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 and unnamed
133 function parameters are included in this numbering. For example, if the
134 entry basic block is not given a label name and all function parameters are
135 named, then it will get number 0.
137 It also shows a convention that we follow in this document. When
138 demonstrating instructions, we will follow an instruction with a comment
139 that defines the type and name of value produced.
147 LLVM programs are composed of ``Module``'s, each of which is a
148 translation unit of the input programs. Each module consists of
149 functions, global variables, and symbol table entries. Modules may be
150 combined together with the LLVM linker, which merges function (and
151 global variable) definitions, resolves forward declarations, and merges
152 symbol table entries. Here is an example of the "hello world" module:
156 ; Declare the string constant as a global constant.
157 @.str = private unnamed_addr constant [13 x i8] c"hello world\0A\00"
159 ; External declaration of the puts function
160 declare i32 @puts(i8* nocapture) nounwind
162 ; Definition of main function
163 define i32 @main() { ; i32()*
164 ; Convert [13 x i8]* to i8 *...
165 %cast210 = getelementptr [13 x i8], [13 x i8]* @.str, i64 0, i64 0
167 ; Call puts function to write out the string to stdout.
168 call i32 @puts(i8* %cast210)
173 !0 = !{i32 42, null, !"string"}
176 This example is made up of a :ref:`global variable <globalvars>` named
177 "``.str``", an external declaration of the "``puts``" function, a
178 :ref:`function definition <functionstructure>` for "``main``" and
179 :ref:`named metadata <namedmetadatastructure>` "``foo``".
181 In general, a module is made up of a list of global values (where both
182 functions and global variables are global values). Global values are
183 represented by a pointer to a memory location (in this case, a pointer
184 to an array of char, and a pointer to a function), and have one of the
185 following :ref:`linkage types <linkage>`.
192 All Global Variables and Functions have one of the following types of
196 Global values with "``private``" linkage are only directly
197 accessible by objects in the current module. In particular, linking
198 code into a module with an private global value may cause the
199 private to be renamed as necessary to avoid collisions. Because the
200 symbol is private to the module, all references can be updated. This
201 doesn't show up in any symbol table in the object file.
203 Similar to private, but the value shows as a local symbol
204 (``STB_LOCAL`` in the case of ELF) in the object file. This
205 corresponds to the notion of the '``static``' keyword in C.
206 ``available_externally``
207 Globals with "``available_externally``" linkage are never emitted
208 into the object file corresponding to the LLVM module. They exist to
209 allow inlining and other optimizations to take place given knowledge
210 of the definition of the global, which is known to be somewhere
211 outside the module. Globals with ``available_externally`` linkage
212 are allowed to be discarded at will, and are otherwise the same as
213 ``linkonce_odr``. This linkage type is only allowed on definitions,
216 Globals with "``linkonce``" linkage are merged with other globals of
217 the same name when linkage occurs. This can be used to implement
218 some forms of inline functions, templates, or other code which must
219 be generated in each translation unit that uses it, but where the
220 body may be overridden with a more definitive definition later.
221 Unreferenced ``linkonce`` globals are allowed to be discarded. Note
222 that ``linkonce`` linkage does not actually allow the optimizer to
223 inline the body of this function into callers because it doesn't
224 know if this definition of the function is the definitive definition
225 within the program or whether it will be overridden by a stronger
226 definition. To enable inlining and other optimizations, use
227 "``linkonce_odr``" linkage.
229 "``weak``" linkage has the same merging semantics as ``linkonce``
230 linkage, except that unreferenced globals with ``weak`` linkage may
231 not be discarded. This is used for globals that are declared "weak"
234 "``common``" linkage is most similar to "``weak``" linkage, but they
235 are used for tentative definitions in C, such as "``int X;``" at
236 global scope. Symbols with "``common``" linkage are merged in the
237 same way as ``weak symbols``, and they may not be deleted if
238 unreferenced. ``common`` symbols may not have an explicit section,
239 must have a zero initializer, and may not be marked
240 ':ref:`constant <globalvars>`'. Functions and aliases may not have
243 .. _linkage_appending:
246 "``appending``" linkage may only be applied to global variables of
247 pointer to array type. When two global variables with appending
248 linkage are linked together, the two global arrays are appended
249 together. This is the LLVM, typesafe, equivalent of having the
250 system linker append together "sections" with identical names when
253 The semantics of this linkage follow the ELF object file model: the
254 symbol is weak until linked, if not linked, the symbol becomes null
255 instead of being an undefined reference.
256 ``linkonce_odr``, ``weak_odr``
257 Some languages allow differing globals to be merged, such as two
258 functions with different semantics. Other languages, such as
259 ``C++``, ensure that only equivalent globals are ever merged (the
260 "one definition rule" --- "ODR"). Such languages can use the
261 ``linkonce_odr`` and ``weak_odr`` linkage types to indicate that the
262 global will only be merged with equivalent globals. These linkage
263 types are otherwise the same as their non-``odr`` versions.
265 If none of the above identifiers are used, the global is externally
266 visible, meaning that it participates in linkage and can be used to
267 resolve external symbol references.
269 It is illegal for a function *declaration* to have any linkage type
270 other than ``external`` or ``extern_weak``.
277 LLVM :ref:`functions <functionstructure>`, :ref:`calls <i_call>` and
278 :ref:`invokes <i_invoke>` can all have an optional calling convention
279 specified for the call. The calling convention of any pair of dynamic
280 caller/callee must match, or the behavior of the program is undefined.
281 The following calling conventions are supported by LLVM, and more may be
284 "``ccc``" - The C calling convention
285 This calling convention (the default if no other calling convention
286 is specified) matches the target C calling conventions. This calling
287 convention supports varargs function calls and tolerates some
288 mismatch in the declared prototype and implemented declaration of
289 the function (as does normal C).
290 "``fastcc``" - The fast calling convention
291 This calling convention attempts to make calls as fast as possible
292 (e.g. by passing things in registers). This calling convention
293 allows the target to use whatever tricks it wants to produce fast
294 code for the target, without having to conform to an externally
295 specified ABI (Application Binary Interface). `Tail calls can only
296 be optimized when this, the GHC or the HiPE convention is
297 used. <CodeGenerator.html#id80>`_ This calling convention does not
298 support varargs and requires the prototype of all callees to exactly
299 match the prototype of the function definition.
300 "``coldcc``" - The cold calling convention
301 This calling convention attempts to make code in the caller as
302 efficient as possible under the assumption that the call is not
303 commonly executed. As such, these calls often preserve all registers
304 so that the call does not break any live ranges in the caller side.
305 This calling convention does not support varargs and requires the
306 prototype of all callees to exactly match the prototype of the
307 function definition. Furthermore the inliner doesn't consider such function
309 "``cc 10``" - GHC convention
310 This calling convention has been implemented specifically for use by
311 the `Glasgow Haskell Compiler (GHC) <http://www.haskell.org/ghc>`_.
312 It passes everything in registers, going to extremes to achieve this
313 by disabling callee save registers. This calling convention should
314 not be used lightly but only for specific situations such as an
315 alternative to the *register pinning* performance technique often
316 used when implementing functional programming languages. At the
317 moment only X86 supports this convention and it has the following
320 - On *X86-32* only supports up to 4 bit type parameters. No
321 floating point types are supported.
322 - On *X86-64* only supports up to 10 bit type parameters and 6
323 floating point parameters.
325 This calling convention supports `tail call
326 optimization <CodeGenerator.html#id80>`_ but requires both the
327 caller and callee are using it.
328 "``cc 11``" - The HiPE calling convention
329 This calling convention has been implemented specifically for use by
330 the `High-Performance Erlang
331 (HiPE) <http://www.it.uu.se/research/group/hipe/>`_ compiler, *the*
332 native code compiler of the `Ericsson's Open Source Erlang/OTP
333 system <http://www.erlang.org/download.shtml>`_. It uses more
334 registers for argument passing than the ordinary C calling
335 convention and defines no callee-saved registers. The calling
336 convention properly supports `tail call
337 optimization <CodeGenerator.html#id80>`_ but requires that both the
338 caller and the callee use it. It uses a *register pinning*
339 mechanism, similar to GHC's convention, for keeping frequently
340 accessed runtime components pinned to specific hardware registers.
341 At the moment only X86 supports this convention (both 32 and 64
343 "``webkit_jscc``" - WebKit's JavaScript calling convention
344 This calling convention has been implemented for `WebKit FTL JIT
345 <https://trac.webkit.org/wiki/FTLJIT>`_. It passes arguments on the
346 stack right to left (as cdecl does), and returns a value in the
347 platform's customary return register.
348 "``anyregcc``" - Dynamic calling convention for code patching
349 This is a special convention that supports patching an arbitrary code
350 sequence in place of a call site. This convention forces the call
351 arguments into registers but allows them to be dynamically
352 allocated. This can currently only be used with calls to
353 llvm.experimental.patchpoint because only this intrinsic records
354 the location of its arguments in a side table. See :doc:`StackMaps`.
355 "``preserve_mostcc``" - The `PreserveMost` calling convention
356 This calling convention attempts to make the code in the caller as
357 unintrusive as possible. This convention behaves identically to the `C`
358 calling convention on how arguments and return values are passed, but it
359 uses a different set of caller/callee-saved registers. This alleviates the
360 burden of saving and recovering a large register set before and after the
361 call in the caller. If the arguments are passed in callee-saved registers,
362 then they will be preserved by the callee across the call. This doesn't
363 apply for values returned in callee-saved registers.
365 - On X86-64 the callee preserves all general purpose registers, except for
366 R11. R11 can be used as a scratch register. Floating-point registers
367 (XMMs/YMMs) are not preserved and need to be saved by the caller.
369 The idea behind this convention is to support calls to runtime functions
370 that have a hot path and a cold path. The hot path is usually a small piece
371 of code that doesn't use many registers. The cold path might need to call out to
372 another function and therefore only needs to preserve the caller-saved
373 registers, which haven't already been saved by the caller. The
374 `PreserveMost` calling convention is very similar to the `cold` calling
375 convention in terms of caller/callee-saved registers, but they are used for
376 different types of function calls. `coldcc` is for function calls that are
377 rarely executed, whereas `preserve_mostcc` function calls are intended to be
378 on the hot path and definitely executed a lot. Furthermore `preserve_mostcc`
379 doesn't prevent the inliner from inlining the function call.
381 This calling convention will be used by a future version of the ObjectiveC
382 runtime and should therefore still be considered experimental at this time.
383 Although this convention was created to optimize certain runtime calls to
384 the ObjectiveC runtime, it is not limited to this runtime and might be used
385 by other runtimes in the future too. The current implementation only
386 supports X86-64, but the intention is to support more architectures in the
388 "``preserve_allcc``" - The `PreserveAll` calling convention
389 This calling convention attempts to make the code in the caller even less
390 intrusive than the `PreserveMost` calling convention. This calling
391 convention also behaves identical to the `C` calling convention on how
392 arguments and return values are passed, but it uses a different set of
393 caller/callee-saved registers. This removes the burden of saving and
394 recovering a large register set before and after the call in the caller. If
395 the arguments are passed in callee-saved registers, then they will be
396 preserved by the callee across the call. This doesn't apply for values
397 returned in callee-saved registers.
399 - On X86-64 the callee preserves all general purpose registers, except for
400 R11. R11 can be used as a scratch register. Furthermore it also preserves
401 all floating-point registers (XMMs/YMMs).
403 The idea behind this convention is to support calls to runtime functions
404 that don't need to call out to any other functions.
406 This calling convention, like the `PreserveMost` calling convention, will be
407 used by a future version of the ObjectiveC runtime and should be considered
408 experimental at this time.
409 "``cxx_fast_tlscc``" - The `CXX_FAST_TLS` calling convention for access functions
410 Clang generates an access function to access C++-style TLS. The access
411 function generally has an entry block, an exit block and an initialization
412 block that is run at the first time. The entry and exit blocks can access
413 a few TLS IR variables, each access will be lowered to a platform-specific
416 This calling convention aims to minimize overhead in the caller by
417 preserving as many registers as possible (all the registers that are
418 perserved on the fast path, composed of the entry and exit blocks).
420 This calling convention behaves identical to the `C` calling convention on
421 how arguments and return values are passed, but it uses a different set of
422 caller/callee-saved registers.
424 Given that each platform has its own lowering sequence, hence its own set
425 of preserved registers, we can't use the existing `PreserveMost`.
427 - On X86-64 the callee preserves all general purpose registers, except for
429 "``cc <n>``" - Numbered convention
430 Any calling convention may be specified by number, allowing
431 target-specific calling conventions to be used. Target specific
432 calling conventions start at 64.
434 More calling conventions can be added/defined on an as-needed basis, to
435 support Pascal conventions or any other well-known target-independent
438 .. _visibilitystyles:
443 All Global Variables and Functions have one of the following visibility
446 "``default``" - Default style
447 On targets that use the ELF object file format, default visibility
448 means that the declaration is visible to other modules and, in
449 shared libraries, means that the declared entity may be overridden.
450 On Darwin, default visibility means that the declaration is visible
451 to other modules. Default visibility corresponds to "external
452 linkage" in the language.
453 "``hidden``" - Hidden style
454 Two declarations of an object with hidden visibility refer to the
455 same object if they are in the same shared object. Usually, hidden
456 visibility indicates that the symbol will not be placed into the
457 dynamic symbol table, so no other module (executable or shared
458 library) can reference it directly.
459 "``protected``" - Protected style
460 On ELF, protected visibility indicates that the symbol will be
461 placed in the dynamic symbol table, but that references within the
462 defining module will bind to the local symbol. That is, the symbol
463 cannot be overridden by another module.
465 A symbol with ``internal`` or ``private`` linkage must have ``default``
473 All Global Variables, Functions and Aliases can have one of the following
477 "``dllimport``" causes the compiler to reference a function or variable via
478 a global pointer to a pointer that is set up by the DLL exporting the
479 symbol. On Microsoft Windows targets, the pointer name is formed by
480 combining ``__imp_`` and the function or variable name.
482 "``dllexport``" causes the compiler to provide a global pointer to a pointer
483 in a DLL, so that it can be referenced with the ``dllimport`` attribute. On
484 Microsoft Windows targets, the pointer name is formed by combining
485 ``__imp_`` and the function or variable name. Since this storage class
486 exists for defining a dll interface, the compiler, assembler and linker know
487 it is externally referenced and must refrain from deleting the symbol.
491 Thread Local Storage Models
492 ---------------------------
494 A variable may be defined as ``thread_local``, which means that it will
495 not be shared by threads (each thread will have a separated copy of the
496 variable). Not all targets support thread-local variables. Optionally, a
497 TLS model may be specified:
500 For variables that are only used within the current shared library.
502 For variables in modules that will not be loaded dynamically.
504 For variables defined in the executable and only used within it.
506 If no explicit model is given, the "general dynamic" model is used.
508 The models correspond to the ELF TLS models; see `ELF Handling For
509 Thread-Local Storage <http://people.redhat.com/drepper/tls.pdf>`_ for
510 more information on under which circumstances the different models may
511 be used. The target may choose a different TLS model if the specified
512 model is not supported, or if a better choice of model can be made.
514 A model can also be specified in an alias, but then it only governs how
515 the alias is accessed. It will not have any effect in the aliasee.
517 For platforms without linker support of ELF TLS model, the -femulated-tls
518 flag can be used to generate GCC compatible emulated TLS code.
525 LLVM IR allows you to specify both "identified" and "literal" :ref:`structure
526 types <t_struct>`. Literal types are uniqued structurally, but identified types
527 are never uniqued. An :ref:`opaque structural type <t_opaque>` can also be used
528 to forward declare a type that is not yet available.
530 An example of an identified structure specification is:
534 %mytype = type { %mytype*, i32 }
536 Prior to the LLVM 3.0 release, identified types were structurally uniqued. Only
537 literal types are uniqued in recent versions of LLVM.
544 Global variables define regions of memory allocated at compilation time
547 Global variable definitions must be initialized.
549 Global variables in other translation units can also be declared, in which
550 case they don't have an initializer.
552 Either global variable definitions or declarations may have an explicit section
553 to be placed in and may have an optional explicit alignment specified.
555 A variable may be defined as a global ``constant``, which indicates that
556 the contents of the variable will **never** be modified (enabling better
557 optimization, allowing the global data to be placed in the read-only
558 section of an executable, etc). Note that variables that need runtime
559 initialization cannot be marked ``constant`` as there is a store to the
562 LLVM explicitly allows *declarations* of global variables to be marked
563 constant, even if the final definition of the global is not. This
564 capability can be used to enable slightly better optimization of the
565 program, but requires the language definition to guarantee that
566 optimizations based on the 'constantness' are valid for the translation
567 units that do not include the definition.
569 As SSA values, global variables define pointer values that are in scope
570 (i.e. they dominate) all basic blocks in the program. Global variables
571 always define a pointer to their "content" type because they describe a
572 region of memory, and all memory objects in LLVM are accessed through
575 Global variables can be marked with ``unnamed_addr`` which indicates
576 that the address is not significant, only the content. Constants marked
577 like this can be merged with other constants if they have the same
578 initializer. Note that a constant with significant address *can* be
579 merged with a ``unnamed_addr`` constant, the result being a constant
580 whose address is significant.
582 A global variable may be declared to reside in a target-specific
583 numbered address space. For targets that support them, address spaces
584 may affect how optimizations are performed and/or what target
585 instructions are used to access the variable. The default address space
586 is zero. The address space qualifier must precede any other attributes.
588 LLVM allows an explicit section to be specified for globals. If the
589 target supports it, it will emit globals to the section specified.
590 Additionally, the global can placed in a comdat if the target has the necessary
593 By default, global initializers are optimized by assuming that global
594 variables defined within the module are not modified from their
595 initial values before the start of the global initializer. This is
596 true even for variables potentially accessible from outside the
597 module, including those with external linkage or appearing in
598 ``@llvm.used`` or dllexported variables. This assumption may be suppressed
599 by marking the variable with ``externally_initialized``.
601 An explicit alignment may be specified for a global, which must be a
602 power of 2. If not present, or if the alignment is set to zero, the
603 alignment of the global is set by the target to whatever it feels
604 convenient. If an explicit alignment is specified, the global is forced
605 to have exactly that alignment. Targets and optimizers are not allowed
606 to over-align the global if the global has an assigned section. In this
607 case, the extra alignment could be observable: for example, code could
608 assume that the globals are densely packed in their section and try to
609 iterate over them as an array, alignment padding would break this
610 iteration. The maximum alignment is ``1 << 29``.
612 Globals can also have a :ref:`DLL storage class <dllstorageclass>`.
614 Variables and aliases can have a
615 :ref:`Thread Local Storage Model <tls_model>`.
619 [@<GlobalVarName> =] [Linkage] [Visibility] [DLLStorageClass] [ThreadLocal]
620 [unnamed_addr] [AddrSpace] [ExternallyInitialized]
621 <global | constant> <Type> [<InitializerConstant>]
622 [, section "name"] [, comdat [($name)]]
623 [, align <Alignment>]
625 For example, the following defines a global in a numbered address space
626 with an initializer, section, and alignment:
630 @G = addrspace(5) constant float 1.0, section "foo", align 4
632 The following example just declares a global variable
636 @G = external global i32
638 The following example defines a thread-local global with the
639 ``initialexec`` TLS model:
643 @G = thread_local(initialexec) global i32 0, align 4
645 .. _functionstructure:
650 LLVM function definitions consist of the "``define``" keyword, an
651 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
652 style <visibility>`, an optional :ref:`DLL storage class <dllstorageclass>`,
653 an optional :ref:`calling convention <callingconv>`,
654 an optional ``unnamed_addr`` attribute, a return type, an optional
655 :ref:`parameter attribute <paramattrs>` for the return type, a function
656 name, a (possibly empty) argument list (each with optional :ref:`parameter
657 attributes <paramattrs>`), optional :ref:`function attributes <fnattrs>`,
658 an optional section, an optional alignment,
659 an optional :ref:`comdat <langref_comdats>`,
660 an optional :ref:`garbage collector name <gc>`, an optional :ref:`prefix <prefixdata>`,
661 an optional :ref:`prologue <prologuedata>`,
662 an optional :ref:`personality <personalityfn>`,
663 an optional list of attached :ref:`metadata <metadata>`,
664 an opening curly brace, a list of basic blocks, and a closing curly brace.
666 LLVM function declarations consist of the "``declare``" keyword, an
667 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
668 style <visibility>`, an optional :ref:`DLL storage class <dllstorageclass>`,
669 an optional :ref:`calling convention <callingconv>`,
670 an optional ``unnamed_addr`` attribute, a return type, an optional
671 :ref:`parameter attribute <paramattrs>` for the return type, a function
672 name, a possibly empty list of arguments, an optional alignment, an optional
673 :ref:`garbage collector name <gc>`, an optional :ref:`prefix <prefixdata>`,
674 and an optional :ref:`prologue <prologuedata>`.
676 A function definition contains a list of basic blocks, forming the CFG (Control
677 Flow Graph) for the function. Each basic block may optionally start with a label
678 (giving the basic block a symbol table entry), contains a list of instructions,
679 and ends with a :ref:`terminator <terminators>` instruction (such as a branch or
680 function return). If an explicit label is not provided, a block is assigned an
681 implicit numbered label, using the next value from the same counter as used for
682 unnamed temporaries (:ref:`see above<identifiers>`). For example, if a function
683 entry block does not have an explicit label, it will be assigned label "%0",
684 then the first unnamed temporary in that block will be "%1", etc.
686 The first basic block in a function is special in two ways: it is
687 immediately executed on entrance to the function, and it is not allowed
688 to have predecessor basic blocks (i.e. there can not be any branches to
689 the entry block of a function). Because the block can have no
690 predecessors, it also cannot have any :ref:`PHI nodes <i_phi>`.
692 LLVM allows an explicit section to be specified for functions. If the
693 target supports it, it will emit functions to the section specified.
694 Additionally, the function can be placed in a COMDAT.
696 An explicit alignment may be specified for a function. If not present,
697 or if the alignment is set to zero, the alignment of the function is set
698 by the target to whatever it feels convenient. If an explicit alignment
699 is specified, the function is forced to have at least that much
700 alignment. All alignments must be a power of 2.
702 If the ``unnamed_addr`` attribute is given, the address is known to not
703 be significant and two identical functions can be merged.
707 define [linkage] [visibility] [DLLStorageClass]
709 <ResultType> @<FunctionName> ([argument list])
710 [unnamed_addr] [fn Attrs] [section "name"] [comdat [($name)]]
711 [align N] [gc] [prefix Constant] [prologue Constant]
712 [personality Constant] (!name !N)* { ... }
714 The argument list is a comma separated sequence of arguments where each
715 argument is of the following form:
719 <type> [parameter Attrs] [name]
727 Aliases, unlike function or variables, don't create any new data. They
728 are just a new symbol and metadata for an existing position.
730 Aliases have a name and an aliasee that is either a global value or a
733 Aliases may have an optional :ref:`linkage type <linkage>`, an optional
734 :ref:`visibility style <visibility>`, an optional :ref:`DLL storage class
735 <dllstorageclass>` and an optional :ref:`tls model <tls_model>`.
739 @<Name> = [Linkage] [Visibility] [DLLStorageClass] [ThreadLocal] [unnamed_addr] alias <AliaseeTy>, <AliaseeTy>* @<Aliasee>
741 The linkage must be one of ``private``, ``internal``, ``linkonce``, ``weak``,
742 ``linkonce_odr``, ``weak_odr``, ``external``. Note that some system linkers
743 might not correctly handle dropping a weak symbol that is aliased.
745 Aliases that are not ``unnamed_addr`` are guaranteed to have the same address as
746 the aliasee expression. ``unnamed_addr`` ones are only guaranteed to point
749 Since aliases are only a second name, some restrictions apply, of which
750 some can only be checked when producing an object file:
752 * The expression defining the aliasee must be computable at assembly
753 time. Since it is just a name, no relocations can be used.
755 * No alias in the expression can be weak as the possibility of the
756 intermediate alias being overridden cannot be represented in an
759 * No global value in the expression can be a declaration, since that
760 would require a relocation, which is not possible.
767 Comdat IR provides access to COFF and ELF object file COMDAT functionality.
769 Comdats have a name which represents the COMDAT key. All global objects that
770 specify this key will only end up in the final object file if the linker chooses
771 that key over some other key. Aliases are placed in the same COMDAT that their
772 aliasee computes to, if any.
774 Comdats have a selection kind to provide input on how the linker should
775 choose between keys in two different object files.
779 $<Name> = comdat SelectionKind
781 The selection kind must be one of the following:
784 The linker may choose any COMDAT key, the choice is arbitrary.
786 The linker may choose any COMDAT key but the sections must contain the
789 The linker will choose the section containing the largest COMDAT key.
791 The linker requires that only section with this COMDAT key exist.
793 The linker may choose any COMDAT key but the sections must contain the
796 Note that the Mach-O platform doesn't support COMDATs and ELF only supports
797 ``any`` as a selection kind.
799 Here is an example of a COMDAT group where a function will only be selected if
800 the COMDAT key's section is the largest:
804 $foo = comdat largest
805 @foo = global i32 2, comdat($foo)
807 define void @bar() comdat($foo) {
811 As a syntactic sugar the ``$name`` can be omitted if the name is the same as
817 @foo = global i32 2, comdat
820 In a COFF object file, this will create a COMDAT section with selection kind
821 ``IMAGE_COMDAT_SELECT_LARGEST`` containing the contents of the ``@foo`` symbol
822 and another COMDAT section with selection kind
823 ``IMAGE_COMDAT_SELECT_ASSOCIATIVE`` which is associated with the first COMDAT
824 section and contains the contents of the ``@bar`` symbol.
826 There are some restrictions on the properties of the global object.
827 It, or an alias to it, must have the same name as the COMDAT group when
829 The contents and size of this object may be used during link-time to determine
830 which COMDAT groups get selected depending on the selection kind.
831 Because the name of the object must match the name of the COMDAT group, the
832 linkage of the global object must not be local; local symbols can get renamed
833 if a collision occurs in the symbol table.
835 The combined use of COMDATS and section attributes may yield surprising results.
842 @g1 = global i32 42, section "sec", comdat($foo)
843 @g2 = global i32 42, section "sec", comdat($bar)
845 From the object file perspective, this requires the creation of two sections
846 with the same name. This is necessary because both globals belong to different
847 COMDAT groups and COMDATs, at the object file level, are represented by
850 Note that certain IR constructs like global variables and functions may
851 create COMDATs in the object file in addition to any which are specified using
852 COMDAT IR. This arises when the code generator is configured to emit globals
853 in individual sections (e.g. when `-data-sections` or `-function-sections`
854 is supplied to `llc`).
856 .. _namedmetadatastructure:
861 Named metadata is a collection of metadata. :ref:`Metadata
862 nodes <metadata>` (but not metadata strings) are the only valid
863 operands for a named metadata.
865 #. Named metadata are represented as a string of characters with the
866 metadata prefix. The rules for metadata names are the same as for
867 identifiers, but quoted names are not allowed. ``"\xx"`` type escapes
868 are still valid, which allows any character to be part of a name.
872 ; Some unnamed metadata nodes, which are referenced by the named metadata.
877 !name = !{!0, !1, !2}
884 The return type and each parameter of a function type may have a set of
885 *parameter attributes* associated with them. Parameter attributes are
886 used to communicate additional information about the result or
887 parameters of a function. Parameter attributes are considered to be part
888 of the function, not of the function type, so functions with different
889 parameter attributes can have the same function type.
891 Parameter attributes are simple keywords that follow the type specified.
892 If multiple parameter attributes are needed, they are space separated.
897 declare i32 @printf(i8* noalias nocapture, ...)
898 declare i32 @atoi(i8 zeroext)
899 declare signext i8 @returns_signed_char()
901 Note that any attributes for the function result (``nounwind``,
902 ``readonly``) come immediately after the argument list.
904 Currently, only the following parameter attributes are defined:
907 This indicates to the code generator that the parameter or return
908 value should be zero-extended to the extent required by the target's
909 ABI (which is usually 32-bits, but is 8-bits for a i1 on x86-64) by
910 the caller (for a parameter) or the callee (for a return value).
912 This indicates to the code generator that the parameter or return
913 value should be sign-extended to the extent required by the target's
914 ABI (which is usually 32-bits) by the caller (for a parameter) or
915 the callee (for a return value).
917 This indicates that this parameter or return value should be treated
918 in a special target-dependent fashion while emitting code for
919 a function call or return (usually, by putting it in a register as
920 opposed to memory, though some targets use it to distinguish between
921 two different kinds of registers). Use of this attribute is
924 This indicates that the pointer parameter should really be passed by
925 value to the function. The attribute implies that a hidden copy of
926 the pointee is made between the caller and the callee, so the callee
927 is unable to modify the value in the caller. This attribute is only
928 valid on LLVM pointer arguments. It is generally used to pass
929 structs and arrays by value, but is also valid on pointers to
930 scalars. The copy is considered to belong to the caller not the
931 callee (for example, ``readonly`` functions should not write to
932 ``byval`` parameters). This is not a valid attribute for return
935 The byval attribute also supports specifying an alignment with the
936 align attribute. It indicates the alignment of the stack slot to
937 form and the known alignment of the pointer specified to the call
938 site. If the alignment is not specified, then the code generator
939 makes a target-specific assumption.
945 The ``inalloca`` argument attribute allows the caller to take the
946 address of outgoing stack arguments. An ``inalloca`` argument must
947 be a pointer to stack memory produced by an ``alloca`` instruction.
948 The alloca, or argument allocation, must also be tagged with the
949 inalloca keyword. Only the last argument may have the ``inalloca``
950 attribute, and that argument is guaranteed to be passed in memory.
952 An argument allocation may be used by a call at most once because
953 the call may deallocate it. The ``inalloca`` attribute cannot be
954 used in conjunction with other attributes that affect argument
955 storage, like ``inreg``, ``nest``, ``sret``, or ``byval``. The
956 ``inalloca`` attribute also disables LLVM's implicit lowering of
957 large aggregate return values, which means that frontend authors
958 must lower them with ``sret`` pointers.
960 When the call site is reached, the argument allocation must have
961 been the most recent stack allocation that is still live, or the
962 results are undefined. It is possible to allocate additional stack
963 space after an argument allocation and before its call site, but it
964 must be cleared off with :ref:`llvm.stackrestore
967 See :doc:`InAlloca` for more information on how to use this
971 This indicates that the pointer parameter specifies the address of a
972 structure that is the return value of the function in the source
973 program. This pointer must be guaranteed by the caller to be valid:
974 loads and stores to the structure may be assumed by the callee
975 not to trap and to be properly aligned. This may only be applied to
976 the first parameter. This is not a valid attribute for return
980 This indicates that the pointer value may be assumed by the optimizer to
981 have the specified alignment.
983 Note that this attribute has additional semantics when combined with the
989 This indicates that objects accessed via pointer values
990 :ref:`based <pointeraliasing>` on the argument or return value are not also
991 accessed, during the execution of the function, via pointer values not
992 *based* on the argument or return value. The attribute on a return value
993 also has additional semantics described below. The caller shares the
994 responsibility with the callee for ensuring that these requirements are met.
995 For further details, please see the discussion of the NoAlias response in
996 :ref:`alias analysis <Must, May, or No>`.
998 Note that this definition of ``noalias`` is intentionally similar
999 to the definition of ``restrict`` in C99 for function arguments.
1001 For function return values, C99's ``restrict`` is not meaningful,
1002 while LLVM's ``noalias`` is. Furthermore, the semantics of the ``noalias``
1003 attribute on return values are stronger than the semantics of the attribute
1004 when used on function arguments. On function return values, the ``noalias``
1005 attribute indicates that the function acts like a system memory allocation
1006 function, returning a pointer to allocated storage disjoint from the
1007 storage for any other object accessible to the caller.
1010 This indicates that the callee does not make any copies of the
1011 pointer that outlive the callee itself. This is not a valid
1012 attribute for return values.
1017 This indicates that the pointer parameter can be excised using the
1018 :ref:`trampoline intrinsics <int_trampoline>`. This is not a valid
1019 attribute for return values and can only be applied to one parameter.
1022 This indicates that the function always returns the argument as its return
1023 value. This is an optimization hint to the code generator when generating
1024 the caller, allowing tail call optimization and omission of register saves
1025 and restores in some cases; it is not checked or enforced when generating
1026 the callee. The parameter and the function return type must be valid
1027 operands for the :ref:`bitcast instruction <i_bitcast>`. This is not a
1028 valid attribute for return values and can only be applied to one parameter.
1031 This indicates that the parameter or return pointer is not null. This
1032 attribute may only be applied to pointer typed parameters. This is not
1033 checked or enforced by LLVM, the caller must ensure that the pointer
1034 passed in is non-null, or the callee must ensure that the returned pointer
1037 ``dereferenceable(<n>)``
1038 This indicates that the parameter or return pointer is dereferenceable. This
1039 attribute may only be applied to pointer typed parameters. A pointer that
1040 is dereferenceable can be loaded from speculatively without a risk of
1041 trapping. The number of bytes known to be dereferenceable must be provided
1042 in parentheses. It is legal for the number of bytes to be less than the
1043 size of the pointee type. The ``nonnull`` attribute does not imply
1044 dereferenceability (consider a pointer to one element past the end of an
1045 array), however ``dereferenceable(<n>)`` does imply ``nonnull`` in
1046 ``addrspace(0)`` (which is the default address space).
1048 ``dereferenceable_or_null(<n>)``
1049 This indicates that the parameter or return value isn't both
1050 non-null and non-dereferenceable (up to ``<n>`` bytes) at the same
1051 time. All non-null pointers tagged with
1052 ``dereferenceable_or_null(<n>)`` are ``dereferenceable(<n>)``.
