniqued by their contents and can never be recursive
2139 or opaque since there is no way to write one. Identified types can be
2140 recursive, can be opaqued, and are never uniqued.
2146 %T1 = type { <type list> } ; Identified normal struct type
2147 %T2 = type <{ <type list> }> ; Identified packed struct type
2151 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2152 | ``{ i32, i32, i32 }`` | A triple of three ``i32`` values |
2153 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2154 | ``{ 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``. |
2155 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2156 | ``<{ i8, i32 }>`` | A packed struct known to be 5 bytes in size. |
2157 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2161 Opaque Structure Types
2162 """"""""""""""""""""""
2166 Opaque structure types are used to represent named structure types that
2167 do not have a body specified. This corresponds (for example) to the C
2168 notion of a forward declared structure.
2179 +--------------+-------------------+
2180 | ``opaque`` | An opaque type. |
2181 +--------------+-------------------+
2188 LLVM has several different basic types of constants. This section
2189 describes them all and their syntax.
2194 **Boolean constants**
2195 The two strings '``true``' and '``false``' are both valid constants
2197 **Integer constants**
2198 Standard integers (such as '4') are constants of the
2199 :ref:`integer <t_integer>` type. Negative numbers may be used with
2201 **Floating point constants**
2202 Floating point constants use standard decimal notation (e.g.
2203 123.421), exponential notation (e.g. 1.23421e+2), or a more precise
2204 hexadecimal notation (see below). The assembler requires the exact
2205 decimal value of a floating-point constant. For example, the
2206 assembler accepts 1.25 but rejects 1.3 because 1.3 is a repeating
2207 decimal in binary. Floating point constants must have a :ref:`floating
2208 point <t_floating>` type.
2209 **Null pointer constants**
2210 The identifier '``null``' is recognized as a null pointer constant
2211 and must be of :ref:`pointer type <t_pointer>`.
2213 The one non-intuitive notation for constants is the hexadecimal form of
2214 floating point constants. For example, the form
2215 '``double 0x432ff973cafa8000``' is equivalent to (but harder to read
2216 than) '``double 4.5e+15``'. The only time hexadecimal floating point
2217 constants are required (and the only time that they are generated by the
2218 disassembler) is when a floating point constant must be emitted but it
2219 cannot be represented as a decimal floating point number in a reasonable
2220 number of digits. For example, NaN's, infinities, and other special
2221 values are represented in their IEEE hexadecimal format so that assembly
2222 and disassembly do not cause any bits to change in the constants.
2224 When using the hexadecimal form, constants of types half, float, and
2225 double are represented using the 16-digit form shown above (which
2226 matches the IEEE754 representation for double); half and float values
2227 must, however, be exactly representable as IEEE 754 half and single
2228 precision, respectively. Hexadecimal format is always used for long
2229 double, and there are three forms of long double. The 80-bit format used
2230 by x86 is represented as ``0xK`` followed by 20 hexadecimal digits. The
2231 128-bit format used by PowerPC (two adjacent doubles) is represented by
2232 ``0xM`` followed by 32 hexadecimal digits. The IEEE 128-bit format is
2233 represented by ``0xL`` followed by 32 hexadecimal digits. Long doubles
2234 will only work if they match the long double format on your target.
2235 The IEEE 16-bit format (half precision) is represented by ``0xH``
2236 followed by 4 hexadecimal digits. All hexadecimal formats are big-endian
2237 (sign bit at the left).
2239 There are no constants of type x86_mmx.
2241 .. _complexconstants:
2246 Complex constants are a (potentially recursive) combination of simple
2247 constants and smaller complex constants.
2249 **Structure constants**
2250 Structure constants are represented with notation similar to
2251 structure type definitions (a comma separated list of elements,
2252 surrounded by braces (``{}``)). For example:
2253 "``{ i32 4, float 17.0, i32* @G }``", where "``@G``" is declared as
2254 "``@G = external global i32``". Structure constants must have
2255 :ref:`structure type <t_struct>`, and the number and types of elements
2256 must match those specified by the type.
2258 Array constants are represented with notation similar to array type
2259 definitions (a comma separated list of elements, surrounded by
2260 square brackets (``[]``)). For example:
2261 "``[ i32 42, i32 11, i32 74 ]``". Array constants must have
2262 :ref:`array type <t_array>`, and the number and types of elements must
2263 match those specified by the type.
2264 **Vector constants**
2265 Vector constants are represented with notation similar to vector
2266 type definitions (a comma separated list of elements, surrounded by
2267 less-than/greater-than's (``<>``)). For example:
2268 "``< i32 42, i32 11, i32 74, i32 100 >``". Vector constants
2269 must have :ref:`vector type <t_vector>`, and the number and types of
2270 elements must match those specified by the type.
2271 **Zero initialization**
2272 The string '``zeroinitializer``' can be used to zero initialize a
2273 value to zero of *any* type, including scalar and
2274 :ref:`aggregate <t_aggregate>` types. This is often used to avoid
2275 having to print large zero initializers (e.g. for large arrays) and
2276 is always exactly equivalent to using explicit zero initializers.
2278 A metadata node is a structure-like constant with :ref:`metadata
2279 type <t_metadata>`. For example:
2280 "``metadata !{ i32 0, metadata !"test" }``". Unlike other
2281 constants that are meant to be interpreted as part of the
2282 instruction stream, metadata is a place to attach additional
2283 information such as debug info.
2285 Global Variable and Function Addresses
2286 --------------------------------------
2288 The addresses of :ref:`global variables <globalvars>` and
2289 :ref:`functions <functionstructure>` are always implicitly valid
2290 (link-time) constants. These constants are explicitly referenced when
2291 the :ref:`identifier for the global <identifiers>` is used and always have
2292 :ref:`pointer <t_pointer>` type. For example, the following is a legal LLVM
2295 .. code-block:: llvm
2299 @Z = global [2 x i32*] [ i32* @X, i32* @Y ]
2306 The string '``undef``' can be used anywhere a constant is expected, and
2307 indicates that the user of the value may receive an unspecified
2308 bit-pattern. Undefined values may be of any type (other than '``label``'
2309 or '``void``') and be used anywhere a constant is permitted.
2311 Undefined values are useful because they indicate to the compiler that
2312 the program is well defined no matter what value is used. This gives the
2313 compiler more freedom to optimize. Here are some examples of
2314 (potentially surprising) transformations that are valid (in pseudo IR):
2316 .. code-block:: llvm
2326 This is safe because all of the output bits are affected by the undef
2327 bits. Any output bit can have a zero or one depending on the input bits.
2329 .. code-block:: llvm
2340 These logical operations have bits that are not always affected by the
2341 input. For example, if ``%X`` has a zero bit, then the output of the
2342 '``and``' operation will always be a zero for that bit, no matter what
2343 the corresponding bit from the '``undef``' is. As such, it is unsafe to
2344 optimize or assume that the result of the '``and``' is '``undef``'.
2345 However, it is safe to assume that all bits of the '``undef``' could be
2346 0, and optimize the '``and``' to 0. Likewise, it is safe to assume that
2347 all the bits of the '``undef``' operand to the '``or``' could be set,
2348 allowing the '``or``' to be folded to -1.
2350 .. code-block:: llvm
2352 %A = select undef, %X, %Y
2353 %B = select undef, 42, %Y
2354 %C = select %X, %Y, undef
2364 This set of examples shows that undefined '``select``' (and conditional
2365 branch) conditions can go *either way*, but they have to come from one
2366 of the two operands. In the ``%A`` example, if ``%X`` and ``%Y`` were
2367 both known to have a clear low bit, then ``%A`` would have to have a
2368 cleared low bit. However, in the ``%C`` example, the optimizer is
2369 allowed to assume that the '``undef``' operand could be the same as
2370 ``%Y``, allowing the whole '``select``' to be eliminated.
2372 .. code-block:: llvm
2374 %A = xor undef, undef
2391 This example points out that two '``undef``' operands are not
2392 necessarily the same. This can be surprising to people (and also matches
2393 C semantics) where they assume that "``X^X``" is always zero, even if
2394 ``X`` is undefined. This isn't true for a number of reasons, but the
2395 short answer is that an '``undef``' "variable" can arbitrarily change
2396 its value over its "live range". This is true because the variable
2397 doesn't actually *have a live range*. Instead, the value is logically
2398 read from arbitrary registers that happen to be around when needed, so
2399 the value is not necessarily consistent over time. In fact, ``%A`` and
2400 ``%C`` need to have the same semantics or the core LLVM "replace all
2401 uses with" concept would not hold.
2403 .. code-block:: llvm
2411 These examples show the crucial difference between an *undefined value*
2412 and *undefined behavior*. An undefined value (like '``undef``') is
2413 allowed to have an arbitrary bit-pattern. This means that the ``%A``
2414 operation can be constant folded to '``undef``', because the '``undef``'
2415 could be an SNaN, and ``fdiv`` is not (currently) defined on SNaN's.
2416 However, in the second example, we can make a more aggressive
2417 assumption: because the ``undef`` is allowed to be an arbitrary value,
2418 we are allowed to assume that it could be zero. Since a divide by zero
2419 has *undefined behavior*, we are allowed to assume that the operation
2420 does not execute at all. This allows us to delete the divide and all
2421 code after it. Because the undefined operation "can't happen", the
2422 optimizer can assume that it occurs in dead code.
2424 .. code-block:: llvm
2426 a: store undef -> %X
2427 b: store %X -> undef
2432 These examples reiterate the ``fdiv`` example: a store *of* an undefined
2433 value can be assumed to not have any effect; we can assume that the
2434 value is overwritten with bits that happen to match what was already
2435 there. However, a store *to* an undefined location could clobber
2436 arbitrary memory, therefore, it has undefined behavior.
2443 Poison values are similar to :ref:`undef values <undefvalues>`, however
2444 they also represent the fact that an instruction or constant expression
2445 that cannot evoke side effects has nevertheless detected a condition
2446 that results in undefined behavior.
2448 There is currently no way of representing a poison value in the IR; they
2449 only exist when produced by operations such as :ref:`add <i_add>` with
2452 Poison value behavior is defined in terms of value *dependence*:
2454 - Values other than :ref:`phi <i_phi>` nodes depend on their operands.
2455 - :ref:`Phi <i_phi>` nodes depend on the operand corresponding to
2456 their dynamic predecessor basic block.
2457 - Function arguments depend on the corresponding actual argument values
2458 in the dynamic callers of their functions.
2459 - :ref:`Call <i_call>` instructions depend on the :ref:`ret <i_ret>`
2460 instructions that dynamically transfer control back to them.
2461 - :ref:`Invoke <i_invoke>` instructions depend on the
2462 :ref:`ret <i_ret>`, :ref:`resume <i_resume>`, or exception-throwing
2463 call instructions that dynamically transfer control back to them.
2464 - Non-volatile loads and stores depend on the most recent stores to all
2465 of the referenced memory addresses, following the order in the IR
2466 (including loads and stores implied by intrinsics such as
2467 :ref:`@llvm.memcpy <int_memcpy>`.)
2468 - An instruction with externally visible side effects depends on the
2469 most recent preceding instruction with externally visible side
2470 effects, following the order in the IR. (This includes :ref:`volatile
2471 operations <volatile>`.)
2472 - An instruction *control-depends* on a :ref:`terminator
2473 instruction <terminators>` if the terminator instruction has
2474 multiple successors and the instruction is always executed when
2475 control transfers to one of the successors, and may not be executed
2476 when control is transferred to another.
2477 - Additionally, an instruction also *control-depends* on a terminator
2478 instruction if the set of instructions it otherwise depends on would
2479 be different if the terminator had transferred control to a different
2481 - Dependence is transitive.
2483 Poison values have the same behavior as :ref:`undef values <undefvalues>`,
2484 with the additional effect that any instruction that has a *dependence*
2485 on a poison value has undefined behavior.
2487 Here are some examples:
2489 .. code-block:: llvm
2492 %poison = sub nuw i32 0, 1 ; Results in a poison value.
2493 %still_poison = and i32 %poison, 0 ; 0, but also poison.
2494 %poison_yet_again = getelementptr i32* @h, i32 %still_poison
2495 store i32 0, i32* %poison_yet_again ; memory at @h[0] is poisoned
2497 store i32 %poison, i32* @g ; Poison value stored to memory.
2498 %poison2 = load i32* @g ; Poison value loaded back from memory.
2500 store volatile i32 %poison, i32* @g ; External observation; undefined behavior.
2502 %narrowaddr = bitcast i32* @g to i16*
2503 %wideaddr = bitcast i32* @g to i64*
2504 %poison3 = load i16* %narrowaddr ; Returns a poison value.
2505 %poison4 = load i64* %wideaddr ; Returns a poison value.
2507 %cmp = icmp slt i32 %poison, 0 ; Returns a poison value.
2508 br i1 %cmp, label %true, label %end ; Branch to either destination.
2511 store volatile i32 0, i32* @g ; This is control-dependent on %cmp, so
2512 ; it has undefined behavior.
2516 %p = phi i32 [ 0, %entry ], [ 1, %true ]
2517 ; Both edges into this PHI are
2518 ; control-dependent on %cmp, so this
2519 ; always results in a poison value.
2521 store volatile i32 0, i32* @g ; This would depend on the store in %true
2522 ; if %cmp is true, or the store in %entry
2523 ; otherwise, so this is undefined behavior.
2525 br i1 %cmp, label %second_true, label %second_end
2526 ; The same branch again, but this time the
2527 ; true block doesn't have side effects.
2534 store volatile i32 0, i32* @g ; This time, the instruction always depends
2535 ; on the store in %end. Also, it is
2536 ; control-equivalent to %end, so this is
2537 ; well-defined (ignoring earlier undefined
2538 ; behavior in this example).
2542 Addresses of Basic Blocks
2543 -------------------------
2545 ``blockaddress(@function, %block)``
2547 The '``blockaddress``' constant computes the address of the specified
2548 basic block in the specified function, and always has an ``i8*`` type.
2549 Taking the address of the entry block is illegal.
2551 This value only has defined behavior when used as an operand to the
2552 ':ref:`indirectbr <i_indirectbr>`' instruction, or for comparisons
2553 against null. Pointer equality tests between labels addresses results in
2554 undefined behavior --- though, again, comparison against null is ok, and
2555 no label is equal to the null pointer. This may be passed around as an
2556 opaque pointer sized value as long as the bits are not inspected. This
2557 allows ``ptrtoint`` and arithmetic to be performed on these values so
2558 long as the original value is reconstituted before the ``indirectbr``
2561 Finally, some targets may provide defined semantics when using the value
2562 as the operand to an inline assembly, but that is target specific.
2566 Constant Expressions
2567 --------------------
2569 Constant expressions are used to allow expressions involving other
2570 constants to be used as constants. Constant expressions may be of any
2571 :ref:`first class <t_firstclass>` type and may involve any LLVM operation
2572 that does not have side effects (e.g. load and call are not supported).
2573 The following is the syntax for constant expressions:
2575 ``trunc (CST to TYPE)``
2576 Truncate a constant to another type. The bit size of CST must be
2577 larger than the bit size of TYPE. Both types must be integers.
2578 ``zext (CST to TYPE)``
2579 Zero extend a constant to another type. The bit size of CST must be
2580 smaller than the bit size of TYPE. Both types must be integers.
2581 ``sext (CST to TYPE)``
2582 Sign extend a constant to another type. The bit size of CST must be
2583 smaller than the bit size of TYPE. Both types must be integers.
2584 ``fptrunc (CST to TYPE)``
2585 Truncate a floating point constant to another floating point type.
2586 The size of CST must be larger than the size of TYPE. Both types
2587 must be floating point.
2588 ``fpext (CST to TYPE)``
2589 Floating point extend a constant to another type. The size of CST
2590 must be smaller or equal to the size of TYPE. Both types must be
2592 ``fptoui (CST to TYPE)``
2593 Convert a floating point constant to the corresponding unsigned
2594 integer constant. TYPE must be a scalar or vector integer type. CST
2595 must be of scalar or vector floating point type. Both CST and TYPE
2596 must be scalars, or vectors of the same number of elements. If the
2597 value won't fit in the integer type, the results are undefined.
2598 ``fptosi (CST to TYPE)``
2599 Convert a floating point constant to the corresponding signed
2600 integer constant. TYPE must be a scalar or vector integer type. CST
2601 must be of scalar or vector floating point type. Both CST and TYPE
2602 must be scalars, or vectors of the same number of elements. If the
2603 value won't fit in the integer type, the results are undefined.
2604 ``uitofp (CST to TYPE)``
2605 Convert an unsigned integer constant to the corresponding floating
2606 point constant. TYPE must be a scalar or vector floating point type.
2607 CST must be of scalar or vector integer type. Both CST and TYPE must
2608 be scalars, or vectors of the same number of elements. If the value
2609 won't fit in the floating point type, the results are undefined.
2610 ``sitofp (CST to TYPE)``
2611 Convert a signed integer constant to the corresponding floating
2612 point constant. TYPE must be a scalar or vector floating point type.
2613 CST must be of scalar or vector integer type. Both CST and TYPE must
2614 be scalars, or vectors of the same number of elements. If the value
2615 won't fit in the floating point type, the results are undefined.
2616 ``ptrtoint (CST to TYPE)``
2617 Convert a pointer typed constant to the corresponding integer
2618 constant. ``TYPE`` must be an integer type. ``CST`` must be of
2619 pointer type. The ``CST`` value is zero extended, truncated, or
2620 unchanged to make it fit in ``TYPE``.
2621 ``inttoptr (CST to TYPE)``
2622 Convert an integer constant to a pointer constant. TYPE must be a
2623 pointer type. CST must be of integer type. The CST value is zero
2624 extended, truncated, or unchanged to make it fit in a pointer size.
2625 This one is *really* dangerous!
2626 ``bitcast (CST to TYPE)``
2627 Convert a constant, CST, to another TYPE. The constraints of the
2628 operands are the same as those for the :ref:`bitcast
2629 instruction <i_bitcast>`.
2630 ``addrspacecast (CST to TYPE)``
2631 Convert a constant pointer or constant vector of pointer, CST, to another
2632 TYPE in a different address space. The constraints of the operands are the
2633 same as those for the :ref:`addrspacecast instruction <i_addrspacecast>`.
2634 ``getelementptr (CSTPTR, IDX0, IDX1, ...)``, ``getelementptr inbounds (CSTPTR, IDX0, IDX1, ...)``
2635 Perform the :ref:`getelementptr operation <i_getelementptr>` on
2636 constants. As with the :ref:`getelementptr <i_getelementptr>`
2637 instruction, the index list may have zero or more indexes, which are
2638 required to make sense for the type of "CSTPTR".
2639 ``select (COND, VAL1, VAL2)``
2640 Perform the :ref:`select operation <i_select>` on constants.
2641 ``icmp COND (VAL1, VAL2)``
2642 Performs the :ref:`icmp operation <i_icmp>` on constants.
2643 ``fcmp COND (VAL1, VAL2)``
2644 Performs the :ref:`fcmp operation <i_fcmp>` on constants.
2645 ``extractelement (VAL, IDX)``
2646 Perform the :ref:`extractelement operation <i_extractelement>` on
2648 ``insertelement (VAL, ELT, IDX)``
2649 Perform the :ref:`insertelement operation <i_insertelement>` on
2651 ``shufflevector (VEC1, VEC2, IDXMASK)``
2652 Perform the :ref:`shufflevector operation <i_shufflevector>` on
2654 ``extractvalue (VAL, IDX0, IDX1, ...)``
2655 Perform the :ref:`extractvalue operation <i_extractvalue>` on
2656 constants. The index list is interpreted in a similar manner as
2657 indices in a ':ref:`getelementptr <i_getelementptr>`' operation. At
2658 least one index value must be specified.
2659 ``insertvalue (VAL, ELT, IDX0, IDX1, ...)``
2660 Perform the :ref:`insertvalue operation <i_insertvalue>` on constants.
2661 The index list is interpreted in a similar manner as indices in a
2662 ':ref:`getelementptr <i_getelementptr>`' operation. At least one index
2663 value must be specified.
2664 ``OPCODE (LHS, RHS)``
2665 Perform the specified operation of the LHS and RHS constants. OPCODE
2666 may be any of the :ref:`binary <binaryops>` or :ref:`bitwise
2667 binary <bitwiseops>` operations. The constraints on operands are
2668 the same as those for the corresponding instruction (e.g. no bitwise
2669 operations on floating point values are allowed).
2676 Inline Assembler Expressions
2677 ----------------------------
2679 LLVM supports inline assembler expressions (as opposed to :ref:`Module-Level
2680 Inline Assembly <moduleasm>`) through the use of a special value. This
2681 value represents the inline assembler as a string (containing the
2682 instructions to emit), a list of operand constraints (stored as a
2683 string), a flag that indicates whether or not the inline asm expression
2684 has side effects, and a flag indicating whether the function containing
2685 the asm needs to align its stack conservatively. An example inline
2686 assembler expression is:
2688 .. code-block:: llvm
2690 i32 (i32) asm "bswap $0", "=r,r"
2692 Inline assembler expressions may **only** be used as the callee operand
2693 of a :ref:`call <i_call>` or an :ref:`invoke <i_invoke>` instruction.
2694 Thus, typically we have:
2696 .. code-block:: llvm
2698 %X = call i32 asm "bswap $0", "=r,r"(i32 %Y)
2700 Inline asms with side effects not visible in the constraint list must be
2701 marked as having side effects. This is done through the use of the
2702 '``sideeffect``' keyword, like so:
2704 .. code-block:: llvm
2706 call void asm sideeffect "eieio", ""()
2708 In some cases inline asms will contain code that will not work unless
2709 the stack is aligned in some way, such as calls or SSE instructions on
2710 x86, yet will not contain code that does that alignment within the asm.
2711 The compiler should make conservative assumptions about what the asm
2712 might contain and should generate its usual stack alignment code in the
2713 prologue if the '``alignstack``' keyword is present:
2715 .. code-block:: llvm
2717 call void asm alignstack "eieio", ""()
2719 Inline asms also support using non-standard assembly dialects. The
2720 assumed dialect is ATT. When the '``inteldialect``' keyword is present,
2721 the inline asm is using the Intel dialect. Currently, ATT and Intel are
2722 the only supported dialects. An example is:
2724 .. code-block:: llvm
2726 call void asm inteldialect "eieio", ""()
2728 If multiple keywords appear the '``sideeffect``' keyword must come
2729 first, the '``alignstack``' keyword second and the '``inteldialect``'
2735 The call instructions that wrap inline asm nodes may have a
2736 "``!srcloc``" MDNode attached to it that contains a list of constant
2737 integers. If present, the code generator will use the integer as the
2738 location cookie value when report errors through the ``LLVMContext``
2739 error reporting mechanisms. This allows a front-end to correlate backend
2740 errors that occur with inline asm back to the source code that produced
2743 .. code-block:: llvm
2745 call void asm sideeffect "something bad", ""(), !srcloc !42
2747 !42 = !{ i32 1234567 }
2749 It is up to the front-end to make sense of the magic numbers it places
2750 in the IR. If the MDNode contains multiple constants, the code generator
2751 will use the one that corresponds to the line of the asm that the error
2756 Metadata Nodes and Metadata Strings
2757 -----------------------------------
2759 LLVM IR allows metadata to be attached to instructions in the program
2760 that can convey extra information about the code to the optimizers and
2761 code generator. One example application of metadata is source-level
2762 debug information. There are two metadata primitives: strings and nodes.
2763 All metadata has the ``metadata`` type and is identified in syntax by a
2764 preceding exclamation point ('``!``').
2766 A metadata string is a string surrounded by double quotes. It can
2767 contain any character by escaping non-printable characters with
2768 "``\xx``" where "``xx``" is the two digit hex code. For example:
2771 Metadata nodes are represented with notation similar to structure
2772 constants (a comma separated list of elements, surrounded by braces and
2773 preceded by an exclamation point). Metadata nodes can have any values as
2774 their operand. For example:
2776 .. code-block:: llvm
2778 !{ metadata !"test\00", i32 10}
2780 A :ref:`named metadata <namedmetadatastructure>` is a collection of
2781 metadata nodes, which can be looked up in the module symbol table. For
2784 .. code-block:: llvm
2786 !foo = metadata !{!4, !3}
2788 Metadata can be used as function arguments. Here ``llvm.dbg.value``
2789 function is using two metadata arguments:
2791 .. code-block:: llvm
2793 call void @llvm.dbg.value(metadata !24, i64 0, metadata !25)
2795 Metadata can be attached with an instruction. Here metadata ``!21`` is
2796 attached to the ``add`` instruction using the ``!dbg`` identifier:
2798 .. code-block:: llvm
2800 %indvar.next = add i64 %indvar, 1, !dbg !21
2802 More information about specific metadata nodes recognized by the
2803 optimizers and code generator is found below.
2808 In LLVM IR, memory does not have types, so LLVM's own type system is not
2809 suitable for doing TBAA. Instead, metadata is added to the IR to
2810 describe a type system of a higher level language. This can be used to
2811 implement typical C/C++ TBAA, but it can also be used to implement
2812 custom alias analysis behavior for other languages.
2814 The current metadata format is very simple. TBAA metadata nodes have up
2815 to three fields, e.g.:
2817 .. code-block:: llvm
2819 !0 = metadata !{ metadata !"an example type tree" }
2820 !1 = metadata !{ metadata !"int", metadata !0 }
2821 !2 = metadata !{ metadata !"float", metadata !0 }
2822 !3 = metadata !{ metadata !"const float", metadata !2, i64 1 }
2824 The first field is an identity field. It can be any value, usually a
2825 metadata string, which uniquely identifies the type. The most important
2826 name in the tree is the name of the root node. Two trees with different
2827 root node names are entirely disjoint, even if they have leaves with
2830 The second field identifies the type's parent node in the tree, or is
2831 null or omitted for a root node. A type is considered to alias all of
2832 its descendants and all of its ancestors in the tree. Also, a type is
2833 considered to alias all types in other trees, so that bitcode produced
2834 from multiple front-ends is handled conservatively.
2836 If the third field is present, it's an integer which if equal to 1
2837 indicates that the type is "constant" (meaning
2838 ``pointsToConstantMemory`` should return true; see `other useful
2839 AliasAnalysis methods <AliasAnalysis.html#OtherItfs>`_).
2841 '``tbaa.struct``' Metadata
2842 ^^^^^^^^^^^^^^^^^^^^^^^^^^
2844 The :ref:`llvm.memcpy <int_memcpy>` is often used to implement
2845 aggregate assignment operations in C and similar languages, however it
2846 is defined to copy a contiguous region of memory, which is more than
2847 strictly necessary for aggregate types which contain holes due to
2848 padding. Also, it doesn't contain any TBAA information about the fields
2851 ``!tbaa.struct`` metadata can describe which memory subregions in a
2852 memcpy are padding and what the TBAA tags of the struct are.
2854 The current metadata format is very simple. ``!tbaa.struct`` metadata
2855 nodes are a list of operands which are in conceptual groups of three.
2856 For each group of three, the first operand gives the byte offset of a
2857 field in bytes, the second gives its size in bytes, and the third gives
2860 .. code-block:: llvm
2862 !4 = metadata !{ i64 0, i64 4, metadata !1, i64 8, i64 4, metadata !2 }
2864 This describes a struct with two fields. The first is at offset 0 bytes
2865 with size 4 bytes, and has tbaa tag !1. The second is at offset 8 bytes
2866 and has size 4 bytes and has tbaa tag !2.
2868 Note that the fields need not be contiguous. In this example, there is a
2869 4 byte gap between the two fields. This gap represents padding which
2870 does not carry useful data and need not be preserved.
2872 '``noalias``' and '``alias.scope``' Metadata
2873 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2875 ``noalias`` and ``alias.scope`` metadata provide the ability to specify generic
2876 noalias memory-access sets. This means that some collection of memory access
2877 instructions (loads, stores, memory-accessing calls, etc.) that carry
2878 ``noalias`` metadata can specifically be specified not to alias with some other
2879 collection of memory access instructions that carry ``alias.scope`` metadata.
2880 Each type of metadata specifies a list of scopes where each scope has an id and
2881 a domain. When evaluating an aliasing query, if for some some domain, the set
2882 of scopes with that domain in one instruction's ``alias.scope`` list is a
2883 subset of (or qual to) the set of scopes for that domain in another
2884 instruction's ``noalias`` list, then the two memory accesses are assumed not to
2887 The metadata identifying each domain is itself a list containing one or two
2888 entries. The first entry is the name of the domain. Note that if the name is a
2889 string then it can be combined accross functions and translation units. A
2890 self-reference can be used to create globally unique domain names. A
2891 descriptive string may optionally be provided as a second list entry.
2893 The metadata identifying each scope is also itself a list containing two or
2894 three entries. The first entry is the name of the scope. Note that if the name
2895 is a string then it can be combined accross functions and translation units. A
2896 self-reference can be used to create globally unique scope names. A metadata
2897 reference to the scope's domain is the second entry. A descriptive string may
2898 optionally be provided as a third list entry.
2902 .. code-block:: llvm
2904 ; Two scope domains:
2905 !0 = metadata !{metadata !0}
2906 !1 = metadata !{metadata !1}
2908 ; Some scopes in these domains:
2909 !2 = metadata !{metadata !2, metadata !0}
2910 !3 = metadata !{metadata !3, metadata !0}
2911 !4 = metadata !{metadata !4, metadata !1}
2914 !5 = metadata !{metadata !4} ; A list containing only scope !4
2915 !6 = metadata !{metadata !4, metadata !3, metadata !2}
2916 !7 = metadata !{metadata !3}
2918 ; These two instructions don't alias:
2919 %0 = load float* %c, align 4, !alias.scope !5
2920 store float %0, float* %arrayidx.i, align 4, !noalias !5
2922 ; These two instructions also don't alias (for domain !1, the set of scopes
2923 ; in the !alias.scope equals that in the !noalias list):
2924 %2 = load float* %c, align 4, !alias.scope !5
2925 store float %2, float* %arrayidx.i2, align 4, !noalias !6
2927 ; These two instructions don't alias (for domain !0, the set of scopes in
2928 ; the !noalias list is not a superset of, or equal to, the scopes in the
2929 ; !alias.scope list):
2930 %2 = load float* %c, align 4, !alias.scope !6
2931 store float %0, float* %arrayidx.i, align 4, !noalias !7
2933 '``fpmath``' Metadata
2934 ^^^^^^^^^^^^^^^^^^^^^
2936 ``fpmath`` metadata may be attached to any instruction of floating point
2937 type. It can be used to express the maximum acceptable error in the
2938 result of that instruction, in ULPs, thus potentially allowing the
2939 compiler to use a more efficient but less accurate method of computing
2940 it. ULP is defined as follows:
2942 If ``x`` is a real number that lies between two finite consecutive
2943 floating-point numbers ``a`` and ``b``, without being equal to one
2944 of them, then ``ulp(x) = |b - a|``, otherwise ``ulp(x)`` is the
2945 distance between the two non-equal finite floating-point numbers
2946 nearest ``x``. Moreover, ``ulp(NaN)`` is ``NaN``.
2948 The metadata node shall consist of a single positive floating point
2949 number representing the maximum relative error, for example:
2951 .. code-block:: llvm
2953 !0 = metadata !{ float 2.5 } ; maximum acceptable inaccuracy is 2.5 ULPs
2955 '``range``' Metadata
2956 ^^^^^^^^^^^^^^^^^^^^
2958 ``range`` metadata may be attached only to ``load``, ``call`` and ``invoke`` of
2959 integer types. It expresses the possible ranges the loaded value or the value
2960 returned by the called function at this call site is in. The ranges are
2961 represented with a flattened list of integers. The loaded value or the value
2962 returned is known to be in the union of the ranges defined by each consecutive
2963 pair. Each pair has the following properties:
2965 - The type must match the type loaded by the instruction.
2966 - The pair ``a,b`` represents the range ``[a,b)``.
2967 - Both ``a`` and ``b`` are constants.
2968 - The range is allowed to wrap.
2969 - The range should not represent the full or empty set. That is,
2972 In addition, the pairs must be in signed order of the lower bound and
2973 they must be non-contiguous.
2977 .. code-block:: llvm
2979 %a = load i8* %x, align 1, !range !0 ; Can only be 0 or 1
2980 %b = load i8* %y, align 1, !range !1 ; Can only be 255 (-1), 0 or 1
2981 %c = call i8 @foo(), !range !2 ; Can only be 0, 1, 3, 4 or 5
2982 %d = invoke i8 @bar() to label %cont
2983 unwind label %lpad, !range !3 ; Can only be -2, -1, 3, 4 or 5
2985 !0 = metadata !{ i8 0, i8 2 }
2986 !1 = metadata !{ i8 255, i8 2 }
2987 !2 = metadata !{ i8 0, i8 2, i8 3, i8 6 }
2988 !3 = metadata !{ i8 -2, i8 0, i8 3, i8 6 }
2993 It is sometimes useful to attach information to loop constructs. Currently,
2994 loop metadata is implemented as metadata attached to the branch instruction
2995 in the loop latch block. This type of metadata refer to a metadata node that is
2996 guaranteed to be separate for each loop. The loop identifier metadata is
2997 specified with the name ``llvm.loop``.
2999 The loop identifier metadata is implemented using a metadata that refers to
3000 itself to avoid merging it with any other identifier metadata, e.g.,
3001 during module linkage or function inlining. That is, each loop should refer
3002 to their own identification metadata even if they reside in separate functions.
3003 The following example contains loop identifier metadata for two separate loop
3006 .. code-block:: llvm
3008 !0 = metadata !{ metadata !0 }
3009 !1 = metadata !{ metadata !1 }
3011 The loop identifier metadata can be used to specify additional
3012 per-loop metadata. Any operands after the first operand can be treated
3013 as user-defined metadata. For example the ``llvm.loop.unroll.count``
3014 suggests an unroll factor to the loop unroller:
3016 .. code-block:: llvm
3018 br i1 %exitcond, label %._crit_edge, label %.lr.ph, !llvm.loop !0
3020 !0 = metadata !{ metadata !0, metadata !1 }
3021 !1 = metadata !{ metadata !"llvm.loop.unroll.count", i32 4 }
3023 '``llvm.loop.vectorize``' and '``llvm.loop.interleave``'
3024 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3026 Metadata prefixed with ``llvm.loop.vectorize`` or ``llvm.loop.interleave`` are
3027 used to control per-loop vectorization and interleaving parameters such as
3028 vectorization width and interleave count. These metadata should be used in
3029 conjunction with ``llvm.loop`` loop identification metadata. The
3030 ``llvm.loop.vectorize`` and ``llvm.loop.interleave`` metadata are only
3031 optimization hints and the optimizer will only interleave and vectorize loops if
3032 it believes it is safe to do so. The ``llvm.mem.parallel_loop_access`` metadata
3033 which contains information about loop-carried memory dependencies can be helpful
3034 in determining the safety of these transformations.
3036 '``llvm.loop.interleave.count``' Metadata
3037 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3039 This metadata suggests an interleave count to the loop interleaver.
3040 The first operand is the string ``llvm.loop.interleave.count`` and the
3041 second operand is an integer specifying the interleave count. For
3044 .. code-block:: llvm
3046 !0 = metadata !{ metadata !"llvm.loop.interleave.count", i32 4 }
3048 Note that setting ``llvm.loop.interleave.count`` to 1 disables interleaving
3049 multiple iterations of the loop. If ``llvm.loop.interleave.count`` is set to 0
3050 then the interleave count will be determined automatically.
3052 '``llvm.loop.vectorize.enable``' Metadata
3053 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3055 This metadata selectively enables or disables vectorization for the loop. The
3056 first operand is the string ``llvm.loop.vectorize.enable`` and the second operand
3057 is a bit. If the bit operand value is 1 vectorization is enabled. A value of
3058 0 disables vectorization:
3060 .. code-block:: llvm
3062 !0 = metadata !{ metadata !"llvm.loop.vectorize.enable", i1 0 }
3063 !1 = metadata !{ metadata !"llvm.loop.vectorize.enable", i1 1 }
3065 '``llvm.loop.vectorize.width``' Metadata
3066 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3068 This metadata sets the target width of the vectorizer. The first
3069 operand is the string ``llvm.loop.vectorize.width`` and the second
3070 operand is an integer specifying the width. For example:
3072 .. code-block:: llvm
3074 !0 = metadata !{ metadata !"llvm.loop.vectorize.width", i32 4 }
3076 Note that setting ``llvm.loop.vectorize.width`` to 1 disables
3077 vectorization of the loop. If ``llvm.loop.vectorize.width`` is set to
3078 0 or if the loop does not have this metadata the width will be
3079 determined automatically.
3081 '``llvm.loop.unroll``'
3082 ^^^^^^^^^^^^^^^^^^^^^^
3084 Metadata prefixed with ``llvm.loop.unroll`` are loop unrolling
3085 optimization hints such as the unroll factor. ``llvm.loop.unroll``
3086 metadata should be used in conjunction with ``llvm.loop`` loop
3087 identification metadata. The ``llvm.loop.unroll`` metadata are only
3088 optimization hints and the unrolling will only be performed if the
3089 optimizer believes it is safe to do so.
3091 '``llvm.loop.unroll.count``' Metadata
3092 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3094 This metadata suggests an unroll factor to the loop unroller. The
3095 first operand is the string ``llvm.loop.unroll.count`` and the second
3096 operand is a positive integer specifying the unroll factor. For
3099 .. code-block:: llvm
3101 !0 = metadata !{ metadata !"llvm.loop.unroll.count", i32 4 }
3103 If the trip count of the loop is less than the unroll count the loop
3104 will be partially unrolled.
3106 '``llvm.loop.unroll.disable``' Metadata
3107 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3109 This metadata either disables loop unrolling. The metadata has a single operand
3110 which is the string ``llvm.loop.unroll.disable``. For example:
3112 .. code-block:: llvm
3114 !0 = metadata !{ metadata !"llvm.loop.unroll.disable" }
3116 '``llvm.loop.unroll.full``' Metadata
3117 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3119 This metadata either suggests that the loop should be unrolled fully. The
3120 metadata has a single operand which is the string ``llvm.loop.unroll.disable``.
3123 .. code-block:: llvm
3125 !0 = metadata !{ metadata !"llvm.loop.unroll.full" }
3130 Metadata types used to annotate memory accesses with information helpful
3131 for optimizations are prefixed with ``llvm.mem``.
3133 '``llvm.mem.parallel_loop_access``' Metadata
3134 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3136 The ``llvm.mem.parallel_loop_access`` metadata refers to a loop identifier,
3137 or metadata containing a list of loop identifiers for nested loops.
3138 The metadata is attached to memory accessing instructions and denotes that
3139 no loop carried memory dependence exist between it and other instructions denoted
3140 with the same loop identifier.
3142 Precisely, given two instructions ``m1`` and ``m2`` that both have the
3143 ``llvm.mem.parallel_loop_access`` metadata, with ``L1`` and ``L2`` being the
3144 set of loops associated with that metadata, respectively, then there is no loop
3145 carried dependence between ``m1`` and ``m2`` for loops in both ``L1`` and
3148 As a special case, if all memory accessing instructions in a loop have
3149 ``llvm.mem.parallel_loop_access`` metadata that refers to that loop, then the
3150 loop has no loop carried memory dependences and is considered to be a parallel
3153 Note that if not all memory access instructions have such metadata referring to
3154 the loop, then the loop is considered not being trivially parallel. Additional
3155 memory dependence analysis is required to make that determination. As a fail
3156 safe mechanism, this causes loops that were originally parallel to be considered
3157 sequential (if optimization passes that are unaware of the parallel semantics
3158 insert new memory instructions into the loop body).
3160 Example of a loop that is considered parallel due to its correct use of
3161 both ``llvm.loop`` and ``llvm.mem.parallel_loop_access``
3162 metadata types that refer to the same loop identifier metadata.
3164 .. code-block:: llvm
3168 %val0 = load i32* %arrayidx, !llvm.mem.parallel_loop_access !0
3170 store i32 %val0, i32* %arrayidx1, !llvm.mem.parallel_loop_access !0
3172 br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !0
3176 !0 = metadata !{ metadata !0 }
3178 It is also possible to have nested parallel loops. In that case the
3179 memory accesses refer to a list of loop identifier metadata nodes instead of
3180 the loop identifier metadata node directly:
3182 .. code-block:: llvm
3186 %val1 = load i32* %arrayidx3, !llvm.mem.parallel_loop_access !2
3188 br label %inner.for.body
3192 %val0 = load i32* %arrayidx1, !llvm.mem.parallel_loop_access !0
3194 store i32 %val0, i32* %arrayidx2, !llvm.mem.parallel_loop_access !0
3196 br i1 %exitcond, label %inner.for.end, label %inner.for.body, !llvm.loop !1
3200 store i32 %val1, i32* %arrayidx4, !llvm.mem.parallel_loop_access !2
3202 br i1 %exitcond, label %outer.for.end, label %outer.for.body, !llvm.loop !2
3204 outer.for.end: ; preds = %for.body
3206 !0 = metadata !{ metadata !1, metadata !2 } ; a list of loop identifiers
3207 !1 = metadata !{ metadata !1 } ; an identifier for the inner loop
3208 !2 = metadata !{ metadata !2 } ; an identifier for the outer loop
3210 Module Flags Metadata
3211 =====================
3213 Information about the module as a whole is difficult to convey to LLVM's
3214 subsystems. The LLVM IR isn't sufficient to transmit this information.
3215 The ``llvm.module.flags`` named metadata exists in order to facilitate
3216 this. These flags are in the form of key / value pairs --- much like a
3217 dictionary --- making it easy for any subsystem who cares about a flag to
3220 The ``llvm.module.flags`` metadata contains a list of metadata triplets.
3221 Each triplet has the following form:
3223 - The first element is a *behavior* flag, which specifies the behavior
3224 when two (or more) modules are merged together, and it encounters two
3225 (or more) metadata with the same ID. The supported behaviors are
3227 - The second element is a metadata string that is a unique ID for the
3228 metadata. Each module may only have one flag entry for each unique ID (not
3229 including entries with the **Require** behavior).
3230 - The third element is the value of the flag.
3232 When two (or more) modules are merged together, the resulting
3233 ``llvm.module.flags`` metadata is the union of the modules' flags. That is, for
3234 each unique metadata ID string, there will be exactly one entry in the merged
3235 modules ``llvm.module.flags`` metadata table, and the value for that entry will
3236 be determined by the merge behavior flag, as described below. The only exception
3237 is that entries with the *Require* behavior are always preserved.
3239 The following behaviors are supported:
3250 Emits an error if two values disagree, otherwise the resulting value
3251 is that of the operands.
3255 Emits a warning if two values disagree. The result value will be the
3256 operand for the flag from the first module being linked.
3260 Adds a requirement that another module flag be present and have a
3261 specified value after linking is performed. The value must be a
3262 metadata pair, where the first element of the pair is the ID of the
3263 module flag to be restricted, and the second element of the pair is
3264 the value the module flag should be restricted to. This behavior can
3265 be used to restrict the allowable results (via triggering of an
3266 error) of linking IDs with the **Override** behavior.
3270 Uses the specified value, regardless of the behavior or value of the
3271 other module. If both modules specify **Override**, but the values
3272 differ, an error will be emitted.
3276 Appends the two values, which are required to be metadata nodes.
3280 Appends the two values, which are required to be metadata
3281 nodes. However, duplicate entries in the second list are dropped
3282 during the append operation.
3284 It is an error for a particular unique flag ID to have multiple behaviors,
3285 except in the case of **Require** (which adds restrictions on another metadata
3286 value) or **Override**.
3288 An example of module flags:
3290 .. code-block:: llvm
3292 !0 = metadata !{ i32 1, metadata !"foo", i32 1 }
3293 !1 = metadata !{ i32 4, metadata !"bar", i32 37 }
3294 !2 = metadata !{ i32 2, metadata !"qux", i32 42 }
3295 !3 = metadata !{ i32 3, metadata !"qux",
3297 metadata !"foo", i32 1
3300 !llvm.module.flags = !{ !0, !1, !2, !3 }
3302 - Metadata ``!0`` has the ID ``!"foo"`` and the value '1'. The behavior
3303 if two or more ``!"foo"`` flags are seen is to emit an error if their
3304 values are not equal.
3306 - Metadata ``!1`` has the ID ``!"bar"`` and the value '37'. The
3307 behavior if two or more ``!"bar"`` flags are seen is to use the value
3310 - Metadata ``!2`` has the ID ``!"qux"`` and the value '42'. The
3311 behavior if two or more ``!"qux"`` flags are seen is to emit a
3312 warning if their values are not equal.
3314 - Metadata ``!3`` has the ID ``!"qux"`` and the value:
3318 metadata !{ metadata !"foo", i32 1 }
3320 The behavior is to emit an error if the ``llvm.module.flags`` does not
3321 contain a flag with the ID ``!"foo"`` that has the value '1' after linking is
3324 Objective-C Garbage Collection Module Flags Metadata
3325 ----------------------------------------------------
3327 On the Mach-O platform, Objective-C stores metadata about garbage
3328 collection in a special section called "image info". The metadata
3329 consists of a version number and a bitmask specifying what types of
3330 garbage collection are supported (if any) by the file. If two or more
3331 modules are linked together their garbage collection metadata needs to
3332 be merged rather than appended together.
3334 The Objective-C garbage collection module flags metadata consists of the
3335 following key-value pairs:
3344 * - ``Objective-C Version``
3345 - **[Required]** --- The Objective-C ABI version. Valid values are 1 and 2.
3347 * - ``Objective-C Image Info Version``
3348 - **[Required]** --- The version of the image info section. Currently
3351 * - ``Objective-C Image Info Section``
3352 - **[Required]** --- The section to place the metadata. Valid values are
3353 ``"__OBJC, __image_info, regular"`` for Objective-C ABI version 1, and
3354 ``"__DATA,__objc_imageinfo, regular, no_dead_strip"`` for
3355 Objective-C ABI version 2.
3357 * - ``Objective-C Garbage Collection``
3358 - **[Required]** --- Specifies whether garbage collection is supported or
3359 not. Valid values are 0, for no garbage collection, and 2, for garbage
3360 collection supported.
3362 * - ``Objective-C GC Only``
3363 - **[Optional]** --- Specifies that only garbage collection is supported.
3364 If present, its value must be 6. This flag requires that the
3365 ``Objective-C Garbage Collection`` flag have the value 2.
3367 Some important flag interactions:
3369 - If a module with ``Objective-C Garbage Collection`` set to 0 is
3370 merged with a module with ``Objective-C Garbage Collection`` set to
3371 2, then the resulting module has the
3372 ``Objective-C Garbage Collection`` flag set to 0.
3373 - A module with ``Objective-C Garbage Collection`` set to 0 cannot be
3374 merged with a module with ``Objective-C GC Only`` set to 6.
3376 Automatic Linker Flags Module Flags Metadata
3377 --------------------------------------------
3379 Some targets support embedding flags to the linker inside individual object
3380 files. Typically this is used in conjunction with language extensions which
3381 allow source files to explicitly declare the libraries they depend on, and have
3382 these automatically be transmitted to the linker via object files.
3384 These flags are encoded in the IR using metadata in the module flags section,
3385 using the ``Linker Options`` key. The merge behavior for this flag is required
3386 to be ``AppendUnique``, and the value for the key is expected to be a metadata
3387 node which should be a list of other metadata nodes, each of which should be a
3388 list of metadata strings defining linker options.
3390 For example, the following metadata section specifies two separate sets of
3391 linker options, presumably to link against ``libz`` and the ``Cocoa``
3394 !0 = metadata !{ i32 6, metadata !"Linker Options",
3396 metadata !{ metadata !"-lz" },
3397 metadata !{ metadata !"-framework", metadata !"Cocoa" } } }
3398 !llvm.module.flags = !{ !0 }
3400 The metadata encoding as lists of lists of options, as opposed to a collapsed
3401 list of options, is chosen so that the IR encoding can use multiple option
3402 strings to specify e.g., a single library, while still having that specifier be
3403 preserved as an atomic element that can be recognized by a target specific
3404 assembly writer or object file emitter.
3406 Each individual option is required to be either a valid option for the target's
3407 linker, or an option that is reserved by the target specific assembly writer or
3408 object file emitter. No other aspect of these options is defined by the IR.
3410 C type width Module Flags Metadata
3411 ----------------------------------
3413 The ARM backend emits a section into each generated object file describing the
3414 options that it was compiled with (in a compiler-independent way) to prevent
3415 linking incompatible objects, and to allow automatic library selection. Some
3416 of these options are not visible at the IR level, namely wchar_t width and enum
3419 To pass this information to the backend, these options are encoded in module
3420 flags metadata, using the following key-value pairs:
3430 - * 0 --- sizeof(wchar_t) == 4
3431 * 1 --- sizeof(wchar_t) == 2
3434 - * 0 --- Enums are at least as large as an ``int``.
3435 * 1 --- Enums are stored in the smallest integer type which can
3436 represent all of its values.
3438 For example, the following metadata section specifies that the module was
3439 compiled with a ``wchar_t`` width of 4 bytes, and the underlying type of an
3440 enum is the smallest type which can represent all of its values::
3442 !llvm.module.flags = !{!0, !1}
3443 !0 = metadata !{i32 1, metadata !"short_wchar", i32 1}
3444 !1 = metadata !{i32 1, metadata !"short_enum", i32 0}
3446 .. _intrinsicglobalvariables:
3448 Intrinsic Global Variables
3449 ==========================
3451 LLVM has a number of "magic" global variables that contain data that
3452 affect code generation or other IR semantics. These are documented here.
3453 All globals of this sort should have a section specified as
3454 "``llvm.metadata``". This section and all globals that start with
3455 "``llvm.``" are reserved for use by LLVM.
3459 The '``llvm.used``' Global Variable
3460 -----------------------------------
3462 The ``@llvm.used`` global is an array which has
3463 :ref:`appending linkage <linkage_appending>`. This array contains a list of
3464 pointers to named global variables, functions and aliases which may optionally
3465 have a pointer cast formed of bitcast or getelementptr. For example, a legal
3468 .. code-block:: llvm
3473 @llvm.used = appending global [2 x i8*] [
3475 i8* bitcast (i32* @Y to i8*)
3476 ], section "llvm.metadata"
3478 If a symbol appears in the ``@llvm.used`` list, then the compiler, assembler,
3479 and linker are required to treat the symbol as if there is a reference to the
3480 symbol that it cannot see (which is why they have to be named). For example, if
3481 a variable has internal linkage and no references other than that from the
3482 ``@llvm.used`` list, it cannot be deleted. This is commonly used to represent
3483 references from inline asms and other things the compiler cannot "see", and
3484 corresponds to "``attribute((used))``" in GNU C.
3486 On some targets, the code generator must emit a directive to the
3487 assembler or object file to prevent the assembler and linker from
3488 molesting the symbol.
3490 .. _gv_llvmcompilerused:
3492 The '``llvm.compiler.used``' Global Variable
3493 --------------------------------------------
3495 The ``@llvm.compiler.used`` directive is the same as the ``@llvm.used``
3496 directive, except that it only prevents the compiler from touching the
3497 symbol. On targets that support it, this allows an intelligent linker to
3498 optimize references to the symbol without being impeded as it would be
3501 This is a rare construct that should only be used in rare circumstances,
3502 and should not be exposed to source languages.
3504 .. _gv_llvmglobalctors:
3506 The '``llvm.global_ctors``' Global Variable
3507 -------------------------------------------
3509 .. code-block:: llvm
3511 %0 = type { i32, void ()*, i8* }
3512 @llvm.global_ctors = appending global [1 x %0] [%0 { i32 65535, void ()* @ctor, i8* @data }]
3514 The ``@llvm.global_ctors`` array contains a list of constructor
3515 functions, priorities, and an optional associated global or function.
3516 The functions referenced by this array will be called in ascending order
3517 of priority (i.e. lowest first) when the module is loaded. The order of
3518 functions with the same priority is not defined.
3520 If the third field is present, non-null, and points to a global variable
3521 or function, the initializer function will only run if the associated
3522 data from the current module is not discarded.
3524 .. _llvmglobaldtors:
3526 The '``llvm.global_dtors``' Global Variable
3527 -------------------------------------------
3529 .. code-block:: llvm
3531 %0 = type { i32, void ()*, i8* }
3532 @llvm.global_dtors = appending global [1 x %0] [%0 { i32 65535, void ()* @dtor, i8* @data }]
3534 The ``@llvm.global_dtors`` array contains a list of destructor
3535 functions, priorities, and an optional associated global or function.
3536 The functions referenced by this array will be called in descending
3537 order of priority (i.e. highest first) when the module is unloaded. The
3538 order of functions with the same priority is not defined.
3540 If the third field is present, non-null, and points to a global variable
3541 or function, the destructor function will only run if the associated
3542 data from the current module is not discarded.