1053 For address space 0 ``dereferenceable_or_null(<n>)`` implies that
1054 a pointer is exactly one of ``dereferenceable(<n>)`` or ``null``,
1055 and in other address spaces ``dereferenceable_or_null(<n>)``
1056 implies that a pointer is at least one of ``dereferenceable(<n>)``
1057 or ``null`` (i.e. it may be both ``null`` and
1058 ``dereferenceable(<n>)``). This attribute may only be applied to
1059 pointer typed parameters.
1063 Garbage Collector Strategy Names
1064 --------------------------------
1066 Each function may specify a garbage collector strategy name, which is simply a
1069 .. code-block:: llvm
1071 define void @f() gc "name" { ... }
1073 The supported values of *name* includes those :ref:`built in to LLVM
1074 <builtin-gc-strategies>` and any provided by loaded plugins. Specifying a GC
1075 strategy will cause the compiler to alter its output in order to support the
1076 named garbage collection algorithm. Note that LLVM itself does not contain a
1077 garbage collector, this functionality is restricted to generating machine code
1078 which can interoperate with a collector provided externally.
1085 Prefix data is data associated with a function which the code
1086 generator will emit immediately before the function's entrypoint.
1087 The purpose of this feature is to allow frontends to associate
1088 language-specific runtime metadata with specific functions and make it
1089 available through the function pointer while still allowing the
1090 function pointer to be called.
1092 To access the data for a given function, a program may bitcast the
1093 function pointer to a pointer to the constant's type and dereference
1094 index -1. This implies that the IR symbol points just past the end of
1095 the prefix data. For instance, take the example of a function annotated
1096 with a single ``i32``,
1098 .. code-block:: llvm
1100 define void @f() prefix i32 123 { ... }
1102 The prefix data can be referenced as,
1104 .. code-block:: llvm
1106 %0 = bitcast void* () @f to i32*
1107 %a = getelementptr inbounds i32, i32* %0, i32 -1
1108 %b = load i32, i32* %a
1110 Prefix data is laid out as if it were an initializer for a global variable
1111 of the prefix data's type. The function will be placed such that the
1112 beginning of the prefix data is aligned. This means that if the size
1113 of the prefix data is not a multiple of the alignment size, the
1114 function's entrypoint will not be aligned. If alignment of the
1115 function's entrypoint is desired, padding must be added to the prefix
1118 A function may have prefix data but no body. This has similar semantics
1119 to the ``available_externally`` linkage in that the data may be used by the
1120 optimizers but will not be emitted in the object file.
1127 The ``prologue`` attribute allows arbitrary code (encoded as bytes) to
1128 be inserted prior to the function body. This can be used for enabling
1129 function hot-patching and instrumentation.
1131 To maintain the semantics of ordinary function calls, the prologue data must
1132 have a particular format. Specifically, it must begin with a sequence of
1133 bytes which decode to a sequence of machine instructions, valid for the
1134 module's target, which transfer control to the point immediately succeeding
1135 the prologue data, without performing any other visible action. This allows
1136 the inliner and other passes to reason about the semantics of the function
1137 definition without needing to reason about the prologue data. Obviously this
1138 makes the format of the prologue data highly target dependent.
1140 A trivial example of valid prologue data for the x86 architecture is ``i8 144``,
1141 which encodes the ``nop`` instruction:
1143 .. code-block:: llvm
1145 define void @f() prologue i8 144 { ... }
1147 Generally prologue data can be formed by encoding a relative branch instruction
1148 which skips the metadata, as in this example of valid prologue data for the
1149 x86_64 architecture, where the first two bytes encode ``jmp .+10``:
1151 .. code-block:: llvm
1153 %0 = type <{ i8, i8, i8* }>
1155 define void @f() prologue %0 <{ i8 235, i8 8, i8* @md}> { ... }
1157 A function may have prologue data but no body. This has similar semantics
1158 to the ``available_externally`` linkage in that the data may be used by the
1159 optimizers but will not be emitted in the object file.
1163 Personality Function
1164 --------------------
1166 The ``personality`` attribute permits functions to specify what function
1167 to use for exception handling.
1174 Attribute groups are groups of attributes that are referenced by objects within
1175 the IR. They are important for keeping ``.ll`` files readable, because a lot of
1176 functions will use the same set of attributes. In the degenerative case of a
1177 ``.ll`` file that corresponds to a single ``.c`` file, the single attribute
1178 group will capture the important command line flags used to build that file.
1180 An attribute group is a module-level object. To use an attribute group, an
1181 object references the attribute group's ID (e.g. ``#37``). An object may refer
1182 to more than one attribute group. In that situation, the attributes from the
1183 different groups are merged.
1185 Here is an example of attribute groups for a function that should always be
1186 inlined, has a stack alignment of 4, and which shouldn't use SSE instructions:
1188 .. code-block:: llvm
1190 ; Target-independent attributes:
1191 attributes #0 = { alwaysinline alignstack=4 }
1193 ; Target-dependent attributes:
1194 attributes #1 = { "no-sse" }
1196 ; Function @f has attributes: alwaysinline, alignstack=4, and "no-sse".
1197 define void @f() #0 #1 { ... }
1204 Function attributes are set to communicate additional information about
1205 a function. Function attributes are considered to be part of the
1206 function, not of the function type, so functions with different function
1207 attributes can have the same function type.
1209 Function attributes are simple keywords that follow the type specified.
1210 If multiple attributes are needed, they are space separated. For
1213 .. code-block:: llvm
1215 define void @f() noinline { ... }
1216 define void @f() alwaysinline { ... }
1217 define void @f() alwaysinline optsize { ... }
1218 define void @f() optsize { ... }
1221 This attribute indicates that, when emitting the prologue and
1222 epilogue, the backend should forcibly align the stack pointer.
1223 Specify the desired alignment, which must be a power of two, in
1226 This attribute indicates that the inliner should attempt to inline
1227 this function into callers whenever possible, ignoring any active
1228 inlining size threshold for this caller.
1230 This indicates that the callee function at a call site should be
1231 recognized as a built-in function, even though the function's declaration
1232 uses the ``nobuiltin`` attribute. This is only valid at call sites for
1233 direct calls to functions that are declared with the ``nobuiltin``
1236 This attribute indicates that this function is rarely called. When
1237 computing edge weights, basic blocks post-dominated by a cold
1238 function call are also considered to be cold; and, thus, given low
1241 This attribute indicates that the callee is dependent on a convergent
1242 thread execution pattern under certain parallel execution models.
1243 Transformations that are execution model agnostic may not make the execution
1244 of a convergent operation control dependent on any additional values.
1246 This attribute indicates that the source code contained a hint that
1247 inlining this function is desirable (such as the "inline" keyword in
1248 C/C++). It is just a hint; it imposes no requirements on the
1251 This attribute indicates that the function should be added to a
1252 jump-instruction table at code-generation time, and that all address-taken
1253 references to this function should be replaced with a reference to the
1254 appropriate jump-instruction-table function pointer. Note that this creates
1255 a new pointer for the original function, which means that code that depends
1256 on function-pointer identity can break. So, any function annotated with
1257 ``jumptable`` must also be ``unnamed_addr``.
1259 This attribute suggests that optimization passes and code generator
1260 passes make choices that keep the code size of this function as small
1261 as possible and perform optimizations that may sacrifice runtime
1262 performance in order to minimize the size of the generated code.
1264 This attribute disables prologue / epilogue emission for the
1265 function. This can have very system-specific consequences.
1267 This indicates that the callee function at a call site is not recognized as
1268 a built-in function. LLVM will retain the original call and not replace it
1269 with equivalent code based on the semantics of the built-in function, unless
1270 the call site uses the ``builtin`` attribute. This is valid at call sites
1271 and on function declarations and definitions.
1273 This attribute indicates that calls to the function cannot be
1274 duplicated. A call to a ``noduplicate`` function may be moved
1275 within its parent function, but may not be duplicated within
1276 its parent function.
1278 A function containing a ``noduplicate`` call may still
1279 be an inlining candidate, provided that the call is not
1280 duplicated by inlining. That implies that the function has
1281 internal linkage and only has one call site, so the original
1282 call is dead after inlining.
1284 This attributes disables implicit floating point instructions.
1286 This attribute indicates that the inliner should never inline this
1287 function in any situation. This attribute may not be used together
1288 with the ``alwaysinline`` attribute.
1290 This attribute suppresses lazy symbol binding for the function. This
1291 may make calls to the function faster, at the cost of extra program
1292 startup time if the function is not called during program startup.
1294 This attribute indicates that the code generator should not use a
1295 red zone, even if the target-specific ABI normally permits it.
1297 This function attribute indicates that the function never returns
1298 normally. This produces undefined behavior at runtime if the
1299 function ever does dynamically return.
1301 This function attribute indicates that the function does not call itself
1302 either directly or indirectly down any possible call path. This produces
1303 undefined behavior at runtime if the function ever does recurse.
1305 This function attribute indicates that the function never raises an
1306 exception. If the function does raise an exception, its runtime
1307 behavior is undefined. However, functions marked nounwind may still
1308 trap or generate asynchronous exceptions. Exception handling schemes
1309 that are recognized by LLVM to handle asynchronous exceptions, such
1310 as SEH, will still provide their implementation defined semantics.
1312 This function attribute indicates that most optimization passes will skip
1313 this function, with the exception of interprocedural optimization passes.
1314 Code generation defaults to the "fast" instruction selector.
1315 This attribute cannot be used together with the ``alwaysinline``
1316 attribute; this attribute is also incompatible
1317 with the ``minsize`` attribute and the ``optsize`` attribute.
1319 This attribute requires the ``noinline`` attribute to be specified on
1320 the function as well, so the function is never inlined into any caller.
1321 Only functions with the ``alwaysinline`` attribute are valid
1322 candidates for inlining into the body of this function.
1324 This attribute suggests that optimization passes and code generator
1325 passes make choices that keep the code size of this function low,
1326 and otherwise do optimizations specifically to reduce code size as
1327 long as they do not significantly impact runtime performance.
1329 On a function, this attribute indicates that the function computes its
1330 result (or decides to unwind an exception) based strictly on its arguments,
1331 without dereferencing any pointer arguments or otherwise accessing
1332 any mutable state (e.g. memory, control registers, etc) visible to
1333 caller functions. It does not write through any pointer arguments
1334 (including ``byval`` arguments) and never changes any state visible
1335 to callers. This means that it cannot unwind exceptions by calling
1336 the ``C++`` exception throwing methods.
1338 On an argument, this attribute indicates that the function does not
1339 dereference that pointer argument, even though it may read or write the
1340 memory that the pointer points to if accessed through other pointers.
1342 On a function, this attribute indicates that the function does not write
1343 through any pointer arguments (including ``byval`` arguments) or otherwise
1344 modify any state (e.g. memory, control registers, etc) visible to
1345 caller functions. It may dereference pointer arguments and read
1346 state that may be set in the caller. A readonly function always
1347 returns the same value (or unwinds an exception identically) when
1348 called with the same set of arguments and global state. It cannot
1349 unwind an exception by calling the ``C++`` exception throwing
1352 On an argument, this attribute indicates that the function does not write
1353 through this pointer argument, even though it may write to the memory that
1354 the pointer points to.
1356 This attribute indicates that the only memory accesses inside function are
1357 loads and stores from objects pointed to by its pointer-typed arguments,
1358 with arbitrary offsets. Or in other words, all memory operations in the
1359 function can refer to memory only using pointers based on its function
1361 Note that ``argmemonly`` can be used together with ``readonly`` attribute
1362 in order to specify that function reads only from its arguments.
1364 This attribute indicates that this function can return twice. The C
1365 ``setjmp`` is an example of such a function. The compiler disables
1366 some optimizations (like tail calls) in the caller of these
1369 This attribute indicates that
1370 `SafeStack <http://clang.llvm.org/docs/SafeStack.html>`_
1371 protection is enabled for this function.
1373 If a function that has a ``safestack`` attribute is inlined into a
1374 function that doesn't have a ``safestack`` attribute or which has an
1375 ``ssp``, ``sspstrong`` or ``sspreq`` attribute, then the resulting
1376 function will have a ``safestack`` attribute.
1377 ``sanitize_address``
1378 This attribute indicates that AddressSanitizer checks
1379 (dynamic address safety analysis) are enabled for this function.
1381 This attribute indicates that MemorySanitizer checks (dynamic detection
1382 of accesses to uninitialized memory) are enabled for this function.
1384 This attribute indicates that ThreadSanitizer checks
1385 (dynamic thread safety analysis) are enabled for this function.
1387 This attribute indicates that the function should emit a stack
1388 smashing protector. It is in the form of a "canary" --- a random value
1389 placed on the stack before the local variables that's checked upon
1390 return from the function to see if it has been overwritten. A
1391 heuristic is used to determine if a function needs stack protectors
1392 or not. The heuristic used will enable protectors for functions with:
1394 - Character arrays larger than ``ssp-buffer-size`` (default 8).
1395 - Aggregates containing character arrays larger than ``ssp-buffer-size``.
1396 - Calls to alloca() with variable sizes or constant sizes greater than
1397 ``ssp-buffer-size``.
1399 Variables that are identified as requiring a protector will be arranged
1400 on the stack such that they are adjacent to the stack protector guard.
1402 If a function that has an ``ssp`` attribute is inlined into a
1403 function that doesn't have an ``ssp`` attribute, then the resulting
1404 function will have an ``ssp`` attribute.
1406 This attribute indicates that the function should *always* emit a
1407 stack smashing protector. This overrides the ``ssp`` function
1410 Variables that are identified as requiring a protector will be arranged
1411 on the stack such that they are adjacent to the stack protector guard.
1412 The specific layout rules are:
1414 #. Large arrays and structures containing large arrays
1415 (``>= ssp-buffer-size``) are closest to the stack protector.
1416 #. Small arrays and structures containing small arrays
1417 (``< ssp-buffer-size``) are 2nd closest to the protector.
1418 #. Variables that have had their address taken are 3rd closest to the
1421 If a function that has an ``sspreq`` attribute is inlined into a
1422 function that doesn't have an ``sspreq`` attribute or which has an
1423 ``ssp`` or ``sspstrong`` attribute, then the resulting function will have
1424 an ``sspreq`` attribute.
1426 This attribute indicates that the function should emit a stack smashing
1427 protector. This attribute causes a strong heuristic to be used when
1428 determining if a function needs stack protectors. The strong heuristic
1429 will enable protectors for functions with:
1431 - Arrays of any size and type
1432 - Aggregates containing an array of any size and type.
1433 - Calls to alloca().
1434 - Local variables that have had their address taken.
1436 Variables that are identified as requiring a protector will be arranged
1437 on the stack such that they are adjacent to the stack protector guard.
1438 The specific layout rules are:
1440 #. Large arrays and structures containing large arrays
1441 (``>= ssp-buffer-size``) are closest to the stack protector.
1442 #. Small arrays and structures containing small arrays
1443 (``< ssp-buffer-size``) are 2nd closest to the protector.
1444 #. Variables that have had their address taken are 3rd closest to the
1447 This overrides the ``ssp`` function attribute.
1449 If a function that has an ``sspstrong`` attribute is inlined into a
1450 function that doesn't have an ``sspstrong`` attribute, then the
1451 resulting function will have an ``sspstrong`` attribute.
1453 This attribute indicates that the function will delegate to some other
1454 function with a tail call. The prototype of a thunk should not be used for
1455 optimization purposes. The caller is expected to cast the thunk prototype to
1456 match the thunk target prototype.
1458 This attribute indicates that the ABI being targeted requires that
1459 an unwind table entry be produced for this function even if we can
1460 show that no exceptions passes by it. This is normally the case for
1461 the ELF x86-64 abi, but it can be disabled for some compilation
1470 Note: operand bundles are a work in progress, and they should be
1471 considered experimental at this time.
1473 Operand bundles are tagged sets of SSA values that can be associated
1474 with certain LLVM instructions (currently only ``call`` s and
1475 ``invoke`` s). In a way they are like metadata, but dropping them is
1476 incorrect and will change program semantics.
1480 operand bundle set ::= '[' operand bundle (, operand bundle )* ']'
1481 operand bundle ::= tag '(' [ bundle operand ] (, bundle operand )* ')'
1482 bundle operand ::= SSA value
1483 tag ::= string constant
1485 Operand bundles are **not** part of a function's signature, and a
1486 given function may be called from multiple places with different kinds
1487 of operand bundles. This reflects the fact that the operand bundles
1488 are conceptually a part of the ``call`` (or ``invoke``), not the
1489 callee being dispatched to.
1491 Operand bundles are a generic mechanism intended to support
1492 runtime-introspection-like functionality for managed languages. While
1493 the exact semantics of an operand bundle depend on the bundle tag,
1494 there are certain limitations to how much the presence of an operand
1495 bundle can influence the semantics of a program. These restrictions
1496 are described as the semantics of an "unknown" operand bundle. As
1497 long as the behavior of an operand bundle is describable within these
1498 restrictions, LLVM does not need to have special knowledge of the
1499 operand bundle to not miscompile programs containing it.
1501 - The bundle operands for an unknown operand bundle escape in unknown
1502 ways before control is transferred to the callee or invokee.
1503 - Calls and invokes with operand bundles have unknown read / write
1504 effect on the heap on entry and exit (even if the call target is
1505 ``readnone`` or ``readonly``), unless they're overriden with
1506 callsite specific attributes.
1507 - An operand bundle at a call site cannot change the implementation
1508 of the called function. Inter-procedural optimizations work as
1509 usual as long as they take into account the first two properties.
1511 More specific types of operand bundles are described below.
1513 Deoptimization Operand Bundles
1514 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1516 Deoptimization operand bundles are characterized by the ``"deopt"``
1517 operand bundle tag. These operand bundles represent an alternate
1518 "safe" continuation for the call site they're attached to, and can be
1519 used by a suitable runtime to deoptimize the compiled frame at the
1520 specified call site. There can be at most one ``"deopt"`` operand
1521 bundle attached to a call site. Exact details of deoptimization is
1522 out of scope for the language reference, but it usually involves
1523 rewriting a compiled frame into a set of interpreted frames.
1525 From the compiler's perspective, deoptimization operand bundles make
1526 the call sites they're attached to at least ``readonly``. They read
1527 through all of their pointer typed operands (even if they're not
1528 otherwise escaped) and the entire visible heap. Deoptimization
1529 operand bundles do not capture their operands except during
1530 deoptimization, in which case control will not be returned to the
1533 The inliner knows how to inline through calls that have deoptimization
1534 operand bundles. Just like inlining through a normal call site
1535 involves composing the normal and exceptional continuations, inlining
1536 through a call site with a deoptimization operand bundle needs to
1537 appropriately compose the "safe" deoptimization continuation. The
1538 inliner does this by prepending the parent's deoptimization
1539 continuation to every deoptimization continuation in the inlined body.
1540 E.g. inlining ``@f`` into ``@g`` in the following example
1542 .. code-block:: llvm
1545 call void @x() ;; no deopt state
1546 call void @y() [ "deopt"(i32 10) ]
1547 call void @y() [ "deopt"(i32 10), "unknown"(i8* null) ]
1552 call void @f() [ "deopt"(i32 20) ]
1558 .. code-block:: llvm
1561 call void @x() ;; still no deopt state
1562 call void @y() [ "deopt"(i32 20, i32 10) ]
1563 call void @y() [ "deopt"(i32 20, i32 10), "unknown"(i8* null) ]
1567 It is the frontend's responsibility to structure or encode the
1568 deoptimization state in a way that syntactically prepending the
1569 caller's deoptimization state to the callee's deoptimization state is
1570 semantically equivalent to composing the caller's deoptimization
1571 continuation after the callee's deoptimization continuation.
1575 Module-Level Inline Assembly
1576 ----------------------------
1578 Modules may contain "module-level inline asm" blocks, which corresponds
1579 to the GCC "file scope inline asm" blocks. These blocks are internally
1580 concatenated by LLVM and treated as a single unit, but may be separated
1581 in the ``.ll`` file if desired. The syntax is very simple:
1583 .. code-block:: llvm
1585 module asm "inline asm code goes here"
1586 module asm "more can go here"
1588 The strings can contain any character by escaping non-printable
1589 characters. The escape sequence used is simply "\\xx" where "xx" is the
1590 two digit hex code for the number.
1592 Note that the assembly string *must* be parseable by LLVM's integrated assembler
1593 (unless it is disabled), even when emitting a ``.s`` file.
1595 .. _langref_datalayout:
1600 A module may specify a target specific data layout string that specifies
1601 how data is to be laid out in memory. The syntax for the data layout is
1604 .. code-block:: llvm
1606 target datalayout = "layout specification"
1608 The *layout specification* consists of a list of specifications
1609 separated by the minus sign character ('-'). Each specification starts
1610 with a letter and may include other information after the letter to
1611 define some aspect of the data layout. The specifications accepted are
1615 Specifies that the target lays out data in big-endian form. That is,
1616 the bits with the most significance have the lowest address
1619 Specifies that the target lays out data in little-endian form. That
1620 is, the bits with the least significance have the lowest address
1623 Specifies the natural alignment of the stack in bits. Alignment
1624 promotion of stack variables is limited to the natural stack
1625 alignment to avoid dynamic stack realignment. The stack alignment
1626 must be a multiple of 8-bits. If omitted, the natural stack
1627 alignment defaults to "unspecified", which does not prevent any
1628 alignment promotions.
1629 ``p[n]:<size>:<abi>:<pref>``
1630 This specifies the *size* of a pointer and its ``<abi>`` and
1631 ``<pref>``\erred alignments for address space ``n``. All sizes are in
1632 bits. The address space, ``n``, is optional, and if not specified,
1633 denotes the default address space 0. The value of ``n`` must be
1634 in the range [1,2^23).
1635 ``i<size>:<abi>:<pref>``
1636 This specifies the alignment for an integer type of a given bit
1637 ``<size>``. The value of ``<size>`` must be in the range [1,2^23).
1638 ``v<size>:<abi>:<pref>``
1639 This specifies the alignment for a vector type of a given bit
1641 ``f<size>:<abi>:<pref>``
1642 This specifies the alignment for a floating point type of a given bit
1643 ``<size>``. Only values of ``<size>`` that are supported by the target
1644 will work. 32 (float) and 64 (double) are supported on all targets; 80
1645 or 128 (different flavors of long double) are also supported on some
1648 This specifies the alignment for an object of aggregate type.
1650 If present, specifies that llvm names are mangled in the output. The
1653 * ``e``: ELF mangling: Private symbols get a ``.L`` prefix.
1654 * ``m``: Mips mangling: Private symbols get a ``$`` prefix.
1655 * ``o``: Mach-O mangling: Private symbols get ``L`` prefix. Other
1656 symbols get a ``_`` prefix.
1657 * ``w``: Windows COFF prefix: Similar to Mach-O, but stdcall and fastcall
1658 functions also get a suffix based on the frame size.
1659 * ``x``: Windows x86 COFF prefix: Similar to Windows COFF, but use a ``_``
1660 prefix for ``__cdecl`` functions.
1661 ``n<size1>:<size2>:<size3>...``
1662 This specifies a set of native integer widths for the target CPU in
1663 bits. For example, it might contain ``n32`` for 32-bit PowerPC,
1664 ``n32:64`` for PowerPC 64, or ``n8:16:32:64`` for X86-64. Elements of
1665 this set are considered to support most general arithmetic operations
1668 On every specification that takes a ``<abi>:<pref>``, specifying the
1669 ``<pref>`` alignment is optional. If omitted, the preceding ``:``
1670 should be omitted too and ``<pref>`` will be equal to ``<abi>``.
1672 When constructing the data layout for a given target, LLVM starts with a
1673 default set of specifications which are then (possibly) overridden by
1674 the specifications in the ``datalayout`` keyword. The default
1675 specifications are given in this list:
1677 - ``E`` - big endian
1678 - ``p:64:64:64`` - 64-bit pointers with 64-bit alignment.
1679 - ``p[n]:64:64:64`` - Other address spaces are assumed to be the
1680 same as the default address space.
1681 - ``S0`` - natural stack alignment is unspecified
1682 - ``i1:8:8`` - i1 is 8-bit (byte) aligned
1683 - ``i8:8:8`` - i8 is 8-bit (byte) aligned
1684 - ``i16:16:16`` - i16 is 16-bit aligned
1685 - ``i32:32:32`` - i32 is 32-bit aligned
1686 - ``i64:32:64`` - i64 has ABI alignment of 32-bits but preferred
1687 alignment of 64-bits
1688 - ``f16:16:16`` - half is 16-bit aligned
1689 - ``f32:32:32`` - float is 32-bit aligned
1690 - ``f64:64:64`` - double is 64-bit aligned
1691 - ``f128:128:128`` - quad is 128-bit aligned
1692 - ``v64:64:64`` - 64-bit vector is 64-bit aligned
1693 - ``v128:128:128`` - 128-bit vector is 128-bit aligned
1694 - ``a:0:64`` - aggregates are 64-bit aligned
1696 When LLVM is determining the alignment for a given type, it uses the
1699 #. If the type sought is an exact match for one of the specifications,
1700 that specification is used.
1701 #. If no match is found, and the type sought is an integer type, then
1702 the smallest integer type that is larger than the bitwidth of the
1703 sought type is used. If none of the specifications are larger than
1704 the bitwidth then the largest integer type is used. For example,
1705 given the default specifications above, the i7 type will use the
1706 alignment of i8 (next largest) while both i65 and i256 will use the
1707 alignment of i64 (largest specified).
1708 #. If no match is found, and the type sought is a vector type, then the
1709 largest vector type that is smaller than the sought vector type will
1710 be used as a fall back. This happens because <128 x double> can be
1711 implemented in terms of 64 <2 x double>, for example.
1713 The function of the data layout string may not be what you expect.
1714 Notably, this is not a specification from the frontend of what alignment
1715 the code generator should use.
1717 Instead, if specified, the target data layout is required to match what
1718 the ultimate *code generator* expects. This string is used by the
1719 mid-level optimizers to improve code, and this only works if it matches
1720 what the ultimate code generator uses. There is no way to generate IR
1721 that does not embed this target-specific detail into the IR. If you
1722 don't specify the string, the default specifications will be used to
1723 generate a Data Layout and the optimization phases will operate
1724 accordingly and introduce target specificity into the IR with respect to
1725 these default specifications.
1732 A module may specify a target triple string that describes the target
1733 host. The syntax for the target triple is simply:
1735 .. code-block:: llvm
1737 target triple = "x86_64-apple-macosx10.7.0"
1739 The *target triple* string consists of a series of identifiers delimited
1740 by the minus sign character ('-'). The canonical forms are:
1744 ARCHITECTURE-VENDOR-OPERATING_SYSTEM
1745 ARCHITECTURE-VENDOR-OPERATING_SYSTEM-ENVIRONMENT
1747 This information is passed along to the backend so that it generates
1748 code for the proper architecture. It's possible to override this on the
1749 command line with the ``-mtriple`` command line option.
1751 .. _pointeraliasing:
1753 Pointer Aliasing Rules
1754 ----------------------
1756 Any memory access must be done through a pointer value associated with
1757 an address range of the memory access, otherwise the behavior is
1758 undefined. Pointer values are associated with address ranges according
1759 to the following rules:
1761 - A pointer value is associated with the addresses associated with any
1762 value it is *based* on.
1763 - An address of a global variable is associated with the address range
1764 of the variable's storage.
1765 - The result value of an allocation instruction is associated with the
1766 address range of the allocated storage.
1767 - A null pointer in the default address-space is associated with no
1769 - An integer constant other than zero or a pointer value returned from
1770 a function not defined within LLVM may be associated with address
1771 ranges allocated through mechanisms other than those provided by
1772 LLVM. Such ranges shall not overlap with any ranges of addresses
1773 allocated by mechanisms provided by LLVM.
1775 A pointer value is *based* on another pointer value according to the
1778 - A pointer value formed from a ``getelementptr`` operation is *based*
1779 on the first value operand of the ``getelementptr``.
1780 - The result value of a ``bitcast`` is *based* on the operand of the
1782 - A pointer value formed by an ``inttoptr`` is *based* on all pointer
1783 values that contribute (directly or indirectly) to the computation of
1784 the pointer's value.
1785 - The "*based* on" relationship is transitive.
1787 Note that this definition of *"based"* is intentionally similar to the
1788 definition of *"based"* in C99, though it is slightly weaker.
1790 LLVM IR does not associate types with memory. The result type of a
1791 ``load`` merely indicates the size and alignment of the memory from
1792 which to load, as well as the interpretation of the value. The first
1793 operand type of a ``store`` similarly only indicates the size and
1794 alignment of the store.
1796 Consequently, type-based alias analysis, aka TBAA, aka
1797 ``-fstrict-aliasing``, is not applicable to general unadorned LLVM IR.
1798 :ref:`Metadata <metadata>` may be used to encode additional information
1799 which specialized optimization passes may use to implement type-based
1804 Volatile Memory Accesses
1805 ------------------------
1807 Certain memory accesses, such as :ref:`load <i_load>`'s,
1808 :ref:`store <i_store>`'s, and :ref:`llvm.memcpy <int_memcpy>`'s may be
1809 marked ``volatile``. The optimizers must not change the number of
1810 volatile operations or change their order of execution relative to other
1811 volatile operations. The optimizers *may* change the order of volatile
1812 operations relative to non-volatile operations. This is not Java's
1813 "volatile" and has no cross-thread synchronization behavior.
1815 IR-level volatile loads and stores cannot safely be optimized into
1816 llvm.memcpy or llvm.memmove intrinsics even when those intrinsics are
1817 flagged volatile. Likewise, the backend should never split or merge
1818 target-legal volatile load/store instructions.
1820 .. admonition:: Rationale
1822 Platforms may rely on volatile loads and stores of natively supported
1823 data width to be executed as single instruction. For example, in C
1824 this holds for an l-value of volatile primitive type with native
1825 hardware support, but not necessarily for aggregate types. The
1826 frontend upholds these expectations, which are intentionally
1827 unspecified in the IR. The rules above ensure that IR transformations
1828 do not violate the frontend's contract with the language.
1832 Memory Model for Concurrent Operations
1833 --------------------------------------
1835 The LLVM IR does not define any way to start parallel threads of
1836 execution or to register signal handlers. Nonetheless, there are
1837 platform-specific ways to create them, and we define LLVM IR's behavior
1838 in their presence. This model is inspired by the C++0x memory model.
1840 For a more informal introduction to this model, see the :doc:`Atomics`.
1842 We define a *happens-before* partial order as the least partial order
1845 - Is a superset of single-thread program order, and
1846 - When a *synchronizes-with* ``b``, includes an edge from ``a`` to
1847 ``b``. *Synchronizes-with* pairs are introduced by platform-specific
1848 techniques, like pthread locks, thread creation, thread joining,
1849 etc., and by atomic instructions. (See also :ref:`Atomic Memory Ordering
1850 Constraints <ordering>`).
1852 Note that program order does not introduce *happens-before* edges
1853 between a thread and signals executing inside that thread.
1855 Every (defined) read operation (load instructions, memcpy, atomic
1856 loads/read-modify-writes, etc.) R reads a series of bytes written by
1857 (defined) write operations (store instructions, atomic
1858 stores/read-modify-writes, memcpy, etc.). For the purposes of this
1859 section, initialized globals are considered to have a write of the
1860 initializer which is atomic and happens before any other read or write
1861 of the memory in question. For each byte of a read R, R\ :sub:`byte`
1862 may see any write to the same byte, except:
1864 - If write\ :sub:`1` happens before write\ :sub:`2`, and
1865 write\ :sub:`2` happens before R\ :sub:`byte`, then
1866 R\ :sub:`byte` does not see write\ :sub:`1`.
1867 - If R\ :sub:`byte` happens before write\ :sub:`3`, then
1868 R\ :sub:`byte` does not see write\ :sub:`3`.
1870 Given that definition, R\ :sub:`byte` is defined as follows:
1872 - If R is volatile, the result is target-dependent. (Volatile is
1873 supposed to give guarantees which can support ``sig_atomic_t`` in
1874 C/C++, and may be used for accesses to addresses that do not behave
1875 like normal memory. It does not generally provide cross-thread
1877 - Otherwise, if there is no write to the same byte that happens before
1878 R\ :sub:`byte`, R\ :sub:`byte` returns ``undef`` for that byte.
1879 - Otherwise, if R\ :sub:`byte` may see exactly one write,
1880 R\ :sub:`byte` returns the value written by that write.
1881 - Otherwise, if R is atomic, and all the writes R\ :sub:`byte` may
1882 see are atomic, it chooses one of the values written. See the :ref:`Atomic
1883 Memory Ordering Constraints <ordering>` section for additional
1884 constraints on how the choice is made.
1885 - Otherwise R\ :sub:`byte` returns ``undef``.
1887 R returns the value composed of the series of bytes it read. This
1888 implies that some bytes within the value may be ``undef`` **without**
1889 the entire value being ``undef``. Note that this only defines the
1890 semantics of the operation; it doesn't mean that targets will emit more
1891 than one instruction to read the series of bytes.
1893 Note that in cases where none of the atomic intrinsics are used, this
1894 model places only one restriction on IR transformations on top of what
1895 is required for single-threaded execution: introducing a store to a byte
1896 which might not otherwise be stored is not allowed in general.
1897 (Specifically, in the case where another thread might write to and read
1898 from an address, introducing a store can change a load that may see
1899 exactly one write into a load that may see multiple writes.)