3544 Instruction Reference
3545 =====================
3547 The LLVM instruction set consists of several different classifications
3548 of instructions: :ref:`terminator instructions <terminators>`, :ref:`binary
3549 instructions <binaryops>`, :ref:`bitwise binary
3550 instructions <bitwiseops>`, :ref:`memory instructions <memoryops>`, and
3551 :ref:`other instructions <otherops>`.
3555 Terminator Instructions
3556 -----------------------
3558 As mentioned :ref:`previously <functionstructure>`, every basic block in a
3559 program ends with a "Terminator" instruction, which indicates which
3560 block should be executed after the current block is finished. These
3561 terminator instructions typically yield a '``void``' value: they produce
3562 control flow, not values (the one exception being the
3563 ':ref:`invoke <i_invoke>`' instruction).
3565 The terminator instructions are: ':ref:`ret <i_ret>`',
3566 ':ref:`br <i_br>`', ':ref:`switch <i_switch>`',
3567 ':ref:`indirectbr <i_indirectbr>`', ':ref:`invoke <i_invoke>`',
3568 ':ref:`resume <i_resume>`', and ':ref:`unreachable <i_unreachable>`'.
3572 '``ret``' Instruction
3573 ^^^^^^^^^^^^^^^^^^^^^
3580 ret <type> <value> ; Return a value from a non-void function
3581 ret void ; Return from void function
3586 The '``ret``' instruction is used to return control flow (and optionally
3587 a value) from a function back to the caller.
3589 There are two forms of the '``ret``' instruction: one that returns a
3590 value and then causes control flow, and one that just causes control
3596 The '``ret``' instruction optionally accepts a single argument, the
3597 return value. The type of the return value must be a ':ref:`first
3598 class <t_firstclass>`' type.
3600 A function is not :ref:`well formed <wellformed>` if it it has a non-void
3601 return type and contains a '``ret``' instruction with no return value or
3602 a return value with a type that does not match its type, or if it has a
3603 void return type and contains a '``ret``' instruction with a return
3609 When the '``ret``' instruction is executed, control flow returns back to
3610 the calling function's context. If the caller is a
3611 ":ref:`call <i_call>`" instruction, execution continues at the
3612 instruction after the call. If the caller was an
3613 ":ref:`invoke <i_invoke>`" instruction, execution continues at the
3614 beginning of the "normal" destination block. If the instruction returns
3615 a value, that value shall set the call or invoke instruction's return
3621 .. code-block:: llvm
3623 ret i32 5 ; Return an integer value of 5
3624 ret void ; Return from a void function
3625 ret { i32, i8 } { i32 4, i8 2 } ; Return a struct of values 4 and 2
3629 '``br``' Instruction
3630 ^^^^^^^^^^^^^^^^^^^^
3637 br i1 <cond>, label <iftrue>, label <iffalse>
3638 br label <dest> ; Unconditional branch
3643 The '``br``' instruction is used to cause control flow to transfer to a
3644 different basic block in the current function. There are two forms of
3645 this instruction, corresponding to a conditional branch and an
3646 unconditional branch.
3651 The conditional branch form of the '``br``' instruction takes a single
3652 '``i1``' value and two '``label``' values. The unconditional form of the
3653 '``br``' instruction takes a single '``label``' value as a target.
3658 Upon execution of a conditional '``br``' instruction, the '``i1``'
3659 argument is evaluated. If the value is ``true``, control flows to the
3660 '``iftrue``' ``label`` argument. If "cond" is ``false``, control flows
3661 to the '``iffalse``' ``label`` argument.
3666 .. code-block:: llvm
3669 %cond = icmp eq i32 %a, %b
3670 br i1 %cond, label %IfEqual, label %IfUnequal
3678 '``switch``' Instruction
3679 ^^^^^^^^^^^^^^^^^^^^^^^^
3686 switch <intty> <value>, label <defaultdest> [ <intty> <val>, label <dest> ... ]
3691 The '``switch``' instruction is used to transfer control flow to one of
3692 several different places. It is a generalization of the '``br``'
3693 instruction, allowing a branch to occur to one of many possible
3699 The '``switch``' instruction uses three parameters: an integer
3700 comparison value '``value``', a default '``label``' destination, and an
3701 array of pairs of comparison value constants and '``label``'s. The table
3702 is not allowed to contain duplicate constant entries.
3707 The ``switch`` instruction specifies a table of values and destinations.
3708 When the '``switch``' instruction is executed, this table is searched
3709 for the given value. If the value is found, control flow is transferred
3710 to the corresponding destination; otherwise, control flow is transferred
3711 to the default destination.
3716 Depending on properties of the target machine and the particular
3717 ``switch`` instruction, this instruction may be code generated in
3718 different ways. For example, it could be generated as a series of
3719 chained conditional branches or with a lookup table.
3724 .. code-block:: llvm
3726 ; Emulate a conditional br instruction
3727 %Val = zext i1 %value to i32
3728 switch i32 %Val, label %truedest [ i32 0, label %falsedest ]
3730 ; Emulate an unconditional br instruction
3731 switch i32 0, label %dest [ ]
3733 ; Implement a jump table:
3734 switch i32 %val, label %otherwise [ i32 0, label %onzero
3736 i32 2, label %ontwo ]
3740 '``indirectbr``' Instruction
3741 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3748 indirectbr <somety>* <address>, [ label <dest1>, label <dest2>, ... ]
3753 The '``indirectbr``' instruction implements an indirect branch to a
3754 label within the current function, whose address is specified by
3755 "``address``". Address must be derived from a
3756 :ref:`blockaddress <blockaddress>` constant.
3761 The '``address``' argument is the address of the label to jump to. The
3762 rest of the arguments indicate the full set of possible destinations
3763 that the address may point to. Blocks are allowed to occur multiple
3764 times in the destination list, though this isn't particularly useful.
3766 This destination list is required so that dataflow analysis has an
3767 accurate understanding of the CFG.
3772 Control transfers to the block specified in the address argument. All
3773 possible destination blocks must be listed in the label list, otherwise
3774 this instruction has undefined behavior. This implies that jumps to
3775 labels defined in other functions have undefined behavior as well.
3780 This is typically implemented with a jump through a register.
3785 .. code-block:: llvm
3787 indirectbr i8* %Addr, [ label %bb1, label %bb2, label %bb3 ]
3791 '``invoke``' Instruction
3792 ^^^^^^^^^^^^^^^^^^^^^^^^
3799 <result> = invoke [cconv] [ret attrs] <ptr to function ty> <function ptr val>(<function args>) [fn attrs]
3800 to label <normal label> unwind label <exception label>
3805 The '``invoke``' instruction causes control to transfer to a specified
3806 function, with the possibility of control flow transfer to either the
3807 '``normal``' label or the '``exception``' label. If the callee function
3808 returns with the "``ret``" instruction, control flow will return to the
3809 "normal" label. If the callee (or any indirect callees) returns via the
3810 ":ref:`resume <i_resume>`" instruction or other exception handling
3811 mechanism, control is interrupted and continued at the dynamically
3812 nearest "exception" label.
3814 The '``exception``' label is a `landing
3815 pad <ExceptionHandling.html#overview>`_ for the exception. As such,
3816 '``exception``' label is required to have the
3817 ":ref:`landingpad <i_landingpad>`" instruction, which contains the
3818 information about the behavior of the program after unwinding happens,
3819 as its first non-PHI instruction. The restrictions on the
3820 "``landingpad``" instruction's tightly couples it to the "``invoke``"
3821 instruction, so that the important information contained within the
3822 "``landingpad``" instruction can't be lost through normal code motion.
3827 This instruction requires several arguments:
3829 #. The optional "cconv" marker indicates which :ref:`calling
3830 convention <callingconv>` the call should use. If none is
3831 specified, the call defaults to using C calling conventions.
3832 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
3833 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
3835 #. '``ptr to function ty``': shall be the signature of the pointer to
3836 function value being invoked. In most cases, this is a direct
3837 function invocation, but indirect ``invoke``'s are just as possible,
3838 branching off an arbitrary pointer to function value.
3839 #. '``function ptr val``': An LLVM value containing a pointer to a
3840 function to be invoked.
3841 #. '``function args``': argument list whose types match the function
3842 signature argument types and parameter attributes. All arguments must
3843 be of :ref:`first class <t_firstclass>` type. If the function signature
3844 indicates the function accepts a variable number of arguments, the
3845 extra arguments can be specified.
3846 #. '``normal label``': the label reached when the called function
3847 executes a '``ret``' instruction.
3848 #. '``exception label``': the label reached when a callee returns via
3849 the :ref:`resume <i_resume>` instruction or other exception handling
3851 #. The optional :ref:`function attributes <fnattrs>` list. Only
3852 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
3853 attributes are valid here.
3858 This instruction is designed to operate as a standard '``call``'
3859 instruction in most regards. The primary difference is that it
3860 establishes an association with a label, which is used by the runtime
3861 library to unwind the stack.
3863 This instruction is used in languages with destructors to ensure that
3864 proper cleanup is performed in the case of either a ``longjmp`` or a
3865 thrown exception. Additionally, this is important for implementation of
3866 '``catch``' clauses in high-level languages that support them.
3868 For the purposes of the SSA form, the definition of the value returned
3869 by the '``invoke``' instruction is deemed to occur on the edge from the
3870 current block to the "normal" label. If the callee unwinds then no
3871 return value is available.
3876 .. code-block:: llvm
3878 %retval = invoke i32 @Test(i32 15) to label %Continue
3879 unwind label %TestCleanup ; i32:retval set
3880 %retval = invoke coldcc i32 %Testfnptr(i32 15) to label %Continue
3881 unwind label %TestCleanup ; i32:retval set
3885 '``resume``' Instruction
3886 ^^^^^^^^^^^^^^^^^^^^^^^^
3893 resume <type> <value>
3898 The '``resume``' instruction is a terminator instruction that has no
3904 The '``resume``' instruction requires one argument, which must have the
3905 same type as the result of any '``landingpad``' instruction in the same
3911 The '``resume``' instruction resumes propagation of an existing
3912 (in-flight) exception whose unwinding was interrupted with a
3913 :ref:`landingpad <i_landingpad>` instruction.
3918 .. code-block:: llvm
3920 resume { i8*, i32 } %exn
3924 '``unreachable``' Instruction
3925 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3937 The '``unreachable``' instruction has no defined semantics. This
3938 instruction is used to inform the optimizer that a particular portion of
3939 the code is not reachable. This can be used to indicate that the code
3940 after a no-return function cannot be reached, and other facts.
3945 The '``unreachable``' instruction has no defined semantics.
3952 Binary operators are used to do most of the computation in a program.
3953 They require two operands of the same type, execute an operation on
3954 them, and produce a single value. The operands might represent multiple
3955 data, as is the case with the :ref:`vector <t_vector>` data type. The
3956 result value has the same type as its operands.
3958 There are several different binary operators:
3962 '``add``' Instruction
3963 ^^^^^^^^^^^^^^^^^^^^^
3970 <result> = add <ty> <op1>, <op2> ; yields ty:result
3971 <result> = add nuw <ty> <op1>, <op2> ; yields ty:result
3972 <result> = add nsw <ty> <op1>, <op2> ; yields ty:result
3973 <result> = add nuw nsw <ty> <op1>, <op2> ; yields ty:result
3978 The '``add``' instruction returns the sum of its two operands.
3983 The two arguments to the '``add``' instruction must be
3984 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3985 arguments must have identical types.
3990 The value produced is the integer sum of the two operands.
3992 If the sum has unsigned overflow, the result returned is the
3993 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
3996 Because LLVM integers use a two's complement representation, this
3997 instruction is appropriate for both signed and unsigned integers.
3999 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
4000 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
4001 result value of the ``add`` is a :ref:`poison value <poisonvalues>` if
4002 unsigned and/or signed overflow, respectively, occurs.
4007 .. code-block:: llvm
4009 <result> = add i32 4, %var ; yields i32:result = 4 + %var
4013 '``fadd``' Instruction
4014 ^^^^^^^^^^^^^^^^^^^^^^
4021 <result> = fadd [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
4026 The '``fadd``' instruction returns the sum of its two operands.
4031 The two arguments to the '``fadd``' instruction must be :ref:`floating
4032 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4033 Both arguments must have identical types.
4038 The value produced is the floating point sum of the two operands. This
4039 instruction can also take any number of :ref:`fast-math flags <fastmath>`,
4040 which are optimization hints to enable otherwise unsafe floating point
4046 .. code-block:: llvm
4048 <result> = fadd float 4.0, %var ; yields float:result = 4.0 + %var
4050 '``sub``' Instruction
4051 ^^^^^^^^^^^^^^^^^^^^^
4058 <result> = sub <ty> <op1>, <op2> ; yields ty:result
4059 <result> = sub nuw <ty> <op1>, <op2> ; yields ty:result
4060 <result> = sub nsw <ty> <op1>, <op2> ; yields ty:result
4061 <result> = sub nuw nsw <ty> <op1>, <op2> ; yields ty:result
4066 The '``sub``' instruction returns the difference of its two operands.
4068 Note that the '``sub``' instruction is used to represent the '``neg``'
4069 instruction present in most other intermediate representations.
4074 The two arguments to the '``sub``' instruction must be
4075 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4076 arguments must have identical types.
4081 The value produced is the integer difference of the two operands.
4083 If the difference has unsigned overflow, the result returned is the
4084 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
4087 Because LLVM integers use a two's complement representation, this
4088 instruction is appropriate for both signed and unsigned integers.
4090 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
4091 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
4092 result value of the ``sub`` is a :ref:`poison value <poisonvalues>` if
4093 unsigned and/or signed overflow, respectively, occurs.
4098 .. code-block:: llvm
4100 <result> = sub i32 4, %var ; yields i32:result = 4 - %var
4101 <result> = sub i32 0, %val ; yields i32:result = -%var
4105 '``fsub``' Instruction
4106 ^^^^^^^^^^^^^^^^^^^^^^
4113 <result> = fsub [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
4118 The '``fsub``' instruction returns the difference of its two operands.
4120 Note that the '``fsub``' instruction is used to represent the '``fneg``'
4121 instruction present in most other intermediate representations.
4126 The two arguments to the '``fsub``' instruction must be :ref:`floating
4127 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4128 Both arguments must have identical types.
4133 The value produced is the floating point difference of the two operands.
4134 This instruction can also take any number of :ref:`fast-math
4135 flags <fastmath>`, which are optimization hints to enable otherwise
4136 unsafe floating point optimizations:
4141 .. code-block:: llvm
4143 <result> = fsub float 4.0, %var ; yields float:result = 4.0 - %var
4144 <result> = fsub float -0.0, %val ; yields float:result = -%var
4146 '``mul``' Instruction
4147 ^^^^^^^^^^^^^^^^^^^^^
4154 <result> = mul <ty> <op1>, <op2> ; yields ty:result
4155 <result> = mul nuw <ty> <op1>, <op2> ; yields ty:result
4156 <result> = mul nsw <ty> <op1>, <op2> ; yields ty:result
4157 <result> = mul nuw nsw <ty> <op1>, <op2> ; yields ty:result
4162 The '``mul``' instruction returns the product of its two operands.
4167 The two arguments to the '``mul``' instruction must be
4168 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4169 arguments must have identical types.
4174 The value produced is the integer product of the two operands.
4176 If the result of the multiplication has unsigned overflow, the result
4177 returned is the mathematical result modulo 2\ :sup:`n`\ , where n is the
4178 bit width of the result.
4180 Because LLVM integers use a two's complement representation, and the
4181 result is the same width as the operands, this instruction returns the
4182 correct result for both signed and unsigned integers. If a full product
4183 (e.g. ``i32`` * ``i32`` -> ``i64``) is needed, the operands should be
4184 sign-extended or zero-extended as appropriate to the width of the full
4187 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
4188 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
4189 result value of the ``mul`` is a :ref:`poison value <poisonvalues>` if
4190 unsigned and/or signed overflow, respectively, occurs.
4195 .. code-block:: llvm
4197 <result> = mul i32 4, %var ; yields i32:result = 4 * %var
4201 '``fmul``' Instruction
4202 ^^^^^^^^^^^^^^^^^^^^^^
4209 <result> = fmul [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
4214 The '``fmul``' instruction returns the product of its two operands.
4219 The two arguments to the '``fmul``' instruction must be :ref:`floating
4220 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4221 Both arguments must have identical types.
4226 The value produced is the floating point product of the two operands.
4227 This instruction can also take any number of :ref:`fast-math
4228 flags <fastmath>`, which are optimization hints to enable otherwise
4229 unsafe floating point optimizations:
4234 .. code-block:: llvm
4236 <result> = fmul float 4.0, %var ; yields float:result = 4.0 * %var
4238 '``udiv``' Instruction
4239 ^^^^^^^^^^^^^^^^^^^^^^
4246 <result> = udiv <ty> <op1>, <op2> ; yields ty:result
4247 <result> = udiv exact <ty> <op1>, <op2> ; yields ty:result
4252 The '``udiv``' instruction returns the quotient of its two operands.
4257 The two arguments to the '``udiv``' instruction must be
4258 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4259 arguments must have identical types.
4264 The value produced is the unsigned integer quotient of the two operands.
4266 Note that unsigned integer division and signed integer division are
4267 distinct operations; for signed integer division, use '``sdiv``'.
4269 Division by zero leads to undefined behavior.
4271 If the ``exact`` keyword is present, the result value of the ``udiv`` is
4272 a :ref:`poison value <poisonvalues>` if %op1 is not a multiple of %op2 (as
4273 such, "((a udiv exact b) mul b) == a").
4278 .. code-block:: llvm
4280 <result> = udiv i32 4, %var ; yields i32:result = 4 / %var
4282 '``sdiv``' Instruction
4283 ^^^^^^^^^^^^^^^^^^^^^^
4290 <result> = sdiv <ty> <op1>, <op2> ; yields ty:result
4291 <result> = sdiv exact <ty> <op1>, <op2> ; yields ty:result
4296 The '``sdiv``' instruction returns the quotient of its two operands.
4301 The two arguments to the '``sdiv``' instruction must be
4302 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4303 arguments must have identical types.
4308 The value produced is the signed integer quotient of the two operands
4309 rounded towards zero.
4311 Note that signed integer division and unsigned integer division are
4312 distinct operations; for unsigned integer division, use '``udiv``'.
4314 Division by zero leads to undefined behavior. Overflow also leads to
4315 undefined behavior; this is a rare case, but can occur, for example, by
4316 doing a 32-bit division of -2147483648 by -1.
4318 If the ``exact`` keyword is present, the result value of the ``sdiv`` is
4319 a :ref:`poison value <poisonvalues>` if the result would be rounded.
4324 .. code-block:: llvm
4326 <result> = sdiv i32 4, %var ; yields i32:result = 4 / %var
4330 '``fdiv``' Instruction
4331 ^^^^^^^^^^^^^^^^^^^^^^
4338 <result> = fdiv [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
4343 The '``fdiv``' instruction returns the quotient of its two operands.
4348 The two arguments to the '``fdiv``' instruction must be :ref:`floating
4349 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4350 Both arguments must have identical types.
4355 The value produced is the floating point quotient of the two operands.
4356 This instruction can also take any number of :ref:`fast-math
4357 flags <fastmath>`, which are optimization hints to enable otherwise
4358 unsafe floating point optimizations:
4363 .. code-block:: llvm
4365 <result> = fdiv float 4.0, %var ; yields float:result = 4.0 / %var
4367 '``urem``' Instruction
4368 ^^^^^^^^^^^^^^^^^^^^^^
4375 <result> = urem <ty> <op1>, <op2> ; yields ty:result
4380 The '``urem``' instruction returns the remainder from the unsigned
4381 division of its two arguments.
4386 The two arguments to the '``urem``' instruction must be
4387 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4388 arguments must have identical types.
4393 This instruction returns the unsigned integer *remainder* of a division.
4394 This instruction always performs an unsigned division to get the
4397 Note that unsigned integer remainder and signed integer remainder are
4398 distinct operations; for signed integer remainder, use '``srem``'.
4400 Taking the remainder of a division by zero leads to undefined behavior.
4405 .. code-block:: llvm
4407 <result> = urem i32 4, %var ; yields i32:result = 4 % %var
4409 '``srem``' Instruction
4410 ^^^^^^^^^^^^^^^^^^^^^^
4417 <result> = srem <ty> <op1>, <op2> ; yields ty:result
4422 The '``srem``' instruction returns the remainder from the signed
4423 division of its two operands. This instruction can also take
4424 :ref:`vector <t_vector>` versions of the values in which case the elements
4430 The two arguments to the '``srem``' instruction must be
4431 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4432 arguments must have identical types.
4437 This instruction returns the *remainder* of a division (where the result
4438 is either zero or has the same sign as the dividend, ``op1``), not the
4439 *modulo* operator (where the result is either zero or has the same sign
4440 as the divisor, ``op2``) of a value. For more information about the
4441 difference, see `The Math
4442 Forum <http://mathforum.org/dr.math/problems/anne.4.28.99.html>`_. For a
4443 table of how this is implemented in various languages, please see
4445 operation <http://en.wikipedia.org/wiki/Modulo_operation>`_.
4447 Note that signed integer remainder and unsigned integer remainder are
4448 distinct operations; for unsigned integer remainder, use '``urem``'.
4450 Taking the remainder of a division by zero leads to undefined behavior.
4451 Overflow also leads to undefined behavior; this is a rare case, but can
4452 occur, for example, by taking the remainder of a 32-bit division of
4453 -2147483648 by -1. (The remainder doesn't actually overflow, but this
4454 rule lets srem be implemented using instructions that return both the
4455 result of the division and the remainder.)
4460 .. code-block:: llvm
4462 <result> = srem i32 4, %var ; yields i32:result = 4 % %var
4466 '``frem``' Instruction
4467 ^^^^^^^^^^^^^^^^^^^^^^
4474 <result> = frem [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
4479 The '``frem``' instruction returns the remainder from the division of
4485 The two arguments to the '``frem``' instruction must be :ref:`floating
4486 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4487 Both arguments must have identical types.
4492 This instruction returns the *remainder* of a division. The remainder
4493 has the same sign as the dividend. This instruction can also take any
4494 number of :ref:`fast-math flags <fastmath>`, which are optimization hints
4495 to enable otherwise unsafe floating point optimizations:
4500 .. code-block:: llvm
4502 <result> = frem float 4.0, %var ; yields float:result = 4.0 % %var
4506 Bitwise Binary Operations
4507 -------------------------
4509 Bitwise binary operators are used to do various forms of bit-twiddling
4510 in a program. They are generally very efficient instructions and can
4511 commonly be strength reduced from other instructions. They require two
4512 operands of the same type, execute an operation on them, and produce a
4513 single value. The resulting value is the same type as its operands.
4515 '``shl``' Instruction
4516 ^^^^^^^^^^^^^^^^^^^^^
4523 <result> = shl <ty> <op1>, <op2> ; yields ty:result
4524 <result> = shl nuw <ty> <op1>, <op2> ; yields ty:result
4525 <result> = shl nsw <ty> <op1>, <op2> ; yields ty:result
4526 <result> = shl nuw nsw <ty> <op1>, <op2> ; yields ty:result
4531 The '``shl``' instruction returns the first operand shifted to the left
4532 a specified number of bits.
4537 Both arguments to the '``shl``' instruction must be the same
4538 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4539 '``op2``' is treated as an unsigned value.
4544 The value produced is ``op1`` \* 2\ :sup:`op2` mod 2\ :sup:`n`,
4545 where ``n`` is the width of the result. If ``op2`` is (statically or
4546 dynamically) negative or equal to or larger than the number of bits in
4547 ``op1``, the result is undefined. If the arguments are vectors, each
4548 vector element of ``op1`` is shifted by the corresponding shift amount
4551 If the ``nuw`` keyword is present, then the shift produces a :ref:`poison
4552 value <poisonvalues>` if it shifts out any non-zero bits. If the
4553 ``nsw`` keyword is present, then the shift produces a :ref:`poison
4554 value <poisonvalues>` if it shifts out any bits that disagree with the
4555 resultant sign bit. As such, NUW/NSW have the same semantics as they
4556 would if the shift were expressed as a mul instruction with the same
4557 nsw/nuw bits in (mul %op1, (shl 1, %op2)).
4562 .. code-block:: llvm
4564 <result> = shl i32 4, %var ; yields i32: 4 << %var
4565 <result> = shl i32 4, 2 ; yields i32: 16
4566 <result> = shl i32 1, 10 ; yields i32: 1024
4567 <result> = shl i32 1, 32 ; undefined
4568 <result> = shl <2 x i32> < i32 1, i32 1>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 2, i32 4>
4570 '``lshr``' Instruction
4571 ^^^^^^^^^^^^^^^^^^^^^^
4578 <result> = lshr <ty> <op1>, <op2> ; yields ty:result
4579 <result> = lshr exact <ty> <op1>, <op2> ; yields ty:result
4584 The '``lshr``' instruction (logical shift right) returns the first
4585 operand shifted to the right a specified number of bits with zero fill.
4590 Both arguments to the '``lshr``' instruction must be the same
4591 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4592 '``op2``' is treated as an unsigned value.
4597 This instruction always performs a logical shift right operation. The
4598 most significant bits of the result will be filled with zero bits after
4599 the shift. If ``op2`` is (statically or dynamically) equal to or larger
4600 than the number of bits in ``op1``, the result is undefined. If the
4601 arguments are vectors, each vector element of ``op1`` is shifted by the
4602 corresponding shift amount in ``op2``.
4604 If the ``exact`` keyword is present, the result value of the ``lshr`` is
4605 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
4611 .. code-block:: llvm
4613 <result> = lshr i32 4, 1 ; yields i32:result = 2
4614 <result> = lshr i32 4, 2 ; yields i32:result = 1
4615 <result> = lshr i8 4, 3 ; yields i8:result = 0
4616 <result> = lshr i8 -2, 1 ; yields i8:result = 0x7F
4617 <result> = lshr i32 1, 32 ; undefined
4618 <result> = lshr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 0x7FFFFFFF, i32 1>
4620 '``ashr``' Instruction
4621 ^^^^^^^^^^^^^^^^^^^^^^
4628 <result> = ashr <ty> <op1>, <op2> ; yields ty:result
4629 <result> = ashr exact <ty> <op1>, <op2> ; yields ty:result
4634 The '``ashr``' instruction (arithmetic shift right) returns the first
4635 operand shifted to the right a specified number of bits with sign
4641 Both arguments to the '``ashr``' instruction must be the same
4642 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4643 '``op2``' is treated as an unsigned value.
4648 This instruction always performs an arithmetic shift right operation,
4649 The most significant bits of the result will be filled with the sign bit
4650 of ``op1``. If ``op2`` is (statically or dynamically) equal to or larger
4651 than the number of bits in ``op1``, the result is undefined. If the
4652 arguments are vectors, each vector element of ``op1`` is shifted by the
4653 corresponding shift amount in ``op2``.
4655 If the ``exact`` keyword is present, the result value of the ``ashr`` is
4656 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
4662 .. code-block:: llvm
4664 <result> = ashr i32 4, 1 ; yields i32:result = 2
4665 <result> = ashr i32 4, 2 ; yields i32:result = 1
4666 <result> = ashr i8 4, 3 ; yields i8:result = 0
4667 <result> = ashr i8 -2, 1 ; yields i8:result = -1
4668 <result> = ashr i32 1, 32 ; undefined
4669 <result> = ashr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 3> ; yields: result=<2 x i32> < i32 -1, i32 0>
4671 '``and``' Instruction
4672 ^^^^^^^^^^^^^^^^^^^^^
4679 <result> = and <ty> <op1>, <op2> ; yields ty:result
4684 The '``and``' instruction returns the bitwise logical and of its two
4690 The two arguments to the '``and``' instruction must be
4691 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4692 arguments must have identical types.
4697 The truth table used for the '``and``' instruction is:
4714 .. code-block:: llvm
4716 <result> = and i32 4, %var ; yields i32:result = 4 & %var
4717 <result> = and i32 15, 40 ; yields i32:result = 8
4718 <result> = and i32 4, 8 ; yields i32:result = 0
4720 '``or``' Instruction
4721 ^^^^^^^^^^^^^^^^^^^^
4728 <result> = or <ty> <op1>, <op2> ; yields ty:result
4733 The '``or``' instruction returns the bitwise logical inclusive or of its
4739 The two arguments to the '``or``' instruction must be
4740 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4741 arguments must have identical types.
4746 The truth table used for the '``or``' instruction is:
4765 <result> = or i32 4, %var ; yields i32:result = 4 | %var
4766 <result> = or i32 15, 40 ; yields i32:result = 47
4767 <result> = or i32 4, 8 ; yields i32:result = 12
4769 '``xor``' Instruction
4770 ^^^^^^^^^^^^^^^^^^^^^
4777 <result> = xor <ty> <op1>, <op2> ; yields ty:result
4782 The '``xor``' instruction returns the bitwise logical exclusive or of
4783 its two operands. The ``xor`` is used to implement the "one's
4784 complement" operation, which is the "~" operator in C.
4789 The two arguments to the '``xor``' instruction must be
4790 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4791 arguments must have identical types.
4796 The truth table used for the '``xor``' instruction is:
4813 .. code-block:: llvm
4815 <result> = xor i32 4, %var ; yields i32:result = 4 ^ %var
4816 <result> = xor i32 15, 40 ; yields i32:result = 39
4817 <result> = xor i32 4, 8 ; yields i32:result = 12
4818 <result> = xor i32 %V, -1 ; yields i32:result = ~%V
4823 LLVM supports several instructions to represent vector operations in a
4824 target-independent manner. These instructions cover the element-access
4825 and vector-specific operations needed to process vectors effectively.
4826 While LLVM does directly support these vector operations, many
4827 sophisticated algorithms will want to use target-specific intrinsics to
4828 take full advantage of a specific target.
4830 .. _i_extractelement:
4832 '``extractelement``' Instruction
4833 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4840 <result> = extractelement <n x <ty>> <val>, <ty2> <idx> ; yields <ty>
4845 The '``extractelement``' instruction extracts a single scalar element
4846 from a vector at a specified index.
4851 The first operand of an '``extractelement``' instruction is a value of
4852 :ref:`vector <t_vector>` type. The second operand is an index indicating
4853 the position from which to extract the element. The index may be a
4854 variable of any integer type.
4859 The result is a scalar of the same type as the element type of ``val``.
4860 Its value is the value at position ``idx`` of ``val``. If ``idx``
4861 exceeds the length of ``val``, the results are undefined.
4866 .. code-block:: llvm
4868 <result> = extractelement <4 x i32> %vec, i32 0 ; yields i32
4870 .. _i_insertelement:
4872 '``insertelement``' Instruction
4873 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4880 <result> = insertelement <n x <ty>> <val>, <ty> <elt>, <ty2> <idx> ; yields <n x <ty>>
4885 The '``insertelement``' instruction inserts a scalar element into a
4886 vector at a specified index.
4891 The first operand of an '``insertelement``' instruction is a value of
4892 :ref:`vector <t_vector>` type. The second operand is a scalar value whose
4893 type must equal the element type of the first operand. The third operand
4894 is an index indicating the position at which to insert the value. The
4895 index may be a variable of any integer type.
4900 The result is a vector of the same type as ``val``. Its element values
4901 are those of ``val`` except at position ``idx``, where it gets the value
4902 ``elt``. If ``idx`` exceeds the length of ``val``, the results are
4908 .. code-block:: llvm
4910 <result> = insertelement <4 x i32> %vec, i32 1, i32 0 ; yields <4 x i32>
4912 .. _i_shufflevector:
4914 '``shufflevector``' Instruction
4915 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4922 <result> = shufflevector <n x <ty>> <v1>, <n x <ty>> <v2>, <m x i32> <mask> ; yields <m x <ty>>
4927 The '``shufflevector``' instruction constructs a permutation of elements
4928 from two input vectors, returning a vector with the same element type as
4929 the input and length that is the same as the shuffle mask.
4934 The first two operands of a '``shufflevector``' instruction are vectors
4935 with the same type. The third argument is a shuffle mask whose element
4936 type is always 'i32'. The result of the instruction is a vector whose
4937 length is the same as the shuffle mask and whose element type is the
4938 same as the element type of the first two operands.
4940 The shuffle mask operand is required to be a constant vector with either
4941 constant integer or undef values.
4946 The elements of the two input vectors are numbered from left to right
4947 across both of the vectors. The shuffle mask operand specifies, for each
4948 element of the result vector, which element of the two input vectors the
4949 result element gets. The element selector may be undef (meaning "don't
4950 care") and the second operand may be undef if performing a shuffle from
4956 .. code-block:: llvm
4958 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4959 <4 x i32> <i32 0, i32 4, i32 1, i32 5> ; yields <4 x i32>
4960 <result> = shufflevector <4 x i32> %v1, <4 x i32> undef,
4961 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32> - Identity shuffle.
4962 <result> = shufflevector <8 x i32> %v1, <8 x i32> undef,
4963 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32>
4964 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4965 <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7 > ; yields <8 x i32>
4967 Aggregate Operations
4968 --------------------
4970 LLVM supports several instructions for working with
4971 :ref:`aggregate <t_aggregate>` values.
4975 '``extractvalue``' Instruction
4976 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4983 <result> = extractvalue <aggregate type> <val>, <idx>{, <idx>}*
4988 The '``extractvalue``' instruction extracts the value of a member field
4989 from an :ref:`aggregate <t_aggregate>` value.
4994 The first operand of an '``extractvalue``' instruction is a value of
4995 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The operands are
4996 constant indices to specify which value to extract in a similar manner
4997 as indices in a '``getelementptr``' instruction.
4999 The major differences to ``getelementptr`` indexing are:
5001 - Since the value being indexed is not a pointer, the first index is
5002 omitted and assumed to be zero.
5003 - At least one index must be specified.
5004 - Not only struct indices but also array indices must be in bounds.
5009 The result is the value at the position in the aggregate specified by
5015 .. code-block:: llvm
5017 <result> = extractvalue {i32, float} %agg, 0 ; yields i32
5021 '``insertvalue``' Instruction
5022 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5029 <result> = insertvalue <aggregate type> <val>, <ty> <elt>, <idx>{, <idx>}* ; yields <aggregate type>
5034 The '``insertvalue``' instruction inserts a value into a member field in
5035 an :ref:`aggregate <t_aggregate>` value.
5040 The first operand of an '``insertvalue``' instruction is a value of
5041 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The second operand is
5042 a first-class value to insert. The following operands are constant
5043 indices indicating the position at which to insert the value in a
5044 similar manner as indices in a '``extractvalue``' instruction. The value
5045 to insert must have the same type as the value identified by the
5051 The result is an aggregate of the same type as ``val``. Its value is
5052 that of ``val`` except that the value at the position specified by the
5053 indices is that of ``elt``.
5058 .. code-block:: llvm
5060 %agg1 = insertvalue {i32, float} undef, i32 1, 0 ; yields {i32 1, float undef}
5061 %agg2 = insertvalue {i32, float} %agg1, float %val, 1 ; yields {i32 1, float %val}
5062 %agg3 = insertvalue {i32, {float}} %agg1, float %val, 1, 0 ; yields {i32 1, float %val}
5066 Memory Access and Addressing Operations
5067 ---------------------------------------
5069 A key design point of an SSA-based representation is how it represents
5070 memory. In LLVM, no memory locations are in SSA form, which makes things
5071 very simple. This section describes how to read, write, and allocate
5076 '``alloca``' Instruction
5077 ^^^^^^^^^^^^^^^^^^^^^^^^
5084 <result> = alloca [inalloca] <type> [, <ty> <NumElements>] [, align <alignment>] ; yields type*:result
5089 The '``alloca``' instruction allocates memory on the stack frame of the
5090 currently executing function, to be automatically released when this
5091 function returns to its caller. The object is always allocated in the
5092 generic address space (address space zero).
5097 The '``alloca``' instruction allocates ``sizeof(<type>)*NumElements``
5098 bytes of memory on the runtime stack, returning a pointer of the
5099 appropriate type to the program. If "NumElements" is specified, it is
5100 the number of elements allocated, otherwise "NumElements" is defaulted
5101 to be one. If a constant alignment is specified, the value result of the
5102 allocation is guaranteed to be aligned to at least that boundary. The
5103 alignment may not be greater than ``1 << 29``. If not specified, or if
5104 zero, the target can choose to align the allocation on any convenient
5105 boundary compatible with the type.
5107 '``type``' may be any sized type.
5112 Memory is allocated; a pointer is returned. The operation is undefined
5113 if there is insufficient stack space for the allocation. '``alloca``'d
5114 memory is automatically released when the function returns. The
5115 '``alloca``' instruction is commonly used to represent automatic
5116 variables that must have an address available. When the function returns
5117 (either with the ``ret`` or ``resume`` instructions), the memory is
5118 reclaimed. Allocating zero bytes is legal, but the result is undefined.
5119 The order in which memory is allocated (ie., which way the stack grows)
5125 .. code-block:: llvm
5127 %ptr = alloca i32 ; yields i32*:ptr
5128 %ptr = alloca i32, i32 4 ; yields i32*:ptr
5129 %ptr = alloca i32, i32 4, align 1024 ; yields i32*:ptr
5130 %ptr = alloca i32, align 1024 ; yields i32*:ptr
5134 '``load``' Instruction
5135 ^^^^^^^^^^^^^^^^^^^^^^
5142 <result> = load [volatile] <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>][, !invariant.load !<index>]
5143 <result> = load atomic [volatile] <ty>* <pointer> [singlethread] <ordering>, align <alignment>
5144 !<index> = !{ i32 1 }
5149 The '``load``' instruction is used to read from memory.
5154 The argument to the ``load`` instruction specifies the memory address
5155 from which to load. The pointer must point to a :ref:`first
5156 class <t_firstclass>` type. If the ``load`` is marked as ``volatile``,
5157 then the optimizer is not allowed to modify the number or order of
5158 execution of this ``load`` with other :ref:`volatile
5159 operations <volatile>`.
5161 If the ``load`` is marked as ``atomic``, it takes an extra
5162 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
5163 ``release`` and ``acq_rel`` orderings are not valid on ``load``
5164 instructions. Atomic loads produce :ref:`defined <memmodel>` results
5165 when they may see multiple atomic stores. The type of the pointee must
5166 be an integer type whose bit width is a power of two greater than or
5167 equal to eight and less than or equal to a target-specific size limit.
5168 ``align`` must be explicitly specified on atomic loads, and the load has
5169 undefined behavior if the alignment is not set to a value which is at
5170 least the size in bytes of the pointee. ``!nontemporal`` does not have
5171 any defined semantics for atomic loads.
5173 The optional constant ``align`` argument specifies the alignment of the
5174 operation (that is, the alignment of the memory address). A value of 0
5175 or an omitted ``align`` argument means that the operation has the ABI
5176 alignment for the target. It is the responsibility of the code emitter
5177 to ensure that the alignment information is correct. Overestimating the
5178 alignment results in undefined behavior. Underestimating the alignment
5179 may produce less efficient code. An alignment of 1 is always safe. The
5180 maximum possible alignment is ``1 << 29``.
5182 The optional ``!nontemporal`` metadata must reference a single
5183 metadata name ``<index>`` corresponding to a metadata node with one
5184 ``i32`` entry of value 1. The existence of the ``!nontemporal``
5185 metadata on the instruction tells the optimizer and code generator
5186 that this load is not expected to be reused in the cache. The code
5187 generator may select special instructions to save cache bandwidth, such
5188 as the ``MOVNT`` instruction on x86.
5190 The optional ``!invariant.load`` metadata must reference a single
5191 metadata name ``<index>`` corresponding to a metadata node with no
5192 entries. The existence of the ``!invariant.load`` metadata on the
5193 instruction tells the optimizer and code generator that this load
5194 address points to memory which does not change value during program
5195 execution. The optimizer may then move this load around, for example, by
5196 hoisting it out of loops using loop invariant code motion.
5201 The location of memory pointed to is loaded. If the value being loaded
5202 is of scalar type then the number of bytes read does not exceed the
5203 minimum number of bytes needed to hold all bits of the type. For
5204 example, loading an ``i24`` reads at most three bytes. When loading a
5205 value of a type like ``i20`` with a size that is not an integral number
5206 of bytes, the result is undefined if the value was not originally
5207 written using a store of the same type.
5212 .. code-block:: llvm
5214 %ptr = alloca i32 ; yields i32*:ptr
5215 store i32 3, i32* %ptr ; yields void
5216 %val = load i32* %ptr ; yields i32:val = i32 3
5220 '``store``' Instruction
5221 ^^^^^^^^^^^^^^^^^^^^^^^
5228 store [volatile] <ty> <value>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>] ; yields void
5229 store atomic [volatile] <ty> <value>, <ty>* <pointer> [singlethread] <ordering>, align <alignment> ; yields void
5234 The '``store``' instruction is used to write to memory.
5239 There are two arguments to the ``store`` instruction: a value to store
5240 and an address at which to store it. The type of the ``<pointer>``
5241 operand must be a pointer to the :ref:`first class <t_firstclass>` type of
5242 the ``<value>`` operand. If the ``store`` is marked as ``volatile``,
5243 then the optimizer is not allowed to modify the number or order of
5244 execution of this ``store`` with other :ref:`volatile
5245 operations <volatile>`.
5247 If the ``store`` is marked as ``atomic``, it takes an extra
5248 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
5249 ``acquire`` and ``acq_rel`` orderings aren't valid on ``store``
5250 instructions. Atomic loads produce :ref:`defined <memmodel>` results
5251 when they may see multiple atomic stores. The type of the pointee must
5252 be an integer type whose bit width is a power of two greater than or
5253 equal to eight and less than or equal to a target-specific size limit.
5254 ``align`` must be explicitly specified on atomic stores, and the store
5255 has undefined behavior if the alignment is not set to a value which is
5256 at least the size in bytes of the pointee. ``!nontemporal`` does not
5257 have any defined semantics for atomic stores.
5259 The optional constant ``align`` argument specifies the alignment of the
5260 operation (that is, the alignment of the memory address). A value of 0
5261 or an omitted ``align`` argument means that the operation has the ABI
5262 alignment for the target. It is the responsibility of the code emitter
5263 to ensure that the alignment information is correct. Overestimating the
5264 alignment results in undefined behavior. Underestimating the
5265 alignment may produce less efficient code. An alignment of 1 is always
5266 safe. The maximum possible alignment is ``1 << 29``.
5268 The optional ``!nontemporal`` metadata must reference a single metadata
5269 name ``<index>`` corresponding to a metadata node with one ``i32`` entry of
5270 value 1. The existence of the ``!nontemporal`` metadata on the instruction
5271 tells the optimizer and code generator that this load is not expected to
5272 be reused in the cache. The code generator may select special
5273 instructions to save cache bandwidth, such as the MOVNT instruction on
5279 The contents of memory are updated to contain ``<value>`` at the
5280 location specified by the ``<pointer>`` operand. If ``<value>`` is
5281 of scalar type then the number of bytes written does not exceed the
5282 minimum number of bytes needed to hold all bits of the type. For
5283 example, storing an ``i24`` writes at most three bytes. When writing a
5284 value of a type like ``i20`` with a size that is not an integral number
5285 of bytes, it is unspecified what happens to the extra bits that do not
5286 belong to the type, but they will typically be overwritten.
5291 .. code-block:: llvm
5293 %ptr = alloca i32 ; yields i32*:ptr
5294 store i32 3, i32* %ptr ; yields void
5295 %val = load i32* %ptr ; yields i32:val = i32 3
5299 '``fence``' Instruction
5300 ^^^^^^^^^^^^^^^^^^^^^^^
5307 fence [singlethread] <ordering> ; yields void
5312 The '``fence``' instruction is used to introduce happens-before edges
5318 '``fence``' instructions take an :ref:`ordering <ordering>` argument which
5319 defines what *synchronizes-with* edges they add. They can only be given
5320 ``acquire``, ``release``, ``acq_rel``, and ``seq_cst`` orderings.
5325 A fence A which has (at least) ``release`` ordering semantics
5326 *synchronizes with* a fence B with (at least) ``acquire`` ordering
5327 semantics if and only if there exist atomic operations X and Y, both
5328 operating on some atomic object M, such that A is sequenced before X, X
5329 modifies M (either directly or through some side effect of a sequence
5330 headed by X), Y is sequenced before B, and Y observes M. This provides a
5331 *happens-before* dependency between A and B. Rather than an explicit
5332 ``fence``, one (but not both) of the atomic operations X or Y might
5333 provide a ``release`` or ``acquire`` (resp.) ordering constraint and
5334 still *synchronize-with* the explicit ``fence`` and establish the
5335 *happens-before* edge.
5337 A ``fence`` which has ``seq_cst`` ordering, in addition to having both
5338 ``acquire`` and ``release`` semantics specified above, participates in
5339 the global program order of other ``seq_cst`` operations and/or fences.
5341 The optional ":ref:`singlethread <singlethread>`" argument specifies
5342 that the fence only synchronizes with other fences in the same thread.
5343 (This is useful for interacting with signal handlers.)
5348 .. code-block:: llvm
5350 fence acquire ; yields void
5351 fence singlethread seq_cst ; yields void
5355 '``cmpxchg``' Instruction
5356 ^^^^^^^^^^^^^^^^^^^^^^^^^
5363 cmpxchg [weak] [volatile] <ty>* <pointer>, <ty> <cmp>, <ty> <new> [singlethread] <success ordering> <failure ordering> ; yields { ty, i1 }
5368 The '``cmpxchg``' instruction is used to atomically modify memory. It
5369 loads a value in memory and compares it to a given value. If they are
5370 equal, it tries to store a new value into the memory.
5375 There are three arguments to the '``cmpxchg``' instruction: an address
5376 to operate on, a value to compare to the value currently be at that
5377 address, and a new value to place at that address if the compared values
5378 are equal. The type of '<cmp>' must be an integer type whose bit width
5379 is a power of two greater than or equal to eight and less than or equal
5380 to a target-specific size limit. '<cmp>' and '<new>' must have the same
5381 type, and the type of '<pointer>' must be a pointer to that type. If the
5382 ``cmpxchg`` is marked as ``volatile``, then the optimizer is not allowed
5383 to modify the number or order of execution of this ``cmpxchg`` with
5384 other :ref:`volatile operations <volatile>`.
5386 The success and failure :ref:`ordering <ordering>` arguments specify how this
5387 ``cmpxchg`` synchronizes with other atomic operations. Both ordering parameters
5388 must be at least ``monotonic``, the ordering constraint on failure must be no
5389 stronger than that on success, and the failure ordering cannot be either
5390 ``release`` or ``acq_rel``.
5392 The optional "``singlethread``" argument declares that the ``cmpxchg``
5393 is only atomic with respect to code (usually signal handlers) running in
5394 the same thread as the ``cmpxchg``. Otherwise the cmpxchg is atomic with
5395 respect to all other code in the system.
5397 The pointer passed into cmpxchg must have alignment greater than or
5398 equal to the size in memory of the operand.
5403 The contents of memory at the location specified by the '``<pointer>``' operand
5404 is read and compared to '``<cmp>``'; if the read value is the equal, the
5405 '``<new>``' is written. The original value at the location is returned, together
5406 with a flag indicating success (true) or failure (false).
5408 If the cmpxchg operation is marked as ``weak`` then a spurious failure is
5409 permitted: the operation may not write ``<new>`` even if the comparison
5412 If the cmpxchg operation is strong (the default), the i1 value is 1 if and only
5413 if the value loaded equals ``cmp``.
5415 A successful ``cmpxchg`` is a read-modify-write instruction for the purpose of
5416 identifying release sequences. A failed ``cmpxchg`` is equivalent to an atomic
5417 load with an ordering parameter determined the second ordering parameter.
5422 .. code-block:: llvm
5425 %orig = atomic load i32* %ptr unordered ; yields i32
5429 %cmp = phi i32 [ %orig, %entry ], [%old, %loop]
5430 %squared = mul i32 %cmp, %cmp
5431 %val_success = cmpxchg i32* %ptr, i32 %cmp, i32 %squared acq_rel monotonic ; yields { i32, i1 }
5432 %value_loaded = extractvalue { i32, i1 } %val_success, 0
5433 %success = extractvalue { i32, i1 } %val_success, 1
5434 br i1 %success, label %done, label %loop
5441 '``atomicrmw``' Instruction
5442 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
5449 atomicrmw [volatile] <operation> <ty>* <pointer>, <ty> <value> [singlethread] <ordering> ; yields ty
5454 The '``atomicrmw``' instruction is used to atomically modify memory.
5459 There are three arguments to the '``atomicrmw``' instruction: an
5460 operation to apply, an address whose value to modify, an argument to the
5461 operation. The operation must be one of the following keywords:
5475 The type of '<value>' must be an integer type whose bit width is a power
5476 of two greater than or equal to eight and less than or equal to a
5477 target-specific size limit. The type of the '``<pointer>``' operand must
5478 be a pointer to that type. If the ``atomicrmw`` is marked as
5479 ``volatile``, then the optimizer is not allowed to modify the number or
5480 order of execution of this ``atomicrmw`` with other :ref:`volatile
5481 operations <volatile>`.
5486 The contents of memory at the location specified by the '``<pointer>``'
5487 operand are atomically read, modified, and written back. The original
5488 value at the location is returned. The modification is specified by the
5491 - xchg: ``*ptr = val``
5492 - add: ``*ptr = *ptr + val``
5493 - sub: ``*ptr = *ptr - val``
5494 - and: ``*ptr = *ptr & val``
5495 - nand: ``*ptr = ~(*ptr & val)``
5496 - or: ``*ptr = *ptr | val``
5497 - xor: ``*ptr = *ptr ^ val``
5498 - max: ``*ptr = *ptr > val ? *ptr : val`` (using a signed comparison)
5499 - min: ``*ptr = *ptr < val ? *ptr : val`` (using a signed comparison)
5500 - umax: ``*ptr = *ptr > val ? *ptr : val`` (using an unsigned
5502 - umin: ``*ptr = *ptr < val ? *ptr : val`` (using an unsigned
5508 .. code-block:: llvm
5510 %old = atomicrmw add i32* %ptr, i32 1 acquire ; yields i32
5512 .. _i_getelementptr:
5514 '``getelementptr``' Instruction
5515 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5522 <result> = getelementptr <pty>* <ptrval>{, <ty> <idx>}*
5523 <result> = getelementptr inbounds <pty>* <ptrval>{, <ty> <idx>}*
5524 <result> = getelementptr <ptr vector> ptrval, <vector index type> idx
5529 The '``getelementptr``' instruction is used to get the address of a
5530 subelement of an :ref:`aggregate <t_aggregate>` data structure. It performs
5531 address calculation only and does not access memory.
5536 The first argument is always a pointer or a vector of pointers, and
5537 forms the basis of the calculation. The remaining arguments are indices
5538 that indicate which of the elements of the aggregate object are indexed.
5539 The interpretation of each index is dependent on the type being indexed
5540 into. The first index always indexes the pointer value given as the
5541 first argument, the second index indexes a value of the type pointed to
5542 (not necessarily the value directly pointed to, since the first index
5543 can be non-zero), etc. The first type indexed into must be a pointer
5544 value, subsequent types can be arrays, vectors, and structs. Note that
5545 subsequent types being indexed into can never be pointers, since that
5546 would require loading the pointer before continuing calculation.
5548 The type of each index argument depends on the type it is indexing into.
5549 When indexing into a (optionally packed) structure, only ``i32`` integer
5550 **constants** are allowed (when using a vector of indices they must all
5551 be the **same** ``i32`` integer constant). When indexing into an array,
5552 pointer or vector, integers of any width are allowed, and they are not
5553 required to be constant. These integers are treated as signed values
5556 For example, let's consider a C code fragment and how it gets compiled
5572 int *foo(struct ST *s) {
5573 return &s[1].Z.B[5][13];
5576 The LLVM code generated by Clang is:
5578 .. code-block:: llvm
5580 %struct.RT = type { i8, [10 x [20 x i32]], i8 }
5581 %struct.ST = type { i32, double, %struct.RT }
5583 define i32* @foo(%struct.ST* %s) nounwind uwtable readnone optsize ssp {
5585 %arrayidx = getelementptr inbounds %struct.ST* %s, i64 1, i32 2, i32 1, i64 5, i64 13
5592 In the example above, the first index is indexing into the
5593 '``%struct.ST*``' type, which is a pointer, yielding a '``%struct.ST``'
5594 = '``{ i32, double, %struct.RT }``' type, a structure. The second index
5595 indexes into the third element of the structure, yielding a
5596 '``%struct.RT``' = '``{ i8 , [10 x [20 x i32]], i8 }``' type, another
5597 structure. The third index indexes into the second element of the
5598 structure, yielding a '``[10 x [20 x i32]]``' type, an array. The two
5599 dimensions of the array are subscripted into, yielding an '``i32``'
5600 type. The '``getelementptr``' instruction returns a pointer to this
5601 element, thus computing a value of '``i32*``' type.