1903 Atomic Memory Ordering Constraints
1904 ----------------------------------
1906 Atomic instructions (:ref:`cmpxchg <i_cmpxchg>`,
1907 :ref:`atomicrmw <i_atomicrmw>`, :ref:`fence <i_fence>`,
1908 :ref:`atomic load <i_load>`, and :ref:`atomic store <i_store>`) take
1909 ordering parameters that determine which other atomic instructions on
1910 the same address they *synchronize with*. These semantics are borrowed
1911 from Java and C++0x, but are somewhat more colloquial. If these
1912 descriptions aren't precise enough, check those specs (see spec
1913 references in the :doc:`atomics guide <Atomics>`).
1914 :ref:`fence <i_fence>` instructions treat these orderings somewhat
1915 differently since they don't take an address. See that instruction's
1916 documentation for details.
1918 For a simpler introduction to the ordering constraints, see the
1922 The set of values that can be read is governed by the happens-before
1923 partial order. A value cannot be read unless some operation wrote
1924 it. This is intended to provide a guarantee strong enough to model
1925 Java's non-volatile shared variables. This ordering cannot be
1926 specified for read-modify-write operations; it is not strong enough
1927 to make them atomic in any interesting way.
1929 In addition to the guarantees of ``unordered``, there is a single
1930 total order for modifications by ``monotonic`` operations on each
1931 address. All modification orders must be compatible with the
1932 happens-before order. There is no guarantee that the modification
1933 orders can be combined to a global total order for the whole program
1934 (and this often will not be possible). The read in an atomic
1935 read-modify-write operation (:ref:`cmpxchg <i_cmpxchg>` and
1936 :ref:`atomicrmw <i_atomicrmw>`) reads the value in the modification
1937 order immediately before the value it writes. If one atomic read
1938 happens before another atomic read of the same address, the later
1939 read must see the same value or a later value in the address's
1940 modification order. This disallows reordering of ``monotonic`` (or
1941 stronger) operations on the same address. If an address is written
1942 ``monotonic``-ally by one thread, and other threads ``monotonic``-ally
1943 read that address repeatedly, the other threads must eventually see
1944 the write. This corresponds to the C++0x/C1x
1945 ``memory_order_relaxed``.
1947 In addition to the guarantees of ``monotonic``, a
1948 *synchronizes-with* edge may be formed with a ``release`` operation.
1949 This is intended to model C++'s ``memory_order_acquire``.
1951 In addition to the guarantees of ``monotonic``, if this operation
1952 writes a value which is subsequently read by an ``acquire``
1953 operation, it *synchronizes-with* that operation. (This isn't a
1954 complete description; see the C++0x definition of a release
1955 sequence.) This corresponds to the C++0x/C1x
1956 ``memory_order_release``.
1957 ``acq_rel`` (acquire+release)
1958 Acts as both an ``acquire`` and ``release`` operation on its
1959 address. This corresponds to the C++0x/C1x ``memory_order_acq_rel``.
1960 ``seq_cst`` (sequentially consistent)
1961 In addition to the guarantees of ``acq_rel`` (``acquire`` for an
1962 operation that only reads, ``release`` for an operation that only
1963 writes), there is a global total order on all
1964 sequentially-consistent operations on all addresses, which is
1965 consistent with the *happens-before* partial order and with the
1966 modification orders of all the affected addresses. Each
1967 sequentially-consistent read sees the last preceding write to the
1968 same address in this global order. This corresponds to the C++0x/C1x
1969 ``memory_order_seq_cst`` and Java volatile.
1973 If an atomic operation is marked ``singlethread``, it only *synchronizes
1974 with* or participates in modification and seq\_cst total orderings with
1975 other operations running in the same thread (for example, in signal
1983 LLVM IR floating-point binary ops (:ref:`fadd <i_fadd>`,
1984 :ref:`fsub <i_fsub>`, :ref:`fmul <i_fmul>`, :ref:`fdiv <i_fdiv>`,
1985 :ref:`frem <i_frem>`, :ref:`fcmp <i_fcmp>`) have the following flags that can
1986 be set to enable otherwise unsafe floating point operations
1989 No NaNs - Allow optimizations to assume the arguments and result are not
1990 NaN. Such optimizations are required to retain defined behavior over
1991 NaNs, but the value of the result is undefined.
1994 No Infs - Allow optimizations to assume the arguments and result are not
1995 +/-Inf. Such optimizations are required to retain defined behavior over
1996 +/-Inf, but the value of the result is undefined.
1999 No Signed Zeros - Allow optimizations to treat the sign of a zero
2000 argument or result as insignificant.
2003 Allow Reciprocal - Allow optimizations to use the reciprocal of an
2004 argument rather than perform division.
2007 Fast - Allow algebraically equivalent transformations that may
2008 dramatically change results in floating point (e.g. reassociate). This
2009 flag implies all the others.
2013 Use-list Order Directives
2014 -------------------------
2016 Use-list directives encode the in-memory order of each use-list, allowing the
2017 order to be recreated. ``<order-indexes>`` is a comma-separated list of
2018 indexes that are assigned to the referenced value's uses. The referenced
2019 value's use-list is immediately sorted by these indexes.
2021 Use-list directives may appear at function scope or global scope. They are not
2022 instructions, and have no effect on the semantics of the IR. When they're at
2023 function scope, they must appear after the terminator of the final basic block.
2025 If basic blocks have their address taken via ``blockaddress()`` expressions,
2026 ``uselistorder_bb`` can be used to reorder their use-lists from outside their
2033 uselistorder <ty> <value>, { <order-indexes> }
2034 uselistorder_bb @function, %block { <order-indexes> }
2040 define void @foo(i32 %arg1, i32 %arg2) {
2042 ; ... instructions ...
2044 ; ... instructions ...
2046 ; At function scope.
2047 uselistorder i32 %arg1, { 1, 0, 2 }
2048 uselistorder label %bb, { 1, 0 }
2052 uselistorder i32* @global, { 1, 2, 0 }
2053 uselistorder i32 7, { 1, 0 }
2054 uselistorder i32 (i32) @bar, { 1, 0 }
2055 uselistorder_bb @foo, %bb, { 5, 1, 3, 2, 0, 4 }
2062 The LLVM type system is one of the most important features of the
2063 intermediate representation. Being typed enables a number of
2064 optimizations to be performed on the intermediate representation
2065 directly, without having to do extra analyses on the side before the
2066 transformation. A strong type system makes it easier to read the
2067 generated code and enables novel analyses and transformations that are
2068 not feasible to perform on normal three address code representations.
2078 The void type does not represent any value and has no size.
2096 The function type can be thought of as a function signature. It consists of a
2097 return type and a list of formal parameter types. The return type of a function
2098 type is a void type or first class type --- except for :ref:`label <t_label>`
2099 and :ref:`metadata <t_metadata>` types.
2105 <returntype> (<parameter list>)
2107 ...where '``<parameter list>``' is a comma-separated list of type
2108 specifiers. Optionally, the parameter list may include a type ``...``, which
2109 indicates that the function takes a variable number of arguments. Variable
2110 argument functions can access their arguments with the :ref:`variable argument
2111 handling intrinsic <int_varargs>` functions. '``<returntype>``' is any type
2112 except :ref:`label <t_label>` and :ref:`metadata <t_metadata>`.
2116 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2117 | ``i32 (i32)`` | function taking an ``i32``, returning an ``i32`` |
2118 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2119 | ``float (i16, i32 *) *`` | :ref:`Pointer <t_pointer>` to a function that takes an ``i16`` and a :ref:`pointer <t_pointer>` to ``i32``, returning ``float``. |
2120 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2121 | ``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. |
2122 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2123 | ``{i32, i32} (i32)`` | A function taking an ``i32``, returning a :ref:`structure <t_struct>` containing two ``i32`` values |
2124 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2131 The :ref:`first class <t_firstclass>` types are perhaps the most important.
2132 Values of these types are the only ones which can be produced by
2140 These are the types that are valid in registers from CodeGen's perspective.
2149 The integer type is a very simple type that simply specifies an
2150 arbitrary bit width for the integer type desired. Any bit width from 1
2151 bit to 2\ :sup:`23`\ -1 (about 8 million) can be specified.
2159 The number of bits the integer will occupy is specified by the ``N``
2165 +----------------+------------------------------------------------+
2166 | ``i1`` | a single-bit integer. |
2167 +----------------+------------------------------------------------+
2168 | ``i32`` | a 32-bit integer. |
2169 +----------------+------------------------------------------------+
2170 | ``i1942652`` | a really big integer of over 1 million bits. |
2171 +----------------+------------------------------------------------+
2175 Floating Point Types
2176 """"""""""""""""""""
2185 - 16-bit floating point value
2188 - 32-bit floating point value
2191 - 64-bit floating point value
2194 - 128-bit floating point value (112-bit mantissa)
2197 - 80-bit floating point value (X87)
2200 - 128-bit floating point value (two 64-bits)
2207 The x86_mmx type represents a value held in an MMX register on an x86
2208 machine. The operations allowed on it are quite limited: parameters and
2209 return values, load and store, and bitcast. User-specified MMX
2210 instructions are represented as intrinsic or asm calls with arguments
2211 and/or results of this type. There are no arrays, vectors or constants
2228 The pointer type is used to specify memory locations. Pointers are
2229 commonly used to reference objects in memory.
2231 Pointer types may have an optional address space attribute defining the
2232 numbered address space where the pointed-to object resides. The default
2233 address space is number zero. The semantics of non-zero address spaces
2234 are target-specific.
2236 Note that LLVM does not permit pointers to void (``void*``) nor does it
2237 permit pointers to labels (``label*``). Use ``i8*`` instead.
2247 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2248 | ``[4 x i32]*`` | A :ref:`pointer <t_pointer>` to :ref:`array <t_array>` of four ``i32`` values. |
2249 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2250 | ``i32 (i32*) *`` | A :ref:`pointer <t_pointer>` to a :ref:`function <t_function>` that takes an ``i32*``, returning an ``i32``. |
2251 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2252 | ``i32 addrspace(5)*`` | A :ref:`pointer <t_pointer>` to an ``i32`` value that resides in address space #5. |
2253 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2262 A vector type is a simple derived type that represents a vector of
2263 elements. Vector types are used when multiple primitive data are
2264 operated in parallel using a single instruction (SIMD). A vector type
2265 requires a size (number of elements) and an underlying primitive data
2266 type. Vector types are considered :ref:`first class <t_firstclass>`.
2272 < <# elements> x <elementtype> >
2274 The number of elements is a constant integer value larger than 0;
2275 elementtype may be any integer, floating point or pointer type. Vectors
2276 of size zero are not allowed.
2280 +-------------------+--------------------------------------------------+
2281 | ``<4 x i32>`` | Vector of 4 32-bit integer values. |
2282 +-------------------+--------------------------------------------------+
2283 | ``<8 x float>`` | Vector of 8 32-bit floating-point values. |
2284 +-------------------+--------------------------------------------------+
2285 | ``<2 x i64>`` | Vector of 2 64-bit integer values. |
2286 +-------------------+--------------------------------------------------+
2287 | ``<4 x i64*>`` | Vector of 4 pointers to 64-bit integer values. |
2288 +-------------------+--------------------------------------------------+
2297 The label type represents code labels.
2312 The token type is used when a value is associated with an instruction
2313 but all uses of the value must not attempt to introspect or obscure it.
2314 As such, it is not appropriate to have a :ref:`phi <i_phi>` or
2315 :ref:`select <i_select>` of type token.
2332 The metadata type represents embedded metadata. No derived types may be
2333 created from metadata except for :ref:`function <t_function>` arguments.
2346 Aggregate Types are a subset of derived types that can contain multiple
2347 member types. :ref:`Arrays <t_array>` and :ref:`structs <t_struct>` are
2348 aggregate types. :ref:`Vectors <t_vector>` are not considered to be
2358 The array type is a very simple derived type that arranges elements
2359 sequentially in memory. The array type requires a size (number of
2360 elements) and an underlying data type.
2366 [<# elements> x <elementtype>]
2368 The number of elements is a constant integer value; ``elementtype`` may
2369 be any type with a size.
2373 +------------------+--------------------------------------+
2374 | ``[40 x i32]`` | Array of 40 32-bit integer values. |
2375 +------------------+--------------------------------------+
2376 | ``[41 x i32]`` | Array of 41 32-bit integer values. |
2377 +------------------+--------------------------------------+
2378 | ``[4 x i8]`` | Array of 4 8-bit integer values. |
2379 +------------------+--------------------------------------+
2381 Here are some examples of multidimensional arrays:
2383 +-----------------------------+----------------------------------------------------------+
2384 | ``[3 x [4 x i32]]`` | 3x4 array of 32-bit integer values. |
2385 +-----------------------------+----------------------------------------------------------+
2386 | ``[12 x [10 x float]]`` | 12x10 array of single precision floating point values. |
2387 +-----------------------------+----------------------------------------------------------+
2388 | ``[2 x [3 x [4 x i16]]]`` | 2x3x4 array of 16-bit integer values. |
2389 +-----------------------------+----------------------------------------------------------+
2391 There is no restriction on indexing beyond the end of the array implied
2392 by a static type (though there are restrictions on indexing beyond the
2393 bounds of an allocated object in some cases). This means that
2394 single-dimension 'variable sized array' addressing can be implemented in
2395 LLVM with a zero length array type. An implementation of 'pascal style
2396 arrays' in LLVM could use the type "``{ i32, [0 x float]}``", for
2406 The structure type is used to represent a collection of data members
2407 together in memory. The elements of a structure may be any type that has
2410 Structures in memory are accessed using '``load``' and '``store``' by
2411 getting a pointer to a field with the '``getelementptr``' instruction.
2412 Structures in registers are accessed using the '``extractvalue``' and
2413 '``insertvalue``' instructions.
2415 Structures may optionally be "packed" structures, which indicate that
2416 the alignment of the struct is one byte, and that there is no padding
2417 between the elements. In non-packed structs, padding between field types
2418 is inserted as defined by the DataLayout string in the module, which is
2419 required to match what the underlying code generator expects.
2421 Structures can either be "literal" or "identified". A literal structure
2422 is defined inline with other types (e.g. ``{i32, i32}*``) whereas
2423 identified types are always defined at the top level with a name.
2424 Literal types are uniqued by their contents and can never be recursive
2425 or opaque since there is no way to write one. Identified types can be
2426 recursive, can be opaqued, and are never uniqued.
2432 %T1 = type { <type list> } ; Identified normal struct type
2433 %T2 = type <{ <type list> }> ; Identified packed struct type
2437 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2438 | ``{ i32, i32, i32 }`` | A triple of three ``i32`` values |
2439 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2440 | ``{ 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``. |
2441 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2442 | ``<{ i8, i32 }>`` | A packed struct known to be 5 bytes in size. |
2443 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2447 Opaque Structure Types
2448 """"""""""""""""""""""
2452 Opaque structure types are used to represent named structure types that
2453 do not have a body specified. This corresponds (for example) to the C
2454 notion of a forward declared structure.
2465 +--------------+-------------------+
2466 | ``opaque`` | An opaque type. |
2467 +--------------+-------------------+
2474 LLVM has several different basic types of constants. This section
2475 describes them all and their syntax.
2480 **Boolean constants**
2481 The two strings '``true``' and '``false``' are both valid constants
2483 **Integer constants**
2484 Standard integers (such as '4') are constants of the
2485 :ref:`integer <t_integer>` type. Negative numbers may be used with
2487 **Floating point constants**
2488 Floating point constants use standard decimal notation (e.g.
2489 123.421), exponential notation (e.g. 1.23421e+2), or a more precise
2490 hexadecimal notation (see below). The assembler requires the exact
2491 decimal value of a floating-point constant. For example, the
2492 assembler accepts 1.25 but rejects 1.3 because 1.3 is a repeating
2493 decimal in binary. Floating point constants must have a :ref:`floating
2494 point <t_floating>` type.
2495 **Null pointer constants**
2496 The identifier '``null``' is recognized as a null pointer constant
2497 and must be of :ref:`pointer type <t_pointer>`.
2499 The identifier '``none``' is recognized as an empty token constant
2500 and must be of :ref:`token type <t_token>`.
2502 The one non-intuitive notation for constants is the hexadecimal form of
2503 floating point constants. For example, the form
2504 '``double 0x432ff973cafa8000``' is equivalent to (but harder to read
2505 than) '``double 4.5e+15``'. The only time hexadecimal floating point
2506 constants are required (and the only time that they are generated by the
2507 disassembler) is when a floating point constant must be emitted but it
2508 cannot be represented as a decimal floating point number in a reasonable
2509 number of digits. For example, NaN's, infinities, and other special
2510 values are represented in their IEEE hexadecimal format so that assembly
2511 and disassembly do not cause any bits to change in the constants.
2513 When using the hexadecimal form, constants of types half, float, and
2514 double are represented using the 16-digit form shown above (which
2515 matches the IEEE754 representation for double); half and float values
2516 must, however, be exactly representable as IEEE 754 half and single
2517 precision, respectively. Hexadecimal format is always used for long
2518 double, and there are three forms of long double. The 80-bit format used
2519 by x86 is represented as ``0xK`` followed by 20 hexadecimal digits. The
2520 128-bit format used by PowerPC (two adjacent doubles) is represented by
2521 ``0xM`` followed by 32 hexadecimal digits. The IEEE 128-bit format is
2522 represented by ``0xL`` followed by 32 hexadecimal digits. Long doubles
2523 will only work if they match the long double format on your target.
2524 The IEEE 16-bit format (half precision) is represented by ``0xH``
2525 followed by 4 hexadecimal digits. All hexadecimal formats are big-endian
2526 (sign bit at the left).
2528 There are no constants of type x86_mmx.
2530 .. _complexconstants:
2535 Complex constants are a (potentially recursive) combination of simple
2536 constants and smaller complex constants.
2538 **Structure constants**
2539 Structure constants are represented with notation similar to
2540 structure type definitions (a comma separated list of elements,
2541 surrounded by braces (``{}``)). For example:
2542 "``{ i32 4, float 17.0, i32* @G }``", where "``@G``" is declared as
2543 "``@G = external global i32``". Structure constants must have
2544 :ref:`structure type <t_struct>`, and the number and types of elements
2545 must match those specified by the type.
2547 Array constants are represented with notation similar to array type
2548 definitions (a comma separated list of elements, surrounded by
2549 square brackets (``[]``)). For example:
2550 "``[ i32 42, i32 11, i32 74 ]``". Array constants must have
2551 :ref:`array type <t_array>`, and the number and types of elements must
2552 match those specified by the type. As a special case, character array
2553 constants may also be represented as a double-quoted string using the ``c``
2554 prefix. For example: "``c"Hello World\0A\00"``".
2555 **Vector constants**
2556 Vector constants are represented with notation similar to vector
2557 type definitions (a comma separated list of elements, surrounded by
2558 less-than/greater-than's (``<>``)). For example:
2559 "``< i32 42, i32 11, i32 74, i32 100 >``". Vector constants
2560 must have :ref:`vector type <t_vector>`, and the number and types of
2561 elements must match those specified by the type.
2562 **Zero initialization**
2563 The string '``zeroinitializer``' can be used to zero initialize a
2564 value to zero of *any* type, including scalar and
2565 :ref:`aggregate <t_aggregate>` types. This is often used to avoid
2566 having to print large zero initializers (e.g. for large arrays) and
2567 is always exactly equivalent to using explicit zero initializers.
2569 A metadata node is a constant tuple without types. For example:
2570 "``!{!0, !{!2, !0}, !"test"}``". Metadata can reference constant values,
2571 for example: "``!{!0, i32 0, i8* @global, i64 (i64)* @function, !"str"}``".
2572 Unlike other typed constants that are meant to be interpreted as part of
2573 the instruction stream, metadata is a place to attach additional
2574 information such as debug info.
2576 Global Variable and Function Addresses
2577 --------------------------------------
2579 The addresses of :ref:`global variables <globalvars>` and
2580 :ref:`functions <functionstructure>` are always implicitly valid
2581 (link-time) constants. These constants are explicitly referenced when
2582 the :ref:`identifier for the global <identifiers>` is used and always have
2583 :ref:`pointer <t_pointer>` type. For example, the following is a legal LLVM
2586 .. code-block:: llvm
2590 @Z = global [2 x i32*] [ i32* @X, i32* @Y ]
2597 The string '``undef``' can be used anywhere a constant is expected, and
2598 indicates that the user of the value may receive an unspecified
2599 bit-pattern. Undefined values may be of any type (other than '``label``'
2600 or '``void``') and be used anywhere a constant is permitted.
2602 Undefined values are useful because they indicate to the compiler that
2603 the program is well defined no matter what value is used. This gives the
2604 compiler more freedom to optimize. Here are some examples of
2605 (potentially surprising) transformations that are valid (in pseudo IR):
2607 .. code-block:: llvm
2617 This is safe because all of the output bits are affected by the undef
2618 bits. Any output bit can have a zero or one depending on the input bits.
2620 .. code-block:: llvm
2631 These logical operations have bits that are not always affected by the
2632 input. For example, if ``%X`` has a zero bit, then the output of the
2633 '``and``' operation will always be a zero for that bit, no matter what
2634 the corresponding bit from the '``undef``' is. As such, it is unsafe to
2635 optimize or assume that the result of the '``and``' is '``undef``'.
2636 However, it is safe to assume that all bits of the '``undef``' could be
2637 0, and optimize the '``and``' to 0. Likewise, it is safe to assume that
2638 all the bits of the '``undef``' operand to the '``or``' could be set,
2639 allowing the '``or``' to be folded to -1.
2641 .. code-block:: llvm
2643 %A = select undef, %X, %Y
2644 %B = select undef, 42, %Y
2645 %C = select %X, %Y, undef
2655 This set of examples shows that undefined '``select``' (and conditional
2656 branch) conditions can go *either way*, but they have to come from one
2657 of the two operands. In the ``%A`` example, if ``%X`` and ``%Y`` were
2658 both known to have a clear low bit, then ``%A`` would have to have a
2659 cleared low bit. However, in the ``%C`` example, the optimizer is
2660 allowed to assume that the '``undef``' operand could be the same as
2661 ``%Y``, allowing the whole '``select``' to be eliminated.
2663 .. code-block:: llvm
2665 %A = xor undef, undef
2682 This example points out that two '``undef``' operands are not
2683 necessarily the same. This can be surprising to people (and also matches
2684 C semantics) where they assume that "``X^X``" is always zero, even if
2685 ``X`` is undefined. This isn't true for a number of reasons, but the
2686 short answer is that an '``undef``' "variable" can arbitrarily change
2687 its value over its "live range". This is true because the variable
2688 doesn't actually *have a live range*. Instead, the value is logically
2689 read from arbitrary registers that happen to be around when needed, so
2690 the value is not necessarily consistent over time. In fact, ``%A`` and
2691 ``%C`` need to have the same semantics or the core LLVM "replace all
2692 uses with" concept would not hold.
2694 .. code-block:: llvm
2702 These examples show the crucial difference between an *undefined value*
2703 and *undefined behavior*. An undefined value (like '``undef``') is
2704 allowed to have an arbitrary bit-pattern. This means that the ``%A``
2705 operation can be constant folded to '``undef``', because the '``undef``'
2706 could be an SNaN, and ``fdiv`` is not (currently) defined on SNaN's.
2707 However, in the second example, we can make a more aggressive
2708 assumption: because the ``undef`` is allowed to be an arbitrary value,
2709 we are allowed to assume that it could be zero. Since a divide by zero
2710 has *undefined behavior*, we are allowed to assume that the operation
2711 does not execute at all. This allows us to delete the divide and all
2712 code after it. Because the undefined operation "can't happen", the
2713 optimizer can assume that it occurs in dead code.
2715 .. code-block:: llvm
2717 a: store undef -> %X
2718 b: store %X -> undef
2723 These examples reiterate the ``fdiv`` example: a store *of* an undefined
2724 value can be assumed to not have any effect; we can assume that the
2725 value is overwritten with bits that happen to match what was already
2726 there. However, a store *to* an undefined location could clobber
2727 arbitrary memory, therefore, it has undefined behavior.
2734 Poison values are similar to :ref:`undef values <undefvalues>`, however
2735 they also represent the fact that an instruction or constant expression
2736 that cannot evoke side effects has nevertheless detected a condition
2737 that results in undefined behavior.
2739 There is currently no way of representing a poison value in the IR; they
2740 only exist when produced by operations such as :ref:`add <i_add>` with
2743 Poison value behavior is defined in terms of value *dependence*:
2745 - Values other than :ref:`phi <i_phi>` nodes depend on their operands.
2746 - :ref:`Phi <i_phi>` nodes depend on the operand corresponding to
2747 their dynamic predecessor basic block.
2748 - Function arguments depend on the corresponding actual argument values
2749 in the dynamic callers of their functions.
2750 - :ref:`Call <i_call>` instructions depend on the :ref:`ret <i_ret>`
2751 instructions that dynamically transfer control back to them.
2752 - :ref:`Invoke <i_invoke>` instructions depend on the
2753 :ref:`ret <i_ret>`, :ref:`resume <i_resume>`, or exception-throwing
2754 call instructions that dynamically transfer control back to them.
2755 - Non-volatile loads and stores depend on the most recent stores to all
2756 of the referenced memory addresses, following the order in the IR
2757 (including loads and stores implied by intrinsics such as
2758 :ref:`@llvm.memcpy <int_memcpy>`.)
2759 - An instruction with externally visible side effects depends on the
2760 most recent preceding instruction with externally visible side
2761 effects, following the order in the IR. (This includes :ref:`volatile
2762 operations <volatile>`.)
2763 - An instruction *control-depends* on a :ref:`terminator
2764 instruction <terminators>` if the terminator instruction has
2765 multiple successors and the instruction is always executed when
2766 control transfers to one of the successors, and may not be executed
2767 when control is transferred to another.
2768 - Additionally, an instruction also *control-depends* on a terminator
2769 instruction if the set of instructions it otherwise depends on would
2770 be different if the terminator had transferred control to a different
2772 - Dependence is transitive.
2774 Poison values have the same behavior as :ref:`undef values <undefvalues>`,
2775 with the additional effect that any instruction that has a *dependence*
2776 on a poison value has undefined behavior.
2778 Here are some examples:
2780 .. code-block:: llvm
2783 %poison = sub nuw i32 0, 1 ; Results in a poison value.
2784 %still_poison = and i32 %poison, 0 ; 0, but also poison.
2785 %poison_yet_again = getelementptr i32, i32* @h, i32 %still_poison
2786 store i32 0, i32* %poison_yet_again ; memory at @h[0] is poisoned
2788 store i32 %poison, i32* @g ; Poison value stored to memory.
2789 %poison2 = load i32, i32* @g ; Poison value loaded back from memory.
2791 store volatile i32 %poison, i32* @g ; External observation; undefined behavior.
2793 %narrowaddr = bitcast i32* @g to i16*
2794 %wideaddr = bitcast i32* @g to i64*
2795 %poison3 = load i16, i16* %narrowaddr ; Returns a poison value.
2796 %poison4 = load i64, i64* %wideaddr ; Returns a poison value.
2798 %cmp = icmp slt i32 %poison, 0 ; Returns a poison value.
2799 br i1 %cmp, label %true, label %end ; Branch to either destination.
2802 store volatile i32 0, i32* @g ; This is control-dependent on %cmp, so
2803 ; it has undefined behavior.
2807 %p = phi i32 [ 0, %entry ], [ 1, %true ]
2808 ; Both edges into this PHI are
2809 ; control-dependent on %cmp, so this
2810 ; always results in a poison value.
2812 store volatile i32 0, i32* @g ; This would depend on the store in %true
2813 ; if %cmp is true, or the store in %entry
2814 ; otherwise, so this is undefined behavior.
2816 br i1 %cmp, label %second_true, label %second_end
2817 ; The same branch again, but this time the
2818 ; true block doesn't have side effects.
2825 store volatile i32 0, i32* @g ; This time, the instruction always depends
2826 ; on the store in %end. Also, it is
2827 ; control-equivalent to %end, so this is
2828 ; well-defined (ignoring earlier undefined
2829 ; behavior in this example).
2833 Addresses of Basic Blocks
2834 -------------------------
2836 ``blockaddress(@function, %block)``
2838 The '``blockaddress``' constant computes the address of the specified
2839 basic block in the specified function, and always has an ``i8*`` type.
2840 Taking the address of the entry block is illegal.
2842 This value only has defined behavior when used as an operand to the
2843 ':ref:`indirectbr <i_indirectbr>`' instruction, or for comparisons
2844 against null. Pointer equality tests between labels addresses results in
2845 undefined behavior --- though, again, comparison against null is ok, and
2846 no label is equal to the null pointer. This may be passed around as an
2847 opaque pointer sized value as long as the bits are not inspected. This
2848 allows ``ptrtoint`` and arithmetic to be performed on these values so
2849 long as the original value is reconstituted before the ``indirectbr``
2852 Finally, some targets may provide defined semantics when using the value
2853 as the operand to an inline assembly, but that is target specific.
2857 Constant Expressions
2858 --------------------
2860 Constant expressions are used to allow expressions involving other
2861 constants to be used as constants. Constant expressions may be of any
2862 :ref:`first class <t_firstclass>` type and may involve any LLVM operation
2863 that does not have side effects (e.g. load and call are not supported).
2864 The following is the syntax for constant expressions:
2866 ``trunc (CST to TYPE)``
2867 Truncate a constant to another type. The bit size of CST must be
2868 larger than the bit size of TYPE. Both types must be integers.
2869 ``zext (CST to TYPE)``
2870 Zero extend a constant to another type. The bit size of CST must be
2871 smaller than the bit size of TYPE. Both types must be integers.
2872 ``sext (CST to TYPE)``
2873 Sign extend a constant to another type. The bit size of CST must be
2874 smaller than the bit size of TYPE. Both types must be integers.
2875 ``fptrunc (CST to TYPE)``
2876 Truncate a floating point constant to another floating point type.
2877 The size of CST must be larger than the size of TYPE. Both types
2878 must be floating point.
2879 ``fpext (CST to TYPE)``
2880 Floating point extend a constant to another type. The size of CST
2881 must be smaller or equal to the size of TYPE. Both types must be
2883 ``fptoui (CST to TYPE)``
2884 Convert a floating point constant to the corresponding unsigned
2885 integer constant. TYPE must be a scalar or vector integer type. CST
2886 must be of scalar or vector floating point type. Both CST and TYPE
2887 must be scalars, or vectors of the same number of elements. If the
2888 value won't fit in the integer type, the results are undefined.
2889 ``fptosi (CST to TYPE)``
2890 Convert a floating point constant to the corresponding signed
2891 integer constant. TYPE must be a scalar or vector integer type. CST
2892 must be of scalar or vector floating point type. Both CST and TYPE
2893 must be scalars, or vectors of the same number of elements. If the
2894 value won't fit in the integer type, the results are undefined.
2895 ``uitofp (CST to TYPE)``
2896 Convert an unsigned integer constant to the corresponding floating
2897 point constant. TYPE must be a scalar or vector floating point type.
2898 CST must be of scalar or vector integer type. Both CST and TYPE must
2899 be scalars, or vectors of the same number of elements. If the value
2900 won't fit in the floating point type, the results are undefined.
2901 ``sitofp (CST to TYPE)``
2902 Convert a signed integer constant to the corresponding floating
2903 point constant. TYPE must be a scalar or vector floating point type.
2904 CST must be of scalar or vector integer type. Both CST and TYPE must
2905 be scalars, or vectors of the same number of elements. If the value
2906 won't fit in the floating point type, the results are undefined.
2907 ``ptrtoint (CST to TYPE)``
2908 Convert a pointer typed constant to the corresponding integer
2909 constant. ``TYPE`` must be an integer type. ``CST`` must be of
2910 pointer type. The ``CST`` value is zero extended, truncated, or
2911 unchanged to make it fit in ``TYPE``.
2912 ``inttoptr (CST to TYPE)``
2913 Convert an integer constant to a pointer constant. TYPE must be a
2914 pointer type. CST must be of integer type. The CST value is zero
2915 extended, truncated, or unchanged to make it fit in a pointer size.
2916 This one is *really* dangerous!
2917 ``bitcast (CST to TYPE)``
2918 Convert a constant, CST, to another TYPE. The constraints of the
2919 operands are the same as those for the :ref:`bitcast
2920 instruction <i_bitcast>`.
2921 ``addrspacecast (CST to TYPE)``
2922 Convert a constant pointer or constant vector of pointer, CST, to another
2923 TYPE in a different address space. The constraints of the operands are the
2924 same as those for the :ref:`addrspacecast instruction <i_addrspacecast>`.
2925 ``getelementptr (TY, CSTPTR, IDX0, IDX1, ...)``, ``getelementptr inbounds (TY, CSTPTR, IDX0, IDX1, ...)``
2926 Perform the :ref:`getelementptr operation <i_getelementptr>` on
2927 constants. As with the :ref:`getelementptr <i_getelementptr>`
2928 instruction, the index list may have zero or more indexes, which are
2929 required to make sense for the type of "pointer to TY".
2930 ``select (COND, VAL1, VAL2)``
2931 Perform the :ref:`select operation <i_select>` on constants.
2932 ``icmp COND (VAL1, VAL2)``
2933 Performs the :ref:`icmp operation <i_icmp>` on constants.
2934 ``fcmp COND (VAL1, VAL2)``
2935 Performs the :ref:`fcmp operation <i_fcmp>` on constants.
2936 ``extractelement (VAL, IDX)``
2937 Perform the :ref:`extractelement operation <i_extractelement>` on
2939 ``insertelement (VAL, ELT, IDX)``
2940 Perform the :ref:`insertelement operation <i_insertelement>` on
2942 ``shufflevector (VEC1, VEC2, IDXMASK)``
2943 Perform the :ref:`shufflevector operation <i_shufflevector>` on
2945 ``extractvalue (VAL, IDX0, IDX1, ...)``
2946 Perform the :ref:`extractvalue operation <i_extractvalue>` on
2947 constants. The index list is interpreted in a similar manner as
2948 indices in a ':ref:`getelementptr <i_getelementptr>`' operation. At
2949 least one index value must be specified.