5603 Note that it is perfectly legal to index partially through a structure,
5604 returning a pointer to an inner element. Because of this, the LLVM code
5605 for the given testcase is equivalent to:
5607 .. code-block:: llvm
5609 define i32* @foo(%struct.ST* %s) {
5610 %t1 = getelementptr %struct.ST* %s, i32 1 ; yields %struct.ST*:%t1
5611 %t2 = getelementptr %struct.ST* %t1, i32 0, i32 2 ; yields %struct.RT*:%t2
5612 %t3 = getelementptr %struct.RT* %t2, i32 0, i32 1 ; yields [10 x [20 x i32]]*:%t3
5613 %t4 = getelementptr [10 x [20 x i32]]* %t3, i32 0, i32 5 ; yields [20 x i32]*:%t4
5614 %t5 = getelementptr [20 x i32]* %t4, i32 0, i32 13 ; yields i32*:%t5
5618 If the ``inbounds`` keyword is present, the result value of the
5619 ``getelementptr`` is a :ref:`poison value <poisonvalues>` if the base
5620 pointer is not an *in bounds* address of an allocated object, or if any
5621 of the addresses that would be formed by successive addition of the
5622 offsets implied by the indices to the base address with infinitely
5623 precise signed arithmetic are not an *in bounds* address of that
5624 allocated object. The *in bounds* addresses for an allocated object are
5625 all the addresses that point into the object, plus the address one byte
5626 past the end. In cases where the base is a vector of pointers the
5627 ``inbounds`` keyword applies to each of the computations element-wise.
5629 If the ``inbounds`` keyword is not present, the offsets are added to the
5630 base address with silently-wrapping two's complement arithmetic. If the
5631 offsets have a different width from the pointer, they are sign-extended
5632 or truncated to the width of the pointer. The result value of the
5633 ``getelementptr`` may be outside the object pointed to by the base
5634 pointer. The result value may not necessarily be used to access memory
5635 though, even if it happens to point into allocated storage. See the
5636 :ref:`Pointer Aliasing Rules <pointeraliasing>` section for more
5639 The getelementptr instruction is often confusing. For some more insight
5640 into how it works, see :doc:`the getelementptr FAQ <GetElementPtr>`.
5645 .. code-block:: llvm
5647 ; yields [12 x i8]*:aptr
5648 %aptr = getelementptr {i32, [12 x i8]}* %saptr, i64 0, i32 1
5650 %vptr = getelementptr {i32, <2 x i8>}* %svptr, i64 0, i32 1, i32 1
5652 %eptr = getelementptr [12 x i8]* %aptr, i64 0, i32 1
5654 %iptr = getelementptr [10 x i32]* @arr, i16 0, i16 0
5656 In cases where the pointer argument is a vector of pointers, each index
5657 must be a vector with the same number of elements. For example:
5659 .. code-block:: llvm
5661 %A = getelementptr <4 x i8*> %ptrs, <4 x i64> %offsets,
5663 Conversion Operations
5664 ---------------------
5666 The instructions in this category are the conversion instructions
5667 (casting) which all take a single operand and a type. They perform
5668 various bit conversions on the operand.
5670 '``trunc .. to``' Instruction
5671 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5678 <result> = trunc <ty> <value> to <ty2> ; yields ty2
5683 The '``trunc``' instruction truncates its operand to the type ``ty2``.
5688 The '``trunc``' instruction takes a value to trunc, and a type to trunc
5689 it to. Both types must be of :ref:`integer <t_integer>` types, or vectors
5690 of the same number of integers. The bit size of the ``value`` must be
5691 larger than the bit size of the destination type, ``ty2``. Equal sized
5692 types are not allowed.
5697 The '``trunc``' instruction truncates the high order bits in ``value``
5698 and converts the remaining bits to ``ty2``. Since the source size must
5699 be larger than the destination size, ``trunc`` cannot be a *no-op cast*.
5700 It will always truncate bits.
5705 .. code-block:: llvm
5707 %X = trunc i32 257 to i8 ; yields i8:1
5708 %Y = trunc i32 123 to i1 ; yields i1:true
5709 %Z = trunc i32 122 to i1 ; yields i1:false
5710 %W = trunc <2 x i16> <i16 8, i16 7> to <2 x i8> ; yields <i8 8, i8 7>
5712 '``zext .. to``' Instruction
5713 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5720 <result> = zext <ty> <value> to <ty2> ; yields ty2
5725 The '``zext``' instruction zero extends its operand to type ``ty2``.
5730 The '``zext``' instruction takes a value to cast, and a type to cast it
5731 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
5732 the same number of integers. The bit size of the ``value`` must be
5733 smaller than the bit size of the destination type, ``ty2``.
5738 The ``zext`` fills the high order bits of the ``value`` with zero bits
5739 until it reaches the size of the destination type, ``ty2``.
5741 When zero extending from i1, the result will always be either 0 or 1.
5746 .. code-block:: llvm
5748 %X = zext i32 257 to i64 ; yields i64:257
5749 %Y = zext i1 true to i32 ; yields i32:1
5750 %Z = zext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5752 '``sext .. to``' Instruction
5753 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5760 <result> = sext <ty> <value> to <ty2> ; yields ty2
5765 The '``sext``' sign extends ``value`` to the type ``ty2``.
5770 The '``sext``' instruction takes a value to cast, and a type to cast it
5771 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
5772 the same number of integers. The bit size of the ``value`` must be
5773 smaller than the bit size of the destination type, ``ty2``.
5778 The '``sext``' instruction performs a sign extension by copying the sign
5779 bit (highest order bit) of the ``value`` until it reaches the bit size
5780 of the type ``ty2``.
5782 When sign extending from i1, the extension always results in -1 or 0.
5787 .. code-block:: llvm
5789 %X = sext i8 -1 to i16 ; yields i16 :65535
5790 %Y = sext i1 true to i32 ; yields i32:-1
5791 %Z = sext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5793 '``fptrunc .. to``' Instruction
5794 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5801 <result> = fptrunc <ty> <value> to <ty2> ; yields ty2
5806 The '``fptrunc``' instruction truncates ``value`` to type ``ty2``.
5811 The '``fptrunc``' instruction takes a :ref:`floating point <t_floating>`
5812 value to cast and a :ref:`floating point <t_floating>` type to cast it to.
5813 The size of ``value`` must be larger than the size of ``ty2``. This
5814 implies that ``fptrunc`` cannot be used to make a *no-op cast*.
5819 The '``fptrunc``' instruction truncates a ``value`` from a larger
5820 :ref:`floating point <t_floating>` type to a smaller :ref:`floating
5821 point <t_floating>` type. If the value cannot fit within the
5822 destination type, ``ty2``, then the results are undefined.
5827 .. code-block:: llvm
5829 %X = fptrunc double 123.0 to float ; yields float:123.0
5830 %Y = fptrunc double 1.0E+300 to float ; yields undefined
5832 '``fpext .. to``' Instruction
5833 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5840 <result> = fpext <ty> <value> to <ty2> ; yields ty2
5845 The '``fpext``' extends a floating point ``value`` to a larger floating
5851 The '``fpext``' instruction takes a :ref:`floating point <t_floating>`
5852 ``value`` to cast, and a :ref:`floating point <t_floating>` type to cast it
5853 to. The source type must be smaller than the destination type.
5858 The '``fpext``' instruction extends the ``value`` from a smaller
5859 :ref:`floating point <t_floating>` type to a larger :ref:`floating
5860 point <t_floating>` type. The ``fpext`` cannot be used to make a
5861 *no-op cast* because it always changes bits. Use ``bitcast`` to make a
5862 *no-op cast* for a floating point cast.
5867 .. code-block:: llvm
5869 %X = fpext float 3.125 to double ; yields double:3.125000e+00
5870 %Y = fpext double %X to fp128 ; yields fp128:0xL00000000000000004000900000000000
5872 '``fptoui .. to``' Instruction
5873 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5880 <result> = fptoui <ty> <value> to <ty2> ; yields ty2
5885 The '``fptoui``' converts a floating point ``value`` to its unsigned
5886 integer equivalent of type ``ty2``.
5891 The '``fptoui``' instruction takes a value to cast, which must be a
5892 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5893 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5894 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5895 type with the same number of elements as ``ty``
5900 The '``fptoui``' instruction converts its :ref:`floating
5901 point <t_floating>` operand into the nearest (rounding towards zero)
5902 unsigned integer value. If the value cannot fit in ``ty2``, the results
5908 .. code-block:: llvm
5910 %X = fptoui double 123.0 to i32 ; yields i32:123
5911 %Y = fptoui float 1.0E+300 to i1 ; yields undefined:1
5912 %Z = fptoui float 1.04E+17 to i8 ; yields undefined:1
5914 '``fptosi .. to``' Instruction
5915 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5922 <result> = fptosi <ty> <value> to <ty2> ; yields ty2
5927 The '``fptosi``' instruction converts :ref:`floating point <t_floating>`
5928 ``value`` to type ``ty2``.
5933 The '``fptosi``' instruction takes a value to cast, which must be a
5934 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5935 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5936 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5937 type with the same number of elements as ``ty``
5942 The '``fptosi``' instruction converts its :ref:`floating
5943 point <t_floating>` operand into the nearest (rounding towards zero)
5944 signed integer value. If the value cannot fit in ``ty2``, the results
5950 .. code-block:: llvm
5952 %X = fptosi double -123.0 to i32 ; yields i32:-123
5953 %Y = fptosi float 1.0E-247 to i1 ; yields undefined:1
5954 %Z = fptosi float 1.04E+17 to i8 ; yields undefined:1
5956 '``uitofp .. to``' Instruction
5957 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5964 <result> = uitofp <ty> <value> to <ty2> ; yields ty2
5969 The '``uitofp``' instruction regards ``value`` as an unsigned integer
5970 and converts that value to the ``ty2`` type.
5975 The '``uitofp``' instruction takes a value to cast, which must be a
5976 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
5977 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
5978 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
5979 type with the same number of elements as ``ty``
5984 The '``uitofp``' instruction interprets its operand as an unsigned
5985 integer quantity and converts it to the corresponding floating point
5986 value. If the value cannot fit in the floating point value, the results
5992 .. code-block:: llvm
5994 %X = uitofp i32 257 to float ; yields float:257.0
5995 %Y = uitofp i8 -1 to double ; yields double:255.0
5997 '``sitofp .. to``' Instruction
5998 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6005 <result> = sitofp <ty> <value> to <ty2> ; yields ty2
6010 The '``sitofp``' instruction regards ``value`` as a signed integer and
6011 converts that value to the ``ty2`` type.
6016 The '``sitofp``' instruction takes a value to cast, which must be a
6017 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
6018 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
6019 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
6020 type with the same number of elements as ``ty``
6025 The '``sitofp``' instruction interprets its operand as a signed integer
6026 quantity and converts it to the corresponding floating point value. If
6027 the value cannot fit in the floating point value, the results are
6033 .. code-block:: llvm
6035 %X = sitofp i32 257 to float ; yields float:257.0
6036 %Y = sitofp i8 -1 to double ; yields double:-1.0
6040 '``ptrtoint .. to``' Instruction
6041 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6048 <result> = ptrtoint <ty> <value> to <ty2> ; yields ty2
6053 The '``ptrtoint``' instruction converts the pointer or a vector of
6054 pointers ``value`` to the integer (or vector of integers) type ``ty2``.
6059 The '``ptrtoint``' instruction takes a ``value`` to cast, which must be
6060 a a value of type :ref:`pointer <t_pointer>` or a vector of pointers, and a
6061 type to cast it to ``ty2``, which must be an :ref:`integer <t_integer>` or
6062 a vector of integers type.
6067 The '``ptrtoint``' instruction converts ``value`` to integer type
6068 ``ty2`` by interpreting the pointer value as an integer and either
6069 truncating or zero extending that value to the size of the integer type.
6070 If ``value`` is smaller than ``ty2`` then a zero extension is done. If
6071 ``value`` is larger than ``ty2`` then a truncation is done. If they are
6072 the same size, then nothing is done (*no-op cast*) other than a type
6078 .. code-block:: llvm
6080 %X = ptrtoint i32* %P to i8 ; yields truncation on 32-bit architecture
6081 %Y = ptrtoint i32* %P to i64 ; yields zero extension on 32-bit architecture
6082 %Z = ptrtoint <4 x i32*> %P to <4 x i64>; yields vector zero extension for a vector of addresses on 32-bit architecture
6086 '``inttoptr .. to``' Instruction
6087 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6094 <result> = inttoptr <ty> <value> to <ty2> ; yields ty2
6099 The '``inttoptr``' instruction converts an integer ``value`` to a
6100 pointer type, ``ty2``.
6105 The '``inttoptr``' instruction takes an :ref:`integer <t_integer>` value to
6106 cast, and a type to cast it to, which must be a :ref:`pointer <t_pointer>`
6112 The '``inttoptr``' instruction converts ``value`` to type ``ty2`` by
6113 applying either a zero extension or a truncation depending on the size
6114 of the integer ``value``. If ``value`` is larger than the size of a
6115 pointer then a truncation is done. If ``value`` is smaller than the size
6116 of a pointer then a zero extension is done. If they are the same size,
6117 nothing is done (*no-op cast*).
6122 .. code-block:: llvm
6124 %X = inttoptr i32 255 to i32* ; yields zero extension on 64-bit architecture
6125 %Y = inttoptr i32 255 to i32* ; yields no-op on 32-bit architecture
6126 %Z = inttoptr i64 0 to i32* ; yields truncation on 32-bit architecture
6127 %Z = inttoptr <4 x i32> %G to <4 x i8*>; yields truncation of vector G to four pointers
6131 '``bitcast .. to``' Instruction
6132 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6139 <result> = bitcast <ty> <value> to <ty2> ; yields ty2
6144 The '``bitcast``' instruction converts ``value`` to type ``ty2`` without
6150 The '``bitcast``' instruction takes a value to cast, which must be a
6151 non-aggregate first class value, and a type to cast it to, which must
6152 also be a non-aggregate :ref:`first class <t_firstclass>` type. The
6153 bit sizes of ``value`` and the destination type, ``ty2``, must be
6154 identical. If the source type is a pointer, the destination type must
6155 also be a pointer of the same size. This instruction supports bitwise
6156 conversion of vectors to integers and to vectors of other types (as
6157 long as they have the same size).
6162 The '``bitcast``' instruction converts ``value`` to type ``ty2``. It
6163 is always a *no-op cast* because no bits change with this
6164 conversion. The conversion is done as if the ``value`` had been stored
6165 to memory and read back as type ``ty2``. Pointer (or vector of
6166 pointers) types may only be converted to other pointer (or vector of
6167 pointers) types with the same address space through this instruction.
6168 To convert pointers to other types, use the :ref:`inttoptr <i_inttoptr>`
6169 or :ref:`ptrtoint <i_ptrtoint>` instructions first.
6174 .. code-block:: llvm
6176 %X = bitcast i8 255 to i8 ; yields i8 :-1
6177 %Y = bitcast i32* %x to sint* ; yields sint*:%x
6178 %Z = bitcast <2 x int> %V to i64; ; yields i64: %V
6179 %Z = bitcast <2 x i32*> %V to <2 x i64*> ; yields <2 x i64*>
6181 .. _i_addrspacecast:
6183 '``addrspacecast .. to``' Instruction
6184 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6191 <result> = addrspacecast <pty> <ptrval> to <pty2> ; yields pty2
6196 The '``addrspacecast``' instruction converts ``ptrval`` from ``pty`` in
6197 address space ``n`` to type ``pty2`` in address space ``m``.
6202 The '``addrspacecast``' instruction takes a pointer or vector of pointer value
6203 to cast and a pointer type to cast it to, which must have a different
6209 The '``addrspacecast``' instruction converts the pointer value
6210 ``ptrval`` to type ``pty2``. It can be a *no-op cast* or a complex
6211 value modification, depending on the target and the address space
6212 pair. Pointer conversions within the same address space must be
6213 performed with the ``bitcast`` instruction. Note that if the address space
6214 conversion is legal then both result and operand refer to the same memory
6220 .. code-block:: llvm
6222 %X = addrspacecast i32* %x to i32 addrspace(1)* ; yields i32 addrspace(1)*:%x
6223 %Y = addrspacecast i32 addrspace(1)* %y to i64 addrspace(2)* ; yields i64 addrspace(2)*:%y
6224 %Z = addrspacecast <4 x i32*> %z to <4 x float addrspace(3)*> ; yields <4 x float addrspace(3)*>:%z
6231 The instructions in this category are the "miscellaneous" instructions,
6232 which defy better classification.
6236 '``icmp``' Instruction
6237 ^^^^^^^^^^^^^^^^^^^^^^
6244 <result> = icmp <cond> <ty> <op1>, <op2> ; yields i1 or <N x i1>:result
6249 The '``icmp``' instruction returns a boolean value or a vector of
6250 boolean values based on comparison of its two integer, integer vector,
6251 pointer, or pointer vector operands.
6256 The '``icmp``' instruction takes three operands. The first operand is
6257 the condition code indicating the kind of comparison to perform. It is
6258 not a value, just a keyword. The possible condition code are:
6261 #. ``ne``: not equal
6262 #. ``ugt``: unsigned greater than
6263 #. ``uge``: unsigned greater or equal
6264 #. ``ult``: unsigned less than
6265 #. ``ule``: unsigned less or equal
6266 #. ``sgt``: signed greater than
6267 #. ``sge``: signed greater or equal
6268 #. ``slt``: signed less than
6269 #. ``sle``: signed less or equal
6271 The remaining two arguments must be :ref:`integer <t_integer>` or
6272 :ref:`pointer <t_pointer>` or integer :ref:`vector <t_vector>` typed. They
6273 must also be identical types.
6278 The '``icmp``' compares ``op1`` and ``op2`` according to the condition
6279 code given as ``cond``. The comparison performed always yields either an
6280 :ref:`i1 <t_integer>` or vector of ``i1`` result, as follows:
6282 #. ``eq``: yields ``true`` if the operands are equal, ``false``
6283 otherwise. No sign interpretation is necessary or performed.
6284 #. ``ne``: yields ``true`` if the operands are unequal, ``false``
6285 otherwise. No sign interpretation is necessary or performed.
6286 #. ``ugt``: interprets the operands as unsigned values and yields
6287 ``true`` if ``op1`` is greater than ``op2``.
6288 #. ``uge``: interprets the operands as unsigned values and yields
6289 ``true`` if ``op1`` is greater than or equal to ``op2``.
6290 #. ``ult``: interprets the operands as unsigned values and yields
6291 ``true`` if ``op1`` is less than ``op2``.
6292 #. ``ule``: interprets the operands as unsigned values and yields
6293 ``true`` if ``op1`` is less than or equal to ``op2``.
6294 #. ``sgt``: interprets the operands as signed values and yields ``true``
6295 if ``op1`` is greater than ``op2``.
6296 #. ``sge``: interprets the operands as signed values and yields ``true``
6297 if ``op1`` is greater than or equal to ``op2``.
6298 #. ``slt``: interprets the operands as signed values and yields ``true``
6299 if ``op1`` is less than ``op2``.
6300 #. ``sle``: interprets the operands as signed values and yields ``true``
6301 if ``op1`` is less than or equal to ``op2``.
6303 If the operands are :ref:`pointer <t_pointer>` typed, the pointer values
6304 are compared as if they were integers.
6306 If the operands are integer vectors, then they are compared element by
6307 element. The result is an ``i1`` vector with the same number of elements
6308 as the values being compared. Otherwise, the result is an ``i1``.
6313 .. code-block:: llvm
6315 <result> = icmp eq i32 4, 5 ; yields: result=false
6316 <result> = icmp ne float* %X, %X ; yields: result=false
6317 <result> = icmp ult i16 4, 5 ; yields: result=true
6318 <result> = icmp sgt i16 4, 5 ; yields: result=false
6319 <result> = icmp ule i16 -4, 5 ; yields: result=false
6320 <result> = icmp sge i16 4, 5 ; yields: result=false
6322 Note that the code generator does not yet support vector types with the
6323 ``icmp`` instruction.
6327 '``fcmp``' Instruction
6328 ^^^^^^^^^^^^^^^^^^^^^^
6335 <result> = fcmp <cond> <ty> <op1>, <op2> ; yields i1 or <N x i1>:result
6340 The '``fcmp``' instruction returns a boolean value or vector of boolean
6341 values based on comparison of its operands.
6343 If the operands are floating point scalars, then the result type is a
6344 boolean (:ref:`i1 <t_integer>`).
6346 If the operands are floating point vectors, then the result type is a
6347 vector of boolean with the same number of elements as the operands being
6353 The '``fcmp``' instruction takes three operands. The first operand is
6354 the condition code indicating the kind of comparison to perform. It is
6355 not a value, just a keyword. The possible condition code are:
6357 #. ``false``: no comparison, always returns false
6358 #. ``oeq``: ordered and equal
6359 #. ``ogt``: ordered and greater than
6360 #. ``oge``: ordered and greater than or equal
6361 #. ``olt``: ordered and less than
6362 #. ``ole``: ordered and less than or equal
6363 #. ``one``: ordered and not equal
6364 #. ``ord``: ordered (no nans)
6365 #. ``ueq``: unordered or equal
6366 #. ``ugt``: unordered or greater than
6367 #. ``uge``: unordered or greater than or equal
6368 #. ``ult``: unordered or less than
6369 #. ``ule``: unordered or less than or equal
6370 #. ``une``: unordered or not equal
6371 #. ``uno``: unordered (either nans)
6372 #. ``true``: no comparison, always returns true
6374 *Ordered* means that neither operand is a QNAN while *unordered* means
6375 that either operand may be a QNAN.
6377 Each of ``val1`` and ``val2`` arguments must be either a :ref:`floating
6378 point <t_floating>` type or a :ref:`vector <t_vector>` of floating point
6379 type. They must have identical types.
6384 The '``fcmp``' instruction compares ``op1`` and ``op2`` according to the
6385 condition code given as ``cond``. If the operands are vectors, then the
6386 vectors are compared element by element. Each comparison performed
6387 always yields an :ref:`i1 <t_integer>` result, as follows:
6389 #. ``false``: always yields ``false``, regardless of operands.
6390 #. ``oeq``: yields ``true`` if both operands are not a QNAN and ``op1``
6391 is equal to ``op2``.
6392 #. ``ogt``: yields ``true`` if both operands are not a QNAN and ``op1``
6393 is greater than ``op2``.
6394 #. ``oge``: yields ``true`` if both operands are not a QNAN and ``op1``
6395 is greater than or equal to ``op2``.
6396 #. ``olt``: yields ``true`` if both operands are not a QNAN and ``op1``
6397 is less than ``op2``.
6398 #. ``ole``: yields ``true`` if both operands are not a QNAN and ``op1``
6399 is less than or equal to ``op2``.
6400 #. ``one``: yields ``true`` if both operands are not a QNAN and ``op1``
6401 is not equal to ``op2``.
6402 #. ``ord``: yields ``true`` if both operands are not a QNAN.
6403 #. ``ueq``: yields ``true`` if either operand is a QNAN or ``op1`` is
6405 #. ``ugt``: yields ``true`` if either operand is a QNAN or ``op1`` is
6406 greater than ``op2``.
6407 #. ``uge``: yields ``true`` if either operand is a QNAN or ``op1`` is
6408 greater than or equal to ``op2``.
6409 #. ``ult``: yields ``true`` if either operand is a QNAN or ``op1`` is
6411 #. ``ule``: yields ``true`` if either operand is a QNAN or ``op1`` is
6412 less than or equal to ``op2``.
6413 #. ``une``: yields ``true`` if either operand is a QNAN or ``op1`` is
6414 not equal to ``op2``.
6415 #. ``uno``: yields ``true`` if either operand is a QNAN.
6416 #. ``true``: always yields ``true``, regardless of operands.
6421 .. code-block:: llvm
6423 <result> = fcmp oeq float 4.0, 5.0 ; yields: result=false
6424 <result> = fcmp one float 4.0, 5.0 ; yields: result=true
6425 <result> = fcmp olt float 4.0, 5.0 ; yields: result=true
6426 <result> = fcmp ueq double 1.0, 2.0 ; yields: result=false
6428 Note that the code generator does not yet support vector types with the
6429 ``fcmp`` instruction.
6433 '``phi``' Instruction
6434 ^^^^^^^^^^^^^^^^^^^^^
6441 <result> = phi <ty> [ <val0>, <label0>], ...
6446 The '``phi``' instruction is used to implement the φ node in the SSA
6447 graph representing the function.
6452 The type of the incoming values is specified with the first type field.
6453 After this, the '``phi``' instruction takes a list of pairs as
6454 arguments, with one pair for each predecessor basic block of the current
6455 block. Only values of :ref:`first class <t_firstclass>` type may be used as
6456 the value arguments to the PHI node. Only labels may be used as the
6459 There must be no non-phi instructions between the start of a basic block
6460 and the PHI instructions: i.e. PHI instructions must be first in a basic
6463 For the purposes of the SSA form, the use of each incoming value is
6464 deemed to occur on the edge from the corresponding predecessor block to
6465 the current block (but after any definition of an '``invoke``'
6466 instruction's return value on the same edge).
6471 At runtime, the '``phi``' instruction logically takes on the value
6472 specified by the pair corresponding to the predecessor basic block that
6473 executed just prior to the current block.
6478 .. code-block:: llvm
6480 Loop: ; Infinite loop that counts from 0 on up...
6481 %indvar = phi i32 [ 0, %LoopHeader ], [ %nextindvar, %Loop ]
6482 %nextindvar = add i32 %indvar, 1
6487 '``select``' Instruction
6488 ^^^^^^^^^^^^^^^^^^^^^^^^
6495 <result> = select selty <cond>, <ty> <val1>, <ty> <val2> ; yields ty
6497 selty is either i1 or {<N x i1>}
6502 The '``select``' instruction is used to choose one value based on a
6503 condition, without IR-level branching.
6508 The '``select``' instruction requires an 'i1' value or a vector of 'i1'
6509 values indicating the condition, and two values of the same :ref:`first
6510 class <t_firstclass>` type. If the val1/val2 are vectors and the
6511 condition is a scalar, then entire vectors are selected, not individual
6517 If the condition is an i1 and it evaluates to 1, the instruction returns
6518 the first value argument; otherwise, it returns the second value
6521 If the condition is a vector of i1, then the value arguments must be
6522 vectors of the same size, and the selection is done element by element.
6527 .. code-block:: llvm
6529 %X = select i1 true, i8 17, i8 42 ; yields i8:17
6533 '``call``' Instruction
6534 ^^^^^^^^^^^^^^^^^^^^^^
6541 <result> = [tail | musttail] call [cconv] [ret attrs] <ty> [<fnty>*] <fnptrval>(<function args>) [fn attrs]
6546 The '``call``' instruction represents a simple function call.
6551 This instruction requires several arguments:
6553 #. The optional ``tail`` and ``musttail`` markers indicate that the optimizers
6554 should perform tail call optimization. The ``tail`` marker is a hint that
6555 `can be ignored <CodeGenerator.html#sibcallopt>`_. The ``musttail`` marker
6556 means that the call must be tail call optimized in order for the program to
6557 be correct. The ``musttail`` marker provides these guarantees:
6559 #. The call will not cause unbounded stack growth if it is part of a
6560 recursive cycle in the call graph.
6561 #. Arguments with the :ref:`inalloca <attr_inalloca>` attribute are
6564 Both markers imply that the callee does not access allocas or varargs from
6565 the caller. Calls marked ``musttail`` must obey the following additional
6568 - The call must immediately precede a :ref:`ret <i_ret>` instruction,
6569 or a pointer bitcast followed by a ret instruction.
6570 - The ret instruction must return the (possibly bitcasted) value
6571 produced by the call or void.
6572 - The caller and callee prototypes must match. Pointer types of
6573 parameters or return types may differ in pointee type, but not
6575 - The calling conventions of the caller and callee must match.
6576 - All ABI-impacting function attributes, such as sret, byval, inreg,
6577 returned, and inalloca, must match.
6579 Tail call optimization for calls marked ``tail`` is guaranteed to occur if
6580 the following conditions are met:
6582 - Caller and callee both have the calling convention ``fastcc``.
6583 - The call is in tail position (ret immediately follows call and ret
6584 uses value of call or is void).
6585 - Option ``-tailcallopt`` is enabled, or
6586 ``llvm::GuaranteedTailCallOpt`` is ``true``.
6587 - `Platform-specific constraints are
6588 met. <CodeGenerator.html#tailcallopt>`_
6590 #. The optional "cconv" marker indicates which :ref:`calling
6591 convention <callingconv>` the call should use. If none is
6592 specified, the call defaults to using C calling conventions. The
6593 calling convention of the call must match the calling convention of
6594 the target function, or else the behavior is undefined.
6595 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
6596 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
6598 #. '``ty``': the type of the call instruction itself which is also the
6599 type of the return value. Functions that return no value are marked
6601 #. '``fnty``': shall be the signature of the pointer to function value
6602 being invoked. The argument types must match the types implied by
6603 this signature. This type can be omitted if the function is not
6604 varargs and if the function type does not return a pointer to a
6606 #. '``fnptrval``': An LLVM value containing a pointer to a function to
6607 be invoked. In most cases, this is a direct function invocation, but
6608 indirect ``call``'s are just as possible, calling an arbitrary pointer
6610 #. '``function args``': argument list whose types match the function
6611 signature argument types and parameter attributes. All arguments must
6612 be of :ref:`first class <t_firstclass>` type. If the function signature
6613 indicates the function accepts a variable number of arguments, the
6614 extra arguments can be specified.
6615 #. The optional :ref:`function attributes <fnattrs>` list. Only
6616 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
6617 attributes are valid here.
6622 The '``call``' instruction is used to cause control flow to transfer to
6623 a specified function, with its incoming arguments bound to the specified
6624 values. Upon a '``ret``' instruction in the called function, control
6625 flow continues with the instruction after the function call, and the
6626 return value of the function is bound to the result argument.
6631 .. code-block:: llvm
6633 %retval = call i32 @test(i32 %argc)
6634 call i32 (i8*, ...)* @printf(i8* %msg, i32 12, i8 42) ; yields i32
6635 %X = tail call i32 @foo() ; yields i32
6636 %Y = tail call fastcc i32 @foo() ; yields i32
6637 call void %foo(i8 97 signext)
6639 %struct.A = type { i32, i8 }
6640 %r = call %struct.A @foo() ; yields { i32, i8 }
6641 %gr = extractvalue %struct.A %r, 0 ; yields i32
6642 %gr1 = extractvalue %struct.A %r, 1 ; yields i8
6643 %Z = call void @foo() noreturn ; indicates that %foo never returns normally
6644 %ZZ = call zeroext i32 @bar() ; Return value is %zero extended
6646 llvm treats calls to some functions with names and arguments that match
6647 the standard C99 library as being the C99 library functions, and may
6648 perform optimizations or generate code for them under that assumption.
6649 This is something we'd like to change in the future to provide better
6650 support for freestanding environments and non-C-based languages.
6654 '``va_arg``' Instruction
6655 ^^^^^^^^^^^^^^^^^^^^^^^^
6662 <resultval> = va_arg <va_list*> <arglist>, <argty>
6667 The '``va_arg``' instruction is used to access arguments passed through
6668 the "variable argument" area of a function call. It is used to implement
6669 the ``va_arg`` macro in C.
6674 This instruction takes a ``va_list*`` value and the type of the
6675 argument. It returns a value of the specified argument type and
6676 increments the ``va_list`` to point to the next argument. The actual
6677 type of ``va_list`` is target specific.
6682 The '``va_arg``' instruction loads an argument of the specified type
6683 from the specified ``va_list`` and causes the ``va_list`` to point to
6684 the next argument. For more information, see the variable argument
6685 handling :ref:`Intrinsic Functions <int_varargs>`.
6687 It is legal for this instruction to be called in a function which does
6688 not take a variable number of arguments, for example, the ``vfprintf``
6691 ``va_arg`` is an LLVM instruction instead of an :ref:`intrinsic
6692 function <intrinsics>` because it takes a type as an argument.
6697 See the :ref:`variable argument processing <int_varargs>` section.
6699 Note that the code generator does not yet fully support va\_arg on many
6700 targets. Also, it does not currently support va\_arg with aggregate
6701 types on any target.
6705 '``landingpad``' Instruction
6706 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6713 <resultval> = landingpad <resultty> personality <type> <pers_fn> <clause>+
6714 <resultval> = landingpad <resultty> personality <type> <pers_fn> cleanup <clause>*
6716 <clause> := catch <type> <value>
6717 <clause> := filter <array constant type> <array constant>
6722 The '``landingpad``' instruction is used by `LLVM's exception handling
6723 system <ExceptionHandling.html#overview>`_ to specify that a basic block
6724 is a landing pad --- one where the exception lands, and corresponds to the
6725 code found in the ``catch`` portion of a ``try``/``catch`` sequence. It
6726 defines values supplied by the personality function (``pers_fn``) upon
6727 re-entry to the function. The ``resultval`` has the type ``resultty``.
6732 This instruction takes a ``pers_fn`` value. This is the personality
6733 function associated with the unwinding mechanism. The optional
6734 ``cleanup`` flag indicates that the landing pad block is a cleanup.
6736 A ``clause`` begins with the clause type --- ``catch`` or ``filter`` --- and
6737 contains the global variable representing the "type" that may be caught
6738 or filtered respectively. Unlike the ``catch`` clause, the ``filter``
6739 clause takes an array constant as its argument. Use
6740 "``[0 x i8**] undef``" for a filter which cannot throw. The
6741 '``landingpad``' instruction must contain *at least* one ``clause`` or
6742 the ``cleanup`` flag.
6747 The '``landingpad``' instruction defines the values which are set by the
6748 personality function (``pers_fn``) upon re-entry to the function, and
6749 therefore the "result type" of the ``landingpad`` instruction. As with
6750 calling conventions, how the personality function results are
6751 represented in LLVM IR is target specific.
6753 The clauses are applied in order from top to bottom. If two
6754 ``landingpad`` instructions are merged together through inlining, the
6755 clauses from the calling function are appended to the list of clauses.
6756 When the call stack is being unwound due to an exception being thrown,
6757 the exception is compared against each ``clause`` in turn. If it doesn't
6758 match any of the clauses, and the ``cleanup`` flag is not set, then
6759 unwinding continues further up the call stack.
6761 The ``landingpad`` instruction has several restrictions:
6763 - A landing pad block is a basic block which is the unwind destination
6764 of an '``invoke``' instruction.
6765 - A landing pad block must have a '``landingpad``' instruction as its
6766 first non-PHI instruction.
6767 - There can be only one '``landingpad``' instruction within the landing
6769 - A basic block that is not a landing pad block may not include a
6770 '``landingpad``' instruction.
6771 - All '``landingpad``' instructions in a function must have the same
6772 personality function.
6777 .. code-block:: llvm
6779 ;; A landing pad which can catch an integer.
6780 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6782 ;; A landing pad that is a cleanup.
6783 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6785 ;; A landing pad which can catch an integer and can only throw a double.
6786 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6788 filter [1 x i8**] [@_ZTId]
6795 LLVM supports the notion of an "intrinsic function". These functions
6796 have well known names and semantics and are required to follow certain
6797 restrictions. Overall, these intrinsics represent an extension mechanism
6798 for the LLVM language that does not require changing all of the
6799 transformations in LLVM when adding to the language (or the bitcode
6800 reader/writer, the parser, etc...).
6802 Intrinsic function names must all start with an "``llvm.``" prefix. This
6803 prefix is reserved in LLVM for intrinsic names; thus, function names may
6804 not begin with this prefix. Intrinsic functions must always be external
6805 functions: you cannot define the body of intrinsic functions. Intrinsic
6806 functions may only be used in call or invoke instructions: it is illegal
6807 to take the address of an intrinsic function. Additionally, because
6808 intrinsic functions are part of the LLVM language, it is required if any
6809 are added that they be documented here.
6811 Some intrinsic functions can be overloaded, i.e., the intrinsic
6812 represents a family of functions that perform the same operation but on
6813 different data types. Because LLVM can represent over 8 million
6814 different integer types, overloading is used commonly to allow an
6815 intrinsic function to operate on any integer type. One or more of the
6816 argument types or the result type can be overloaded to accept any
6817 integer type. Argument types may also be defined as exactly matching a
6818 previous argument's type or the result type. This allows an intrinsic
6819 function which accepts multiple arguments, but needs all of them to be
6820 of the same type, to only be overloaded with respect to a single
6821 argument or the result.
6823 Overloaded intrinsics will have the names of its overloaded argument
6824 types encoded into its function name, each preceded by a period. Only
6825 those types which are overloaded result in a name suffix. Arguments
6826 whose type is matched against another type do not. For example, the
6827 ``llvm.ctpop`` function can take an integer of any width and returns an
6828 integer of exactly the same integer width. This leads to a family of
6829 functions such as ``i8 @llvm.ctpop.i8(i8 %val)`` and
6830 ``i29 @llvm.ctpop.i29(i29 %val)``. Only one type, the return type, is
6831 overloaded, and only one type suffix is required. Because the argument's
6832 type is matched against the return type, it does not require its own
6835 To learn how to add an intrinsic function, please see the `Extending
6836 LLVM Guide <ExtendingLLVM.html>`_.
6840 Variable Argument Handling Intrinsics
6841 -------------------------------------
6843 Variable argument support is defined in LLVM with the
6844 :ref:`va_arg <i_va_arg>` instruction and these three intrinsic
6845 functions. These functions are related to the similarly named macros
6846 defined in the ``<stdarg.h>`` header file.
6848 All of these functions operate on arguments that use a target-specific
6849 value type "``va_list``". The LLVM assembly language reference manual
6850 does not define what this type is, so all transformations should be
6851 prepared to handle these functions regardless of the type used.
6853 This example shows how the :ref:`va_arg <i_va_arg>` instruction and the
6854 variable argument handling intrinsic functions are used.
6856 .. code-block:: llvm
6858 define i32 @test(i32 %X, ...) {
6859 ; Initialize variable argument processing
6861 %ap2 = bitcast i8** %ap to i8*
6862 call void @llvm.va_start(i8* %ap2)
6864 ; Read a single integer argument
6865 %tmp = va_arg i8** %ap, i32
6867 ; Demonstrate usage of llvm.va_copy and llvm.va_end
6869 %aq2 = bitcast i8** %aq to i8*
6870 call void @llvm.va_copy(i8* %aq2, i8* %ap2)
6871 call void @llvm.va_end(i8* %aq2)
6873 ; Stop processing of arguments.
6874 call void @llvm.va_end(i8* %ap2)
6878 declare void @llvm.va_start(i8*)
6879 declare void @llvm.va_copy(i8*, i8*)
6880 declare void @llvm.va_end(i8*)
6884 '``llvm.va_start``' Intrinsic
6885 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6892 declare void @llvm.va_start(i8* <arglist>)
6897 The '``llvm.va_start``' intrinsic initializes ``*<arglist>`` for
6898 subsequent use by ``va_arg``.
6903 The argument is a pointer to a ``va_list`` element to initialize.
6908 The '``llvm.va_start``' intrinsic works just like the ``va_start`` macro
6909 available in C. In a target-dependent way, it initializes the
6910 ``va_list`` element to which the argument points, so that the next call
6911 to ``va_arg`` will produce the first variable argument passed to the
6912 function. Unlike the C ``va_start`` macro, this intrinsic does not need
6913 to know the last argument of the function as the compiler can figure
6916 '``llvm.va_end``' Intrinsic
6917 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6924 declare void @llvm.va_end(i8* <arglist>)
6929 The '``llvm.va_end``' intrinsic destroys ``*<arglist>``, which has been
6930 initialized previously with ``llvm.va_start`` or ``llvm.va_copy``.
6935 The argument is a pointer to a ``va_list`` to destroy.
6940 The '``llvm.va_end``' intrinsic works just like the ``va_end`` macro
6941 available in C. In a target-dependent way, it destroys the ``va_list``
6942 element to which the argument points. Calls to
6943 :ref:`llvm.va_start <int_va_start>` and
6944 :ref:`llvm.va_copy <int_va_copy>` must be matched exactly with calls to
6949 '``llvm.va_copy``' Intrinsic
6950 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6957 declare void @llvm.va_copy(i8* <destarglist>, i8* <srcarglist>)
6962 The '``llvm.va_copy``' intrinsic copies the current argument position
6963 from the source argument list to the destination argument list.
6968 The first argument is a pointer to a ``va_list`` element to initialize.
6969 The second argument is a pointer to a ``va_list`` element to copy from.
6974 The '``llvm.va_copy``' intrinsic works just like the ``va_copy`` macro
6975 available in C. In a target-dependent way, it copies the source
6976 ``va_list`` element into the destination ``va_list`` element. This
6977 intrinsic is necessary because the `` llvm.va_start`` intrinsic may be
6978 arbitrarily complex and require, for example, memory allocation.
6980 Accurate Garbage Collection Intrinsics
6981 --------------------------------------
6983 LLVM support for `Accurate Garbage Collection <GarbageCollection.html>`_
6984 (GC) requires the implementation and generation of these intrinsics.
6985 These intrinsics allow identification of :ref:`GC roots on the
6986 stack <int_gcroot>`, as well as garbage collector implementations that
6987 require :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers.
6988 Front-ends for type-safe garbage collected languages should generate
6989 these intrinsics to make use of the LLVM garbage collectors. For more
6990 details, see `Accurate Garbage Collection with
6991 LLVM <GarbageCollection.html>`_.
6993 The garbage collection intrinsics only operate on objects in the generic
6994 address space (address space zero).
6998 '``llvm.gcroot``' Intrinsic
6999 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7006 declare void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
7011 The '``llvm.gcroot``' intrinsic declares the existence of a GC root to
7012 the code generator, and allows some metadata to be associated with it.
7017 The first argument specifies the address of a stack object that contains
7018 the root pointer. The second pointer (which must be either a constant or
7019 a global value address) contains the meta-data to be associated with the
7025 At runtime, a call to this intrinsic stores a null pointer into the
7026 "ptrloc" location. At compile-time, the code generator generates
7027 information to allow the runtime to find the pointer at GC safe points.
7028 The '``llvm.gcroot``' intrinsic may only be used in a function which
7029 :ref:`specifies a GC algorithm <gc>`.
7033 '``llvm.gcread``' Intrinsic
7034 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7041 declare i8* @llvm.gcread(i8* %ObjPtr, i8** %Ptr)
7046 The '``llvm.gcread``' intrinsic identifies reads of references from heap
7047 locations, allowing garbage collector implementations that require read
7053 The second argument is the address to read from, which should be an
7054 address allocated from the garbage collector. The first object is a
7055 pointer to the start of the referenced object, if needed by the language
7056 runtime (otherwise null).
7061 The '``llvm.gcread``' intrinsic has the same semantics as a load
7062 instruction, but may be replaced with substantially more complex code by
7063 the garbage collector runtime, as needed. The '``llvm.gcread``'
7064 intrinsic may only be used in a function which :ref:`specifies a GC
7069 '``llvm.gcwrite``' Intrinsic
7070 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7077 declare void @llvm.gcwrite(i8* %P1, i8* %Obj, i8** %P2)
7082 The '``llvm.gcwrite``' intrinsic identifies writes of references to heap
7083 locations, allowing garbage collector implementations that require write
7084 barriers (such as generational or reference counting collectors).
7089 The first argument is the reference to store, the second is the start of
7090 the object to store it to, and the third is the address of the field of
7091 Obj to store to. If the runtime does not require a pointer to the
7092 object, Obj may be null.
7097 The '``llvm.gcwrite``' intrinsic has the same semantics as a store
7098 instruction, but may be replaced with substantially more complex code by
7099 the garbage collector runtime, as needed. The '``llvm.gcwrite``'
7100 intrinsic may only be used in a function which :ref:`specifies a GC
7103 Code Generator Intrinsics
7104 -------------------------
7106 These intrinsics are provided by LLVM to expose special features that
7107 may only be implemented with code generator support.
7109 '``llvm.returnaddress``' Intrinsic
7110 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7117 declare i8 *@llvm.returnaddress(i32 <level>)
7122 The '``llvm.returnaddress``' intrinsic attempts to compute a
7123 target-specific value indicating the return address of the current
7124 function or one of its callers.
7129 The argument to this intrinsic indicates which function to return the
7130 address for. Zero indicates the calling function, one indicates its
7131 caller, etc. The argument is **required** to be a constant integer
7137 The '``llvm.returnaddress``' intrinsic either returns a pointer
7138 indicating the return address of the specified call frame, or zero if it
7139 cannot be identified. The value returned by this intrinsic is likely to
7140 be incorrect or 0 for arguments other than zero, so it should only be
7141 used for debugging purposes.
7143 Note that calling this intrinsic does not prevent function inlining or
7144 other aggressive transformations, so the value returned may not be that
7145 of the obvious source-language caller.
7147 '``llvm.frameaddress``' Intrinsic
7148 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7155 declare i8* @llvm.frameaddress(i32 <level>)
7160 The '``llvm.frameaddress``' intrinsic attempts to return the
7161 target-specific frame pointer value for the specified stack frame.
7166 The argument to this intrinsic indicates which function to return the
7167 frame pointer for. Zero indicates the calling function, one indicates
7168 its caller, etc. The argument is **required** to be a constant integer
7174 The '``llvm.frameaddress``' intrinsic either returns a pointer
7175 indicating the frame address of the specified call frame, or zero if it
7176 cannot be identified. The value returned by this intrinsic is likely to
7177 be incorrect or 0 for arguments other than zero, so it should only be
7178 used for debugging purposes.
7180 Note that calling this intrinsic does not prevent function inlining or
7181 other aggressive transformations, so the value returned may not be that
7182 of the obvious source-language caller.
7184 .. _int_read_register:
7185 .. _int_write_register:
7187 '``llvm.read_register``' and '``llvm.write_register``' Intrinsics
7188 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7195 declare i32 @llvm.read_register.i32(metadata)
7196 declare i64 @llvm.read_register.i64(metadata)
7197 declare void @llvm.write_register.i32(metadata, i32 @value)
7198 declare void @llvm.write_register.i64(metadata, i64 @value)
7199 !0 = metadata !{metadata !"sp\00"}
7204 The '``llvm.read_register``' and '``llvm.write_register``' intrinsics
7205 provides access to the named register. The register must be valid on
7206 the architecture being compiled to. The type needs to be compatible
7207 with the register being read.
7212 The '``llvm.read_register``' intrinsic returns the current value of the
7213 register, where possible. The '``llvm.write_register``' intrinsic sets
7214 the current value of the register, where possible.
7216 This is useful to implement named register global variables that need
7217 to always be mapped to a specific register, as is common practice on
7218 bare-metal programs including OS kernels.
7220 The compiler doesn't check for register availability or use of the used
7221 register in surrounding code, including inline assembly. Because of that,
7222 allocatable registers are not supported.
7224 Warning: So far it only works with the stack pointer on selected
7225 architectures (ARM, AArch64, PowerPC and x86_64). Significant amount of
7226 work is needed to support other registers and even more so, allocatable
7231 '``llvm.stacksave``' Intrinsic
7232 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7239 declare i8* @llvm.stacksave()
7244 The '``llvm.stacksave``' intrinsic is used to remember the current state
7245 of the function stack, for use with
7246 :ref:`llvm.stackrestore <int_stackrestore>`. This is useful for
7247 implementing language features like scoped automatic variable sized
7253 This intrinsic returns a opaque pointer value that can be passed to
7254 :ref:`llvm.stackrestore <int_stackrestore>`. When an
7255 ``llvm.stackrestore`` intrinsic is executed with a value saved from
7256 ``llvm.stacksave``, it effectively restores the state of the stack to
7257 the state it was in when the ``llvm.stacksave`` intrinsic executed. In
7258 practice, this pops any :ref:`alloca <i_alloca>` blocks from the stack that
7259 were allocated after the ``llvm.stacksave`` was executed.
7261 .. _int_stackrestore:
7263 '``llvm.stackrestore``' Intrinsic
7264 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7271 declare void @llvm.stackrestore(i8* %ptr)
7276 The '``llvm.stackrestore``' intrinsic is used to restore the state of
7277 the function stack to the state it was in when the corresponding
7278 :ref:`llvm.stacksave <int_stacksave>` intrinsic executed. This is
7279 useful for implementing language features like scoped automatic variable
7280 sized arrays in C99.
7285 See the description for :ref:`llvm.stacksave <int_stacksave>`.
7287 '``llvm.prefetch``' Intrinsic
7288 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7295 declare void @llvm.prefetch(i8* <address>, i32 <rw>, i32 <locality>, i32 <cache type>)
7300 The '``llvm.prefetch``' intrinsic is a hint to the code generator to
7301 insert a prefetch instruction if supported; otherwise, it is a noop.
7302 Prefetches have no effect on the behavior of the program but can change
7303 its performance characteristics.
7308 ``address`` is the address to be prefetched, ``rw`` is the specifier
7309 determining if the fetch should be for a read (0) or write (1), and
7310 ``locality`` is a temporal locality specifier ranging from (0) - no
7311 locality, to (3) - extremely local keep in cache. The ``cache type``
7312 specifies whether the prefetch is performed on the data (1) or
7313 instruction (0) cache. The ``rw``, ``locality`` and ``cache type``
7314 arguments must be constant integers.
7319 This intrinsic does not modify the behavior of the program. In
7320 particular, prefetches cannot trap and do not produce a value. On
7321 targets that support this intrinsic, the prefetch can provide hints to
7322 the processor cache for better performance.
7324 '``llvm.pcmarker``' Intrinsic
7325 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7332 declare void @llvm.pcmarker(i32 <id>)
7337 The '``llvm.pcmarker``' intrinsic is a method to export a Program
7338 Counter (PC) in a region of code to simulators and other tools. The
7339 method is target specific, but it is expected that the marker will use
7340 exported symbols to transmit the PC of the marker. The marker makes no
7341 guarantees that it will remain with any specific instruction after
7342 optimizations. It is possible that the presence of a marker will inhibit
7343 optimizations. The intended use is to be inserted after optimizations to
7344 allow correlations of simulation runs.
7349 ``id`` is a numerical id identifying the marker.
7354 This intrinsic does not modify the behavior of the program. Backends
7355 that do not support this intrinsic may ignore it.
7357 '``llvm.readcyclecounter``' Intrinsic
7358 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7365 declare i64 @llvm.readcyclecounter()
7370 The '``llvm.readcyclecounter``' intrinsic provides access to the cycle
7371 counter register (or similar low latency, high accuracy clocks) on those
7372 targets that support it. On X86, it should map to RDTSC. On Alpha, it
7373 should map to RPCC. As the backing counters overflow quickly (on the
7374 order of 9 seconds on alpha), this should only be used for small
7380 When directly supported, reading the cycle counter should not modify any
7381 memory. Implementations are allowed to either return a application
7382 specific value or a system wide value. On backends without support, this
7383 is lowered to a constant 0.
7385 Note that runtime support may be conditional on the privilege-level code is
7386 running at and the host platform.
7388 '``llvm.clear_cache``' Intrinsic
7389 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7396 declare void @llvm.clear_cache(i8*, i8*)
7401 The '``llvm.clear_cache``' intrinsic ensures visibility of modifications
7402 in the specified range to the execution unit of the processor. On
7403 targets with non-unified instruction and data cache, the implementation
7404 flushes the instruction cache.
7409 On platforms with coherent instruction and data caches (e.g. x86), this
7410 intrinsic is a nop. On platforms with non-coherent instruction and data
7411 cache (e.g. ARM, MIPS), the intrinsic is lowered either to appropriate
7412 instructions or a system call, if cache flushing requires special
7415 The default behavior is to emit a call to ``__clear_cache`` from the run
7418 This instrinsic does *not* empty the instruction pipeline. Modifications
7419 of the current function are outside the scope of the intrinsic.
7421 Standard C Library Intrinsics
7422 -----------------------------
7424 LLVM provides intrinsics for a few important standard C library
7425 functions. These intrinsics allow source-language front-ends to pass
7426 information about the alignment of the pointer arguments to the code
7427 generator, providing opportunity for more efficient code generation.
7431 '``llvm.memcpy``' Intrinsic
7432 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7437 This is an overloaded intrinsic. You can use ``llvm.memcpy`` on any
7438 integer bit width and for different address spaces. Not all targets
7439 support all bit widths however.
7443 declare void @llvm.memcpy.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
7444 i32 <len>, i32 <align>, i1 <isvolatile>)
7445 declare void @llvm.memcpy.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
7446 i64 <len>, i32 <align>, i1 <isvolatile>)
7451 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
7452 source location to the destination location.