2950 ``insertvalue (VAL, ELT, IDX0, IDX1, ...)``
2951 Perform the :ref:`insertvalue operation <i_insertvalue>` on constants.
2952 The index list is interpreted in a similar manner as indices in a
2953 ':ref:`getelementptr <i_getelementptr>`' operation. At least one index
2954 value must be specified.
2955 ``OPCODE (LHS, RHS)``
2956 Perform the specified operation of the LHS and RHS constants. OPCODE
2957 may be any of the :ref:`binary <binaryops>` or :ref:`bitwise
2958 binary <bitwiseops>` operations. The constraints on operands are
2959 the same as those for the corresponding instruction (e.g. no bitwise
2960 operations on floating point values are allowed).
2967 Inline Assembler Expressions
2968 ----------------------------
2970 LLVM supports inline assembler expressions (as opposed to :ref:`Module-Level
2971 Inline Assembly <moduleasm>`) through the use of a special value. This value
2972 represents the inline assembler as a template string (containing the
2973 instructions to emit), a list of operand constraints (stored as a string), a
2974 flag that indicates whether or not the inline asm expression has side effects,
2975 and a flag indicating whether the function containing the asm needs to align its
2976 stack conservatively.
2978 The template string supports argument substitution of the operands using "``$``"
2979 followed by a number, to indicate substitution of the given register/memory
2980 location, as specified by the constraint string. "``${NUM:MODIFIER}``" may also
2981 be used, where ``MODIFIER`` is a target-specific annotation for how to print the
2982 operand (See :ref:`inline-asm-modifiers`).
2984 A literal "``$``" may be included by using "``$$``" in the template. To include
2985 other special characters into the output, the usual "``\XX``" escapes may be
2986 used, just as in other strings. Note that after template substitution, the
2987 resulting assembly string is parsed by LLVM's integrated assembler unless it is
2988 disabled -- even when emitting a ``.s`` file -- and thus must contain assembly
2989 syntax known to LLVM.
2991 LLVM's support for inline asm is modeled closely on the requirements of Clang's
2992 GCC-compatible inline-asm support. Thus, the feature-set and the constraint and
2993 modifier codes listed here are similar or identical to those in GCC's inline asm
2994 support. However, to be clear, the syntax of the template and constraint strings
2995 described here is *not* the same as the syntax accepted by GCC and Clang, and,
2996 while most constraint letters are passed through as-is by Clang, some get
2997 translated to other codes when converting from the C source to the LLVM
3000 An example inline assembler expression is:
3002 .. code-block:: llvm
3004 i32 (i32) asm "bswap $0", "=r,r"
3006 Inline assembler expressions may **only** be used as the callee operand
3007 of a :ref:`call <i_call>` or an :ref:`invoke <i_invoke>` instruction.
3008 Thus, typically we have:
3010 .. code-block:: llvm
3012 %X = call i32 asm "bswap $0", "=r,r"(i32 %Y)
3014 Inline asms with side effects not visible in the constraint list must be
3015 marked as having side effects. This is done through the use of the
3016 '``sideeffect``' keyword, like so:
3018 .. code-block:: llvm
3020 call void asm sideeffect "eieio", ""()
3022 In some cases inline asms will contain code that will not work unless
3023 the stack is aligned in some way, such as calls or SSE instructions on
3024 x86, yet will not contain code that does that alignment within the asm.
3025 The compiler should make conservative assumptions about what the asm
3026 might contain and should generate its usual stack alignment code in the
3027 prologue if the '``alignstack``' keyword is present:
3029 .. code-block:: llvm
3031 call void asm alignstack "eieio", ""()
3033 Inline asms also support using non-standard assembly dialects. The
3034 assumed dialect is ATT. When the '``inteldialect``' keyword is present,
3035 the inline asm is using the Intel dialect. Currently, ATT and Intel are
3036 the only supported dialects. An example is:
3038 .. code-block:: llvm
3040 call void asm inteldialect "eieio", ""()
3042 If multiple keywords appear the '``sideeffect``' keyword must come
3043 first, the '``alignstack``' keyword second and the '``inteldialect``'
3046 Inline Asm Constraint String
3047 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3049 The constraint list is a comma-separated string, each element containing one or
3050 more constraint codes.
3052 For each element in the constraint list an appropriate register or memory
3053 operand will be chosen, and it will be made available to assembly template
3054 string expansion as ``$0`` for the first constraint in the list, ``$1`` for the
3057 There are three different types of constraints, which are distinguished by a
3058 prefix symbol in front of the constraint code: Output, Input, and Clobber. The
3059 constraints must always be given in that order: outputs first, then inputs, then
3060 clobbers. They cannot be intermingled.
3062 There are also three different categories of constraint codes:
3064 - Register constraint. This is either a register class, or a fixed physical
3065 register. This kind of constraint will allocate a register, and if necessary,
3066 bitcast the argument or result to the appropriate type.
3067 - Memory constraint. This kind of constraint is for use with an instruction
3068 taking a memory operand. Different constraints allow for different addressing
3069 modes used by the target.
3070 - Immediate value constraint. This kind of constraint is for an integer or other
3071 immediate value which can be rendered directly into an instruction. The
3072 various target-specific constraints allow the selection of a value in the
3073 proper range for the instruction you wish to use it with.
3078 Output constraints are specified by an "``=``" prefix (e.g. "``=r``"). This
3079 indicates that the assembly will write to this operand, and the operand will
3080 then be made available as a return value of the ``asm`` expression. Output
3081 constraints do not consume an argument from the call instruction. (Except, see
3082 below about indirect outputs).
3084 Normally, it is expected that no output locations are written to by the assembly
3085 expression until *all* of the inputs have been read. As such, LLVM may assign
3086 the same register to an output and an input. If this is not safe (e.g. if the
3087 assembly contains two instructions, where the first writes to one output, and
3088 the second reads an input and writes to a second output), then the "``&``"
3089 modifier must be used (e.g. "``=&r``") to specify that the output is an
3090 "early-clobber" output. Marking an ouput as "early-clobber" ensures that LLVM
3091 will not use the same register for any inputs (other than an input tied to this
3097 Input constraints do not have a prefix -- just the constraint codes. Each input
3098 constraint will consume one argument from the call instruction. It is not
3099 permitted for the asm to write to any input register or memory location (unless
3100 that input is tied to an output). Note also that multiple inputs may all be
3101 assigned to the same register, if LLVM can determine that they necessarily all
3102 contain the same value.
3104 Instead of providing a Constraint Code, input constraints may also "tie"
3105 themselves to an output constraint, by providing an integer as the constraint
3106 string. Tied inputs still consume an argument from the call instruction, and
3107 take up a position in the asm template numbering as is usual -- they will simply
3108 be constrained to always use the same register as the output they've been tied
3109 to. For example, a constraint string of "``=r,0``" says to assign a register for
3110 output, and use that register as an input as well (it being the 0'th
3113 It is permitted to tie an input to an "early-clobber" output. In that case, no
3114 *other* input may share the same register as the input tied to the early-clobber
3115 (even when the other input has the same value).
3117 You may only tie an input to an output which has a register constraint, not a
3118 memory constraint. Only a single input may be tied to an output.
3120 There is also an "interesting" feature which deserves a bit of explanation: if a
3121 register class constraint allocates a register which is too small for the value
3122 type operand provided as input, the input value will be split into multiple
3123 registers, and all of them passed to the inline asm.
3125 However, this feature is often not as useful as you might think.
3127 Firstly, the registers are *not* guaranteed to be consecutive. So, on those
3128 architectures that have instructions which operate on multiple consecutive
3129 instructions, this is not an appropriate way to support them. (e.g. the 32-bit
3130 SparcV8 has a 64-bit load, which instruction takes a single 32-bit register. The
3131 hardware then loads into both the named register, and the next register. This
3132 feature of inline asm would not be useful to support that.)
3134 A few of the targets provide a template string modifier allowing explicit access
3135 to the second register of a two-register operand (e.g. MIPS ``L``, ``M``, and
3136 ``D``). On such an architecture, you can actually access the second allocated
3137 register (yet, still, not any subsequent ones). But, in that case, you're still
3138 probably better off simply splitting the value into two separate operands, for
3139 clarity. (e.g. see the description of the ``A`` constraint on X86, which,
3140 despite existing only for use with this feature, is not really a good idea to
3143 Indirect inputs and outputs
3144 """""""""""""""""""""""""""
3146 Indirect output or input constraints can be specified by the "``*``" modifier
3147 (which goes after the "``=``" in case of an output). This indicates that the asm
3148 will write to or read from the contents of an *address* provided as an input
3149 argument. (Note that in this way, indirect outputs act more like an *input* than
3150 an output: just like an input, they consume an argument of the call expression,
3151 rather than producing a return value. An indirect output constraint is an
3152 "output" only in that the asm is expected to write to the contents of the input
3153 memory location, instead of just read from it).
3155 This is most typically used for memory constraint, e.g. "``=*m``", to pass the
3156 address of a variable as a value.
3158 It is also possible to use an indirect *register* constraint, but only on output
3159 (e.g. "``=*r``"). This will cause LLVM to allocate a register for an output
3160 value normally, and then, separately emit a store to the address provided as
3161 input, after the provided inline asm. (It's not clear what value this
3162 functionality provides, compared to writing the store explicitly after the asm
3163 statement, and it can only produce worse code, since it bypasses many
3164 optimization passes. I would recommend not using it.)
3170 A clobber constraint is indicated by a "``~``" prefix. A clobber does not
3171 consume an input operand, nor generate an output. Clobbers cannot use any of the
3172 general constraint code letters -- they may use only explicit register
3173 constraints, e.g. "``~{eax}``". The one exception is that a clobber string of
3174 "``~{memory}``" indicates that the assembly writes to arbitrary undeclared
3175 memory locations -- not only the memory pointed to by a declared indirect
3181 After a potential prefix comes constraint code, or codes.
3183 A Constraint Code is either a single letter (e.g. "``r``"), a "``^``" character
3184 followed by two letters (e.g. "``^wc``"), or "``{``" register-name "``}``"
3187 The one and two letter constraint codes are typically chosen to be the same as
3188 GCC's constraint codes.
3190 A single constraint may include one or more than constraint code in it, leaving
3191 it up to LLVM to choose which one to use. This is included mainly for
3192 compatibility with the translation of GCC inline asm coming from clang.
3194 There are two ways to specify alternatives, and either or both may be used in an
3195 inline asm constraint list:
3197 1) Append the codes to each other, making a constraint code set. E.g. "``im``"
3198 or "``{eax}m``". This means "choose any of the options in the set". The
3199 choice of constraint is made independently for each constraint in the
3202 2) Use "``|``" between constraint code sets, creating alternatives. Every
3203 constraint in the constraint list must have the same number of alternative
3204 sets. With this syntax, the same alternative in *all* of the items in the
3205 constraint list will be chosen together.
3207 Putting those together, you might have a two operand constraint string like
3208 ``"rm|r,ri|rm"``. This indicates that if operand 0 is ``r`` or ``m``, then
3209 operand 1 may be one of ``r`` or ``i``. If operand 0 is ``r``, then operand 1
3210 may be one of ``r`` or ``m``. But, operand 0 and 1 cannot both be of type m.
3212 However, the use of either of the alternatives features is *NOT* recommended, as
3213 LLVM is not able to make an intelligent choice about which one to use. (At the
3214 point it currently needs to choose, not enough information is available to do so
3215 in a smart way.) Thus, it simply tries to make a choice that's most likely to
3216 compile, not one that will be optimal performance. (e.g., given "``rm``", it'll
3217 always choose to use memory, not registers). And, if given multiple registers,
3218 or multiple register classes, it will simply choose the first one. (In fact, it
3219 doesn't currently even ensure explicitly specified physical registers are
3220 unique, so specifying multiple physical registers as alternatives, like
3221 ``{r11}{r12},{r11}{r12}``, will assign r11 to both operands, not at all what was
3224 Supported Constraint Code List
3225 """"""""""""""""""""""""""""""
3227 The constraint codes are, in general, expected to behave the same way they do in
3228 GCC. LLVM's support is often implemented on an 'as-needed' basis, to support C
3229 inline asm code which was supported by GCC. A mismatch in behavior between LLVM
3230 and GCC likely indicates a bug in LLVM.
3232 Some constraint codes are typically supported by all targets:
3234 - ``r``: A register in the target's general purpose register class.
3235 - ``m``: A memory address operand. It is target-specific what addressing modes
3236 are supported, typical examples are register, or register + register offset,
3237 or register + immediate offset (of some target-specific size).
3238 - ``i``: An integer constant (of target-specific width). Allows either a simple
3239 immediate, or a relocatable value.
3240 - ``n``: An integer constant -- *not* including relocatable values.
3241 - ``s``: An integer constant, but allowing *only* relocatable values.
3242 - ``X``: Allows an operand of any kind, no constraint whatsoever. Typically
3243 useful to pass a label for an asm branch or call.
3245 .. FIXME: but that surely isn't actually okay to jump out of an asm
3246 block without telling llvm about the control transfer???)
3248 - ``{register-name}``: Requires exactly the named physical register.
3250 Other constraints are target-specific:
3254 - ``z``: An immediate integer 0. Outputs ``WZR`` or ``XZR``, as appropriate.
3255 - ``I``: An immediate integer valid for an ``ADD`` or ``SUB`` instruction,
3256 i.e. 0 to 4095 with optional shift by 12.
3257 - ``J``: An immediate integer that, when negated, is valid for an ``ADD`` or
3258 ``SUB`` instruction, i.e. -1 to -4095 with optional left shift by 12.
3259 - ``K``: An immediate integer that is valid for the 'bitmask immediate 32' of a
3260 logical instruction like ``AND``, ``EOR``, or ``ORR`` with a 32-bit register.
3261 - ``L``: An immediate integer that is valid for the 'bitmask immediate 64' of a
3262 logical instruction like ``AND``, ``EOR``, or ``ORR`` with a 64-bit register.
3263 - ``M``: An immediate integer for use with the ``MOV`` assembly alias on a
3264 32-bit register. This is a superset of ``K``: in addition to the bitmask
3265 immediate, also allows immediate integers which can be loaded with a single
3266 ``MOVZ`` or ``MOVL`` instruction.
3267 - ``N``: An immediate integer for use with the ``MOV`` assembly alias on a
3268 64-bit register. This is a superset of ``L``.
3269 - ``Q``: Memory address operand must be in a single register (no
3270 offsets). (However, LLVM currently does this for the ``m`` constraint as
3272 - ``r``: A 32 or 64-bit integer register (W* or X*).
3273 - ``w``: A 32, 64, or 128-bit floating-point/SIMD register.
3274 - ``x``: A lower 128-bit floating-point/SIMD register (``V0`` to ``V15``).
3278 - ``r``: A 32 or 64-bit integer register.
3279 - ``[0-9]v``: The 32-bit VGPR register, number 0-9.
3280 - ``[0-9]s``: The 32-bit SGPR register, number 0-9.
3285 - ``Q``, ``Um``, ``Un``, ``Uq``, ``Us``, ``Ut``, ``Uv``, ``Uy``: Memory address
3286 operand. Treated the same as operand ``m``, at the moment.
3288 ARM and ARM's Thumb2 mode:
3290 - ``j``: An immediate integer between 0 and 65535 (valid for ``MOVW``)
3291 - ``I``: An immediate integer valid for a data-processing instruction.
3292 - ``J``: An immediate integer between -4095 and 4095.
3293 - ``K``: An immediate integer whose bitwise inverse is valid for a
3294 data-processing instruction. (Can be used with template modifier "``B``" to
3295 print the inverted value).
3296 - ``L``: An immediate integer whose negation is valid for a data-processing
3297 instruction. (Can be used with template modifier "``n``" to print the negated
3299 - ``M``: A power of two or a integer between 0 and 32.
3300 - ``N``: Invalid immediate constraint.
3301 - ``O``: Invalid immediate constraint.
3302 - ``r``: A general-purpose 32-bit integer register (``r0-r15``).
3303 - ``l``: In Thumb2 mode, low 32-bit GPR registers (``r0-r7``). In ARM mode, same
3305 - ``h``: In Thumb2 mode, a high 32-bit GPR register (``r8-r15``). In ARM mode,
3307 - ``w``: A 32, 64, or 128-bit floating-point/SIMD register: ``s0-s31``,
3308 ``d0-d31``, or ``q0-q15``.
3309 - ``x``: A 32, 64, or 128-bit floating-point/SIMD register: ``s0-s15``,
3310 ``d0-d7``, or ``q0-q3``.
3311 - ``t``: A floating-point/SIMD register, only supports 32-bit values:
3316 - ``I``: An immediate integer between 0 and 255.
3317 - ``J``: An immediate integer between -255 and -1.
3318 - ``K``: An immediate integer between 0 and 255, with optional left-shift by
3320 - ``L``: An immediate integer between -7 and 7.
3321 - ``M``: An immediate integer which is a multiple of 4 between 0 and 1020.
3322 - ``N``: An immediate integer between 0 and 31.
3323 - ``O``: An immediate integer which is a multiple of 4 between -508 and 508.
3324 - ``r``: A low 32-bit GPR register (``r0-r7``).
3325 - ``l``: A low 32-bit GPR register (``r0-r7``).
3326 - ``h``: A high GPR register (``r0-r7``).
3327 - ``w``: A 32, 64, or 128-bit floating-point/SIMD register: ``s0-s31``,
3328 ``d0-d31``, or ``q0-q15``.
3329 - ``x``: A 32, 64, or 128-bit floating-point/SIMD register: ``s0-s15``,
3330 ``d0-d7``, or ``q0-q3``.
3331 - ``t``: A floating-point/SIMD register, only supports 32-bit values:
3337 - ``o``, ``v``: A memory address operand, treated the same as constraint ``m``,
3339 - ``r``: A 32 or 64-bit register.
3343 - ``r``: An 8 or 16-bit register.
3347 - ``I``: An immediate signed 16-bit integer.
3348 - ``J``: An immediate integer zero.
3349 - ``K``: An immediate unsigned 16-bit integer.
3350 - ``L``: An immediate 32-bit integer, where the lower 16 bits are 0.
3351 - ``N``: An immediate integer between -65535 and -1.
3352 - ``O``: An immediate signed 15-bit integer.
3353 - ``P``: An immediate integer between 1 and 65535.
3354 - ``m``: A memory address operand. In MIPS-SE mode, allows a base address
3355 register plus 16-bit immediate offset. In MIPS mode, just a base register.
3356 - ``R``: A memory address operand. In MIPS-SE mode, allows a base address
3357 register plus a 9-bit signed offset. In MIPS mode, the same as constraint
3359 - ``ZC``: A memory address operand, suitable for use in a ``pref``, ``ll``, or
3360 ``sc`` instruction on the given subtarget (details vary).
3361 - ``r``, ``d``, ``y``: A 32 or 64-bit GPR register.
3362 - ``f``: A 32 or 64-bit FPU register (``F0-F31``), or a 128-bit MSA register
3363 (``W0-W31``). In the case of MSA registers, it is recommended to use the ``w``
3364 argument modifier for compatibility with GCC.
3365 - ``c``: A 32-bit or 64-bit GPR register suitable for indirect jump (always
3367 - ``l``: The ``lo`` register, 32 or 64-bit.
3372 - ``b``: A 1-bit integer register.
3373 - ``c`` or ``h``: A 16-bit integer register.
3374 - ``r``: A 32-bit integer register.
3375 - ``l`` or ``N``: A 64-bit integer register.
3376 - ``f``: A 32-bit float register.
3377 - ``d``: A 64-bit float register.
3382 - ``I``: An immediate signed 16-bit integer.
3383 - ``J``: An immediate unsigned 16-bit integer, shifted left 16 bits.
3384 - ``K``: An immediate unsigned 16-bit integer.
3385 - ``L``: An immediate signed 16-bit integer, shifted left 16 bits.
3386 - ``M``: An immediate integer greater than 31.
3387 - ``N``: An immediate integer that is an exact power of 2.
3388 - ``O``: The immediate integer constant 0.
3389 - ``P``: An immediate integer constant whose negation is a signed 16-bit
3391 - ``es``, ``o``, ``Q``, ``Z``, ``Zy``: A memory address operand, currently
3392 treated the same as ``m``.
3393 - ``r``: A 32 or 64-bit integer register.
3394 - ``b``: A 32 or 64-bit integer register, excluding ``R0`` (that is:
3396 - ``f``: A 32 or 64-bit float register (``F0-F31``), or when QPX is enabled, a
3397 128 or 256-bit QPX register (``Q0-Q31``; aliases the ``F`` registers).
3398 - ``v``: For ``4 x f32`` or ``4 x f64`` types, when QPX is enabled, a
3399 128 or 256-bit QPX register (``Q0-Q31``), otherwise a 128-bit
3400 altivec vector register (``V0-V31``).
3402 .. FIXME: is this a bug that v accepts QPX registers? I think this
3403 is supposed to only use the altivec vector registers?
3405 - ``y``: Condition register (``CR0-CR7``).
3406 - ``wc``: An individual CR bit in a CR register.
3407 - ``wa``, ``wd``, ``wf``: Any 128-bit VSX vector register, from the full VSX
3408 register set (overlapping both the floating-point and vector register files).
3409 - ``ws``: A 32 or 64-bit floating point register, from the full VSX register
3414 - ``I``: An immediate 13-bit signed integer.
3415 - ``r``: A 32-bit integer register.
3419 - ``I``: An immediate unsigned 8-bit integer.
3420 - ``J``: An immediate unsigned 12-bit integer.
3421 - ``K``: An immediate signed 16-bit integer.
3422 - ``L``: An immediate signed 20-bit integer.
3423 - ``M``: An immediate integer 0x7fffffff.
3424 - ``Q``, ``R``, ``S``, ``T``: A memory address operand, treated the same as
3425 ``m``, at the moment.
3426 - ``r`` or ``d``: A 32, 64, or 128-bit integer register.
3427 - ``a``: A 32, 64, or 128-bit integer address register (excludes R0, which in an
3428 address context evaluates as zero).
3429 - ``h``: A 32-bit value in the high part of a 64bit data register
3431 - ``f``: A 32, 64, or 128-bit floating point register.
3435 - ``I``: An immediate integer between 0 and 31.
3436 - ``J``: An immediate integer between 0 and 64.
3437 - ``K``: An immediate signed 8-bit integer.
3438 - ``L``: An immediate integer, 0xff or 0xffff or (in 64-bit mode only)
3440 - ``M``: An immediate integer between 0 and 3.
3441 - ``N``: An immediate unsigned 8-bit integer.
3442 - ``O``: An immediate integer between 0 and 127.
3443 - ``e``: An immediate 32-bit signed integer.
3444 - ``Z``: An immediate 32-bit unsigned integer.
3445 - ``o``, ``v``: Treated the same as ``m``, at the moment.
3446 - ``q``: An 8, 16, 32, or 64-bit register which can be accessed as an 8-bit
3447 ``l`` integer register. On X86-32, this is the ``a``, ``b``, ``c``, and ``d``
3448 registers, and on X86-64, it is all of the integer registers.
3449 - ``Q``: An 8, 16, 32, or 64-bit register which can be accessed as an 8-bit
3450 ``h`` integer register. This is the ``a``, ``b``, ``c``, and ``d`` registers.
3451 - ``r`` or ``l``: An 8, 16, 32, or 64-bit integer register.
3452 - ``R``: An 8, 16, 32, or 64-bit "legacy" integer register -- one which has
3453 existed since i386, and can be accessed without the REX prefix.
3454 - ``f``: A 32, 64, or 80-bit '387 FPU stack pseudo-register.
3455 - ``y``: A 64-bit MMX register, if MMX is enabled.
3456 - ``x``: If SSE is enabled: a 32 or 64-bit scalar operand, or 128-bit vector
3457 operand in a SSE register. If AVX is also enabled, can also be a 256-bit
3458 vector operand in an AVX register. If AVX-512 is also enabled, can also be a
3459 512-bit vector operand in an AVX512 register, Otherwise, an error.
3460 - ``Y``: The same as ``x``, if *SSE2* is enabled, otherwise an error.
3461 - ``A``: Special case: allocates EAX first, then EDX, for a single operand (in
3462 32-bit mode, a 64-bit integer operand will get split into two registers). It
3463 is not recommended to use this constraint, as in 64-bit mode, the 64-bit
3464 operand will get allocated only to RAX -- if two 32-bit operands are needed,
3465 you're better off splitting it yourself, before passing it to the asm
3470 - ``r``: A 32-bit integer register.
3473 .. _inline-asm-modifiers:
3475 Asm template argument modifiers
3476 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3478 In the asm template string, modifiers can be used on the operand reference, like
3481 The modifiers are, in general, expected to behave the same way they do in
3482 GCC. LLVM's support is often implemented on an 'as-needed' basis, to support C
3483 inline asm code which was supported by GCC. A mismatch in behavior between LLVM
3484 and GCC likely indicates a bug in LLVM.
3488 - ``c``: Print an immediate integer constant unadorned, without
3489 the target-specific immediate punctuation (e.g. no ``$`` prefix).
3490 - ``n``: Negate and print immediate integer constant unadorned, without the
3491 target-specific immediate punctuation (e.g. no ``$`` prefix).
3492 - ``l``: Print as an unadorned label, without the target-specific label
3493 punctuation (e.g. no ``$`` prefix).
3497 - ``w``: Print a GPR register with a ``w*`` name instead of ``x*`` name. E.g.,
3498 instead of ``x30``, print ``w30``.
3499 - ``x``: Print a GPR register with a ``x*`` name. (this is the default, anyhow).
3500 - ``b``, ``h``, ``s``, ``d``, ``q``: Print a floating-point/SIMD register with a
3501 ``b*``, ``h*``, ``s*``, ``d*``, or ``q*`` name, rather than the default of
3510 - ``a``: Print an operand as an address (with ``[`` and ``]`` surrounding a
3514 - ``y``: Print a VFP single-precision register as an indexed double (e.g. print
3515 as ``d4[1]`` instead of ``s9``)
3516 - ``B``: Bitwise invert and print an immediate integer constant without ``#``
3518 - ``L``: Print the low 16-bits of an immediate integer constant.
3519 - ``M``: Print as a register set suitable for ldm/stm. Also prints *all*
3520 register operands subsequent to the specified one (!), so use carefully.
3521 - ``Q``: Print the low-order register of a register-pair, or the low-order
3522 register of a two-register operand.
3523 - ``R``: Print the high-order register of a register-pair, or the high-order
3524 register of a two-register operand.
3525 - ``H``: Print the second register of a register-pair. (On a big-endian system,
3526 ``H`` is equivalent to ``Q``, and on little-endian system, ``H`` is equivalent
3529 .. FIXME: H doesn't currently support printing the second register
3530 of a two-register operand.
3532 - ``e``: Print the low doubleword register of a NEON quad register.
3533 - ``f``: Print the high doubleword register of a NEON quad register.
3534 - ``m``: Print the base register of a memory operand without the ``[`` and ``]``
3539 - ``L``: Print the second register of a two-register operand. Requires that it
3540 has been allocated consecutively to the first.
3542 .. FIXME: why is it restricted to consecutive ones? And there's
3543 nothing that ensures that happens, is there?
3545 - ``I``: Print the letter 'i' if the operand is an integer constant, otherwise
3546 nothing. Used to print 'addi' vs 'add' instructions.
3550 No additional modifiers.
3554 - ``X``: Print an immediate integer as hexadecimal
3555 - ``x``: Print the low 16 bits of an immediate integer as hexadecimal.
3556 - ``d``: Print an immediate integer as decimal.
3557 - ``m``: Subtract one and print an immediate integer as decimal.
3558 - ``z``: Print $0 if an immediate zero, otherwise print normally.
3559 - ``L``: Print the low-order register of a two-register operand, or prints the
3560 address of the low-order word of a double-word memory operand.
3562 .. FIXME: L seems to be missing memory operand support.
3564 - ``M``: Print the high-order register of a two-register operand, or prints the
3565 address of the high-order word of a double-word memory operand.
3567 .. FIXME: M seems to be missing memory operand support.
3569 - ``D``: Print the second register of a two-register operand, or prints the
3570 second word of a double-word memory operand. (On a big-endian system, ``D`` is
3571 equivalent to ``L``, and on little-endian system, ``D`` is equivalent to
3573 - ``w``: No effect. Provided for compatibility with GCC which requires this
3574 modifier in order to print MSA registers (``W0-W31``) with the ``f``
3583 - ``L``: Print the second register of a two-register operand. Requires that it
3584 has been allocated consecutively to the first.
3586 .. FIXME: why is it restricted to consecutive ones? And there's
3587 nothing that ensures that happens, is there?
3589 - ``I``: Print the letter 'i' if the operand is an integer constant, otherwise
3590 nothing. Used to print 'addi' vs 'add' instructions.
3591 - ``y``: For a memory operand, prints formatter for a two-register X-form
3592 instruction. (Currently always prints ``r0,OPERAND``).
3593 - ``U``: Prints 'u' if the memory operand is an update form, and nothing
3594 otherwise. (NOTE: LLVM does not support update form, so this will currently
3595 always print nothing)
3596 - ``X``: Prints 'x' if the memory operand is an indexed form. (NOTE: LLVM does
3597 not support indexed form, so this will currently always print nothing)
3605 SystemZ implements only ``n``, and does *not* support any of the other
3606 target-independent modifiers.
3610 - ``c``: Print an unadorned integer or symbol name. (The latter is
3611 target-specific behavior for this typically target-independent modifier).
3612 - ``A``: Print a register name with a '``*``' before it.
3613 - ``b``: Print an 8-bit register name (e.g. ``al``); do nothing on a memory
3615 - ``h``: Print the upper 8-bit register name (e.g. ``ah``); do nothing on a
3617 - ``w``: Print the 16-bit register name (e.g. ``ax``); do nothing on a memory
3619 - ``k``: Print the 32-bit register name (e.g. ``eax``); do nothing on a memory
3621 - ``q``: Print the 64-bit register name (e.g. ``rax``), if 64-bit registers are
3622 available, otherwise the 32-bit register name; do nothing on a memory operand.
3623 - ``n``: Negate and print an unadorned integer, or, for operands other than an
3624 immediate integer (e.g. a relocatable symbol expression), print a '-' before
3625 the operand. (The behavior for relocatable symbol expressions is a
3626 target-specific behavior for this typically target-independent modifier)
3627 - ``H``: Print a memory reference with additional offset +8.
3628 - ``P``: Print a memory reference or operand for use as the argument of a call
3629 instruction. (E.g. omit ``(rip)``, even though it's PC-relative.)
3633 No additional modifiers.
3639 The call instructions that wrap inline asm nodes may have a
3640 "``!srcloc``" MDNode attached to it that contains a list of constant
3641 integers. If present, the code generator will use the integer as the
3642 location cookie value when report errors through the ``LLVMContext``
3643 error reporting mechanisms. This allows a front-end to correlate backend
3644 errors that occur with inline asm back to the source code that produced
3647 .. code-block:: llvm
3649 call void asm sideeffect "something bad", ""(), !srcloc !42
3651 !42 = !{ i32 1234567 }
3653 It is up to the front-end to make sense of the magic numbers it places
3654 in the IR. If the MDNode contains multiple constants, the code generator
3655 will use the one that corresponds to the line of the asm that the error
3663 LLVM IR allows metadata to be attached to instructions in the program
3664 that can convey extra information about the code to the optimizers and
3665 code generator. One example application of metadata is source-level
3666 debug information. There are two metadata primitives: strings and nodes.
3668 Metadata does not have a type, and is not a value. If referenced from a
3669 ``call`` instruction, it uses the ``metadata`` type.
3671 All metadata are identified in syntax by a exclamation point ('``!``').
3673 .. _metadata-string:
3675 Metadata Nodes and Metadata Strings
3676 -----------------------------------
3678 A metadata string is a string surrounded by double quotes. It can
3679 contain any character by escaping non-printable characters with
3680 "``\xx``" where "``xx``" is the two digit hex code. For example:
3683 Metadata nodes are represented with notation similar to structure
3684 constants (a comma separated list of elements, surrounded by braces and
3685 preceded by an exclamation point). Metadata nodes can have any values as
3686 their operand. For example:
3688 .. code-block:: llvm
3690 !{ !"test\00", i32 10}
3692 Metadata nodes that aren't uniqued use the ``distinct`` keyword. For example:
3694 .. code-block:: llvm
3696 !0 = distinct !{!"test\00", i32 10}
3698 ``distinct`` nodes are useful when nodes shouldn't be merged based on their
3699 content. They can also occur when transformations cause uniquing collisions
3700 when metadata operands change.
3702 A :ref:`named metadata <namedmetadatastructure>` is a collection of
3703 metadata nodes, which can be looked up in the module symbol table. For
3706 .. code-block:: llvm
3710 Metadata can be used as function arguments. Here ``llvm.dbg.value``
3711 function is using two metadata arguments:
3713 .. code-block:: llvm
3715 call void @llvm.dbg.value(metadata !24, i64 0, metadata !25)
3717 Metadata can be attached to an instruction. Here metadata ``!21`` is attached
3718 to the ``add`` instruction using the ``!dbg`` identifier:
3720 .. code-block:: llvm
3722 %indvar.next = add i64 %indvar, 1, !dbg !21
3724 Metadata can also be attached to a function definition. Here metadata ``!22``
3725 is attached to the ``foo`` function using the ``!dbg`` identifier:
3727 .. code-block:: llvm
3729 define void @foo() !dbg !22 {
3733 More information about specific metadata nodes recognized by the
3734 optimizers and code generator is found below.
3736 .. _specialized-metadata:
3738 Specialized Metadata Nodes
3739 ^^^^^^^^^^^^^^^^^^^^^^^^^^
3741 Specialized metadata nodes are custom data structures in metadata (as opposed
3742 to generic tuples). Their fields are labelled, and can be specified in any
3745 These aren't inherently debug info centric, but currently all the specialized
3746 metadata nodes are related to debug info.
3753 ``DICompileUnit`` nodes represent a compile unit. The ``enums:``,
3754 ``retainedTypes:``, ``subprograms:``, ``globals:``, ``imports:`` and ``macros:``
3755 fields are tuples containing the debug info to be emitted along with the compile
3756 unit, regardless of code optimizations (some nodes are only emitted if there are
3757 references to them from instructions).
3759 .. code-block:: llvm
3761 !0 = !DICompileUnit(language: DW_LANG_C99, file: !1, producer: "clang",
3762 isOptimized: true, flags: "-O2", runtimeVersion: 2,
3763 splitDebugFilename: "abc.debug", emissionKind: 1,
3764 enums: !2, retainedTypes: !3, subprograms: !4,
3765 globals: !5, imports: !6, macros: !7, dwoId: 0x0abcd)
3767 Compile unit descriptors provide the root scope for objects declared in a
3768 specific compilation unit. File descriptors are defined using this scope.
3769 These descriptors are collected by a named metadata ``!llvm.dbg.cu``. They
3770 keep track of subprograms, global variables, type information, and imported
3771 entities (declarations and namespaces).
3778 ``DIFile`` nodes represent files. The ``filename:`` can include slashes.
3780 .. code-block:: llvm
3782 !0 = !DIFile(filename: "path/to/file", directory: "/path/to/dir")
3784 Files are sometimes used in ``scope:`` fields, and are the only valid target
3785 for ``file:`` fields.
3792 ``DIBasicType`` nodes represent primitive types, such as ``int``, ``bool`` and
3793 ``float``. ``tag:`` defaults to ``DW_TAG_base_type``.
3795 .. code-block:: llvm
3797 !0 = !DIBasicType(name: "unsigned char", size: 8, align: 8,
3798 encoding: DW_ATE_unsigned_char)
3799 !1 = !DIBasicType(tag: DW_TAG_unspecified_type, name: "decltype(nullptr)")
3801 The ``encoding:`` describes the details of the type. Usually it's one of the
3804 .. code-block:: llvm
3810 DW_ATE_signed_char = 6
3812 DW_ATE_unsigned_char = 8
3814 .. _DISubroutineType:
3819 ``DISubroutineType`` nodes represent subroutine types. Their ``types:`` field
3820 refers to a tuple; the first operand is the return type, while the rest are the
3821 types of the formal arguments in order. If the first operand is ``null``, that
3822 represents a function with no return value (such as ``void foo() {}`` in C++).
3824 .. code-block:: llvm
3826 !0 = !BasicType(name: "int", size: 32, align: 32, DW_ATE_signed)
3827 !1 = !BasicType(name: "char", size: 8, align: 8, DW_ATE_signed_char)
3828 !2 = !DISubroutineType(types: !{null, !0, !1}) ; void (int, char)
3835 ``DIDerivedType`` nodes represent types derived from other types, such as
3838 .. code-block:: llvm
3840 !0 = !DIBasicType(name: "unsigned char", size: 8, align: 8,
3841 encoding: DW_ATE_unsigned_char)
3842 !1 = !DIDerivedType(tag: DW_TAG_pointer_type, baseType: !0, size: 32,
3845 The following ``tag:`` values are valid:
3847 .. code-block:: llvm
3849 DW_TAG_formal_parameter = 5
3851 DW_TAG_pointer_type = 15
3852 DW_TAG_reference_type = 16
3854 DW_TAG_ptr_to_member_type = 31
3855 DW_TAG_const_type = 38
3856 DW_TAG_volatile_type = 53
3857 DW_TAG_restrict_type = 55
3859 ``DW_TAG_member`` is used to define a member of a :ref:`composite type
3860 <DICompositeType>` or :ref:`subprogram <DISubprogram>`. The type of the member
3861 is the ``baseType:``. The ``offset:`` is the member's bit offset.
3862 ``DW_TAG_formal_parameter`` is used to define a member which is a formal
3863 argument of a subprogram.
3865 ``DW_TAG_typedef`` is used to provide a name for the ``baseType:``.
3867 ``DW_TAG_pointer_type``, ``DW_TAG_reference_type``, ``DW_TAG_const_type``,
3868 ``DW_TAG_volatile_type`` and ``DW_TAG_restrict_type`` are used to qualify the
3871 Note that the ``void *`` type is expressed as a type derived from NULL.
3873 .. _DICompositeType:
3878 ``DICompositeType`` nodes represent types composed of other types, like
3879 structures and unions. ``elements:`` points to a tuple of the composed types.
3881 If the source language supports ODR, the ``identifier:`` field gives the unique
3882 identifier used for type merging between modules. When specified, other types
3883 can refer to composite types indirectly via a :ref:`metadata string
3884 <metadata-string>` that matches their identifier.
3886 .. code-block:: llvm
3888 !0 = !DIEnumerator(name: "SixKind", value: 7)
3889 !1 = !DIEnumerator(name: "SevenKind", value: 7)
3890 !2 = !DIEnumerator(name: "NegEightKind", value: -8)
3891 !3 = !DICompositeType(tag: DW_TAG_enumeration_type, name: "Enum", file: !12,
3892 line: 2, size: 32, align: 32, identifier: "_M4Enum",
3893 elements: !{!0, !1, !2})
3895 The following ``tag:`` values are valid:
3897 .. code-block:: llvm
3899 DW_TAG_array_type = 1
3900 DW_TAG_class_type = 2
3901 DW_TAG_enumeration_type = 4
3902 DW_TAG_structure_type = 19
3903 DW_TAG_union_type = 23
3904 DW_TAG_subroutine_type = 21
3905 DW_TAG_inheritance = 28
3908 For ``DW_TAG_array_type``, the ``elements:`` should be :ref:`subrange
3909 descriptors <DISubrange>`, each representing the range of subscripts at that
3910 level of indexing. The ``DIFlagVector`` flag to ``flags:`` indicates that an
3911 array type is a native packed vector.
3913 For ``DW_TAG_enumeration_type``, the ``elements:`` should be :ref:`enumerator
3914 descriptors <DIEnumerator>`, each representing the definition of an enumeration
3915 value for the set. All enumeration type descriptors are collected in the
3916 ``enums:`` field of the :ref:`compile unit <DICompileUnit>`.
3918 For ``DW_TAG_structure_type``, ``DW_TAG_class_type``, and
3919 ``DW_TAG_union_type``, the ``elements:`` should be :ref:`derived types
3920 <DIDerivedType>` with ``tag: DW_TAG_member`` or ``tag: DW_TAG_inheritance``.
3927 ``DISubrange`` nodes are the elements for ``DW_TAG_array_type`` variants of
3928 :ref:`DICompositeType`. ``count: -1`` indicates an empty array.
3930 .. code-block:: llvm
3932 !0 = !DISubrange(count: 5, lowerBound: 0) ; array counting from 0
3933 !1 = !DISubrange(count: 5, lowerBound: 1) ; array counting from 1
3934 !2 = !DISubrange(count: -1) ; empty array.
3941 ``DIEnumerator`` nodes are the elements for ``DW_TAG_enumeration_type``
3942 variants of :ref:`DICompositeType`.
3944 .. code-block:: llvm
3946 !0 = !DIEnumerator(name: "SixKind", value: 7)
3947 !1 = !DIEnumerator(name: "SevenKind", value: 7)
3948 !2 = !DIEnumerator(name: "NegEightKind", value: -8)
3950 DITemplateTypeParameter
3951 """""""""""""""""""""""
3953 ``DITemplateTypeParameter`` nodes represent type parameters to generic source
3954 language constructs. They are used (optionally) in :ref:`DICompositeType` and
3955 :ref:`DISubprogram` ``templateParams:`` fields.
3957 .. code-block:: llvm
3959 !0 = !DITemplateTypeParameter(name: "Ty", type: !1)
3961 DITemplateValueParameter
3962 """"""""""""""""""""""""
3964 ``DITemplateValueParameter`` nodes represent value parameters to generic source
3965 language constructs. ``tag:`` defaults to ``DW_TAG_template_value_parameter``,
3966 but if specified can also be set to ``DW_TAG_GNU_template_template_param`` or
3967 ``DW_TAG_GNU_template_param_pack``. They are used (optionally) in
3968 :ref:`DICompositeType` and :ref:`DISubprogram` ``templateParams:`` fields.
3970 .. code-block:: llvm
3972 !0 = !DITemplateValueParameter(name: "Ty", type: !1, value: i32 7)
3977 ``DINamespace`` nodes represent namespaces in the source language.
3979 .. code-block:: llvm
3981 !0 = !DINamespace(name: "myawesomeproject", scope: !1, file: !2, line: 7)
3986 ``DIGlobalVariable`` nodes represent global variables in the source language.
3988 .. code-block:: llvm
3990 !0 = !DIGlobalVariable(name: "foo", linkageName: "foo", scope: !1,
3991 file: !2, line: 7, type: !3, isLocal: true,
3992 isDefinition: false, variable: i32* @foo,
3995 All global variables should be referenced by the `globals:` field of a
3996 :ref:`compile unit <DICompileUnit>`.
4003 ``DISubprogram`` nodes represent functions from the source language. A
4004 ``DISubprogram`` may be attached to a function definition using ``!dbg``
4005 metadata. The ``variables:`` field points at :ref:`variables <DILocalVariable>`
4006 that must be retained, even if their IR counterparts are optimized out of
4007 the IR. The ``type:`` field must point at an :ref:`DISubroutineType`.
4009 .. code-block:: llvm
4011 define void @_Z3foov() !dbg !0 {
4015 !0 = distinct !DISubprogram(name: "foo", linkageName: "_Zfoov", scope: !1,
4016 file: !2, line: 7, type: !3, isLocal: true,
4017 isDefinition: false, scopeLine: 8,
4019 virtuality: DW_VIRTUALITY_pure_virtual,
4020 virtualIndex: 10, flags: DIFlagPrototyped,
4021 isOptimized: true, templateParams: !5,
4022 declaration: !6, variables: !7)
4029 ``DILexicalBlock`` nodes describe nested blocks within a :ref:`subprogram
4030 <DISubprogram>`. The line number and column numbers are used to distinguish
4031 two lexical blocks at same depth. They are valid targets for ``scope:``
4034 .. code-block:: llvm
4036 !0 = distinct !DILexicalBlock(scope: !1, file: !2, line: 7, column: 35)
4038 Usually lexical blocks are ``distinct`` to prevent node merging based on
4041 .. _DILexicalBlockFile:
4046 ``DILexicalBlockFile`` nodes are used to discriminate between sections of a
4047 :ref:`lexical block <DILexicalBlock>`. The ``file:`` field can be changed to
4048 indicate textual inclusion, or the ``discriminator:`` field can be used to
4049 discriminate between control flow within a single block in the source language.
4051 .. code-block:: llvm
4053 !0 = !DILexicalBlock(scope: !3, file: !4, line: 7, column: 35)
4054 !1 = !DILexicalBlockFile(scope: !0, file: !4, discriminator: 0)
4055 !2 = !DILexicalBlockFile(scope: !0, file: !4, discriminator: 1)
4062 ``DILocation`` nodes represent source debug locations. The ``scope:`` field is
4063 mandatory, and points at an :ref:`DILexicalBlockFile`, an
4064 :ref:`DILexicalBlock`, or an :ref:`DISubprogram`.
4066 .. code-block:: llvm
4068 !0 = !DILocation(line: 2900, column: 42, scope: !1, inlinedAt: !2)
4070 .. _DILocalVariable:
4075 ``DILocalVariable`` nodes represent local variables in the source language. If
4076 the ``arg:`` field is set to non-zero, then this variable is a subprogram
4077 parameter, and it will be included in the ``variables:`` field of its
4078 :ref:`DISubprogram`.
4080 .. code-block:: llvm
4082 !0 = !DILocalVariable(name: "this", arg: 1, scope: !3, file: !2, line: 7,
4083 type: !3, flags: DIFlagArtificial)
4084 !1 = !DILocalVariable(name: "x", arg: 2, scope: !4, file: !2, line: 7,
4086 !2 = !DILocalVariable(name: "y", scope: !5, file: !2, line: 7, type: !3)
4091 ``DIExpression`` nodes represent DWARF expression sequences. They are used in
4092 :ref:`debug intrinsics<dbg_intrinsics>` (such as ``llvm.dbg.declare``) to
4093 describe how the referenced LLVM variable relates to the source language
4096 The current supported vocabulary is limited:
4098 - ``DW_OP_deref`` dereferences the working expression.
4099 - ``DW_OP_plus, 93`` adds ``93`` to the working expression.
4100 - ``DW_OP_bit_piece, 16, 8`` specifies the offset and size (``16`` and ``8``
4101 here, respectively) of the variable piece from the working expression.
4103 .. code-block:: llvm
4105 !0 = !DIExpression(DW_OP_deref)
4106 !1 = !DIExpression(DW_OP_plus, 3)
4107 !2 = !DIExpression(DW_OP_bit_piece, 3, 7)
4108 !3 = !DIExpression(DW_OP_deref, DW_OP_plus, 3, DW_OP_bit_piece, 3, 7)
4113 ``DIObjCProperty`` nodes represent Objective-C property nodes.
4115 .. code-block:: llvm
4117 !3 = !DIObjCProperty(name: "foo", file: !1, line: 7, setter: "setFoo",
4118 getter: "getFoo", attributes: 7, type: !2)
4123 ``DIImportedEntity`` nodes represent entities (such as modules) imported into a
4126 .. code-block:: llvm
4128 !2 = !DIImportedEntity(tag: DW_TAG_imported_module, name: "foo", scope: !0,
4129 entity: !1, line: 7)
4134 ``DIMacro`` nodes represent definition or undefinition of a macro identifiers.
4135 The ``name:`` field is the macro identifier, followed by macro parameters when
4136 definining a function-like macro, and the ``value`` field is the token-string
4137 used to expand the macro identifier.
4139 .. code-block:: llvm
4141 !2 = !DIMacro(macinfo: DW_MACINFO_define, line: 7, name: "foo(x)",
4143 !3 = !DIMacro(macinfo: DW_MACINFO_undef, line: 30, name: "foo")
4148 ``DIMacroFile`` nodes represent inclusion of source files.
4149 The ``nodes:`` field is a list of ``DIMacro`` and ``DIMacroFile`` nodes that
4150 appear in the included source file.
4152 .. code-block:: llvm
4154 !2 = !DIMacroFile(macinfo: DW_MACINFO_start_file, line: 7, file: !2,
4160 In LLVM IR, memory does not have types, so LLVM's own type system is not
4161 suitable for doing TBAA. Instead, metadata is added to the IR to
4162 describe a type system of a higher level language. This can be used to
4163 implement typical C/C++ TBAA, but it can also be used to implement
4164 custom alias analysis behavior for other languages.
4166 The current metadata format is very simple. TBAA metadata nodes have up
4167 to three fields, e.g.:
4169 .. code-block:: llvm
4171 !0 = !{ !"an example type tree" }
4172 !1 = !{ !"int", !0 }
4173 !2 = !{ !"float", !0 }
4174 !3 = !{ !"const float", !2, i64 1 }
4176 The first field is an identity field. It can be any value, usually a
4177 metadata string, which uniquely identifies the type. The most important
4178 name in the tree is the name of the root node. Two trees with different
4179 root node names are entirely disjoint, even if they have leaves with
4182 The second field identifies the type's parent node in the tree, or is
4183 null or omitted for a root node. A type is considered to alias all of
4184 its descendants and all of its ancestors in the tree. Also, a type is
4185 considered to alias all types in other trees, so that bitcode produced
4186 from multiple front-ends is handled conservatively.
4188 If the third field is present, it's an integer which if equal to 1
4189 indicates that the type is "constant" (meaning
4190 ``pointsToConstantMemory`` should return true; see `other useful
4191 AliasAnalysis methods <AliasAnalysis.html#OtherItfs>`_).
4193 '``tbaa.struct``' Metadata
4194 ^^^^^^^^^^^^^^^^^^^^^^^^^^
4196 The :ref:`llvm.memcpy <int_memcpy>` is often used to implement
4197 aggregate assignment operations in C and similar languages, however it
4198 is defined to copy a contiguous region of memory, which is more than
4199 strictly necessary for aggregate types which contain holes due to
4200 padding. Also, it doesn't contain any TBAA information about the fields
4203 ``!tbaa.struct`` metadata can describe which memory subregions in a
4204 memcpy are padding and what the TBAA tags of the struct are.
4206 The current metadata format is very simple. ``!tbaa.struct`` metadata
4207 nodes are a list of operands which are in conceptual groups of three.
4208 For each group of three, the first operand gives the byte offset of a
4209 field in bytes, the second gives its size in bytes, and the third gives
4212 .. code-block:: llvm
4214 !4 = !{ i64 0, i64 4, !1, i64 8, i64 4, !2 }
4216 This describes a struct with two fields. The first is at offset 0 bytes
4217 with size 4 bytes, and has tbaa tag !1. The second is at offset 8 bytes
4218 and has size 4 bytes and has tbaa tag !2.
4220 Note that the fields need not be contiguous. In this example, there is a
4221 4 byte gap between the two fields. This gap represents padding which
4222 does not carry useful data and need not be preserved.
4224 '``noalias``' and '``alias.scope``' Metadata
4225 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4227 ``noalias`` and ``alias.scope`` metadata provide the ability to specify generic
4228 noalias memory-access sets. This means that some collection of memory access
4229 instructions (loads, stores, memory-accessing calls, etc.) that carry
4230 ``noalias`` metadata can specifically be specified not to alias with some other
4231 collection of memory access instructions that carry ``alias.scope`` metadata.
4232 Each type of metadata specifies a list of scopes where each scope has an id and
4233 a domain. When evaluating an aliasing query, if for some domain, the set
4234 of scopes with that domain in one instruction's ``alias.scope`` list is a
4235 subset of (or equal to) the set of scopes for that domain in another
4236 instruction's ``noalias`` list, then the two memory accesses are assumed not to
4239 The metadata identifying each domain is itself a list containing one or two
4240 entries. The first entry is the name of the domain. Note that if the name is a
4241 string then it can be combined across functions and translation units. A
4242 self-reference can be used to create globally unique domain names. A
4243 descriptive string may optionally be provided as a second list entry.
4245 The metadata identifying each scope is also itself a list containing two or
4246 three entries. The first entry is the name of the scope. Note that if the name
4247 is a string then it can be combined across functions and translation units. A
4248 self-reference can be used to create globally unique scope names. A metadata
4249 reference to the scope's domain is the second entry. A descriptive string may
4250 optionally be provided as a third list entry.
4254 .. code-block:: llvm
4256 ; Two scope domains:
4260 ; Some scopes in these domains:
4266 !5 = !{!4} ; A list containing only scope !4
4270 ; These two instructions don't alias:
4271 %0 = load float, float* %c, align 4, !alias.scope !5
4272 store float %0, float* %arrayidx.i, align 4, !noalias !5
4274 ; These two instructions also don't alias (for domain !1, the set of scopes
4275 ; in the !alias.scope equals that in the !noalias list):
4276 %2 = load float, float* %c, align 4, !alias.scope !5
4277 store float %2, float* %arrayidx.i2, align 4, !noalias !6
4279 ; These two instructions may alias (for domain !0, the set of scopes in
4280 ; the !noalias list is not a superset of, or equal to, the scopes in the
4281 ; !alias.scope list):
4282 %2 = load float, float* %c, align 4, !alias.scope !6
4283 store float %0, float* %arrayidx.i, align 4, !noalias !7
4285 '``fpmath``' Metadata
4286 ^^^^^^^^^^^^^^^^^^^^^
4288 ``fpmath`` metadata may be attached to any instruction of floating point
4289 type. It can be used to express the maximum acceptable error in the
4290 result of that instruction, in ULPs, thus potentially allowing the
4291 compiler to use a more efficient but less accurate method of computing
4292 it. ULP is defined as follows:
4294 If ``x`` is a real number that lies between two finite consecutive
4295 floating-point numbers ``a`` and ``b``, without being equal to one
4296 of them, then ``ulp(x) = |b - a|``, otherwise ``ulp(x)`` is the
4297 distance between the two non-equal finite floating-point numbers
4298 nearest ``x``. Moreover, ``ulp(NaN)`` is ``NaN``.
4300 The metadata node shall consist of a single positive floating point
4301 number representing the maximum relative error, for example:
4303 .. code-block:: llvm
4305 !0 = !{ float 2.5 } ; maximum acceptable inaccuracy is 2.5 ULPs
4309 '``range``' Metadata
4310 ^^^^^^^^^^^^^^^^^^^^
4312 ``range`` metadata may be attached only to ``load``, ``call`` and ``invoke`` of
4313 integer types. It expresses the possible ranges the loaded value or the value
4314 returned by the called function at this call site is in. The ranges are
4315 represented with a flattened list of integers. The loaded value or the value
4316 returned is known to be in the union of the ranges defined by each consecutive
4317 pair. Each pair has the following properties:
4319 - The type must match the type loaded by the instruction.
4320 - The pair ``a,b`` represents the range ``[a,b)``.
4321 - Both ``a`` and ``b`` are constants.
4322 - The range is allowed to wrap.
4323 - The range should not represent the full or empty set. That is,
4326 In addition, the pairs must be in signed order of the lower bound and
4327 they must be non-contiguous.
4331 .. code-block:: llvm
4333 %a = load i8, i8* %x, align 1, !range !0 ; Can only be 0 or 1
4334 %b = load i8, i8* %y, align 1, !range !1 ; Can only be 255 (-1), 0 or 1
4335 %c = call i8 @foo(), !range !2 ; Can only be 0, 1, 3, 4 or 5
4336 %d = invoke i8 @bar() to label %cont
4337 unwind label %lpad, !range !3 ; Can only be -2, -1, 3, 4 or 5
4339 !0 = !{ i8 0, i8 2 }
4340 !1 = !{ i8 255, i8 2 }
4341 !2 = !{ i8 0, i8 2, i8 3, i8 6 }
4342 !3 = !{ i8 -2, i8 0, i8 3, i8 6 }
4344 '``unpredictable``' Metadata
4345 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4347 ``unpredictable`` metadata may be attached to any branch or switch
4348 instruction. It can be used to express the unpredictability of control
4349 flow. Similar to the llvm.expect intrinsic, it may be used to alter
4350 optimizations related to compare and branch instructions. The metadata
4351 is treated as a boolean value; if it exists, it signals that the branch
4352 or switch that it is attached to is completely unpredictable.
4357 It is sometimes useful to attach information to loop constructs. Currently,
4358 loop metadata is implemented as metadata attached to the branch instruction
4359 in the loop latch block. This type of metadata refer to a metadata node that is
4360 guaranteed to be separate for each loop. The loop identifier metadata is
4361 specified with the name ``llvm.loop``.
4363 The loop identifier metadata is implemented using a metadata that refers to
4364 itself to avoid merging it with any other identifier metadata, e.g.,
4365 during module linkage or function inlining. That is, each loop should refer
4366 to their own identification metadata even if they reside in separate functions.
4367 The following example contains loop identifier metadata for two separate loop
4370 .. code-block:: llvm
4375 The loop identifier metadata can be used to specify additional
4376 per-loop metadata. Any operands after the first operand can be treated
4377 as user-defined metadata. For example the ``llvm.loop.unroll.count``
4378 suggests an unroll factor to the loop unroller:
4380 .. code-block:: llvm
4382 br i1 %exitcond, label %._crit_edge, label %.lr.ph, !llvm.loop !0
4385 !1 = !{!"llvm.loop.unroll.count", i32 4}
4387 '``llvm.loop.vectorize``' and '``llvm.loop.interleave``'
4388 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4390 Metadata prefixed with ``llvm.loop.vectorize`` or ``llvm.loop.interleave`` are
4391 used to control per-loop vectorization and interleaving parameters such as
4392 vectorization width and interleave count. These metadata should be used in
4393 conjunction with ``llvm.loop`` loop identification metadata. The
4394 ``llvm.loop.vectorize`` and ``llvm.loop.interleave`` metadata are only
4395 optimization hints and the optimizer will only interleave and vectorize loops if
4396 it believes it is safe to do so. The ``llvm.mem.parallel_loop_access`` metadata
4397 which contains information about loop-carried memory dependencies can be helpful
4398 in determining the safety of these transformations.
4400 '``llvm.loop.interleave.count``' Metadata
4401 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4403 This metadata suggests an interleave count to the loop interleaver.
4404 The first operand is the string ``llvm.loop.interleave.count`` and the
4405 second operand is an integer specifying the interleave count. For
4408 .. code-block:: llvm
4410 !0 = !{!"llvm.loop.interleave.count", i32 4}
4412 Note that setting ``llvm.loop.interleave.count`` to 1 disables interleaving
4413 multiple iterations of the loop. If ``llvm.loop.interleave.count`` is set to 0
4414 then the interleave count will be determined automatically.
4416 '``llvm.loop.vectorize.enable``' Metadata
4417 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4419 This metadata selectively enables or disables vectorization for the loop. The
4420 first operand is the string ``llvm.loop.vectorize.enable`` and the second operand
4421 is a bit. If the bit operand value is 1 vectorization is enabled. A value of
4422 0 disables vectorization:
4424 .. code-block:: llvm
4426 !0 = !{!"llvm.loop.vectorize.enable", i1 0}
4427 !1 = !{!"llvm.loop.vectorize.enable", i1 1}
4429 '``llvm.loop.vectorize.width``' Metadata
4430 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4432 This metadata sets the target width of the vectorizer. The first
4433 operand is the string ``llvm.loop.vectorize.width`` and the second
4434 operand is an integer specifying the width. For example:
4436 .. code-block:: llvm
4438 !0 = !{!"llvm.loop.vectorize.width", i32 4}
4440 Note that setting ``llvm.loop.vectorize.width`` to 1 disables
4441 vectorization of the loop. If ``llvm.loop.vectorize.width`` is set to
4442 0 or if the loop does not have this metadata the width will be
4443 determined automatically.
4445 '``llvm.loop.unroll``'
4446 ^^^^^^^^^^^^^^^^^^^^^^
4448 Metadata prefixed with ``llvm.loop.unroll`` are loop unrolling
4449 optimization hints such as the unroll factor. ``llvm.loop.unroll``
4450 metadata should be used in conjunction with ``llvm.loop`` loop
4451 identification metadata. The ``llvm.loop.unroll`` metadata are only
4452 optimization hints and the unrolling will only be performed if the
4453 optimizer believes it is safe to do so.
4455 '``llvm.loop.unroll.count``' Metadata
4456 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4458 This metadata suggests an unroll factor to the loop unroller. The
4459 first operand is the string ``llvm.loop.unroll.count`` and the second
4460 operand is a positive integer specifying the unroll factor. For
4463 .. code-block:: llvm
4465 !0 = !{!"llvm.loop.unroll.count", i32 4}
4467 If the trip count of the loop is less than the unroll count the loop
4468 will be partially unrolled.
4470 '``llvm.loop.unroll.disable``' Metadata
4471 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4473 This metadata disables loop unrolling. The metadata has a single operand
4474 which is the string ``llvm.loop.unroll.disable``. For example:
4476 .. code-block:: llvm
4478 !0 = !{!"llvm.loop.unroll.disable"}
4480 '``llvm.loop.unroll.runtime.disable``' Metadata
4481 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4483 This metadata disables runtime loop unrolling. The metadata has a single
4484 operand which is the string ``llvm.loop.unroll.runtime.disable``. For example:
4486 .. code-block:: llvm
4488 !0 = !{!"llvm.loop.unroll.runtime.disable"}
4490 '``llvm.loop.unroll.enable``' Metadata
4491 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4493 This metadata suggests that the loop should be fully unrolled if the trip count
4494 is known at compile time and partially unrolled if the trip count is not known
4495 at compile time. The metadata has a single operand which is the string
4496 ``llvm.loop.unroll.enable``. For example:
4498 .. code-block:: llvm
4500 !0 = !{!"llvm.loop.unroll.enable"}
4502 '``llvm.loop.unroll.full``' Metadata
4503 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4505 This metadata suggests that the loop should be unrolled fully. The
4506 metadata has a single operand which is the string ``llvm.loop.unroll.full``.
4509 .. code-block:: llvm
4511 !0 = !{!"llvm.loop.unroll.full"}
4516 Metadata types used to annotate memory accesses with information helpful
4517 for optimizations are prefixed with ``llvm.mem``.
4519 '``llvm.mem.parallel_loop_access``' Metadata
4520 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4522 The ``llvm.mem.parallel_loop_access`` metadata refers to a loop identifier,
4523 or metadata containing a list of loop identifiers for nested loops.
4524 The metadata is attached to memory accessing instructions and denotes that
4525 no loop carried memory dependence exist between it and other instructions denoted
4526 with the same loop identifier.
4528 Precisely, given two instructions ``m1`` and ``m2`` that both have the
4529 ``llvm.mem.parallel_loop_access`` metadata, with ``L1`` and ``L2`` being the
4530 set of loops associated with that metadata, respectively, then there is no loop
4531 carried dependence between ``m1`` and ``m2`` for loops in both ``L1`` and
4534 As a special case, if all memory accessing instructions in a loop have
4535 ``llvm.mem.parallel_loop_access`` metadata that refers to that loop, then the
4536 loop has no loop carried memory dependences and is considered to be a parallel
4539 Note that if not all memory access instructions have such metadata referring to
4540 the loop, then the loop is considered not being trivially parallel. Additional
4541 memory dependence analysis is required to make that determination. As a fail
4542 safe mechanism, this causes loops that were originally parallel to be considered
4543 sequential (if optimization passes that are unaware of the parallel semantics
4544 insert new memory instructions into the loop body).
4546 Example of a loop that is considered parallel due to its correct use of
4547 both ``llvm.loop`` and ``llvm.mem.parallel_loop_access``
4548 metadata types that refer to the same loop identifier metadata.
4550 .. code-block:: llvm
4554 %val0 = load i32, i32* %arrayidx, !llvm.mem.parallel_loop_access !0
4556 store i32 %val0, i32* %arrayidx1, !llvm.mem.parallel_loop_access !0
4558 br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !0
4564 It is also possible to have nested parallel loops. In that case the
4565 memory accesses refer to a list of loop identifier metadata nodes instead of
4566 the loop identifier metadata node directly:
4568 .. code-block:: llvm
4572 %val1 = load i32, i32* %arrayidx3, !llvm.mem.parallel_loop_access !2
4574 br label %inner.for.body
4578 %val0 = load i32, i32* %arrayidx1, !llvm.mem.parallel_loop_access !0
4580 store i32 %val0, i32* %arrayidx2, !llvm.mem.parallel_loop_access !0
4582 br i1 %exitcond, label %inner.for.end, label %inner.for.body, !llvm.loop !1
4586 store i32 %val1, i32* %arrayidx4, !llvm.mem.parallel_loop_access !2
4588 br i1 %exitcond, label %outer.for.end, label %outer.for.body, !llvm.loop !2
4590 outer.for.end: ; preds = %for.body
4592 !0 = !{!1, !2} ; a list of loop identifiers
4593 !1 = !{!1} ; an identifier for the inner loop
4594 !2 = !{!2} ; an identifier for the outer loop
4599 The ``llvm.bitsets`` global metadata is used to implement
4600 :doc:`bitsets <BitSets>`.
4602 '``invariant.group``' Metadata
4603 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4605 The ``invariant.group`` metadata may be attached to ``load``/``store`` instructions.
4606 The existence of the ``invariant.group`` metadata on the instruction tells
4607 the optimizer that every ``load`` and ``store`` to the same pointer operand
4608 within the same invariant group can be assumed to load or store the same
4609 value (but see the ``llvm.invariant.group.barrier`` intrinsic which affects
4610 when two pointers are considered the same).
4614 .. code-block:: llvm
4616 @unknownPtr = external global i8
4619 store i8 42, i8* %ptr, !invariant.group !0
4620 call void @foo(i8* %ptr)
4622 %a = load i8, i8* %ptr, !invariant.group !0 ; Can assume that value under %ptr didn't change
4623 call void @foo(i8* %ptr)
4624 %b = load i8, i8* %ptr, !invariant.group !1 ; Can't assume anything, because group changed
4626 %newPtr = call i8* @getPointer(i8* %ptr)
4627 %c = load i8, i8* %newPtr, !invariant.group !0 ; Can't assume anything, because we only have information about %ptr
4629 %unknownValue = load i8, i8* @unknownPtr
4630 store i8 %unknownValue, i8* %ptr, !invariant.group !0 ; Can assume that %unknownValue == 42
4632 call void @foo(i8* %ptr)
4633 %newPtr2 = call i8* @llvm.invariant.group.barrier(i8* %ptr)
4634 %d = load i8, i8* %newPtr2, !invariant.group !0 ; Can't step through invariant.group.barrier to get value of %ptr
4637 declare void @foo(i8*)
4638 declare i8* @getPointer(i8*)
4639 declare i8* @llvm.invariant.group.barrier(i8*)
4641 !0 = !{!"magic ptr"}
4642 !1 = !{!"other ptr"}
4646 Module Flags Metadata
4647 =====================
4649 Information about the module as a whole is difficult to convey to LLVM's
4650 subsystems. The LLVM IR isn't sufficient to transmit this information.