7454 Note that, unlike the standard libc function, the ``llvm.memcpy.*``
7455 intrinsics do not return a value, takes extra alignment/isvolatile
7456 arguments and the pointers can be in specified address spaces.
7461 The first argument is a pointer to the destination, the second is a
7462 pointer to the source. The third argument is an integer argument
7463 specifying the number of bytes to copy, the fourth argument is the
7464 alignment of the source and destination locations, and the fifth is a
7465 boolean indicating a volatile access.
7467 If the call to this intrinsic has an alignment value that is not 0 or 1,
7468 then the caller guarantees that both the source and destination pointers
7469 are aligned to that boundary.
7471 If the ``isvolatile`` parameter is ``true``, the ``llvm.memcpy`` call is
7472 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
7473 very cleanly specified and it is unwise to depend on it.
7478 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
7479 source location to the destination location, which are not allowed to
7480 overlap. It copies "len" bytes of memory over. If the argument is known
7481 to be aligned to some boundary, this can be specified as the fourth
7482 argument, otherwise it should be set to 0 or 1 (both meaning no alignment).
7484 '``llvm.memmove``' Intrinsic
7485 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7490 This is an overloaded intrinsic. You can use llvm.memmove on any integer
7491 bit width and for different address space. Not all targets support all
7496 declare void @llvm.memmove.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
7497 i32 <len>, i32 <align>, i1 <isvolatile>)
7498 declare void @llvm.memmove.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
7499 i64 <len>, i32 <align>, i1 <isvolatile>)
7504 The '``llvm.memmove.*``' intrinsics move a block of memory from the
7505 source location to the destination location. It is similar to the
7506 '``llvm.memcpy``' intrinsic but allows the two memory locations to
7509 Note that, unlike the standard libc function, the ``llvm.memmove.*``
7510 intrinsics do not return a value, takes extra alignment/isvolatile
7511 arguments and the pointers can be in specified address spaces.
7516 The first argument is a pointer to the destination, the second is a
7517 pointer to the source. The third argument is an integer argument
7518 specifying the number of bytes to copy, the fourth argument is the
7519 alignment of the source and destination locations, and the fifth is a
7520 boolean indicating a volatile access.
7522 If the call to this intrinsic has an alignment value that is not 0 or 1,
7523 then the caller guarantees that the source and destination pointers are
7524 aligned to that boundary.
7526 If the ``isvolatile`` parameter is ``true``, the ``llvm.memmove`` call
7527 is a :ref:`volatile operation <volatile>`. The detailed access behavior is
7528 not very cleanly specified and it is unwise to depend on it.
7533 The '``llvm.memmove.*``' intrinsics copy a block of memory from the
7534 source location to the destination location, which may overlap. It
7535 copies "len" bytes of memory over. If the argument is known to be
7536 aligned to some boundary, this can be specified as the fourth argument,
7537 otherwise it should be set to 0 or 1 (both meaning no alignment).
7539 '``llvm.memset.*``' Intrinsics
7540 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7545 This is an overloaded intrinsic. You can use llvm.memset on any integer
7546 bit width and for different address spaces. However, not all targets
7547 support all bit widths.
7551 declare void @llvm.memset.p0i8.i32(i8* <dest>, i8 <val>,
7552 i32 <len>, i32 <align>, i1 <isvolatile>)
7553 declare void @llvm.memset.p0i8.i64(i8* <dest>, i8 <val>,
7554 i64 <len>, i32 <align>, i1 <isvolatile>)
7559 The '``llvm.memset.*``' intrinsics fill a block of memory with a
7560 particular byte value.
7562 Note that, unlike the standard libc function, the ``llvm.memset``
7563 intrinsic does not return a value and takes extra alignment/volatile
7564 arguments. Also, the destination can be in an arbitrary address space.
7569 The first argument is a pointer to the destination to fill, the second
7570 is the byte value with which to fill it, the third argument is an
7571 integer argument specifying the number of bytes to fill, and the fourth
7572 argument is the known alignment of the destination location.
7574 If the call to this intrinsic has an alignment value that is not 0 or 1,
7575 then the caller guarantees that the destination pointer is aligned to
7578 If the ``isvolatile`` parameter is ``true``, the ``llvm.memset`` call is
7579 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
7580 very cleanly specified and it is unwise to depend on it.
7585 The '``llvm.memset.*``' intrinsics fill "len" bytes of memory starting
7586 at the destination location. If the argument is known to be aligned to
7587 some boundary, this can be specified as the fourth argument, otherwise
7588 it should be set to 0 or 1 (both meaning no alignment).
7590 '``llvm.sqrt.*``' Intrinsic
7591 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7596 This is an overloaded intrinsic. You can use ``llvm.sqrt`` on any
7597 floating point or vector of floating point type. Not all targets support
7602 declare float @llvm.sqrt.f32(float %Val)
7603 declare double @llvm.sqrt.f64(double %Val)
7604 declare x86_fp80 @llvm.sqrt.f80(x86_fp80 %Val)
7605 declare fp128 @llvm.sqrt.f128(fp128 %Val)
7606 declare ppc_fp128 @llvm.sqrt.ppcf128(ppc_fp128 %Val)
7611 The '``llvm.sqrt``' intrinsics return the sqrt of the specified operand,
7612 returning the same value as the libm '``sqrt``' functions would. Unlike
7613 ``sqrt`` in libm, however, ``llvm.sqrt`` has undefined behavior for
7614 negative numbers other than -0.0 (which allows for better optimization,
7615 because there is no need to worry about errno being set).
7616 ``llvm.sqrt(-0.0)`` is defined to return -0.0 like IEEE sqrt.
7621 The argument and return value are floating point numbers of the same
7627 This function returns the sqrt of the specified operand if it is a
7628 nonnegative floating point number.
7630 '``llvm.powi.*``' Intrinsic
7631 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7636 This is an overloaded intrinsic. You can use ``llvm.powi`` on any
7637 floating point or vector of floating point type. Not all targets support
7642 declare float @llvm.powi.f32(float %Val, i32 %power)
7643 declare double @llvm.powi.f64(double %Val, i32 %power)
7644 declare x86_fp80 @llvm.powi.f80(x86_fp80 %Val, i32 %power)
7645 declare fp128 @llvm.powi.f128(fp128 %Val, i32 %power)
7646 declare ppc_fp128 @llvm.powi.ppcf128(ppc_fp128 %Val, i32 %power)
7651 The '``llvm.powi.*``' intrinsics return the first operand raised to the
7652 specified (positive or negative) power. The order of evaluation of
7653 multiplications is not defined. When a vector of floating point type is
7654 used, the second argument remains a scalar integer value.
7659 The second argument is an integer power, and the first is a value to
7660 raise to that power.
7665 This function returns the first value raised to the second power with an
7666 unspecified sequence of rounding operations.
7668 '``llvm.sin.*``' Intrinsic
7669 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7674 This is an overloaded intrinsic. You can use ``llvm.sin`` on any
7675 floating point or vector of floating point type. Not all targets support
7680 declare float @llvm.sin.f32(float %Val)
7681 declare double @llvm.sin.f64(double %Val)
7682 declare x86_fp80 @llvm.sin.f80(x86_fp80 %Val)
7683 declare fp128 @llvm.sin.f128(fp128 %Val)
7684 declare ppc_fp128 @llvm.sin.ppcf128(ppc_fp128 %Val)
7689 The '``llvm.sin.*``' intrinsics return the sine of the operand.
7694 The argument and return value are floating point numbers of the same
7700 This function returns the sine of the specified operand, returning the
7701 same values as the libm ``sin`` functions would, and handles error
7702 conditions in the same way.
7704 '``llvm.cos.*``' Intrinsic
7705 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7710 This is an overloaded intrinsic. You can use ``llvm.cos`` on any
7711 floating point or vector of floating point type. Not all targets support
7716 declare float @llvm.cos.f32(float %Val)
7717 declare double @llvm.cos.f64(double %Val)
7718 declare x86_fp80 @llvm.cos.f80(x86_fp80 %Val)
7719 declare fp128 @llvm.cos.f128(fp128 %Val)
7720 declare ppc_fp128 @llvm.cos.ppcf128(ppc_fp128 %Val)
7725 The '``llvm.cos.*``' intrinsics return the cosine of the operand.
7730 The argument and return value are floating point numbers of the same
7736 This function returns the cosine of the specified operand, returning the
7737 same values as the libm ``cos`` functions would, and handles error
7738 conditions in the same way.
7740 '``llvm.pow.*``' Intrinsic
7741 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7746 This is an overloaded intrinsic. You can use ``llvm.pow`` on any
7747 floating point or vector of floating point type. Not all targets support
7752 declare float @llvm.pow.f32(float %Val, float %Power)
7753 declare double @llvm.pow.f64(double %Val, double %Power)
7754 declare x86_fp80 @llvm.pow.f80(x86_fp80 %Val, x86_fp80 %Power)
7755 declare fp128 @llvm.pow.f128(fp128 %Val, fp128 %Power)
7756 declare ppc_fp128 @llvm.pow.ppcf128(ppc_fp128 %Val, ppc_fp128 Power)
7761 The '``llvm.pow.*``' intrinsics return the first operand raised to the
7762 specified (positive or negative) power.
7767 The second argument is a floating point power, and the first is a value
7768 to raise to that power.
7773 This function returns the first value raised to the second power,
7774 returning the same values as the libm ``pow`` functions would, and
7775 handles error conditions in the same way.
7777 '``llvm.exp.*``' Intrinsic
7778 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7783 This is an overloaded intrinsic. You can use ``llvm.exp`` on any
7784 floating point or vector of floating point type. Not all targets support
7789 declare float @llvm.exp.f32(float %Val)
7790 declare double @llvm.exp.f64(double %Val)
7791 declare x86_fp80 @llvm.exp.f80(x86_fp80 %Val)
7792 declare fp128 @llvm.exp.f128(fp128 %Val)
7793 declare ppc_fp128 @llvm.exp.ppcf128(ppc_fp128 %Val)
7798 The '``llvm.exp.*``' intrinsics perform the exp function.
7803 The argument and return value are floating point numbers of the same
7809 This function returns the same values as the libm ``exp`` functions
7810 would, and handles error conditions in the same way.
7812 '``llvm.exp2.*``' Intrinsic
7813 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7818 This is an overloaded intrinsic. You can use ``llvm.exp2`` on any
7819 floating point or vector of floating point type. Not all targets support
7824 declare float @llvm.exp2.f32(float %Val)
7825 declare double @llvm.exp2.f64(double %Val)
7826 declare x86_fp80 @llvm.exp2.f80(x86_fp80 %Val)
7827 declare fp128 @llvm.exp2.f128(fp128 %Val)
7828 declare ppc_fp128 @llvm.exp2.ppcf128(ppc_fp128 %Val)
7833 The '``llvm.exp2.*``' intrinsics perform the exp2 function.
7838 The argument and return value are floating point numbers of the same
7844 This function returns the same values as the libm ``exp2`` functions
7845 would, and handles error conditions in the same way.
7847 '``llvm.log.*``' Intrinsic
7848 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7853 This is an overloaded intrinsic. You can use ``llvm.log`` on any
7854 floating point or vector of floating point type. Not all targets support
7859 declare float @llvm.log.f32(float %Val)
7860 declare double @llvm.log.f64(double %Val)
7861 declare x86_fp80 @llvm.log.f80(x86_fp80 %Val)
7862 declare fp128 @llvm.log.f128(fp128 %Val)
7863 declare ppc_fp128 @llvm.log.ppcf128(ppc_fp128 %Val)
7868 The '``llvm.log.*``' intrinsics perform the log function.
7873 The argument and return value are floating point numbers of the same
7879 This function returns the same values as the libm ``log`` functions
7880 would, and handles error conditions in the same way.
7882 '``llvm.log10.*``' Intrinsic
7883 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7888 This is an overloaded intrinsic. You can use ``llvm.log10`` on any
7889 floating point or vector of floating point type. Not all targets support
7894 declare float @llvm.log10.f32(float %Val)
7895 declare double @llvm.log10.f64(double %Val)
7896 declare x86_fp80 @llvm.log10.f80(x86_fp80 %Val)
7897 declare fp128 @llvm.log10.f128(fp128 %Val)
7898 declare ppc_fp128 @llvm.log10.ppcf128(ppc_fp128 %Val)
7903 The '``llvm.log10.*``' intrinsics perform the log10 function.
7908 The argument and return value are floating point numbers of the same
7914 This function returns the same values as the libm ``log10`` functions
7915 would, and handles error conditions in the same way.
7917 '``llvm.log2.*``' Intrinsic
7918 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7923 This is an overloaded intrinsic. You can use ``llvm.log2`` on any
7924 floating point or vector of floating point type. Not all targets support
7929 declare float @llvm.log2.f32(float %Val)
7930 declare double @llvm.log2.f64(double %Val)
7931 declare x86_fp80 @llvm.log2.f80(x86_fp80 %Val)
7932 declare fp128 @llvm.log2.f128(fp128 %Val)
7933 declare ppc_fp128 @llvm.log2.ppcf128(ppc_fp128 %Val)
7938 The '``llvm.log2.*``' intrinsics perform the log2 function.
7943 The argument and return value are floating point numbers of the same
7949 This function returns the same values as the libm ``log2`` functions
7950 would, and handles error conditions in the same way.
7952 '``llvm.fma.*``' Intrinsic
7953 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7958 This is an overloaded intrinsic. You can use ``llvm.fma`` on any
7959 floating point or vector of floating point type. Not all targets support
7964 declare float @llvm.fma.f32(float %a, float %b, float %c)
7965 declare double @llvm.fma.f64(double %a, double %b, double %c)
7966 declare x86_fp80 @llvm.fma.f80(x86_fp80 %a, x86_fp80 %b, x86_fp80 %c)
7967 declare fp128 @llvm.fma.f128(fp128 %a, fp128 %b, fp128 %c)
7968 declare ppc_fp128 @llvm.fma.ppcf128(ppc_fp128 %a, ppc_fp128 %b, ppc_fp128 %c)
7973 The '``llvm.fma.*``' intrinsics perform the fused multiply-add
7979 The argument and return value are floating point numbers of the same
7985 This function returns the same values as the libm ``fma`` functions
7986 would, and does not set errno.
7988 '``llvm.fabs.*``' Intrinsic
7989 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7994 This is an overloaded intrinsic. You can use ``llvm.fabs`` on any
7995 floating point or vector of floating point type. Not all targets support
8000 declare float @llvm.fabs.f32(float %Val)
8001 declare double @llvm.fabs.f64(double %Val)
8002 declare x86_fp80 @llvm.fabs.f80(x86_fp80 %Val)
8003 declare fp128 @llvm.fabs.f128(fp128 %Val)
8004 declare ppc_fp128 @llvm.fabs.ppcf128(ppc_fp128 %Val)
8009 The '``llvm.fabs.*``' intrinsics return the absolute value of the
8015 The argument and return value are floating point numbers of the same
8021 This function returns the same values as the libm ``fabs`` functions
8022 would, and handles error conditions in the same way.
8024 '``llvm.copysign.*``' Intrinsic
8025 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8030 This is an overloaded intrinsic. You can use ``llvm.copysign`` on any
8031 floating point or vector of floating point type. Not all targets support
8036 declare float @llvm.copysign.f32(float %Mag, float %Sgn)
8037 declare double @llvm.copysign.f64(double %Mag, double %Sgn)
8038 declare x86_fp80 @llvm.copysign.f80(x86_fp80 %Mag, x86_fp80 %Sgn)
8039 declare fp128 @llvm.copysign.f128(fp128 %Mag, fp128 %Sgn)
8040 declare ppc_fp128 @llvm.copysign.ppcf128(ppc_fp128 %Mag, ppc_fp128 %Sgn)
8045 The '``llvm.copysign.*``' intrinsics return a value with the magnitude of the
8046 first operand and the sign of the second operand.
8051 The arguments and return value are floating point numbers of the same
8057 This function returns the same values as the libm ``copysign``
8058 functions would, and handles error conditions in the same way.
8060 '``llvm.floor.*``' Intrinsic
8061 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8066 This is an overloaded intrinsic. You can use ``llvm.floor`` on any
8067 floating point or vector of floating point type. Not all targets support
8072 declare float @llvm.floor.f32(float %Val)
8073 declare double @llvm.floor.f64(double %Val)
8074 declare x86_fp80 @llvm.floor.f80(x86_fp80 %Val)
8075 declare fp128 @llvm.floor.f128(fp128 %Val)
8076 declare ppc_fp128 @llvm.floor.ppcf128(ppc_fp128 %Val)
8081 The '``llvm.floor.*``' intrinsics return the floor of the operand.
8086 The argument and return value are floating point numbers of the same
8092 This function returns the same values as the libm ``floor`` functions
8093 would, and handles error conditions in the same way.
8095 '``llvm.ceil.*``' Intrinsic
8096 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8101 This is an overloaded intrinsic. You can use ``llvm.ceil`` on any
8102 floating point or vector of floating point type. Not all targets support
8107 declare float @llvm.ceil.f32(float %Val)
8108 declare double @llvm.ceil.f64(double %Val)
8109 declare x86_fp80 @llvm.ceil.f80(x86_fp80 %Val)
8110 declare fp128 @llvm.ceil.f128(fp128 %Val)
8111 declare ppc_fp128 @llvm.ceil.ppcf128(ppc_fp128 %Val)
8116 The '``llvm.ceil.*``' intrinsics return the ceiling of the operand.
8121 The argument and return value are floating point numbers of the same
8127 This function returns the same values as the libm ``ceil`` functions
8128 would, and handles error conditions in the same way.
8130 '``llvm.trunc.*``' Intrinsic
8131 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8136 This is an overloaded intrinsic. You can use ``llvm.trunc`` on any
8137 floating point or vector of floating point type. Not all targets support
8142 declare float @llvm.trunc.f32(float %Val)
8143 declare double @llvm.trunc.f64(double %Val)
8144 declare x86_fp80 @llvm.trunc.f80(x86_fp80 %Val)
8145 declare fp128 @llvm.trunc.f128(fp128 %Val)
8146 declare ppc_fp128 @llvm.trunc.ppcf128(ppc_fp128 %Val)
8151 The '``llvm.trunc.*``' intrinsics returns the operand rounded to the
8152 nearest integer not larger in magnitude than the operand.
8157 The argument and return value are floating point numbers of the same
8163 This function returns the same values as the libm ``trunc`` functions
8164 would, and handles error conditions in the same way.
8166 '``llvm.rint.*``' Intrinsic
8167 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8172 This is an overloaded intrinsic. You can use ``llvm.rint`` on any
8173 floating point or vector of floating point type. Not all targets support
8178 declare float @llvm.rint.f32(float %Val)
8179 declare double @llvm.rint.f64(double %Val)
8180 declare x86_fp80 @llvm.rint.f80(x86_fp80 %Val)
8181 declare fp128 @llvm.rint.f128(fp128 %Val)
8182 declare ppc_fp128 @llvm.rint.ppcf128(ppc_fp128 %Val)
8187 The '``llvm.rint.*``' intrinsics returns the operand rounded to the
8188 nearest integer. It may raise an inexact floating-point exception if the
8189 operand isn't an integer.
8194 The argument and return value are floating point numbers of the same
8200 This function returns the same values as the libm ``rint`` functions
8201 would, and handles error conditions in the same way.
8203 '``llvm.nearbyint.*``' Intrinsic
8204 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8209 This is an overloaded intrinsic. You can use ``llvm.nearbyint`` on any
8210 floating point or vector of floating point type. Not all targets support
8215 declare float @llvm.nearbyint.f32(float %Val)
8216 declare double @llvm.nearbyint.f64(double %Val)
8217 declare x86_fp80 @llvm.nearbyint.f80(x86_fp80 %Val)
8218 declare fp128 @llvm.nearbyint.f128(fp128 %Val)
8219 declare ppc_fp128 @llvm.nearbyint.ppcf128(ppc_fp128 %Val)
8224 The '``llvm.nearbyint.*``' intrinsics returns the operand rounded to the
8230 The argument and return value are floating point numbers of the same
8236 This function returns the same values as the libm ``nearbyint``
8237 functions would, and handles error conditions in the same way.
8239 '``llvm.round.*``' Intrinsic
8240 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8245 This is an overloaded intrinsic. You can use ``llvm.round`` on any
8246 floating point or vector of floating point type. Not all targets support
8251 declare float @llvm.round.f32(float %Val)
8252 declare double @llvm.round.f64(double %Val)
8253 declare x86_fp80 @llvm.round.f80(x86_fp80 %Val)
8254 declare fp128 @llvm.round.f128(fp128 %Val)
8255 declare ppc_fp128 @llvm.round.ppcf128(ppc_fp128 %Val)
8260 The '``llvm.round.*``' intrinsics returns the operand rounded to the
8266 The argument and return value are floating point numbers of the same
8272 This function returns the same values as the libm ``round``
8273 functions would, and handles error conditions in the same way.
8275 Bit Manipulation Intrinsics
8276 ---------------------------
8278 LLVM provides intrinsics for a few important bit manipulation
8279 operations. These allow efficient code generation for some algorithms.
8281 '``llvm.bswap.*``' Intrinsics
8282 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8287 This is an overloaded intrinsic function. You can use bswap on any
8288 integer type that is an even number of bytes (i.e. BitWidth % 16 == 0).
8292 declare i16 @llvm.bswap.i16(i16 <id>)
8293 declare i32 @llvm.bswap.i32(i32 <id>)
8294 declare i64 @llvm.bswap.i64(i64 <id>)
8299 The '``llvm.bswap``' family of intrinsics is used to byte swap integer
8300 values with an even number of bytes (positive multiple of 16 bits).
8301 These are useful for performing operations on data that is not in the
8302 target's native byte order.
8307 The ``llvm.bswap.i16`` intrinsic returns an i16 value that has the high
8308 and low byte of the input i16 swapped. Similarly, the ``llvm.bswap.i32``
8309 intrinsic returns an i32 value that has the four bytes of the input i32
8310 swapped, so that if the input bytes are numbered 0, 1, 2, 3 then the
8311 returned i32 will have its bytes in 3, 2, 1, 0 order. The
8312 ``llvm.bswap.i48``, ``llvm.bswap.i64`` and other intrinsics extend this
8313 concept to additional even-byte lengths (6 bytes, 8 bytes and more,
8316 '``llvm.ctpop.*``' Intrinsic
8317 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8322 This is an overloaded intrinsic. You can use llvm.ctpop on any integer
8323 bit width, or on any vector with integer elements. Not all targets
8324 support all bit widths or vector types, however.
8328 declare i8 @llvm.ctpop.i8(i8 <src>)
8329 declare i16 @llvm.ctpop.i16(i16 <src>)
8330 declare i32 @llvm.ctpop.i32(i32 <src>)
8331 declare i64 @llvm.ctpop.i64(i64 <src>)
8332 declare i256 @llvm.ctpop.i256(i256 <src>)
8333 declare <2 x i32> @llvm.ctpop.v2i32(<2 x i32> <src>)
8338 The '``llvm.ctpop``' family of intrinsics counts the number of bits set
8344 The only argument is the value to be counted. The argument may be of any
8345 integer type, or a vector with integer elements. The return type must
8346 match the argument type.
8351 The '``llvm.ctpop``' intrinsic counts the 1's in a variable, or within
8352 each element of a vector.
8354 '``llvm.ctlz.*``' Intrinsic
8355 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8360 This is an overloaded intrinsic. You can use ``llvm.ctlz`` on any
8361 integer bit width, or any vector whose elements are integers. Not all
8362 targets support all bit widths or vector types, however.
8366 declare i8 @llvm.ctlz.i8 (i8 <src>, i1 <is_zero_undef>)
8367 declare i16 @llvm.ctlz.i16 (i16 <src>, i1 <is_zero_undef>)
8368 declare i32 @llvm.ctlz.i32 (i32 <src>, i1 <is_zero_undef>)
8369 declare i64 @llvm.ctlz.i64 (i64 <src>, i1 <is_zero_undef>)
8370 declare i256 @llvm.ctlz.i256(i256 <src>, i1 <is_zero_undef>)
8371 declase <2 x i32> @llvm.ctlz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
8376 The '``llvm.ctlz``' family of intrinsic functions counts the number of
8377 leading zeros in a variable.
8382 The first argument is the value to be counted. This argument may be of
8383 any integer type, or a vectory with integer element type. The return
8384 type must match the first argument type.
8386 The second argument must be a constant and is a flag to indicate whether
8387 the intrinsic should ensure that a zero as the first argument produces a
8388 defined result. Historically some architectures did not provide a
8389 defined result for zero values as efficiently, and many algorithms are
8390 now predicated on avoiding zero-value inputs.
8395 The '``llvm.ctlz``' intrinsic counts the leading (most significant)
8396 zeros in a variable, or within each element of the vector. If
8397 ``src == 0`` then the result is the size in bits of the type of ``src``
8398 if ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
8399 ``llvm.ctlz(i32 2) = 30``.
8401 '``llvm.cttz.*``' Intrinsic
8402 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8407 This is an overloaded intrinsic. You can use ``llvm.cttz`` on any
8408 integer bit width, or any vector of integer elements. Not all targets
8409 support all bit widths or vector types, however.
8413 declare i8 @llvm.cttz.i8 (i8 <src>, i1 <is_zero_undef>)
8414 declare i16 @llvm.cttz.i16 (i16 <src>, i1 <is_zero_undef>)
8415 declare i32 @llvm.cttz.i32 (i32 <src>, i1 <is_zero_undef>)
8416 declare i64 @llvm.cttz.i64 (i64 <src>, i1 <is_zero_undef>)
8417 declare i256 @llvm.cttz.i256(i256 <src>, i1 <is_zero_undef>)
8418 declase <2 x i32> @llvm.cttz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
8423 The '``llvm.cttz``' family of intrinsic functions counts the number of
8429 The first argument is the value to be counted. This argument may be of
8430 any integer type, or a vectory with integer element type. The return
8431 type must match the first argument type.
8433 The second argument must be a constant and is a flag to indicate whether
8434 the intrinsic should ensure that a zero as the first argument produces a
8435 defined result. Historically some architectures did not provide a
8436 defined result for zero values as efficiently, and many algorithms are
8437 now predicated on avoiding zero-value inputs.
8442 The '``llvm.cttz``' intrinsic counts the trailing (least significant)
8443 zeros in a variable, or within each element of a vector. If ``src == 0``
8444 then the result is the size in bits of the type of ``src`` if
8445 ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
8446 ``llvm.cttz(2) = 1``.