4651 The ``llvm.module.flags`` named metadata exists in order to facilitate
4652 this. These flags are in the form of key / value pairs --- much like a
4653 dictionary --- making it easy for any subsystem who cares about a flag to
4656 The ``llvm.module.flags`` metadata contains a list of metadata triplets.
4657 Each triplet has the following form:
4659 - The first element is a *behavior* flag, which specifies the behavior
4660 when two (or more) modules are merged together, and it encounters two
4661 (or more) metadata with the same ID. The supported behaviors are
4663 - The second element is a metadata string that is a unique ID for the
4664 metadata. Each module may only have one flag entry for each unique ID (not
4665 including entries with the **Require** behavior).
4666 - The third element is the value of the flag.
4668 When two (or more) modules are merged together, the resulting
4669 ``llvm.module.flags`` metadata is the union of the modules' flags. That is, for
4670 each unique metadata ID string, there will be exactly one entry in the merged
4671 modules ``llvm.module.flags`` metadata table, and the value for that entry will
4672 be determined by the merge behavior flag, as described below. The only exception
4673 is that entries with the *Require* behavior are always preserved.
4675 The following behaviors are supported:
4686 Emits an error if two values disagree, otherwise the resulting value
4687 is that of the operands.
4691 Emits a warning if two values disagree. The result value will be the
4692 operand for the flag from the first module being linked.
4696 Adds a requirement that another module flag be present and have a
4697 specified value after linking is performed. The value must be a
4698 metadata pair, where the first element of the pair is the ID of the
4699 module flag to be restricted, and the second element of the pair is
4700 the value the module flag should be restricted to. This behavior can
4701 be used to restrict the allowable results (via triggering of an
4702 error) of linking IDs with the **Override** behavior.
4706 Uses the specified value, regardless of the behavior or value of the
4707 other module. If both modules specify **Override**, but the values
4708 differ, an error will be emitted.
4712 Appends the two values, which are required to be metadata nodes.
4716 Appends the two values, which are required to be metadata
4717 nodes. However, duplicate entries in the second list are dropped
4718 during the append operation.
4720 It is an error for a particular unique flag ID to have multiple behaviors,
4721 except in the case of **Require** (which adds restrictions on another metadata
4722 value) or **Override**.
4724 An example of module flags:
4726 .. code-block:: llvm
4728 !0 = !{ i32 1, !"foo", i32 1 }
4729 !1 = !{ i32 4, !"bar", i32 37 }
4730 !2 = !{ i32 2, !"qux", i32 42 }
4731 !3 = !{ i32 3, !"qux",
4736 !llvm.module.flags = !{ !0, !1, !2, !3 }
4738 - Metadata ``!0`` has the ID ``!"foo"`` and the value '1'. The behavior
4739 if two or more ``!"foo"`` flags are seen is to emit an error if their
4740 values are not equal.
4742 - Metadata ``!1`` has the ID ``!"bar"`` and the value '37'. The
4743 behavior if two or more ``!"bar"`` flags are seen is to use the value
4746 - Metadata ``!2`` has the ID ``!"qux"`` and the value '42'. The
4747 behavior if two or more ``!"qux"`` flags are seen is to emit a
4748 warning if their values are not equal.
4750 - Metadata ``!3`` has the ID ``!"qux"`` and the value:
4756 The behavior is to emit an error if the ``llvm.module.flags`` does not
4757 contain a flag with the ID ``!"foo"`` that has the value '1' after linking is
4760 Objective-C Garbage Collection Module Flags Metadata
4761 ----------------------------------------------------
4763 On the Mach-O platform, Objective-C stores metadata about garbage
4764 collection in a special section called "image info". The metadata
4765 consists of a version number and a bitmask specifying what types of
4766 garbage collection are supported (if any) by the file. If two or more
4767 modules are linked together their garbage collection metadata needs to
4768 be merged rather than appended together.
4770 The Objective-C garbage collection module flags metadata consists of the
4771 following key-value pairs:
4780 * - ``Objective-C Version``
4781 - **[Required]** --- The Objective-C ABI version. Valid values are 1 and 2.
4783 * - ``Objective-C Image Info Version``
4784 - **[Required]** --- The version of the image info section. Currently
4787 * - ``Objective-C Image Info Section``
4788 - **[Required]** --- The section to place the metadata. Valid values are
4789 ``"__OBJC, __image_info, regular"`` for Objective-C ABI version 1, and
4790 ``"__DATA,__objc_imageinfo, regular, no_dead_strip"`` for
4791 Objective-C ABI version 2.
4793 * - ``Objective-C Garbage Collection``
4794 - **[Required]** --- Specifies whether garbage collection is supported or
4795 not. Valid values are 0, for no garbage collection, and 2, for garbage
4796 collection supported.
4798 * - ``Objective-C GC Only``
4799 - **[Optional]** --- Specifies that only garbage collection is supported.
4800 If present, its value must be 6. This flag requires that the
4801 ``Objective-C Garbage Collection`` flag have the value 2.
4803 Some important flag interactions:
4805 - If a module with ``Objective-C Garbage Collection`` set to 0 is
4806 merged with a module with ``Objective-C Garbage Collection`` set to
4807 2, then the resulting module has the
4808 ``Objective-C Garbage Collection`` flag set to 0.
4809 - A module with ``Objective-C Garbage Collection`` set to 0 cannot be
4810 merged with a module with ``Objective-C GC Only`` set to 6.
4812 Automatic Linker Flags Module Flags Metadata
4813 --------------------------------------------
4815 Some targets support embedding flags to the linker inside individual object
4816 files. Typically this is used in conjunction with language extensions which
4817 allow source files to explicitly declare the libraries they depend on, and have
4818 these automatically be transmitted to the linker via object files.
4820 These flags are encoded in the IR using metadata in the module flags section,
4821 using the ``Linker Options`` key. The merge behavior for this flag is required
4822 to be ``AppendUnique``, and the value for the key is expected to be a metadata
4823 node which should be a list of other metadata nodes, each of which should be a
4824 list of metadata strings defining linker options.
4826 For example, the following metadata section specifies two separate sets of
4827 linker options, presumably to link against ``libz`` and the ``Cocoa``
4830 !0 = !{ i32 6, !"Linker Options",
4833 !{ !"-framework", !"Cocoa" } } }
4834 !llvm.module.flags = !{ !0 }
4836 The metadata encoding as lists of lists of options, as opposed to a collapsed
4837 list of options, is chosen so that the IR encoding can use multiple option
4838 strings to specify e.g., a single library, while still having that specifier be
4839 preserved as an atomic element that can be recognized by a target specific
4840 assembly writer or object file emitter.
4842 Each individual option is required to be either a valid option for the target's
4843 linker, or an option that is reserved by the target specific assembly writer or
4844 object file emitter. No other aspect of these options is defined by the IR.
4846 C type width Module Flags Metadata
4847 ----------------------------------
4849 The ARM backend emits a section into each generated object file describing the
4850 options that it was compiled with (in a compiler-independent way) to prevent
4851 linking incompatible objects, and to allow automatic library selection. Some
4852 of these options are not visible at the IR level, namely wchar_t width and enum
4855 To pass this information to the backend, these options are encoded in module
4856 flags metadata, using the following key-value pairs:
4866 - * 0 --- sizeof(wchar_t) == 4
4867 * 1 --- sizeof(wchar_t) == 2
4870 - * 0 --- Enums are at least as large as an ``int``.
4871 * 1 --- Enums are stored in the smallest integer type which can
4872 represent all of its values.
4874 For example, the following metadata section specifies that the module was
4875 compiled with a ``wchar_t`` width of 4 bytes, and the underlying type of an
4876 enum is the smallest type which can represent all of its values::
4878 !llvm.module.flags = !{!0, !1}
4879 !0 = !{i32 1, !"short_wchar", i32 1}
4880 !1 = !{i32 1, !"short_enum", i32 0}
4882 .. _intrinsicglobalvariables:
4884 Intrinsic Global Variables
4885 ==========================
4887 LLVM has a number of "magic" global variables that contain data that
4888 affect code generation or other IR semantics. These are documented here.
4889 All globals of this sort should have a section specified as
4890 "``llvm.metadata``". This section and all globals that start with
4891 "``llvm.``" are reserved for use by LLVM.
4895 The '``llvm.used``' Global Variable
4896 -----------------------------------
4898 The ``@llvm.used`` global is an array which has
4899 :ref:`appending linkage <linkage_appending>`. This array contains a list of
4900 pointers to named global variables, functions and aliases which may optionally
4901 have a pointer cast formed of bitcast or getelementptr. For example, a legal
4904 .. code-block:: llvm
4909 @llvm.used = appending global [2 x i8*] [
4911 i8* bitcast (i32* @Y to i8*)
4912 ], section "llvm.metadata"
4914 If a symbol appears in the ``@llvm.used`` list, then the compiler, assembler,
4915 and linker are required to treat the symbol as if there is a reference to the
4916 symbol that it cannot see (which is why they have to be named). For example, if
4917 a variable has internal linkage and no references other than that from the
4918 ``@llvm.used`` list, it cannot be deleted. This is commonly used to represent
4919 references from inline asms and other things the compiler cannot "see", and
4920 corresponds to "``attribute((used))``" in GNU C.
4922 On some targets, the code generator must emit a directive to the
4923 assembler or object file to prevent the assembler and linker from
4924 molesting the symbol.
4926 .. _gv_llvmcompilerused:
4928 The '``llvm.compiler.used``' Global Variable
4929 --------------------------------------------
4931 The ``@llvm.compiler.used`` directive is the same as the ``@llvm.used``
4932 directive, except that it only prevents the compiler from touching the
4933 symbol. On targets that support it, this allows an intelligent linker to
4934 optimize references to the symbol without being impeded as it would be
4937 This is a rare construct that should only be used in rare circumstances,
4938 and should not be exposed to source languages.
4940 .. _gv_llvmglobalctors:
4942 The '``llvm.global_ctors``' Global Variable
4943 -------------------------------------------
4945 .. code-block:: llvm
4947 %0 = type { i32, void ()*, i8* }
4948 @llvm.global_ctors = appending global [1 x %0] [%0 { i32 65535, void ()* @ctor, i8* @data }]
4950 The ``@llvm.global_ctors`` array contains a list of constructor
4951 functions, priorities, and an optional associated global or function.
4952 The functions referenced by this array will be called in ascending order
4953 of priority (i.e. lowest first) when the module is loaded. The order of
4954 functions with the same priority is not defined.
4956 If the third field is present, non-null, and points to a global variable
4957 or function, the initializer function will only run if the associated
4958 data from the current module is not discarded.
4960 .. _llvmglobaldtors:
4962 The '``llvm.global_dtors``' Global Variable
4963 -------------------------------------------
4965 .. code-block:: llvm
4967 %0 = type { i32, void ()*, i8* }
4968 @llvm.global_dtors = appending global [1 x %0] [%0 { i32 65535, void ()* @dtor, i8* @data }]
4970 The ``@llvm.global_dtors`` array contains a list of destructor
4971 functions, priorities, and an optional associated global or function.
4972 The functions referenced by this array will be called in descending
4973 order of priority (i.e. highest first) when the module is unloaded. The
4974 order of functions with the same priority is not defined.
4976 If the third field is present, non-null, and points to a global variable
4977 or function, the destructor function will only run if the associated
4978 data from the current module is not discarded.
4980 Instruction Reference
4981 =====================
4983 The LLVM instruction set consists of several different classifications
4984 of instructions: :ref:`terminator instructions <terminators>`, :ref:`binary
4985 instructions <binaryops>`, :ref:`bitwise binary
4986 instructions <bitwiseops>`, :ref:`memory instructions <memoryops>`, and
4987 :ref:`other instructions <otherops>`.
4991 Terminator Instructions
4992 -----------------------
4994 As mentioned :ref:`previously <functionstructure>`, every basic block in a
4995 program ends with a "Terminator" instruction, which indicates which
4996 block should be executed after the current block is finished. These
4997 terminator instructions typically yield a '``void``' value: they produce
4998 control flow, not values (the one exception being the
4999 ':ref:`invoke <i_invoke>`' instruction).
5001 The terminator instructions are: ':ref:`ret <i_ret>`',
5002 ':ref:`br <i_br>`', ':ref:`switch <i_switch>`',
5003 ':ref:`indirectbr <i_indirectbr>`', ':ref:`invoke <i_invoke>`',
5004 ':ref:`resume <i_resume>`', ':ref:`catchpad <i_catchpad>`',
5005 ':ref:`catchendpad <i_catchendpad>`',
5006 ':ref:`catchret <i_catchret>`',
5007 ':ref:`cleanupendpad <i_cleanupendpad>`',
5008 ':ref:`cleanupret <i_cleanupret>`',
5009 ':ref:`terminatepad <i_terminatepad>`',
5010 and ':ref:`unreachable <i_unreachable>`'.
5014 '``ret``' Instruction
5015 ^^^^^^^^^^^^^^^^^^^^^
5022 ret <type> <value> ; Return a value from a non-void function
5023 ret void ; Return from void function
5028 The '``ret``' instruction is used to return control flow (and optionally
5029 a value) from a function back to the caller.
5031 There are two forms of the '``ret``' instruction: one that returns a
5032 value and then causes control flow, and one that just causes control
5038 The '``ret``' instruction optionally accepts a single argument, the
5039 return value. The type of the return value must be a ':ref:`first
5040 class <t_firstclass>`' type.
5042 A function is not :ref:`well formed <wellformed>` if it it has a non-void
5043 return type and contains a '``ret``' instruction with no return value or
5044 a return value with a type that does not match its type, or if it has a
5045 void return type and contains a '``ret``' instruction with a return
5051 When the '``ret``' instruction is executed, control flow returns back to
5052 the calling function's context. If the caller is a
5053 ":ref:`call <i_call>`" instruction, execution continues at the
5054 instruction after the call. If the caller was an
5055 ":ref:`invoke <i_invoke>`" instruction, execution continues at the
5056 beginning of the "normal" destination block. If the instruction returns
5057 a value, that value shall set the call or invoke instruction's return
5063 .. code-block:: llvm
5065 ret i32 5 ; Return an integer value of 5
5066 ret void ; Return from a void function
5067 ret { i32, i8 } { i32 4, i8 2 } ; Return a struct of values 4 and 2
5071 '``br``' Instruction
5072 ^^^^^^^^^^^^^^^^^^^^
5079 br i1 <cond>, label <iftrue>, label <iffalse>
5080 br label <dest> ; Unconditional branch
5085 The '``br``' instruction is used to cause control flow to transfer to a
5086 different basic block in the current function. There are two forms of
5087 this instruction, corresponding to a conditional branch and an
5088 unconditional branch.
5093 The conditional branch form of the '``br``' instruction takes a single
5094 '``i1``' value and two '``label``' values. The unconditional form of the
5095 '``br``' instruction takes a single '``label``' value as a target.
5100 Upon execution of a conditional '``br``' instruction, the '``i1``'
5101 argument is evaluated. If the value is ``true``, control flows to the
5102 '``iftrue``' ``label`` argument. If "cond" is ``false``, control flows
5103 to the '``iffalse``' ``label`` argument.
5108 .. code-block:: llvm
5111 %cond = icmp eq i32 %a, %b
5112 br i1 %cond, label %IfEqual, label %IfUnequal
5120 '``switch``' Instruction
5121 ^^^^^^^^^^^^^^^^^^^^^^^^
5128 switch <intty> <value>, label <defaultdest> [ <intty> <val>, label <dest> ... ]
5133 The '``switch``' instruction is used to transfer control flow to one of
5134 several different places. It is a generalization of the '``br``'
5135 instruction, allowing a branch to occur to one of many possible
5141 The '``switch``' instruction uses three parameters: an integer
5142 comparison value '``value``', a default '``label``' destination, and an
5143 array of pairs of comparison value constants and '``label``'s. The table
5144 is not allowed to contain duplicate constant entries.
5149 The ``switch`` instruction specifies a table of values and destinations.
5150 When the '``switch``' instruction is executed, this table is searched
5151 for the given value. If the value is found, control flow is transferred
5152 to the corresponding destination; otherwise, control flow is transferred
5153 to the default destination.
5158 Depending on properties of the target machine and the particular
5159 ``switch`` instruction, this instruction may be code generated in
5160 different ways. For example, it could be generated as a series of
5161 chained conditional branches or with a lookup table.
5166 .. code-block:: llvm
5168 ; Emulate a conditional br instruction
5169 %Val = zext i1 %value to i32
5170 switch i32 %Val, label %truedest [ i32 0, label %falsedest ]
5172 ; Emulate an unconditional br instruction
5173 switch i32 0, label %dest [ ]
5175 ; Implement a jump table:
5176 switch i32 %val, label %otherwise [ i32 0, label %onzero
5178 i32 2, label %ontwo ]
5182 '``indirectbr``' Instruction
5183 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5190 indirectbr <somety>* <address>, [ label <dest1>, label <dest2>, ... ]
5195 The '``indirectbr``' instruction implements an indirect branch to a
5196 label within the current function, whose address is specified by
5197 "``address``". Address must be derived from a
5198 :ref:`blockaddress <blockaddress>` constant.
5203 The '``address``' argument is the address of the label to jump to. The
5204 rest of the arguments indicate the full set of possible destinations
5205 that the address may point to. Blocks are allowed to occur multiple
5206 times in the destination list, though this isn't particularly useful.
5208 This destination list is required so that dataflow analysis has an
5209 accurate understanding of the CFG.
5214 Control transfers to the block specified in the address argument. All
5215 possible destination blocks must be listed in the label list, otherwise
5216 this instruction has undefined behavior. This implies that jumps to
5217 labels defined in other functions have undefined behavior as well.
5222 This is typically implemented with a jump through a register.
5227 .. code-block:: llvm
5229 indirectbr i8* %Addr, [ label %bb1, label %bb2, label %bb3 ]
5233 '``invoke``' Instruction
5234 ^^^^^^^^^^^^^^^^^^^^^^^^
5241 <result> = invoke [cconv] [ret attrs] <ptr to function ty> <function ptr val>(<function args>) [fn attrs]
5242 [operand bundles] to label <normal label> unwind label <exception label>
5247 The '``invoke``' instruction causes control to transfer to a specified
5248 function, with the possibility of control flow transfer to either the
5249 '``normal``' label or the '``exception``' label. If the callee function
5250 returns with the "``ret``" instruction, control flow will return to the
5251 "normal" label. If the callee (or any indirect callees) returns via the
5252 ":ref:`resume <i_resume>`" instruction or other exception handling
5253 mechanism, control is interrupted and continued at the dynamically
5254 nearest "exception" label.
5256 The '``exception``' label is a `landing
5257 pad <ExceptionHandling.html#overview>`_ for the exception. As such,
5258 '``exception``' label is required to have the
5259 ":ref:`landingpad <i_landingpad>`" instruction, which contains the
5260 information about the behavior of the program after unwinding happens,
5261 as its first non-PHI instruction. The restrictions on the
5262 "``landingpad``" instruction's tightly couples it to the "``invoke``"
5263 instruction, so that the important information contained within the
5264 "``landingpad``" instruction can't be lost through normal code motion.
5269 This instruction requires several arguments:
5271 #. The optional "cconv" marker indicates which :ref:`calling
5272 convention <callingconv>` the call should use. If none is
5273 specified, the call defaults to using C calling conventions.
5274 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
5275 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
5277 #. '``ptr to function ty``': shall be the signature of the pointer to
5278 function value being invoked. In most cases, this is a direct
5279 function invocation, but indirect ``invoke``'s are just as possible,
5280 branching off an arbitrary pointer to function value.
5281 #. '``function ptr val``': An LLVM value containing a pointer to a
5282 function to be invoked.
5283 #. '``function args``': argument list whose types match the function
5284 signature argument types and parameter attributes. All arguments must
5285 be of :ref:`first class <t_firstclass>` type. If the function signature
5286 indicates the function accepts a variable number of arguments, the
5287 extra arguments can be specified.
5288 #. '``normal label``': the label reached when the called function
5289 executes a '``ret``' instruction.
5290 #. '``exception label``': the label reached when a callee returns via
5291 the :ref:`resume <i_resume>` instruction or other exception handling
5293 #. The optional :ref:`function attributes <fnattrs>` list. Only
5294 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
5295 attributes are valid here.
5296 #. The optional :ref:`operand bundles <opbundles>` list.
5301 This instruction is designed to operate as a standard '``call``'
5302 instruction in most regards. The primary difference is that it
5303 establishes an association with a label, which is used by the runtime
5304 library to unwind the stack.
5306 This instruction is used in languages with destructors to ensure that
5307 proper cleanup is performed in the case of either a ``longjmp`` or a
5308 thrown exception. Additionally, this is important for implementation of
5309 '``catch``' clauses in high-level languages that support them.
5311 For the purposes of the SSA form, the definition of the value returned
5312 by the '``invoke``' instruction is deemed to occur on the edge from the
5313 current block to the "normal" label. If the callee unwinds then no
5314 return value is available.
5319 .. code-block:: llvm
5321 %retval = invoke i32 @Test(i32 15) to label %Continue
5322 unwind label %TestCleanup ; i32:retval set
5323 %retval = invoke coldcc i32 %Testfnptr(i32 15) to label %Continue
5324 unwind label %TestCleanup ; i32:retval set
5328 '``resume``' Instruction
5329 ^^^^^^^^^^^^^^^^^^^^^^^^
5336 resume <type> <value>
5341 The '``resume``' instruction is a terminator instruction that has no
5347 The '``resume``' instruction requires one argument, which must have the
5348 same type as the result of any '``landingpad``' instruction in the same
5354 The '``resume``' instruction resumes propagation of an existing
5355 (in-flight) exception whose unwinding was interrupted with a
5356 :ref:`landingpad <i_landingpad>` instruction.
5361 .. code-block:: llvm
5363 resume { i8*, i32 } %exn
5367 '``catchpad``' Instruction
5368 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5375 <resultval> = catchpad [<args>*]
5376 to label <normal label> unwind label <exception label>
5381 The '``catchpad``' instruction is used by `LLVM's exception handling
5382 system <ExceptionHandling.html#overview>`_ to specify that a basic block
5383 is a catch block --- one where a personality routine attempts to transfer
5384 control to catch an exception.
5385 The ``args`` correspond to whatever information the personality
5386 routine requires to know if this is an appropriate place to catch the
5387 exception. Control is transfered to the ``exception`` label if the
5388 ``catchpad`` is not an appropriate handler for the in-flight exception.
5389 The ``normal`` label should contain the code found in the ``catch``
5390 portion of a ``try``/``catch`` sequence. The ``resultval`` has the type
5391 :ref:`token <t_token>` and is used to match the ``catchpad`` to
5392 corresponding :ref:`catchrets <i_catchret>`.
5397 The instruction takes a list of arbitrary values which are interpreted
5398 by the :ref:`personality function <personalityfn>`.
5400 The ``catchpad`` must be provided a ``normal`` label to transfer control
5401 to if the ``catchpad`` matches the exception and an ``exception``
5402 label to transfer control to if it doesn't.
5407 When the call stack is being unwound due to an exception being thrown,
5408 the exception is compared against the ``args``. If it doesn't match,
5409 then control is transfered to the ``exception`` basic block.
5410 As with calling conventions, how the personality function results are
5411 represented in LLVM IR is target specific.
5413 The ``catchpad`` instruction has several restrictions:
5415 - A catch block is a basic block which is the unwind destination of
5416 an exceptional instruction.
5417 - A catch block must have a '``catchpad``' instruction as its
5418 first non-PHI instruction.
5419 - A catch block's ``exception`` edge must refer to a catch block or a
5421 - There can be only one '``catchpad``' instruction within the
5423 - A basic block that is not a catch block may not include a
5424 '``catchpad``' instruction.
5425 - A catch block which has another catch block as a predecessor may not have
5426 any other predecessors.
5427 - It is undefined behavior for control to transfer from a ``catchpad`` to a
5428 ``ret`` without first executing a ``catchret`` that consumes the
5429 ``catchpad`` or unwinding through its ``catchendpad``.
5430 - It is undefined behavior for control to transfer from a ``catchpad`` to
5431 itself without first executing a ``catchret`` that consumes the
5432 ``catchpad`` or unwinding through its ``catchendpad``.
5437 .. code-block:: llvm
5439 ;; A catch block which can catch an integer.
5440 %tok = catchpad [i8** @_ZTIi]
5441 to label %int.handler unwind label %terminate
5445 '``catchendpad``' Instruction
5446 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5453 catchendpad unwind label <nextaction>
5454 catchendpad unwind to caller
5459 The '``catchendpad``' instruction is used by `LLVM's exception handling
5460 system <ExceptionHandling.html#overview>`_ to communicate to the
5461 :ref:`personality function <personalityfn>` which invokes are associated
5462 with a chain of :ref:`catchpad <i_catchpad>` instructions; propagating an
5463 exception out of a catch handler is represented by unwinding through its
5464 ``catchendpad``. Unwinding to the outer scope when a chain of catch handlers
5465 do not handle an exception is also represented by unwinding through their
5468 The ``nextaction`` label indicates where control should transfer to if
5469 none of the ``catchpad`` instructions are suitable for catching the
5470 in-flight exception.
5472 If a ``nextaction`` label is not present, the instruction unwinds out of
5473 its parent function. The
5474 :ref:`personality function <personalityfn>` will continue processing
5475 exception handling actions in the caller.
5480 The instruction optionally takes a label, ``nextaction``, indicating
5481 where control should transfer to if none of the preceding
5482 ``catchpad`` instructions are suitable for the in-flight exception.
5487 When the call stack is being unwound due to an exception being thrown
5488 and none of the constituent ``catchpad`` instructions match, then
5489 control is transfered to ``nextaction`` if it is present. If it is not
5490 present, control is transfered to the caller.
5492 The ``catchendpad`` instruction has several restrictions:
5494 - A catch-end block is a basic block which is the unwind destination of
5495 an exceptional instruction.
5496 - A catch-end block must have a '``catchendpad``' instruction as its
5497 first non-PHI instruction.
5498 - There can be only one '``catchendpad``' instruction within the
5500 - A basic block that is not a catch-end block may not include a
5501 '``catchendpad``' instruction.
5502 - Exactly one catch block may unwind to a ``catchendpad``.
5503 - It is undefined behavior to execute a ``catchendpad`` if none of the
5504 '``catchpad``'s chained to it have been executed.
5505 - It is undefined behavior to execute a ``catchendpad`` twice without an
5506 intervening execution of one or more of the '``catchpad``'s chained to it.
5507 - It is undefined behavior to execute a ``catchendpad`` if, after the most
5508 recent execution of the normal successor edge of any ``catchpad`` chained
5509 to it, some ``catchret`` consuming that ``catchpad`` has already been
5511 - It is undefined behavior to execute a ``catchendpad`` if, after the most
5512 recent execution of the normal successor edge of any ``catchpad`` chained
5513 to it, any other ``catchpad`` or ``cleanuppad`` has been executed but has
5514 not had a corresponding
5515 ``catchret``/``cleanupret``/``catchendpad``/``cleanupendpad`` executed.
5520 .. code-block:: llvm
5522 catchendpad unwind label %terminate
5523 catchendpad unwind to caller
5527 '``catchret``' Instruction
5528 ^^^^^^^^^^^^^^^^^^^^^^^^^^
5535 catchret <value> to label <normal>
5540 The '``catchret``' instruction is a terminator instruction that has a
5547 The first argument to a '``catchret``' indicates which ``catchpad`` it
5548 exits. It must be a :ref:`catchpad <i_catchpad>`.
5549 The second argument to a '``catchret``' specifies where control will
5555 The '``catchret``' instruction ends the existing (in-flight) exception
5556 whose unwinding was interrupted with a
5557 :ref:`catchpad <i_catchpad>` instruction.
5558 The :ref:`personality function <personalityfn>` gets a chance to execute
5559 arbitrary code to, for example, run a C++ destructor.
5560 Control then transfers to ``normal``.
5561 It may be passed an optional, personality specific, value.
5563 It is undefined behavior to execute a ``catchret`` whose ``catchpad`` has
5566 It is undefined behavior to execute a ``catchret`` if, after the most recent
5567 execution of its ``catchpad``, some ``catchret`` or ``catchendpad`` linked
5568 to the same ``catchpad`` has already been executed.
5570 It is undefined behavior to execute a ``catchret`` if, after the most recent
5571 execution of its ``catchpad``, any other ``catchpad`` or ``cleanuppad`` has
5572 been executed but has not had a corresponding
5573 ``catchret``/``cleanupret``/``catchendpad``/``cleanupendpad`` executed.
5578 .. code-block:: llvm
5580 catchret %catch label %continue
5582 .. _i_cleanupendpad:
5584 '``cleanupendpad``' Instruction
5585 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5592 cleanupendpad <value> unwind label <nextaction>
5593 cleanupendpad <value> unwind to caller
5598 The '``cleanupendpad``' instruction is used by `LLVM's exception handling
5599 system <ExceptionHandling.html#overview>`_ to communicate to the
5600 :ref:`personality function <personalityfn>` which invokes are associated
5601 with a :ref:`cleanuppad <i_cleanuppad>` instructions; propagating an exception
5602 out of a cleanup is represented by unwinding through its ``cleanupendpad``.
5604 The ``nextaction`` label indicates where control should unwind to next, in the
5605 event that a cleanup is exited by means of an(other) exception being raised.
5607 If a ``nextaction`` label is not present, the instruction unwinds out of
5608 its parent function. The
5609 :ref:`personality function <personalityfn>` will continue processing
5610 exception handling actions in the caller.
5615 The '``cleanupendpad``' instruction requires one argument, which indicates
5616 which ``cleanuppad`` it exits, and must be a :ref:`cleanuppad <i_cleanuppad>`.
5617 It also has an optional successor, ``nextaction``, indicating where control
5623 When and exception propagates to a ``cleanupendpad``, control is transfered to
5624 ``nextaction`` if it is present. If it is not present, control is transfered to
5627 The ``cleanupendpad`` instruction has several restrictions:
5629 - A cleanup-end block is a basic block which is the unwind destination of
5630 an exceptional instruction.
5631 - A cleanup-end block must have a '``cleanupendpad``' instruction as its
5632 first non-PHI instruction.
5633 - There can be only one '``cleanupendpad``' instruction within the
5635 - A basic block that is not a cleanup-end block may not include a
5636 '``cleanupendpad``' instruction.
5637 - It is undefined behavior to execute a ``cleanupendpad`` whose ``cleanuppad``
5638 has not been executed.
5639 - It is undefined behavior to execute a ``cleanupendpad`` if, after the most
5640 recent execution of its ``cleanuppad``, some ``cleanupret`` or ``cleanupendpad``
5641 consuming the same ``cleanuppad`` has already been executed.
5642 - It is undefined behavior to execute a ``cleanupendpad`` if, after the most
5643 recent execution of its ``cleanuppad``, any other ``cleanuppad`` or
5644 ``catchpad`` has been executed but has not had a corresponding
5645 ``cleanupret``/``catchret``/``cleanupendpad``/``catchendpad`` executed.
5650 .. code-block:: llvm
5652 cleanupendpad %cleanup unwind label %terminate
5653 cleanupendpad %cleanup unwind to caller
5657 '``cleanupret``' Instruction
5658 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5665 cleanupret <value> unwind label <continue>
5666 cleanupret <value> unwind to caller
5671 The '``cleanupret``' instruction is a terminator instruction that has
5672 an optional successor.
5678 The '``cleanupret``' instruction requires one argument, which indicates
5679 which ``cleanuppad`` it exits, and must be a :ref:`cleanuppad <i_cleanuppad>`.
5680 It also has an optional successor, ``continue``.
5685 The '``cleanupret``' instruction indicates to the
5686 :ref:`personality function <personalityfn>` that one
5687 :ref:`cleanuppad <i_cleanuppad>` it transferred control to has ended.
5688 It transfers control to ``continue`` or unwinds out of the function.
5690 It is undefined behavior to execute a ``cleanupret`` whose ``cleanuppad`` has
5693 It is undefined behavior to execute a ``cleanupret`` if, after the most recent
5694 execution of its ``cleanuppad``, some ``cleanupret`` or ``cleanupendpad``
5695 consuming the same ``cleanuppad`` has already been executed.
5697 It is undefined behavior to execute a ``cleanupret`` if, after the most recent
5698 execution of its ``cleanuppad``, any other ``cleanuppad`` or ``catchpad`` has
5699 been executed but has not had a corresponding
5700 ``cleanupret``/``catchret``/``cleanupendpad``/``catchendpad`` executed.
5705 .. code-block:: llvm
5707 cleanupret %cleanup unwind to caller
5708 cleanupret %cleanup unwind label %continue
5712 '``terminatepad``' Instruction
5713 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5720 terminatepad [<args>*] unwind label <exception label>
5721 terminatepad [<args>*] unwind to caller
5726 The '``terminatepad``' instruction is used by `LLVM's exception handling
5727 system <ExceptionHandling.html#overview>`_ to specify that a basic block
5728 is a terminate block --- one where a personality routine may decide to
5729 terminate the program.
5730 The ``args`` correspond to whatever information the personality
5731 routine requires to know if this is an appropriate place to terminate the
5732 program. Control is transferred to the ``exception`` label if the
5733 personality routine decides not to terminate the program for the
5734 in-flight exception.
5739 The instruction takes a list of arbitrary values which are interpreted
5740 by the :ref:`personality function <personalityfn>`.
5742 The ``terminatepad`` may be given an ``exception`` label to
5743 transfer control to if the in-flight exception matches the ``args``.
5748 When the call stack is being unwound due to an exception being thrown,
5749 the exception is compared against the ``args``. If it matches,
5750 then control is transfered to the ``exception`` basic block. Otherwise,
5751 the program is terminated via personality-specific means. Typically,
5752 the first argument to ``terminatepad`` specifies what function the
5753 personality should defer to in order to terminate the program.
5755 The ``terminatepad`` instruction has several restrictions:
5757 - A terminate block is a basic block which is the unwind destination of
5758 an exceptional instruction.
5759 - A terminate block must have a '``terminatepad``' instruction as its
5760 first non-PHI instruction.
5761 - There can be only one '``terminatepad``' instruction within the
5763 - A basic block that is not a terminate block may not include a
5764 '``terminatepad``' instruction.
5769 .. code-block:: llvm
5771 ;; A terminate block which only permits integers.
5772 terminatepad [i8** @_ZTIi] unwind label %continue
5776 '``unreachable``' Instruction
5777 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5789 The '``unreachable``' instruction has no defined semantics. This
5790 instruction is used to inform the optimizer that a particular portion of
5791 the code is not reachable. This can be used to indicate that the code
5792 after a no-return function cannot be reached, and other facts.
5797 The '``unreachable``' instruction has no defined semantics.
5804 Binary operators are used to do most of the computation in a program.
5805 They require two operands of the same type, execute an operation on
5806 them, and produce a single value. The operands might represent multiple
5807 data, as is the case with the :ref:`vector <t_vector>` data type. The
5808 result value has the same type as its operands.
5810 There are several different binary operators:
5814 '``add``' Instruction
5815 ^^^^^^^^^^^^^^^^^^^^^
5822 <result> = add <ty> <op1>, <op2> ; yields ty:result
5823 <result> = add nuw <ty> <op1>, <op2> ; yields ty:result
5824 <result> = add nsw <ty> <op1>, <op2> ; yields ty:result
5825 <result> = add nuw nsw <ty> <op1>, <op2> ; yields ty:result
5830 The '``add``' instruction returns the sum of its two operands.
5835 The two arguments to the '``add``' instruction must be
5836 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
5837 arguments must have identical types.
5842 The value produced is the integer sum of the two operands.
5844 If the sum has unsigned overflow, the result returned is the
5845 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
5848 Because LLVM integers use a two's complement representation, this
5849 instruction is appropriate for both signed and unsigned integers.
5851 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
5852 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
5853 result value of the ``add`` is a :ref:`poison value <poisonvalues>` if
5854 unsigned and/or signed overflow, respectively, occurs.
5859 .. code-block:: llvm
5861 <result> = add i32 4, %var ; yields i32:result = 4 + %var
5865 '``fadd``' Instruction
5866 ^^^^^^^^^^^^^^^^^^^^^^
5873 <result> = fadd [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
5878 The '``fadd``' instruction returns the sum of its two operands.
5883 The two arguments to the '``fadd``' instruction must be :ref:`floating
5884 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
5885 Both arguments must have identical types.
5890 The value produced is the floating point sum of the two operands. This
5891 instruction can also take any number of :ref:`fast-math flags <fastmath>`,
5892 which are optimization hints to enable otherwise unsafe floating point
5898 .. code-block:: llvm
5900 <result> = fadd float 4.0, %var ; yields float:result = 4.0 + %var
5902 '``sub``' Instruction
5903 ^^^^^^^^^^^^^^^^^^^^^
5910 <result> = sub <ty> <op1>, <op2> ; yields ty:result
5911 <result> = sub nuw <ty> <op1>, <op2> ; yields ty:result
5912 <result> = sub nsw <ty> <op1>, <op2> ; yields ty:result
5913 <result> = sub nuw nsw <ty> <op1>, <op2> ; yields ty:result
5918 The '``sub``' instruction returns the difference of its two operands.
5920 Note that the '``sub``' instruction is used to represent the '``neg``'
5921 instruction present in most other intermediate representations.
5926 The two arguments to the '``sub``' instruction must be
5927 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
5928 arguments must have identical types.
5933 The value produced is the integer difference of the two operands.
5935 If the difference has unsigned overflow, the result returned is the
5936 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
5939 Because LLVM integers use a two's complement representation, this
5940 instruction is appropriate for both signed and unsigned integers.
5942 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
5943 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
5944 result value of the ``sub`` is a :ref:`poison value <poisonvalues>` if
5945 unsigned and/or signed overflow, respectively, occurs.
5950 .. code-block:: llvm
5952 <result> = sub i32 4, %var ; yields i32:result = 4 - %var
5953 <result> = sub i32 0, %val ; yields i32:result = -%var
5957 '``fsub``' Instruction
5958 ^^^^^^^^^^^^^^^^^^^^^^
5965 <result> = fsub [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
5970 The '``fsub``' instruction returns the difference of its two operands.
5972 Note that the '``fsub``' instruction is used to represent the '``fneg``'
5973 instruction present in most other intermediate representations.
5978 The two arguments to the '``fsub``' instruction must be :ref:`floating
5979 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
5980 Both arguments must have identical types.
5985 The value produced is the floating point difference of the two operands.
5986 This instruction can also take any number of :ref:`fast-math
5987 flags <fastmath>`, which are optimization hints to enable otherwise
5988 unsafe floating point optimizations:
5993 .. code-block:: llvm
5995 <result> = fsub float 4.0, %var ; yields float:result = 4.0 - %var
5996 <result> = fsub float -0.0, %val ; yields float:result = -%var
5998 '``mul``' Instruction
5999 ^^^^^^^^^^^^^^^^^^^^^
6006 <result> = mul <ty> <op1>, <op2> ; yields ty:result
6007 <result> = mul nuw <ty> <op1>, <op2> ; yields ty:result
6008 <result> = mul nsw <ty> <op1>, <op2> ; yields ty:result
6009 <result> = mul nuw nsw <ty> <op1>, <op2> ; yields ty:result
6014 The '``mul``' instruction returns the product of its two operands.
6019 The two arguments to the '``mul``' instruction must be
6020 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6021 arguments must have identical types.
6026 The value produced is the integer product of the two operands.
6028 If the result of the multiplication has unsigned overflow, the result
6029 returned is the mathematical result modulo 2\ :sup:`n`\ , where n is the
6030 bit width of the result.
6032 Because LLVM integers use a two's complement representation, and the
6033 result is the same width as the operands, this instruction returns the
6034 correct result for both signed and unsigned integers. If a full product
6035 (e.g. ``i32`` * ``i32`` -> ``i64``) is needed, the operands should be
6036 sign-extended or zero-extended as appropriate to the width of the full
6039 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
6040 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
6041 result value of the ``mul`` is a :ref:`poison value <poisonvalues>` if
6042 unsigned and/or signed overflow, respectively, occurs.
6047 .. code-block:: llvm
6049 <result> = mul i32 4, %var ; yields i32:result = 4 * %var
6053 '``fmul``' Instruction
6054 ^^^^^^^^^^^^^^^^^^^^^^
6061 <result> = fmul [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
6066 The '``fmul``' instruction returns the product of its two operands.
6071 The two arguments to the '``fmul``' instruction must be :ref:`floating
6072 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
6073 Both arguments must have identical types.
6078 The value produced is the floating point product of the two operands.
6079 This instruction can also take any number of :ref:`fast-math
6080 flags <fastmath>`, which are optimization hints to enable otherwise
6081 unsafe floating point optimizations:
6086 .. code-block:: llvm
6088 <result> = fmul float 4.0, %var ; yields float:result = 4.0 * %var
6090 '``udiv``' Instruction
6091 ^^^^^^^^^^^^^^^^^^^^^^
6098 <result> = udiv <ty> <op1>, <op2> ; yields ty:result
6099 <result> = udiv exact <ty> <op1>, <op2> ; yields ty:result
6104 The '``udiv``' instruction returns the quotient of its two operands.
6109 The two arguments to the '``udiv``' instruction must be
6110 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6111 arguments must have identical types.
6116 The value produced is the unsigned integer quotient of the two operands.
6118 Note that unsigned integer division and signed integer division are
6119 distinct operations; for signed integer division, use '``sdiv``'.
6121 Division by zero leads to undefined behavior.
6123 If the ``exact`` keyword is present, the result value of the ``udiv`` is
6124 a :ref:`poison value <poisonvalues>` if %op1 is not a multiple of %op2 (as
6125 such, "((a udiv exact b) mul b) == a").
6130 .. code-block:: llvm
6132 <result> = udiv i32 4, %var ; yields i32:result = 4 / %var
6134 '``sdiv``' Instruction
6135 ^^^^^^^^^^^^^^^^^^^^^^
6142 <result> = sdiv <ty> <op1>, <op2> ; yields ty:result
6143 <result> = sdiv exact <ty> <op1>, <op2> ; yields ty:result
6148 The '``sdiv``' instruction returns the quotient of its two operands.
6153 The two arguments to the '``sdiv``' instruction must be
6154 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6155 arguments must have identical types.
6160 The value produced is the signed integer quotient of the two operands
6161 rounded towards zero.
6163 Note that signed integer division and unsigned integer division are
6164 distinct operations; for unsigned integer division, use '``udiv``'.
6166 Division by zero leads to undefined behavior. Overflow also leads to
6167 undefined behavior; this is a rare case, but can occur, for example, by
6168 doing a 32-bit division of -2147483648 by -1.
6170 If the ``exact`` keyword is present, the result value of the ``sdiv`` is
6171 a :ref:`poison value <poisonvalues>` if the result would be rounded.
6176 .. code-block:: llvm
6178 <result> = sdiv i32 4, %var ; yields i32:result = 4 / %var
6182 '``fdiv``' Instruction
6183 ^^^^^^^^^^^^^^^^^^^^^^
6190 <result> = fdiv [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
6195 The '``fdiv``' instruction returns the quotient of its two operands.
6200 The two arguments to the '``fdiv``' instruction must be :ref:`floating
6201 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
6202 Both arguments must have identical types.
6207 The value produced is the floating point quotient of the two operands.
6208 This instruction can also take any number of :ref:`fast-math
6209 flags <fastmath>`, which are optimization hints to enable otherwise
6210 unsafe floating point optimizations:
6215 .. code-block:: llvm
6217 <result> = fdiv float 4.0, %var ; yields float:result = 4.0 / %var
6219 '``urem``' Instruction
6220 ^^^^^^^^^^^^^^^^^^^^^^
6227 <result> = urem <ty> <op1>, <op2> ; yields ty:result
6232 The '``urem``' instruction returns the remainder from the unsigned
6233 division of its two arguments.
6238 The two arguments to the '``urem``' instruction must be
6239 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6240 arguments must have identical types.
6245 This instruction returns the unsigned integer *remainder* of a division.
6246 This instruction always performs an unsigned division to get the
6249 Note that unsigned integer remainder and signed integer remainder are
6250 distinct operations; for signed integer remainder, use '``srem``'.
6252 Taking the remainder of a division by zero leads to undefined behavior.
6257 .. code-block:: llvm
6259 <result> = urem i32 4, %var ; yields i32:result = 4 % %var
6261 '``srem``' Instruction
6262 ^^^^^^^^^^^^^^^^^^^^^^
6269 <result> = srem <ty> <op1>, <op2> ; yields ty:result
6274 The '``srem``' instruction returns the remainder from the signed
6275 division of its two operands. This instruction can also take
6276 :ref:`vector <t_vector>` versions of the values in which case the elements
6282 The two arguments to the '``srem``' instruction must be
6283 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6284 arguments must have identical types.
6289 This instruction returns the *remainder* of a division (where the result
6290 is either zero or has the same sign as the dividend, ``op1``), not the
6291 *modulo* operator (where the result is either zero or has the same sign
6292 as the divisor, ``op2``) of a value. For more information about the
6293 difference, see `The Math
6294 Forum <http://mathforum.org/dr.math/problems/anne.4.28.99.html>`_. For a
6295 table of how this is implemented in various languages, please see
6297 operation <http://en.wikipedia.org/wiki/Modulo_operation>`_.
6299 Note that signed integer remainder and unsigned integer remainder are
6300 distinct operations; for unsigned integer remainder, use '``urem``'.
6302 Taking the remainder of a division by zero leads to undefined behavior.
6303 Overflow also leads to undefined behavior; this is a rare case, but can
6304 occur, for example, by taking the remainder of a 32-bit division of
6305 -2147483648 by -1. (The remainder doesn't actually overflow, but this
6306 rule lets srem be implemented using instructions that return both the
6307 result of the division and the remainder.)
6312 .. code-block:: llvm
6314 <result> = srem i32 4, %var ; yields i32:result = 4 % %var
6318 '``frem``' Instruction
6319 ^^^^^^^^^^^^^^^^^^^^^^
6326 <result> = frem [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
6331 The '``frem``' instruction returns the remainder from the division of
6337 The two arguments to the '``frem``' instruction must be :ref:`floating
6338 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
6339 Both arguments must have identical types.
6344 This instruction returns the *remainder* of a division. The remainder
6345 has the same sign as the dividend. This instruction can also take any
6346 number of :ref:`fast-math flags <fastmath>`, which are optimization hints
6347 to enable otherwise unsafe floating point optimizations:
6352 .. code-block:: llvm
6354 <result> = frem float 4.0, %var ; yields float:result = 4.0 % %var
6358 Bitwise Binary Operations
6359 -------------------------
6361 Bitwise binary operators are used to do various forms of bit-twiddling
6362 in a program. They are generally very efficient instructions and can
6363 commonly be strength reduced from other instructions. They require two
6364 operands of the same type, execute an operation on them, and produce a
6365 single value. The resulting value is the same type as its operands.
6367 '``shl``' Instruction
6368 ^^^^^^^^^^^^^^^^^^^^^
6375 <result> = shl <ty> <op1>, <op2> ; yields ty:result
6376 <result> = shl nuw <ty> <op1>, <op2> ; yields ty:result
6377 <result> = shl nsw <ty> <op1>, <op2> ; yields ty:result
6378 <result> = shl nuw nsw <ty> <op1>, <op2> ; yields ty:result
6383 The '``shl``' instruction returns the first operand shifted to the left
6384 a specified number of bits.
6389 Both arguments to the '``shl``' instruction must be the same
6390 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
6391 '``op2``' is treated as an unsigned value.
6396 The value produced is ``op1`` \* 2\ :sup:`op2` mod 2\ :sup:`n`,
6397 where ``n`` is the width of the result. If ``op2`` is (statically or
6398 dynamically) equal to or larger than the number of bits in
6399 ``op1``, the result is undefined. If the arguments are vectors, each
6400 vector element of ``op1`` is shifted by the corresponding shift amount
6403 If the ``nuw`` keyword is present, then the shift produces a :ref:`poison
6404 value <poisonvalues>` if it shifts out any non-zero bits. If the
6405 ``nsw`` keyword is present, then the shift produces a :ref:`poison
6406 value <poisonvalues>` if it shifts out any bits that disagree with the
6407 resultant sign bit. As such, NUW/NSW have the same semantics as they
6408 would if the shift were expressed as a mul instruction with the same
6409 nsw/nuw bits in (mul %op1, (shl 1, %op2)).
6414 .. code-block:: llvm
6416 <result> = shl i32 4, %var ; yields i32: 4 << %var
6417 <result> = shl i32 4, 2 ; yields i32: 16
6418 <result> = shl i32 1, 10 ; yields i32: 1024
6419 <result> = shl i32 1, 32 ; undefined
6420 <result> = shl <2 x i32> < i32 1, i32 1>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 2, i32 4>
6422 '``lshr``' Instruction
6423 ^^^^^^^^^^^^^^^^^^^^^^
6430 <result> = lshr <ty> <op1>, <op2> ; yields ty:result
6431 <result> = lshr exact <ty> <op1>, <op2> ; yields ty:result
6436 The '``lshr``' instruction (logical shift right) returns the first
6437 operand shifted to the right a specified number of bits with zero fill.
6442 Both arguments to the '``lshr``' instruction must be the same
6443 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
6444 '``op2``' is treated as an unsigned value.
6449 This instruction always performs a logical shift right operation. The
6450 most significant bits of the result will be filled with zero bits after
6451 the shift. If ``op2`` is (statically or dynamically) equal to or larger
6452 than the number of bits in ``op1``, the result is undefined. If the
6453 arguments are vectors, each vector element of ``op1`` is shifted by the
6454 corresponding shift amount in ``op2``.
6456 If the ``exact`` keyword is present, the result value of the ``lshr`` is
6457 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
6463 .. code-block:: llvm
6465 <result> = lshr i32 4, 1 ; yields i32:result = 2
6466 <result> = lshr i32 4, 2 ; yields i32:result = 1
6467 <result> = lshr i8 4, 3 ; yields i8:result = 0
6468 <result> = lshr i8 -2, 1 ; yields i8:result = 0x7F
6469 <result> = lshr i32 1, 32 ; undefined
6470 <result> = lshr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 0x7FFFFFFF, i32 1>
6472 '``ashr``' Instruction
6473 ^^^^^^^^^^^^^^^^^^^^^^
6480 <result> = ashr <ty> <op1>, <op2> ; yields ty:result
6481 <result> = ashr exact <ty> <op1>, <op2> ; yields ty:result
6486 The '``ashr``' instruction (arithmetic shift right) returns the first
6487 operand shifted to the right a specified number of bits with sign
6493 Both arguments to the '``ashr``' instruction must be the same
6494 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
6495 '``op2``' is treated as an unsigned value.
6500 This instruction always performs an arithmetic shift right operation,
6501 The most significant bits of the result will be filled with the sign bit
6502 of ``op1``. If ``op2`` is (statically or dynamically) equal to or larger
6503 than the number of bits in ``op1``, the result is undefined. If the
6504 arguments are vectors, each vector element of ``op1`` is shifted by the
6505 corresponding shift amount in ``op2``.
6507 If the ``exact`` keyword is present, the result value of the ``ashr`` is
6508 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
6514 .. code-block:: llvm
6516 <result> = ashr i32 4, 1 ; yields i32:result = 2
6517 <result> = ashr i32 4, 2 ; yields i32:result = 1
6518 <result> = ashr i8 4, 3 ; yields i8:result = 0
6519 <result> = ashr i8 -2, 1 ; yields i8:result = -1
6520 <result> = ashr i32 1, 32 ; undefined
6521 <result> = ashr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 3> ; yields: result=<2 x i32> < i32 -1, i32 0>
6523 '``and``' Instruction
6524 ^^^^^^^^^^^^^^^^^^^^^
6531 <result> = and <ty> <op1>, <op2> ; yields ty:result
6536 The '``and``' instruction returns the bitwise logical and of its two
6542 The two arguments to the '``and``' instruction must be
6543 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6544 arguments must have identical types.
6549 The truth table used for the '``and``' instruction is:
6566 .. code-block:: llvm
6568 <result> = and i32 4, %var ; yields i32:result = 4 & %var
6569 <result> = and i32 15, 40 ; yields i32:result = 8
6570 <result> = and i32 4, 8 ; yields i32:result = 0
6572 '``or``' Instruction
6573 ^^^^^^^^^^^^^^^^^^^^
6580 <result> = or <ty> <op1>, <op2> ; yields ty:result
6585 The '``or``' instruction returns the bitwise logical inclusive or of its
6591 The two arguments to the '``or``' instruction must be
6592 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6593 arguments must have identical types.
6598 The truth table used for the '``or``' instruction is:
6617 <result> = or i32 4, %var ; yields i32:result = 4 | %var
6618 <result> = or i32 15, 40 ; yields i32:result = 47
6619 <result> = or i32 4, 8 ; yields i32:result = 12
6621 '``xor``' Instruction
6622 ^^^^^^^^^^^^^^^^^^^^^
6629 <result> = xor <ty> <op1>, <op2> ; yields ty:result
6634 The '``xor``' instruction returns the bitwise logical exclusive or of
6635 its two operands. The ``xor`` is used to implement the "one's
6636 complement" operation, which is the "~" operator in C.
6641 The two arguments to the '``xor``' instruction must be
6642 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6643 arguments must have identical types.
6648 The truth table used for the '``xor``' instruction is:
6665 .. code-block:: llvm
6667 <result> = xor i32 4, %var ; yields i32:result = 4 ^ %var
6668 <result> = xor i32 15, 40 ; yields i32:result = 39
6669 <result> = xor i32 4, 8 ; yields i32:result = 12
6670 <result> = xor i32 %V, -1 ; yields i32:result = ~%V
6675 LLVM supports several instructions to represent vector operations in a
6676 target-independent manner. These instructions cover the element-access
6677 and vector-specific operations needed to process vectors effectively.
6678 While LLVM does directly support these vector operations, many
6679 sophisticated algorithms will want to use target-specific intrinsics to
6680 take full advantage of a specific target.
6682 .. _i_extractelement:
6684 '``extractelement``' Instruction
6685 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6692 <result> = extractelement <n x <ty>> <val>, <ty2> <idx> ; yields <ty>
6697 The '``extractelement``' instruction extracts a single scalar element
6698 from a vector at a specified index.
6703 The first operand of an '``extractelement``' instruction is a value of
6704 :ref:`vector <t_vector>` type. The second operand is an index indicating
6705 the position from which to extract the element. The index may be a
6706 variable of any integer type.
6711 The result is a scalar of the same type as the element type of ``val``.
6712 Its value is the value at position ``idx`` of ``val``. If ``idx``
6713 exceeds the length of ``val``, the results are undefined.
6718 .. code-block:: llvm
6720 <result> = extractelement <4 x i32> %vec, i32 0 ; yields i32
6722 .. _i_insertelement:
6724 '``insertelement``' Instruction
6725 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6732 <result> = insertelement <n x <ty>> <val>, <ty> <elt>, <ty2> <idx> ; yields <n x <ty>>
6737 The '``insertelement``' instruction inserts a scalar element into a
6738 vector at a specified index.
6743 The first operand of an '``insertelement``' instruction is a value of
6744 :ref:`vector <t_vector>` type. The second operand is a scalar value whose
6745 type must equal the element type of the first operand. The third operand
6746 is an index indicating the position at which to insert the value. The
6747 index may be a variable of any integer type.
6752 The result is a vector of the same type as ``val``. Its element values
6753 are those of ``val`` except at position ``idx``, where it gets the value
6754 ``elt``. If ``idx`` exceeds the length of ``val``, the results are
6760 .. code-block:: llvm
6762 <result> = insertelement <4 x i32> %vec, i32 1, i32 0 ; yields <4 x i32>
6764 .. _i_shufflevector:
6766 '``shufflevector``' Instruction
6767 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6774 <result> = shufflevector <n x <ty>> <v1>, <n x <ty>> <v2>, <m x i32> <mask> ; yields <m x <ty>>
6779 The '``shufflevector``' instruction constructs a permutation of elements
6780 from two input vectors, returning a vector with the same element type as
6781 the input and length that is the same as the shuffle mask.
6786 The first two operands of a '``shufflevector``' instruction are vectors
6787 with the same type. The third argument is a shuffle mask whose element
6788 type is always 'i32'. The result of the instruction is a vector whose
6789 length is the same as the shuffle mask and whose element type is the
6790 same as the element type of the first two operands.
6792 The shuffle mask operand is required to be a constant vector with either
6793 constant integer or undef values.
6798 The elements of the two input vectors are numbered from left to right
6799 across both of the vectors. The shuffle mask operand specifies, for each
6800 element of the result vector, which element of the two input vectors the
6801 result element gets. The element selector may be undef (meaning "don't
6802 care") and the second operand may be undef if performing a shuffle from
6808 .. code-block:: llvm
6810 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
6811 <4 x i32> <i32 0, i32 4, i32 1, i32 5> ; yields <4 x i32>
6812 <result> = shufflevector <4 x i32> %v1, <4 x i32> undef,
6813 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32> - Identity shuffle.
6814 <result> = shufflevector <8 x i32> %v1, <8 x i32> undef,
6815 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32>
6816 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
6817 <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7 > ; yields <8 x i32>
6819 Aggregate Operations
6820 --------------------
6822 LLVM supports several instructions for working with
6823 :ref:`aggregate <t_aggregate>` values.
6827 '``extractvalue``' Instruction
6828 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6835 <result> = extractvalue <aggregate type> <val>, <idx>{, <idx>}*
6840 The '``extractvalue``' instruction extracts the value of a member field
6841 from an :ref:`aggregate <t_aggregate>` value.
6846 The first operand of an '``extractvalue``' instruction is a value of
6847 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The other operands are
6848 constant indices to specify which value to extract in a similar manner
6849 as indices in a '``getelementptr``' instruction.
6851 The major differences to ``getelementptr`` indexing are:
6853 - Since the value being indexed is not a pointer, the first index is
6854 omitted and assumed to be zero.
6855 - At least one index must be specified.
6856 - Not only struct indices but also array indices must be in bounds.
6861 The result is the value at the position in the aggregate specified by
6867 .. code-block:: llvm
6869 <result> = extractvalue {i32, float} %agg, 0 ; yields i32
6873 '``insertvalue``' Instruction
6874 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6881 <result> = insertvalue <aggregate type> <val>, <ty> <elt>, <idx>{, <idx>}* ; yields <aggregate type>
6886 The '``insertvalue``' instruction inserts a value into a member field in
6887 an :ref:`aggregate <t_aggregate>` value.
6892 The first operand of an '``insertvalue``' instruction is a value of
6893 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The second operand is
6894 a first-class value to insert. The following operands are constant
6895 indices indicating the position at which to insert the value in a
6896 similar manner as indices in a '``extractvalue``' instruction. The value
6897 to insert must have the same type as the value identified by the
6903 The result is an aggregate of the same type as ``val``. Its value is
6904 that of ``val`` except that the value at the position specified by the
6905 indices is that of ``elt``.
6910 .. code-block:: llvm
6912 %agg1 = insertvalue {i32, float} undef, i32 1, 0 ; yields {i32 1, float undef}
6913 %agg2 = insertvalue {i32, float} %agg1, float %val, 1 ; yields {i32 1, float %val}
6914 %agg3 = insertvalue {i32, {float}} undef, float %val, 1, 0 ; yields {i32 undef, {float %val}}
6918 Memory Access and Addressing Operations
6919 ---------------------------------------
6921 A key design point of an SSA-based representation is how it represents
6922 memory. In LLVM, no memory locations are in SSA form, which makes things
6923 very simple. This section describes how to read, write, and allocate
6928 '``alloca``' Instruction
6929 ^^^^^^^^^^^^^^^^^^^^^^^^
6936 <result> = alloca [inalloca] <type> [, <ty> <NumElements>] [, align <alignment>] ; yields type*:result
6941 The '``alloca``' instruction allocates memory on the stack frame of the
6942 currently executing function, to be automatically released when this
6943 function returns to its caller. The object is always allocated in the
6944 generic address space (address space zero).
6949 The '``alloca``' instruction allocates ``sizeof(<type>)*NumElements``
6950 bytes of memory on the runtime stack, returning a pointer of the
6951 appropriate type to the program. If "NumElements" is specified, it is
6952 the number of elements allocated, otherwise "NumElements" is defaulted
6953 to be one. If a constant alignment is specified, the value result of the
6954 allocation is guaranteed to be aligned to at least that boundary. The
6955 alignment may not be greater than ``1 << 29``. If not specified, or if
6956 zero, the target can choose to align the allocation on any convenient
6957 boundary compatible with the type.
6959 '``type``' may be any sized type.
6964 Memory is allocated; a pointer is returned. The operation is undefined
6965 if there is insufficient stack space for the allocation. '``alloca``'d
6966 memory is automatically released when the function returns. The
6967 '``alloca``' instruction is commonly used to represent automatic
6968 variables that must have an address available. When the function returns
6969 (either with the ``ret`` or ``resume`` instructions), the memory is
6970 reclaimed. Allocating zero bytes is legal, but the result is undefined.
6971 The order in which memory is allocated (ie., which way the stack grows)
6977 .. code-block:: llvm
6979 %ptr = alloca i32 ; yields i32*:ptr
6980 %ptr = alloca i32, i32 4 ; yields i32*:ptr
6981 %ptr = alloca i32, i32 4, align 1024 ; yields i32*:ptr
6982 %ptr = alloca i32, align 1024 ; yields i32*:ptr
6986 '``load``' Instruction
6987 ^^^^^^^^^^^^^^^^^^^^^^
6994 <result> = load [volatile] <ty>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>][, !invariant.load !<index>][, !invariant.group !<index>][, !nonnull !<index>][, !dereferenceable !<deref_bytes_node>][, !dereferenceable_or_null !<deref_bytes_node>][, !align !<align_node>]
6995 <result> = load atomic [volatile] <ty>* <pointer> [singlethread] <ordering>, align <alignment> [, !invariant.group !<index>]
6996 !<index> = !{ i32 1 }
6997 !<deref_bytes_node> = !{i64 <dereferenceable_bytes>}
6998 !<align_node> = !{ i64 <value_alignment> }
7003 The '``load``' instruction is used to read from memory.
7008 The argument to the ``load`` instruction specifies the memory address
7009 from which to load. The type specified must be a :ref:`first
7010 class <t_firstclass>` type. If the ``load`` is marked as ``volatile``,
7011 then the optimizer is not allowed to modify the number or order of
7012 execution of this ``load`` with other :ref:`volatile
7013 operations <volatile>`.
7015 If the ``load`` is marked as ``atomic``, it takes an extra
7016 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
7017 ``release`` and ``acq_rel`` orderings are not valid on ``load``
7018 instructions. Atomic loads produce :ref:`defined <memmodel>` results
7019 when they may see multiple atomic stores. The type of the pointee must
7020 be an integer type whose bit width is a power of two greater than or
7021 equal to eight and less than or equal to a target-specific size limit.
7022 ``align`` must be explicitly specified on atomic loads, and the load has
7023 undefined behavior if the alignment is not set to a value which is at
7024 least the size in bytes of the pointee. ``!nontemporal`` does not have
7025 any defined semantics for atomic loads.
7027 The optional constant ``align`` argument specifies the alignment of the
7028 operation (that is, the alignment of the memory address). A value of 0
7029 or an omitted ``align`` argument means that the operation has the ABI
7030 alignment for the target. It is the responsibility of the code emitter
7031 to ensure that the alignment information is correct. Overestimating the
7032 alignment results in undefined behavior. Underestimating the alignment
7033 may produce less efficient code. An alignment of 1 is always safe. The
7034 maximum possible alignment is ``1 << 29``.
7036 The optional ``!nontemporal`` metadata must reference a single
7037 metadata name ``<index>`` corresponding to a metadata node with one
7038 ``i32`` entry of value 1. The existence of the ``!nontemporal``
7039 metadata on the instruction tells the optimizer and code generator
7040 that this load is not expected to be reused in the cache. The code
7041 generator may select special instructions to save cache bandwidth, such
7042 as the ``MOVNT`` instruction on x86.
7044 The optional ``!invariant.load`` metadata must reference a single
7045 metadata name ``<index>`` corresponding to a metadata node with no
7046 entries. The existence of the ``!invariant.load`` metadata on the
7047 instruction tells the optimizer and code generator that the address
7048 operand to this load points to memory which can be assumed unchanged.
7049 Being invariant does not imply that a location is dereferenceable,
7050 but it does imply that once the location is known dereferenceable
7051 its value is henceforth unchanging.
7053 The optional ``!invariant.group`` metadata must reference a single metadata name
7054 ``<index>`` corresponding to a metadata node. See ``invariant.group`` metadata.
7056 The optional ``!nonnull`` metadata must reference a single
7057 metadata name ``<index>`` corresponding to a metadata node with no
7058 entries. The existence of the ``!nonnull`` metadata on the
7059 instruction tells the optimizer that the value loaded is known to
7060 never be null. This is analogous to the ``nonnull`` attribute
7061 on parameters and return values. This metadata can only be applied
7062 to loads of a pointer type.
7064 The optional ``!dereferenceable`` metadata must reference a single metadata
7065 name ``<deref_bytes_node>`` corresponding to a metadata node with one ``i64``
7066 entry. The existence of the ``!dereferenceable`` metadata on the instruction
7067 tells the optimizer that the value loaded is known to be dereferenceable.
7068 The number of bytes known to be dereferenceable is specified by the integer
7069 value in the metadata node. This is analogous to the ''dereferenceable''
7070 attribute on parameters and return values. This metadata can only be applied
7071 to loads of a pointer type.
7073 The optional ``!dereferenceable_or_null`` metadata must reference a single
7074 metadata name ``<deref_bytes_node>`` corresponding to a metadata node with one
7075 ``i64`` entry. The existence of the ``!dereferenceable_or_null`` metadata on the
7076 instruction tells the optimizer that the value loaded is known to be either
7077 dereferenceable or null.
7078 The number of bytes known to be dereferenceable is specified by the integer
7079 value in the metadata node. This is analogous to the ''dereferenceable_or_null''
7080 attribute on parameters and return values. This metadata can only be applied
7081 to loads of a pointer type.
7083 The optional ``!align`` metadata must reference a single metadata name
7084 ``<align_node>`` corresponding to a metadata node with one ``i64`` entry.
7085 The existence of the ``!align`` metadata on the instruction tells the
7086 optimizer that the value loaded is known to be aligned to a boundary specified
7087 by the integer value in the metadata node. The alignment must be a power of 2.
7088 This is analogous to the ''align'' attribute on parameters and return values.
7089 This metadata can only be applied to loads of a pointer type.
7094 The location of memory pointed to is loaded. If the value being loaded
7095 is of scalar type then the number of bytes read does not exceed the
7096 minimum number of bytes needed to hold all bits of the type. For
7097 example, loading an ``i24`` reads at most three bytes. When loading a
7098 value of a type like ``i20`` with a size that is not an integral number
7099 of bytes, the result is undefined if the value was not originally
7100 written using a store of the same type.