8448 Arithmetic with Overflow Intrinsics
8449 -----------------------------------
8451 LLVM provides intrinsics for some arithmetic with overflow operations.
8453 '``llvm.sadd.with.overflow.*``' Intrinsics
8454 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8459 This is an overloaded intrinsic. You can use ``llvm.sadd.with.overflow``
8460 on any integer bit width.
8464 declare {i16, i1} @llvm.sadd.with.overflow.i16(i16 %a, i16 %b)
8465 declare {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
8466 declare {i64, i1} @llvm.sadd.with.overflow.i64(i64 %a, i64 %b)
8471 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
8472 a signed addition of the two arguments, and indicate whether an overflow
8473 occurred during the signed summation.
8478 The arguments (%a and %b) and the first element of the result structure
8479 may be of integer types of any bit width, but they must have the same
8480 bit width. The second element of the result structure must be of type
8481 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
8487 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
8488 a signed addition of the two variables. They return a structure --- the
8489 first element of which is the signed summation, and the second element
8490 of which is a bit specifying if the signed summation resulted in an
8496 .. code-block:: llvm
8498 %res = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
8499 %sum = extractvalue {i32, i1} %res, 0
8500 %obit = extractvalue {i32, i1} %res, 1
8501 br i1 %obit, label %overflow, label %normal
8503 '``llvm.uadd.with.overflow.*``' Intrinsics
8504 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8509 This is an overloaded intrinsic. You can use ``llvm.uadd.with.overflow``
8510 on any integer bit width.
8514 declare {i16, i1} @llvm.uadd.with.overflow.i16(i16 %a, i16 %b)
8515 declare {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
8516 declare {i64, i1} @llvm.uadd.with.overflow.i64(i64 %a, i64 %b)
8521 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
8522 an unsigned addition of the two arguments, and indicate whether a carry
8523 occurred during the unsigned summation.
8528 The arguments (%a and %b) and the first element of the result structure
8529 may be of integer types of any bit width, but they must have the same
8530 bit width. The second element of the result structure must be of type
8531 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8537 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
8538 an unsigned addition of the two arguments. They return a structure --- the
8539 first element of which is the sum, and the second element of which is a
8540 bit specifying if the unsigned summation resulted in a carry.
8545 .. code-block:: llvm
8547 %res = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
8548 %sum = extractvalue {i32, i1} %res, 0
8549 %obit = extractvalue {i32, i1} %res, 1
8550 br i1 %obit, label %carry, label %normal
8552 '``llvm.ssub.with.overflow.*``' Intrinsics
8553 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8558 This is an overloaded intrinsic. You can use ``llvm.ssub.with.overflow``
8559 on any integer bit width.
8563 declare {i16, i1} @llvm.ssub.with.overflow.i16(i16 %a, i16 %b)
8564 declare {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
8565 declare {i64, i1} @llvm.ssub.with.overflow.i64(i64 %a, i64 %b)
8570 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
8571 a signed subtraction of the two arguments, and indicate whether an
8572 overflow occurred during the signed subtraction.
8577 The arguments (%a and %b) and the first element of the result structure
8578 may be of integer types of any bit width, but they must have the same
8579 bit width. The second element of the result structure must be of type
8580 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
8586 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
8587 a signed subtraction of the two arguments. They return a structure --- the
8588 first element of which is the subtraction, and the second element of
8589 which is a bit specifying if the signed subtraction resulted in an
8595 .. code-block:: llvm
8597 %res = call {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
8598 %sum = extractvalue {i32, i1} %res, 0
8599 %obit = extractvalue {i32, i1} %res, 1
8600 br i1 %obit, label %overflow, label %normal
8602 '``llvm.usub.with.overflow.*``' Intrinsics
8603 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8608 This is an overloaded intrinsic. You can use ``llvm.usub.with.overflow``
8609 on any integer bit width.
8613 declare {i16, i1} @llvm.usub.with.overflow.i16(i16 %a, i16 %b)
8614 declare {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
8615 declare {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
8620 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
8621 an unsigned subtraction of the two arguments, and indicate whether an
8622 overflow occurred during the unsigned subtraction.
8627 The arguments (%a and %b) and the first element of the result structure
8628 may be of integer types of any bit width, but they must have the same
8629 bit width. The second element of the result structure must be of type
8630 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8636 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
8637 an unsigned subtraction of the two arguments. They return a structure ---
8638 the first element of which is the subtraction, and the second element of
8639 which is a bit specifying if the unsigned subtraction resulted in an
8645 .. code-block:: llvm
8647 %res = call {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
8648 %sum = extractvalue {i32, i1} %res, 0
8649 %obit = extractvalue {i32, i1} %res, 1
8650 br i1 %obit, label %overflow, label %normal
8652 '``llvm.smul.with.overflow.*``' Intrinsics
8653 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8658 This is an overloaded intrinsic. You can use ``llvm.smul.with.overflow``
8659 on any integer bit width.
8663 declare {i16, i1} @llvm.smul.with.overflow.i16(i16 %a, i16 %b)
8664 declare {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
8665 declare {i64, i1} @llvm.smul.with.overflow.i64(i64 %a, i64 %b)
8670 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
8671 a signed multiplication of the two arguments, and indicate whether an
8672 overflow occurred during the signed multiplication.
8677 The arguments (%a and %b) and the first element of the result structure
8678 may be of integer types of any bit width, but they must have the same
8679 bit width. The second element of the result structure must be of type
8680 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
8686 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
8687 a signed multiplication of the two arguments. They return a structure ---
8688 the first element of which is the multiplication, and the second element
8689 of which is a bit specifying if the signed multiplication resulted in an
8695 .. code-block:: llvm
8697 %res = call {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
8698 %sum = extractvalue {i32, i1} %res, 0
8699 %obit = extractvalue {i32, i1} %res, 1
8700 br i1 %obit, label %overflow, label %normal
8702 '``llvm.umul.with.overflow.*``' Intrinsics
8703 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8708 This is an overloaded intrinsic. You can use ``llvm.umul.with.overflow``
8709 on any integer bit width.
8713 declare {i16, i1} @llvm.umul.with.overflow.i16(i16 %a, i16 %b)
8714 declare {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
8715 declare {i64, i1} @llvm.umul.with.overflow.i64(i64 %a, i64 %b)
8720 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
8721 a unsigned multiplication of the two arguments, and indicate whether an
8722 overflow occurred during the unsigned multiplication.
8727 The arguments (%a and %b) and the first element of the result structure
8728 may be of integer types of any bit width, but they must have the same
8729 bit width. The second element of the result structure must be of type
8730 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8736 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
8737 an unsigned multiplication of the two arguments. They return a structure ---
8738 the first element of which is the multiplication, and the second
8739 element of which is a bit specifying if the unsigned multiplication
8740 resulted in an overflow.
8745 .. code-block:: llvm
8747 %res = call {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
8748 %sum = extractvalue {i32, i1} %res, 0
8749 %obit = extractvalue {i32, i1} %res, 1
8750 br i1 %obit, label %overflow, label %normal
8752 Specialised Arithmetic Intrinsics
8753 ---------------------------------
8755 '``llvm.fmuladd.*``' Intrinsic
8756 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8763 declare float @llvm.fmuladd.f32(float %a, float %b, float %c)
8764 declare double @llvm.fmuladd.f64(double %a, double %b, double %c)
8769 The '``llvm.fmuladd.*``' intrinsic functions represent multiply-add
8770 expressions that can be fused if the code generator determines that (a) the
8771 target instruction set has support for a fused operation, and (b) that the
8772 fused operation is more efficient than the equivalent, separate pair of mul
8773 and add instructions.
8778 The '``llvm.fmuladd.*``' intrinsics each take three arguments: two
8779 multiplicands, a and b, and an addend c.
8788 %0 = call float @llvm.fmuladd.f32(%a, %b, %c)
8790 is equivalent to the expression a \* b + c, except that rounding will
8791 not be performed between the multiplication and addition steps if the
8792 code generator fuses the operations. Fusion is not guaranteed, even if
8793 the target platform supports it. If a fused multiply-add is required the
8794 corresponding llvm.fma.\* intrinsic function should be used
8795 instead. This never sets errno, just as '``llvm.fma.*``'.
8800 .. code-block:: llvm
8802 %r2 = call float @llvm.fmuladd.f32(float %a, float %b, float %c) ; yields float:r2 = (a * b) + c
8804 Half Precision Floating Point Intrinsics
8805 ----------------------------------------
8807 For most target platforms, half precision floating point is a
8808 storage-only format. This means that it is a dense encoding (in memory)
8809 but does not support computation in the format.
8811 This means that code must first load the half-precision floating point
8812 value as an i16, then convert it to float with
8813 :ref:`llvm.convert.from.fp16 <int_convert_from_fp16>`. Computation can
8814 then be performed on the float value (including extending to double
8815 etc). To store the value back to memory, it is first converted to float
8816 if needed, then converted to i16 with
8817 :ref:`llvm.convert.to.fp16 <int_convert_to_fp16>`, then storing as an
8820 .. _int_convert_to_fp16:
8822 '``llvm.convert.to.fp16``' Intrinsic
8823 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8830 declare i16 @llvm.convert.to.fp16.f32(float %a)
8831 declare i16 @llvm.convert.to.fp16.f64(double %a)
8836 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion from a
8837 conventional floating point type to half precision floating point format.
8842 The intrinsic function contains single argument - the value to be
8848 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion from a
8849 conventional floating point format to half precision floating point format. The
8850 return value is an ``i16`` which contains the converted number.
8855 .. code-block:: llvm
8857 %res = call i16 @llvm.convert.to.fp16.f32(float %a)
8858 store i16 %res, i16* @x, align 2
8860 .. _int_convert_from_fp16:
8862 '``llvm.convert.from.fp16``' Intrinsic
8863 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8870 declare float @llvm.convert.from.fp16.f32(i16 %a)
8871 declare double @llvm.convert.from.fp16.f64(i16 %a)
8876 The '``llvm.convert.from.fp16``' intrinsic function performs a
8877 conversion from half precision floating point format to single precision
8878 floating point format.
8883 The intrinsic function contains single argument - the value to be
8889 The '``llvm.convert.from.fp16``' intrinsic function performs a
8890 conversion from half single precision floating point format to single
8891 precision floating point format. The input half-float value is
8892 represented by an ``i16`` value.
8897 .. code-block:: llvm
8899 %a = load i16* @x, align 2
8900 %res = call float @llvm.convert.from.fp16(i16 %a)
8905 The LLVM debugger intrinsics (which all start with ``llvm.dbg.``
8906 prefix), are described in the `LLVM Source Level
8907 Debugging <SourceLevelDebugging.html#format_common_intrinsics>`_
8910 Exception Handling Intrinsics
8911 -----------------------------
8913 The LLVM exception handling intrinsics (which all start with
8914 ``llvm.eh.`` prefix), are described in the `LLVM Exception
8915 Handling <ExceptionHandling.html#format_common_intrinsics>`_ document.
8919 Trampoline Intrinsics
8920 ---------------------
8922 These intrinsics make it possible to excise one parameter, marked with
8923 the :ref:`nest <nest>` attribute, from a function. The result is a
8924 callable function pointer lacking the nest parameter - the caller does
8925 not need to provide a value for it. Instead, the value to use is stored
8926 in advance in a "trampoline", a block of memory usually allocated on the
8927 stack, which also contains code to splice the nest value into the
8928 argument list. This is used to implement the GCC nested function address
8931 For example, if the function is ``i32 f(i8* nest %c, i32 %x, i32 %y)``
8932 then the resulting function pointer has signature ``i32 (i32, i32)*``.
8933 It can be created as follows:
8935 .. code-block:: llvm
8937 %tramp = alloca [10 x i8], align 4 ; size and alignment only correct for X86
8938 %tramp1 = getelementptr [10 x i8]* %tramp, i32 0, i32 0
8939 call i8* @llvm.init.trampoline(i8* %tramp1, i8* bitcast (i32 (i8*, i32, i32)* @f to i8*), i8* %nval)
8940 %p = call i8* @llvm.adjust.trampoline(i8* %tramp1)
8941 %fp = bitcast i8* %p to i32 (i32, i32)*
8943 The call ``%val = call i32 %fp(i32 %x, i32 %y)`` is then equivalent to
8944 ``%val = call i32 %f(i8* %nval, i32 %x, i32 %y)``.
8948 '``llvm.init.trampoline``' Intrinsic
8949 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8956 declare void @llvm.init.trampoline(i8* <tramp>, i8* <func>, i8* <nval>)
8961 This fills the memory pointed to by ``tramp`` with executable code,
8962 turning it into a trampoline.
8967 The ``llvm.init.trampoline`` intrinsic takes three arguments, all
8968 pointers. The ``tramp`` argument must point to a sufficiently large and
8969 sufficiently aligned block of memory; this memory is written to by the
8970 intrinsic. Note that the size and the alignment are target-specific -
8971 LLVM currently provides no portable way of determining them, so a
8972 front-end that generates this intrinsic needs to have some
8973 target-specific knowledge. The ``func`` argument must hold a function
8974 bitcast to an ``i8*``.
8979 The block of memory pointed to by ``tramp`` is filled with target
8980 dependent code, turning it into a function. Then ``tramp`` needs to be
8981 passed to :ref:`llvm.adjust.trampoline <int_at>` to get a pointer which can
8982 be :ref:`bitcast (to a new function) and called <int_trampoline>`. The new
8983 function's signature is the same as that of ``func`` with any arguments
8984 marked with the ``nest`` attribute removed. At most one such ``nest``
8985 argument is allowed, and it must be of pointer type. Calling the new
8986 function is equivalent to calling ``func`` with the same argument list,
8987 but with ``nval`` used for the missing ``nest`` argument. If, after
8988 calling ``llvm.init.trampoline``, the memory pointed to by ``tramp`` is
8989 modified, then the effect of any later call to the returned function
8990 pointer is undefined.
8994 '``llvm.adjust.trampoline``' Intrinsic
8995 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9002 declare i8* @llvm.adjust.trampoline(i8* <tramp>)
9007 This performs any required machine-specific adjustment to the address of
9008 a trampoline (passed as ``tramp``).
9013 ``tramp`` must point to a block of memory which already has trampoline
9014 code filled in by a previous call to
9015 :ref:`llvm.init.trampoline <int_it>`.
9020 On some architectures the address of the code to be executed needs to be
9021 different than the address where the trampoline is actually stored. This
9022 intrinsic returns the executable address corresponding to ``tramp``
9023 after performing the required machine specific adjustments. The pointer
9024 returned can then be :ref:`bitcast and executed <int_trampoline>`.
9029 This class of intrinsics provides information about the lifetime of
9030 memory objects and ranges where variables are immutable.
9034 '``llvm.lifetime.start``' Intrinsic
9035 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9042 declare void @llvm.lifetime.start(i64 <size>, i8* nocapture <ptr>)
9047 The '``llvm.lifetime.start``' intrinsic specifies the start of a memory
9053 The first argument is a constant integer representing the size of the
9054 object, or -1 if it is variable sized. The second argument is a pointer
9060 This intrinsic indicates that before this point in the code, the value
9061 of the memory pointed to by ``ptr`` is dead. This means that it is known
9062 to never be used and has an undefined value. A load from the pointer
9063 that precedes this intrinsic can be replaced with ``'undef'``.
9067 '``llvm.lifetime.end``' Intrinsic
9068 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9075 declare void @llvm.lifetime.end(i64 <size>, i8* nocapture <ptr>)
9080 The '``llvm.lifetime.end``' intrinsic specifies the end of a memory
9086 The first argument is a constant integer representing the size of the
9087 object, or -1 if it is variable sized. The second argument is a pointer
9093 This intrinsic indicates that after this point in the code, the value of
9094 the memory pointed to by ``ptr`` is dead. This means that it is known to
9095 never be used and has an undefined value. Any stores into the memory
9096 object following this intrinsic may be removed as dead.
9098 '``llvm.invariant.start``' Intrinsic
9099 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9106 declare {}* @llvm.invariant.start(i64 <size>, i8* nocapture <ptr>)
9111 The '``llvm.invariant.start``' intrinsic specifies that the contents of
9112 a memory object will not change.
9117 The first argument is a constant integer representing the size of the
9118 object, or -1 if it is variable sized. The second argument is a pointer
9124 This intrinsic indicates that until an ``llvm.invariant.end`` that uses
9125 the return value, the referenced memory location is constant and
9128 '``llvm.invariant.end``' Intrinsic
9129 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9136 declare void @llvm.invariant.end({}* <start>, i64 <size>, i8* nocapture <ptr>)
9141 The '``llvm.invariant.end``' intrinsic specifies that the contents of a
9142 memory object are mutable.
9147 The first argument is the matching ``llvm.invariant.start`` intrinsic.
9148 The second argument is a constant integer representing the size of the
9149 object, or -1 if it is variable sized and the third argument is a
9150 pointer to the object.
9155 This intrinsic indicates that the memory is mutable again.
9160 This class of intrinsics is designed to be generic and has no specific
9163 '``llvm.var.annotation``' Intrinsic
9164 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9171 declare void @llvm.var.annotation(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
9176 The '``llvm.var.annotation``' intrinsic.
9181 The first argument is a pointer to a value, the second is a pointer to a
9182 global string, the third is a pointer to a global string which is the
9183 source file name, and the last argument is the line number.
9188 This intrinsic allows annotation of local variables with arbitrary
9189 strings. This can be useful for special purpose optimizations that want
9190 to look for these annotations. These have no other defined use; they are
9191 ignored by code generation and optimization.
9193 '``llvm.ptr.annotation.*``' Intrinsic
9194 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9199 This is an overloaded intrinsic. You can use '``llvm.ptr.annotation``' on a
9200 pointer to an integer of any width. *NOTE* you must specify an address space for
9201 the pointer. The identifier for the default address space is the integer
9206 declare i8* @llvm.ptr.annotation.p<address space>i8(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
9207 declare i16* @llvm.ptr.annotation.p<address space>i16(i16* <val>, i8* <str>, i8* <str>, i32 <int>)
9208 declare i32* @llvm.ptr.annotation.p<address space>i32(i32* <val>, i8* <str>, i8* <str>, i32 <int>)
9209 declare i64* @llvm.ptr.annotation.p<address space>i64(i64* <val>, i8* <str>, i8* <str>, i32 <int>)
9210 declare i256* @llvm.ptr.annotation.p<address space>i256(i256* <val>, i8* <str>, i8* <str>, i32 <int>)
9215 The '``llvm.ptr.annotation``' intrinsic.
9220 The first argument is a pointer to an integer value of arbitrary bitwidth
9221 (result of some expression), the second is a pointer to a global string, the
9222 third is a pointer to a global string which is the source file name, and the
9223 last argument is the line number. It returns the value of the first argument.
9228 This intrinsic allows annotation of a pointer to an integer with arbitrary
9229 strings. This can be useful for special purpose optimizations that want to look
9230 for these annotations. These have no other defined use; they are ignored by code
9231 generation and optimization.
9233 '``llvm.annotation.*``' Intrinsic
9234 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9239 This is an overloaded intrinsic. You can use '``llvm.annotation``' on
9240 any integer bit width.
9244 declare i8 @llvm.annotation.i8(i8 <val>, i8* <str>, i8* <str>, i32 <int>)
9245 declare i16 @llvm.annotation.i16(i16 <val>, i8* <str>, i8* <str>, i32 <int>)
9246 declare i32 @llvm.annotation.i32(i32 <val>, i8* <str>, i8* <str>, i32 <int>)
9247 declare i64 @llvm.annotation.i64(i64 <val>, i8* <str>, i8* <str>, i32 <int>)
9248 declare i256 @llvm.annotation.i256(i256 <val>, i8* <str>, i8* <str>, i32 <int>)
9253 The '``llvm.annotation``' intrinsic.
9258 The first argument is an integer value (result of some expression), the
9259 second is a pointer to a global string, the third is a pointer to a
9260 global string which is the source file name, and the last argument is
9261 the line number. It returns the value of the first argument.
9266 This intrinsic allows annotations to be put on arbitrary expressions
9267 with arbitrary strings. This can be useful for special purpose
9268 optimizations that want to look for these annotations. These have no
9269 other defined use; they are ignored by code generation and optimization.
9271 '``llvm.trap``' Intrinsic
9272 ^^^^^^^^^^^^^^^^^^^^^^^^^
9279 declare void @llvm.trap() noreturn nounwind
9284 The '``llvm.trap``' intrinsic.
9294 This intrinsic is lowered to the target dependent trap instruction. If
9295 the target does not have a trap instruction, this intrinsic will be
9296 lowered to a call of the ``abort()`` function.
9298 '``llvm.debugtrap``' Intrinsic
9299 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9306 declare void @llvm.debugtrap() nounwind
9311 The '``llvm.debugtrap``' intrinsic.
9321 This intrinsic is lowered to code which is intended to cause an
9322 execution trap with the intention of requesting the attention of a
9325 '``llvm.stackprotector``' Intrinsic
9326 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9333 declare void @llvm.stackprotector(i8* <guard>, i8** <slot>)
9338 The ``llvm.stackprotector`` intrinsic takes the ``guard`` and stores it
9339 onto the stack at ``slot``. The stack slot is adjusted to ensure that it
9340 is placed on the stack before local variables.
9345 The ``llvm.stackprotector`` intrinsic requires two pointer arguments.
9346 The first argument is the value loaded from the stack guard
9347 ``@__stack_chk_guard``. The second variable is an ``alloca`` that has
9348 enough space to hold the value of the guard.
9353 This intrinsic causes the prologue/epilogue inserter to force the position of
9354 the ``AllocaInst`` stack slot to be before local variables on the stack. This is
9355 to ensure that if a local variable on the stack is overwritten, it will destroy
9356 the value of the guard. When the function exits, the guard on the stack is
9357 checked against the original guard by ``llvm.stackprotectorcheck``. If they are
9358 different, then ``llvm.stackprotectorcheck`` causes the program to abort by
9359 calling the ``__stack_chk_fail()`` function.
9361 '``llvm.stackprotectorcheck``' Intrinsic
9362 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9369 declare void @llvm.stackprotectorcheck(i8** <guard>)
9374 The ``llvm.stackprotectorcheck`` intrinsic compares ``guard`` against an already
9375 created stack protector and if they are not equal calls the
9376 ``__stack_chk_fail()`` function.
9381 The ``llvm.stackprotectorcheck`` intrinsic requires one pointer argument, the
9382 the variable ``@__stack_chk_guard``.
9387 This intrinsic is provided to perform the stack protector check by comparing
9388 ``guard`` with the stack slot created by ``llvm.stackprotector`` and if the
9389 values do not match call the ``__stack_chk_fail()`` function.
9391 The reason to provide this as an IR level intrinsic instead of implementing it
9392 via other IR operations is that in order to perform this operation at the IR
9393 level without an intrinsic, one would need to create additional basic blocks to
9394 handle the success/failure cases. This makes it difficult to stop the stack
9395 protector check from disrupting sibling tail calls in Codegen. With this
9396 intrinsic, we are able to generate the stack protector basic blocks late in
9397 codegen after the tail call decision has occurred.
9399 '``llvm.objectsize``' Intrinsic
9400 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9407 declare i32 @llvm.objectsize.i32(i8* <object>, i1 <min>)
9408 declare i64 @llvm.objectsize.i64(i8* <object>, i1 <min>)
9413 The ``llvm.objectsize`` intrinsic is designed to provide information to
9414 the optimizers to determine at compile time whether a) an operation
9415 (like memcpy) will overflow a buffer that corresponds to an object, or
9416 b) that a runtime check for overflow isn't necessary. An object in this
9417 context means an allocation of a specific class, structure, array, or
9423 The ``llvm.objectsize`` intrinsic takes two arguments. The first
9424 argument is a pointer to or into the ``object``. The second argument is
9425 a boolean and determines whether ``llvm.objectsize`` returns 0 (if true)
9426 or -1 (if false) when the object size is unknown. The second argument
9427 only accepts constants.
9432 The ``llvm.objectsize`` intrinsic is lowered to a constant representing
9433 the size of the object concerned. If the size cannot be determined at
9434 compile time, ``llvm.objectsize`` returns ``i32/i64 -1 or 0`` (depending
9435 on the ``min`` argument).
9437 '``llvm.expect``' Intrinsic
9438 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9443 This is an overloaded intrinsic. You can use ``llvm.expect`` on any
9448 declare i1 @llvm.expect.i1(i1 <val>, i1 <expected_val>)
9449 declare i32 @llvm.expect.i32(i32 <val>, i32 <expected_val>)
9450 declare i64 @llvm.expect.i64(i64 <val>, i64 <expected_val>)
9455 The ``llvm.expect`` intrinsic provides information about expected (the
9456 most probable) value of ``val``, which can be used by optimizers.
9461 The ``llvm.expect`` intrinsic takes two arguments. The first argument is
9462 a value. The second argument is an expected value, this needs to be a
9463 constant value, variables are not allowed.
9468 This intrinsic is lowered to the ``val``.
9470 '``llvm.assume``' Intrinsic
9471 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9478 declare void @llvm.assume(i1 %cond)
9483 The ``llvm.assume`` allows the optimizer to assume that the provided
9484 condition is true. This information can then be used in simplifying other parts
9490 The condition which the optimizer may assume is always true.
9495 The intrinsic allows the optimizer to assume that the provided condition is
9496 always true whenever the control flow reaches the intrinsic call. No code is
9497 generated for this intrinsic, and instructions that contribute only to the
9498 provided condition are not used for code generation. If the condition is
9499 violated during execution, the behavior is undefined.
9501 Please note that optimizer might limit the transformations performed on values
9502 used by the ``llvm.assume`` intrinsic in order to preserve the instructions
9503 only used to form the intrinsic's input argument. This might prove undesirable
9504 if the extra information provided by the ``llvm.assume`` intrinsic does cause
9505 sufficient overall improvement in code quality. For this reason,
9506 ``llvm.assume`` should not be used to document basic mathematical invariants
9507 that the optimizer can otherwise deduce or facts that are of little use to the
9510 '``llvm.donothing``' Intrinsic
9511 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9518 declare void @llvm.donothing() nounwind readnone
9523 The ``llvm.donothing`` intrinsic doesn't perform any operation. It's the
9524 only intrinsic that can be called with an invoke instruction.
9534 This intrinsic does nothing, and it's removed by optimizers and ignored
9537 Stack Map Intrinsics
9538 --------------------
9540 LLVM provides experimental intrinsics to support runtime patching
9541 mechanisms commonly desired in dynamic language JITs. These intrinsics
9542 are described in :doc:`StackMaps`.