7105 .. code-block:: llvm
7107 %ptr = alloca i32 ; yields i32*:ptr
7108 store i32 3, i32* %ptr ; yields void
7109 %val = load i32, i32* %ptr ; yields i32:val = i32 3
7113 '``store``' Instruction
7114 ^^^^^^^^^^^^^^^^^^^^^^^
7121 store [volatile] <ty> <value>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>][, !invariant.group !<index>] ; yields void
7122 store atomic [volatile] <ty> <value>, <ty>* <pointer> [singlethread] <ordering>, align <alignment> [, !invariant.group !<index>] ; yields void
7127 The '``store``' instruction is used to write to memory.
7132 There are two arguments to the ``store`` instruction: a value to store
7133 and an address at which to store it. The type of the ``<pointer>``
7134 operand must be a pointer to the :ref:`first class <t_firstclass>` type of
7135 the ``<value>`` operand. If the ``store`` is marked as ``volatile``,
7136 then the optimizer is not allowed to modify the number or order of
7137 execution of this ``store`` with other :ref:`volatile
7138 operations <volatile>`.
7140 If the ``store`` is marked as ``atomic``, it takes an extra
7141 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
7142 ``acquire`` and ``acq_rel`` orderings aren't valid on ``store``
7143 instructions. Atomic loads produce :ref:`defined <memmodel>` results
7144 when they may see multiple atomic stores. The type of the pointee must
7145 be an integer type whose bit width is a power of two greater than or
7146 equal to eight and less than or equal to a target-specific size limit.
7147 ``align`` must be explicitly specified on atomic stores, and the store
7148 has undefined behavior if the alignment is not set to a value which is
7149 at least the size in bytes of the pointee. ``!nontemporal`` does not
7150 have any defined semantics for atomic stores.
7152 The optional constant ``align`` argument specifies the alignment of the
7153 operation (that is, the alignment of the memory address). A value of 0
7154 or an omitted ``align`` argument means that the operation has the ABI
7155 alignment for the target. It is the responsibility of the code emitter
7156 to ensure that the alignment information is correct. Overestimating the
7157 alignment results in undefined behavior. Underestimating the
7158 alignment may produce less efficient code. An alignment of 1 is always
7159 safe. The maximum possible alignment is ``1 << 29``.
7161 The optional ``!nontemporal`` metadata must reference a single metadata
7162 name ``<index>`` corresponding to a metadata node with one ``i32`` entry of
7163 value 1. The existence of the ``!nontemporal`` metadata on the instruction
7164 tells the optimizer and code generator that this load is not expected to
7165 be reused in the cache. The code generator may select special
7166 instructions to save cache bandwidth, such as the MOVNT instruction on
7169 The optional ``!invariant.group`` metadata must reference a
7170 single metadata name ``<index>``. See ``invariant.group`` metadata.
7175 The contents of memory are updated to contain ``<value>`` at the
7176 location specified by the ``<pointer>`` operand. If ``<value>`` is
7177 of scalar type then the number of bytes written does not exceed the
7178 minimum number of bytes needed to hold all bits of the type. For
7179 example, storing an ``i24`` writes at most three bytes. When writing a
7180 value of a type like ``i20`` with a size that is not an integral number
7181 of bytes, it is unspecified what happens to the extra bits that do not
7182 belong to the type, but they will typically be overwritten.
7187 .. code-block:: llvm
7189 %ptr = alloca i32 ; yields i32*:ptr
7190 store i32 3, i32* %ptr ; yields void
7191 %val = load i32, i32* %ptr ; yields i32:val = i32 3
7195 '``fence``' Instruction
7196 ^^^^^^^^^^^^^^^^^^^^^^^
7203 fence [singlethread] <ordering> ; yields void
7208 The '``fence``' instruction is used to introduce happens-before edges
7214 '``fence``' instructions take an :ref:`ordering <ordering>` argument which
7215 defines what *synchronizes-with* edges they add. They can only be given
7216 ``acquire``, ``release``, ``acq_rel``, and ``seq_cst`` orderings.
7221 A fence A which has (at least) ``release`` ordering semantics
7222 *synchronizes with* a fence B with (at least) ``acquire`` ordering
7223 semantics if and only if there exist atomic operations X and Y, both
7224 operating on some atomic object M, such that A is sequenced before X, X
7225 modifies M (either directly or through some side effect of a sequence
7226 headed by X), Y is sequenced before B, and Y observes M. This provides a
7227 *happens-before* dependency between A and B. Rather than an explicit
7228 ``fence``, one (but not both) of the atomic operations X or Y might
7229 provide a ``release`` or ``acquire`` (resp.) ordering constraint and
7230 still *synchronize-with* the explicit ``fence`` and establish the
7231 *happens-before* edge.
7233 A ``fence`` which has ``seq_cst`` ordering, in addition to having both
7234 ``acquire`` and ``release`` semantics specified above, participates in
7235 the global program order of other ``seq_cst`` operations and/or fences.
7237 The optional ":ref:`singlethread <singlethread>`" argument specifies
7238 that the fence only synchronizes with other fences in the same thread.
7239 (This is useful for interacting with signal handlers.)
7244 .. code-block:: llvm
7246 fence acquire ; yields void
7247 fence singlethread seq_cst ; yields void
7251 '``cmpxchg``' Instruction
7252 ^^^^^^^^^^^^^^^^^^^^^^^^^
7259 cmpxchg [weak] [volatile] <ty>* <pointer>, <ty> <cmp>, <ty> <new> [singlethread] <success ordering> <failure ordering> ; yields { ty, i1 }
7264 The '``cmpxchg``' instruction is used to atomically modify memory. It
7265 loads a value in memory and compares it to a given value. If they are
7266 equal, it tries to store a new value into the memory.
7271 There are three arguments to the '``cmpxchg``' instruction: an address
7272 to operate on, a value to compare to the value currently be at that
7273 address, and a new value to place at that address if the compared values
7274 are equal. The type of '<cmp>' must be an integer type whose bit width
7275 is a power of two greater than or equal to eight and less than or equal
7276 to a target-specific size limit. '<cmp>' and '<new>' must have the same
7277 type, and the type of '<pointer>' must be a pointer to that type. If the
7278 ``cmpxchg`` is marked as ``volatile``, then the optimizer is not allowed
7279 to modify the number or order of execution of this ``cmpxchg`` with
7280 other :ref:`volatile operations <volatile>`.
7282 The success and failure :ref:`ordering <ordering>` arguments specify how this
7283 ``cmpxchg`` synchronizes with other atomic operations. Both ordering parameters
7284 must be at least ``monotonic``, the ordering constraint on failure must be no
7285 stronger than that on success, and the failure ordering cannot be either
7286 ``release`` or ``acq_rel``.
7288 The optional "``singlethread``" argument declares that the ``cmpxchg``
7289 is only atomic with respect to code (usually signal handlers) running in
7290 the same thread as the ``cmpxchg``. Otherwise the cmpxchg is atomic with
7291 respect to all other code in the system.
7293 The pointer passed into cmpxchg must have alignment greater than or
7294 equal to the size in memory of the operand.
7299 The contents of memory at the location specified by the '``<pointer>``' operand
7300 is read and compared to '``<cmp>``'; if the read value is the equal, the
7301 '``<new>``' is written. The original value at the location is returned, together
7302 with a flag indicating success (true) or failure (false).
7304 If the cmpxchg operation is marked as ``weak`` then a spurious failure is
7305 permitted: the operation may not write ``<new>`` even if the comparison
7308 If the cmpxchg operation is strong (the default), the i1 value is 1 if and only
7309 if the value loaded equals ``cmp``.
7311 A successful ``cmpxchg`` is a read-modify-write instruction for the purpose of
7312 identifying release sequences. A failed ``cmpxchg`` is equivalent to an atomic
7313 load with an ordering parameter determined the second ordering parameter.
7318 .. code-block:: llvm
7321 %orig = atomic load i32, i32* %ptr unordered ; yields i32
7325 %cmp = phi i32 [ %orig, %entry ], [%old, %loop]
7326 %squared = mul i32 %cmp, %cmp
7327 %val_success = cmpxchg i32* %ptr, i32 %cmp, i32 %squared acq_rel monotonic ; yields { i32, i1 }
7328 %value_loaded = extractvalue { i32, i1 } %val_success, 0
7329 %success = extractvalue { i32, i1 } %val_success, 1
7330 br i1 %success, label %done, label %loop
7337 '``atomicrmw``' Instruction
7338 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7345 atomicrmw [volatile] <operation> <ty>* <pointer>, <ty> <value> [singlethread] <ordering> ; yields ty
7350 The '``atomicrmw``' instruction is used to atomically modify memory.
7355 There are three arguments to the '``atomicrmw``' instruction: an
7356 operation to apply, an address whose value to modify, an argument to the
7357 operation. The operation must be one of the following keywords:
7371 The type of '<value>' must be an integer type whose bit width is a power
7372 of two greater than or equal to eight and less than or equal to a
7373 target-specific size limit. The type of the '``<pointer>``' operand must
7374 be a pointer to that type. If the ``atomicrmw`` is marked as
7375 ``volatile``, then the optimizer is not allowed to modify the number or
7376 order of execution of this ``atomicrmw`` with other :ref:`volatile
7377 operations <volatile>`.
7382 The contents of memory at the location specified by the '``<pointer>``'
7383 operand are atomically read, modified, and written back. The original
7384 value at the location is returned. The modification is specified by the
7387 - xchg: ``*ptr = val``
7388 - add: ``*ptr = *ptr + val``
7389 - sub: ``*ptr = *ptr - val``
7390 - and: ``*ptr = *ptr & val``
7391 - nand: ``*ptr = ~(*ptr & val)``
7392 - or: ``*ptr = *ptr | val``
7393 - xor: ``*ptr = *ptr ^ val``
7394 - max: ``*ptr = *ptr > val ? *ptr : val`` (using a signed comparison)
7395 - min: ``*ptr = *ptr < val ? *ptr : val`` (using a signed comparison)
7396 - umax: ``*ptr = *ptr > val ? *ptr : val`` (using an unsigned
7398 - umin: ``*ptr = *ptr < val ? *ptr : val`` (using an unsigned
7404 .. code-block:: llvm
7406 %old = atomicrmw add i32* %ptr, i32 1 acquire ; yields i32
7408 .. _i_getelementptr:
7410 '``getelementptr``' Instruction
7411 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7418 <result> = getelementptr <ty>, <ty>* <ptrval>{, <ty> <idx>}*
7419 <result> = getelementptr inbounds <ty>, <ty>* <ptrval>{, <ty> <idx>}*
7420 <result> = getelementptr <ty>, <ptr vector> <ptrval>, <vector index type> <idx>
7425 The '``getelementptr``' instruction is used to get the address of a
7426 subelement of an :ref:`aggregate <t_aggregate>` data structure. It performs
7427 address calculation only and does not access memory. The instruction can also
7428 be used to calculate a vector of such addresses.
7433 The first argument is always a type used as the basis for the calculations.
7434 The second argument is always a pointer or a vector of pointers, and is the
7435 base address to start from. The remaining arguments are indices
7436 that indicate which of the elements of the aggregate object are indexed.
7437 The interpretation of each index is dependent on the type being indexed
7438 into. The first index always indexes the pointer value given as the
7439 first argument, the second index indexes a value of the type pointed to
7440 (not necessarily the value directly pointed to, since the first index
7441 can be non-zero), etc. The first type indexed into must be a pointer
7442 value, subsequent types can be arrays, vectors, and structs. Note that
7443 subsequent types being indexed into can never be pointers, since that
7444 would require loading the pointer before continuing calculation.
7446 The type of each index argument depends on the type it is indexing into.
7447 When indexing into a (optionally packed) structure, only ``i32`` integer
7448 **constants** are allowed (when using a vector of indices they must all
7449 be the **same** ``i32`` integer constant). When indexing into an array,
7450 pointer or vector, integers of any width are allowed, and they are not
7451 required to be constant. These integers are treated as signed values
7454 For example, let's consider a C code fragment and how it gets compiled
7470 int *foo(struct ST *s) {
7471 return &s[1].Z.B[5][13];
7474 The LLVM code generated by Clang is:
7476 .. code-block:: llvm
7478 %struct.RT = type { i8, [10 x [20 x i32]], i8 }
7479 %struct.ST = type { i32, double, %struct.RT }
7481 define i32* @foo(%struct.ST* %s) nounwind uwtable readnone optsize ssp {
7483 %arrayidx = getelementptr inbounds %struct.ST, %struct.ST* %s, i64 1, i32 2, i32 1, i64 5, i64 13
7490 In the example above, the first index is indexing into the
7491 '``%struct.ST*``' type, which is a pointer, yielding a '``%struct.ST``'
7492 = '``{ i32, double, %struct.RT }``' type, a structure. The second index
7493 indexes into the third element of the structure, yielding a
7494 '``%struct.RT``' = '``{ i8 , [10 x [20 x i32]], i8 }``' type, another
7495 structure. The third index indexes into the second element of the
7496 structure, yielding a '``[10 x [20 x i32]]``' type, an array. The two
7497 dimensions of the array are subscripted into, yielding an '``i32``'
7498 type. The '``getelementptr``' instruction returns a pointer to this
7499 element, thus computing a value of '``i32*``' type.
7501 Note that it is perfectly legal to index partially through a structure,
7502 returning a pointer to an inner element. Because of this, the LLVM code
7503 for the given testcase is equivalent to:
7505 .. code-block:: llvm
7507 define i32* @foo(%struct.ST* %s) {
7508 %t1 = getelementptr %struct.ST, %struct.ST* %s, i32 1 ; yields %struct.ST*:%t1
7509 %t2 = getelementptr %struct.ST, %struct.ST* %t1, i32 0, i32 2 ; yields %struct.RT*:%t2
7510 %t3 = getelementptr %struct.RT, %struct.RT* %t2, i32 0, i32 1 ; yields [10 x [20 x i32]]*:%t3
7511 %t4 = getelementptr [10 x [20 x i32]], [10 x [20 x i32]]* %t3, i32 0, i32 5 ; yields [20 x i32]*:%t4
7512 %t5 = getelementptr [20 x i32], [20 x i32]* %t4, i32 0, i32 13 ; yields i32*:%t5
7516 If the ``inbounds`` keyword is present, the result value of the
7517 ``getelementptr`` is a :ref:`poison value <poisonvalues>` if the base
7518 pointer is not an *in bounds* address of an allocated object, or if any
7519 of the addresses that would be formed by successive addition of the
7520 offsets implied by the indices to the base address with infinitely
7521 precise signed arithmetic are not an *in bounds* address of that
7522 allocated object. The *in bounds* addresses for an allocated object are
7523 all the addresses that point into the object, plus the address one byte
7524 past the end. In cases where the base is a vector of pointers the
7525 ``inbounds`` keyword applies to each of the computations element-wise.
7527 If the ``inbounds`` keyword is not present, the offsets are added to the
7528 base address with silently-wrapping two's complement arithmetic. If the
7529 offsets have a different width from the pointer, they are sign-extended
7530 or truncated to the width of the pointer. The result value of the
7531 ``getelementptr`` may be outside the object pointed to by the base
7532 pointer. The result value may not necessarily be used to access memory
7533 though, even if it happens to point into allocated storage. See the
7534 :ref:`Pointer Aliasing Rules <pointeraliasing>` section for more
7537 The getelementptr instruction is often confusing. For some more insight
7538 into how it works, see :doc:`the getelementptr FAQ <GetElementPtr>`.
7543 .. code-block:: llvm
7545 ; yields [12 x i8]*:aptr
7546 %aptr = getelementptr {i32, [12 x i8]}, {i32, [12 x i8]}* %saptr, i64 0, i32 1
7548 %vptr = getelementptr {i32, <2 x i8>}, {i32, <2 x i8>}* %svptr, i64 0, i32 1, i32 1
7550 %eptr = getelementptr [12 x i8], [12 x i8]* %aptr, i64 0, i32 1
7552 %iptr = getelementptr [10 x i32], [10 x i32]* @arr, i16 0, i16 0
7557 The ``getelementptr`` returns a vector of pointers, instead of a single address,
7558 when one or more of its arguments is a vector. In such cases, all vector
7559 arguments should have the same number of elements, and every scalar argument
7560 will be effectively broadcast into a vector during address calculation.
7562 .. code-block:: llvm
7564 ; All arguments are vectors:
7565 ; A[i] = ptrs[i] + offsets[i]*sizeof(i8)
7566 %A = getelementptr i8, <4 x i8*> %ptrs, <4 x i64> %offsets
7568 ; Add the same scalar offset to each pointer of a vector:
7569 ; A[i] = ptrs[i] + offset*sizeof(i8)
7570 %A = getelementptr i8, <4 x i8*> %ptrs, i64 %offset
7572 ; Add distinct offsets to the same pointer:
7573 ; A[i] = ptr + offsets[i]*sizeof(i8)
7574 %A = getelementptr i8, i8* %ptr, <4 x i64> %offsets
7576 ; In all cases described above the type of the result is <4 x i8*>
7578 The two following instructions are equivalent:
7580 .. code-block:: llvm
7582 getelementptr %struct.ST, <4 x %struct.ST*> %s, <4 x i64> %ind1,
7583 <4 x i32> <i32 2, i32 2, i32 2, i32 2>,
7584 <4 x i32> <i32 1, i32 1, i32 1, i32 1>,
7586 <4 x i64> <i64 13, i64 13, i64 13, i64 13>
7588 getelementptr %struct.ST, <4 x %struct.ST*> %s, <4 x i64> %ind1,
7589 i32 2, i32 1, <4 x i32> %ind4, i64 13
7591 Let's look at the C code, where the vector version of ``getelementptr``
7596 // Let's assume that we vectorize the following loop:
7597 double *A, B; int *C;
7598 for (int i = 0; i < size; ++i) {
7602 .. code-block:: llvm
7604 ; get pointers for 8 elements from array B
7605 %ptrs = getelementptr double, double* %B, <8 x i32> %C
7606 ; load 8 elements from array B into A
7607 %A = call <8 x double> @llvm.masked.gather.v8f64(<8 x double*> %ptrs,
7608 i32 8, <8 x i1> %mask, <8 x double> %passthru)
7610 Conversion Operations
7611 ---------------------
7613 The instructions in this category are the conversion instructions
7614 (casting) which all take a single operand and a type. They perform
7615 various bit conversions on the operand.
7617 '``trunc .. to``' Instruction
7618 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7625 <result> = trunc <ty> <value> to <ty2> ; yields ty2
7630 The '``trunc``' instruction truncates its operand to the type ``ty2``.
7635 The '``trunc``' instruction takes a value to trunc, and a type to trunc
7636 it to. Both types must be of :ref:`integer <t_integer>` types, or vectors
7637 of the same number of integers. The bit size of the ``value`` must be
7638 larger than the bit size of the destination type, ``ty2``. Equal sized
7639 types are not allowed.
7644 The '``trunc``' instruction truncates the high order bits in ``value``
7645 and converts the remaining bits to ``ty2``. Since the source size must
7646 be larger than the destination size, ``trunc`` cannot be a *no-op cast*.
7647 It will always truncate bits.
7652 .. code-block:: llvm
7654 %X = trunc i32 257 to i8 ; yields i8:1
7655 %Y = trunc i32 123 to i1 ; yields i1:true
7656 %Z = trunc i32 122 to i1 ; yields i1:false
7657 %W = trunc <2 x i16> <i16 8, i16 7> to <2 x i8> ; yields <i8 8, i8 7>
7659 '``zext .. to``' Instruction
7660 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7667 <result> = zext <ty> <value> to <ty2> ; yields ty2
7672 The '``zext``' instruction zero extends its operand to type ``ty2``.
7677 The '``zext``' instruction takes a value to cast, and a type to cast it
7678 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
7679 the same number of integers. The bit size of the ``value`` must be
7680 smaller than the bit size of the destination type, ``ty2``.
7685 The ``zext`` fills the high order bits of the ``value`` with zero bits
7686 until it reaches the size of the destination type, ``ty2``.
7688 When zero extending from i1, the result will always be either 0 or 1.
7693 .. code-block:: llvm
7695 %X = zext i32 257 to i64 ; yields i64:257
7696 %Y = zext i1 true to i32 ; yields i32:1
7697 %Z = zext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
7699 '``sext .. to``' Instruction
7700 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7707 <result> = sext <ty> <value> to <ty2> ; yields ty2
7712 The '``sext``' sign extends ``value`` to the type ``ty2``.
7717 The '``sext``' instruction takes a value to cast, and a type to cast it
7718 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
7719 the same number of integers. The bit size of the ``value`` must be
7720 smaller than the bit size of the destination type, ``ty2``.
7725 The '``sext``' instruction performs a sign extension by copying the sign
7726 bit (highest order bit) of the ``value`` until it reaches the bit size
7727 of the type ``ty2``.
7729 When sign extending from i1, the extension always results in -1 or 0.
7734 .. code-block:: llvm
7736 %X = sext i8 -1 to i16 ; yields i16 :65535
7737 %Y = sext i1 true to i32 ; yields i32:-1
7738 %Z = sext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
7740 '``fptrunc .. to``' Instruction
7741 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7748 <result> = fptrunc <ty> <value> to <ty2> ; yields ty2
7753 The '``fptrunc``' instruction truncates ``value`` to type ``ty2``.
7758 The '``fptrunc``' instruction takes a :ref:`floating point <t_floating>`
7759 value to cast and a :ref:`floating point <t_floating>` type to cast it to.
7760 The size of ``value`` must be larger than the size of ``ty2``. This
7761 implies that ``fptrunc`` cannot be used to make a *no-op cast*.
7766 The '``fptrunc``' instruction casts a ``value`` from a larger
7767 :ref:`floating point <t_floating>` type to a smaller :ref:`floating
7768 point <t_floating>` type. If the value cannot fit (i.e. overflows) within the
7769 destination type, ``ty2``, then the results are undefined. If the cast produces
7770 an inexact result, how rounding is performed (e.g. truncation, also known as
7771 round to zero) is undefined.
7776 .. code-block:: llvm
7778 %X = fptrunc double 123.0 to float ; yields float:123.0
7779 %Y = fptrunc double 1.0E+300 to float ; yields undefined
7781 '``fpext .. to``' Instruction
7782 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7789 <result> = fpext <ty> <value> to <ty2> ; yields ty2
7794 The '``fpext``' extends a floating point ``value`` to a larger floating
7800 The '``fpext``' instruction takes a :ref:`floating point <t_floating>`
7801 ``value`` to cast, and a :ref:`floating point <t_floating>` type to cast it
7802 to. The source type must be smaller than the destination type.
7807 The '``fpext``' instruction extends the ``value`` from a smaller
7808 :ref:`floating point <t_floating>` type to a larger :ref:`floating
7809 point <t_floating>` type. The ``fpext`` cannot be used to make a
7810 *no-op cast* because it always changes bits. Use ``bitcast`` to make a
7811 *no-op cast* for a floating point cast.
7816 .. code-block:: llvm
7818 %X = fpext float 3.125 to double ; yields double:3.125000e+00
7819 %Y = fpext double %X to fp128 ; yields fp128:0xL00000000000000004000900000000000
7821 '``fptoui .. to``' Instruction
7822 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7829 <result> = fptoui <ty> <value> to <ty2> ; yields ty2
7834 The '``fptoui``' converts a floating point ``value`` to its unsigned
7835 integer equivalent of type ``ty2``.
7840 The '``fptoui``' instruction takes a value to cast, which must be a
7841 scalar or vector :ref:`floating point <t_floating>` value, and a type to
7842 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
7843 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
7844 type with the same number of elements as ``ty``
7849 The '``fptoui``' instruction converts its :ref:`floating
7850 point <t_floating>` operand into the nearest (rounding towards zero)
7851 unsigned integer value. If the value cannot fit in ``ty2``, the results
7857 .. code-block:: llvm
7859 %X = fptoui double 123.0 to i32 ; yields i32:123
7860 %Y = fptoui float 1.0E+300 to i1 ; yields undefined:1
7861 %Z = fptoui float 1.04E+17 to i8 ; yields undefined:1
7863 '``fptosi .. to``' Instruction
7864 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7871 <result> = fptosi <ty> <value> to <ty2> ; yields ty2
7876 The '``fptosi``' instruction converts :ref:`floating point <t_floating>`
7877 ``value`` to type ``ty2``.
7882 The '``fptosi``' instruction takes a value to cast, which must be a
7883 scalar or vector :ref:`floating point <t_floating>` value, and a type to
7884 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
7885 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
7886 type with the same number of elements as ``ty``
7891 The '``fptosi``' instruction converts its :ref:`floating
7892 point <t_floating>` operand into the nearest (rounding towards zero)
7893 signed integer value. If the value cannot fit in ``ty2``, the results
7899 .. code-block:: llvm
7901 %X = fptosi double -123.0 to i32 ; yields i32:-123
7902 %Y = fptosi float 1.0E-247 to i1 ; yields undefined:1
7903 %Z = fptosi float 1.04E+17 to i8 ; yields undefined:1
7905 '``uitofp .. to``' Instruction
7906 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7913 <result> = uitofp <ty> <value> to <ty2> ; yields ty2
7918 The '``uitofp``' instruction regards ``value`` as an unsigned integer
7919 and converts that value to the ``ty2`` type.
7924 The '``uitofp``' instruction takes a value to cast, which must be a
7925 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
7926 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
7927 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
7928 type with the same number of elements as ``ty``
7933 The '``uitofp``' instruction interprets its operand as an unsigned
7934 integer quantity and converts it to the corresponding floating point
7935 value. If the value cannot fit in the floating point value, the results
7941 .. code-block:: llvm
7943 %X = uitofp i32 257 to float ; yields float:257.0
7944 %Y = uitofp i8 -1 to double ; yields double:255.0
7946 '``sitofp .. to``' Instruction
7947 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7954 <result> = sitofp <ty> <value> to <ty2> ; yields ty2
7959 The '``sitofp``' instruction regards ``value`` as a signed integer and
7960 converts that value to the ``ty2`` type.
7965 The '``sitofp``' instruction takes a value to cast, which must be a
7966 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
7967 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
7968 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
7969 type with the same number of elements as ``ty``
7974 The '``sitofp``' instruction interprets its operand as a signed integer
7975 quantity and converts it to the corresponding floating point value. If
7976 the value cannot fit in the floating point value, the results are
7982 .. code-block:: llvm
7984 %X = sitofp i32 257 to float ; yields float:257.0
7985 %Y = sitofp i8 -1 to double ; yields double:-1.0
7989 '``ptrtoint .. to``' Instruction
7990 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7997 <result> = ptrtoint <ty> <value> to <ty2> ; yields ty2
8002 The '``ptrtoint``' instruction converts the pointer or a vector of
8003 pointers ``value`` to the integer (or vector of integers) type ``ty2``.
8008 The '``ptrtoint``' instruction takes a ``value`` to cast, which must be
8009 a value of type :ref:`pointer <t_pointer>` or a vector of pointers, and a
8010 type to cast it to ``ty2``, which must be an :ref:`integer <t_integer>` or
8011 a vector of integers type.
8016 The '``ptrtoint``' instruction converts ``value`` to integer type
8017 ``ty2`` by interpreting the pointer value as an integer and either
8018 truncating or zero extending that value to the size of the integer type.
8019 If ``value`` is smaller than ``ty2`` then a zero extension is done. If
8020 ``value`` is larger than ``ty2`` then a truncation is done. If they are
8021 the same size, then nothing is done (*no-op cast*) other than a type
8027 .. code-block:: llvm
8029 %X = ptrtoint i32* %P to i8 ; yields truncation on 32-bit architecture
8030 %Y = ptrtoint i32* %P to i64 ; yields zero extension on 32-bit architecture
8031 %Z = ptrtoint <4 x i32*> %P to <4 x i64>; yields vector zero extension for a vector of addresses on 32-bit architecture
8035 '``inttoptr .. to``' Instruction
8036 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8043 <result> = inttoptr <ty> <value> to <ty2> ; yields ty2
8048 The '``inttoptr``' instruction converts an integer ``value`` to a
8049 pointer type, ``ty2``.
8054 The '``inttoptr``' instruction takes an :ref:`integer <t_integer>` value to
8055 cast, and a type to cast it to, which must be a :ref:`pointer <t_pointer>`
8061 The '``inttoptr``' instruction converts ``value`` to type ``ty2`` by
8062 applying either a zero extension or a truncation depending on the size
8063 of the integer ``value``. If ``value`` is larger than the size of a
8064 pointer then a truncation is done. If ``value`` is smaller than the size
8065 of a pointer then a zero extension is done. If they are the same size,
8066 nothing is done (*no-op cast*).
8071 .. code-block:: llvm
8073 %X = inttoptr i32 255 to i32* ; yields zero extension on 64-bit architecture
8074 %Y = inttoptr i32 255 to i32* ; yields no-op on 32-bit architecture
8075 %Z = inttoptr i64 0 to i32* ; yields truncation on 32-bit architecture
8076 %Z = inttoptr <4 x i32> %G to <4 x i8*>; yields truncation of vector G to four pointers
8080 '``bitcast .. to``' Instruction
8081 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8088 <result> = bitcast <ty> <value> to <ty2> ; yields ty2
8093 The '``bitcast``' instruction converts ``value`` to type ``ty2`` without
8099 The '``bitcast``' instruction takes a value to cast, which must be a
8100 non-aggregate first class value, and a type to cast it to, which must
8101 also be a non-aggregate :ref:`first class <t_firstclass>` type. The
8102 bit sizes of ``value`` and the destination type, ``ty2``, must be
8103 identical. If the source type is a pointer, the destination type must
8104 also be a pointer of the same size. This instruction supports bitwise
8105 conversion of vectors to integers and to vectors of other types (as
8106 long as they have the same size).
8111 The '``bitcast``' instruction converts ``value`` to type ``ty2``. It
8112 is always a *no-op cast* because no bits change with this
8113 conversion. The conversion is done as if the ``value`` had been stored
8114 to memory and read back as type ``ty2``. Pointer (or vector of
8115 pointers) types may only be converted to other pointer (or vector of
8116 pointers) types with the same address space through this instruction.
8117 To convert pointers to other types, use the :ref:`inttoptr <i_inttoptr>`
8118 or :ref:`ptrtoint <i_ptrtoint>` instructions first.
8123 .. code-block:: llvm
8125 %X = bitcast i8 255 to i8 ; yields i8 :-1
8126 %Y = bitcast i32* %x to sint* ; yields sint*:%x
8127 %Z = bitcast <2 x int> %V to i64; ; yields i64: %V
8128 %Z = bitcast <2 x i32*> %V to <2 x i64*> ; yields <2 x i64*>
8130 .. _i_addrspacecast:
8132 '``addrspacecast .. to``' Instruction
8133 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8140 <result> = addrspacecast <pty> <ptrval> to <pty2> ; yields pty2
8145 The '``addrspacecast``' instruction converts ``ptrval`` from ``pty`` in
8146 address space ``n`` to type ``pty2`` in address space ``m``.
8151 The '``addrspacecast``' instruction takes a pointer or vector of pointer value
8152 to cast and a pointer type to cast it to, which must have a different
8158 The '``addrspacecast``' instruction converts the pointer value
8159 ``ptrval`` to type ``pty2``. It can be a *no-op cast* or a complex
8160 value modification, depending on the target and the address space
8161 pair. Pointer conversions within the same address space must be
8162 performed with the ``bitcast`` instruction. Note that if the address space
8163 conversion is legal then both result and operand refer to the same memory
8169 .. code-block:: llvm
8171 %X = addrspacecast i32* %x to i32 addrspace(1)* ; yields i32 addrspace(1)*:%x
8172 %Y = addrspacecast i32 addrspace(1)* %y to i64 addrspace(2)* ; yields i64 addrspace(2)*:%y
8173 %Z = addrspacecast <4 x i32*> %z to <4 x float addrspace(3)*> ; yields <4 x float addrspace(3)*>:%z
8180 The instructions in this category are the "miscellaneous" instructions,
8181 which defy better classification.
8185 '``icmp``' Instruction
8186 ^^^^^^^^^^^^^^^^^^^^^^
8193 <result> = icmp <cond> <ty> <op1>, <op2> ; yields i1 or <N x i1>:result
8198 The '``icmp``' instruction returns a boolean value or a vector of
8199 boolean values based on comparison of its two integer, integer vector,
8200 pointer, or pointer vector operands.
8205 The '``icmp``' instruction takes three operands. The first operand is
8206 the condition code indicating the kind of comparison to perform. It is
8207 not a value, just a keyword. The possible condition code are:
8210 #. ``ne``: not equal
8211 #. ``ugt``: unsigned greater than
8212 #. ``uge``: unsigned greater or equal
8213 #. ``ult``: unsigned less than
8214 #. ``ule``: unsigned less or equal
8215 #. ``sgt``: signed greater than
8216 #. ``sge``: signed greater or equal
8217 #. ``slt``: signed less than
8218 #. ``sle``: signed less or equal
8220 The remaining two arguments must be :ref:`integer <t_integer>` or
8221 :ref:`pointer <t_pointer>` or integer :ref:`vector <t_vector>` typed. They
8222 must also be identical types.
8227 The '``icmp``' compares ``op1`` and ``op2`` according to the condition
8228 code given as ``cond``. The comparison performed always yields either an
8229 :ref:`i1 <t_integer>` or vector of ``i1`` result, as follows:
8231 #. ``eq``: yields ``true`` if the operands are equal, ``false``
8232 otherwise. No sign interpretation is necessary or performed.
8233 #. ``ne``: yields ``true`` if the operands are unequal, ``false``
8234 otherwise. No sign interpretation is necessary or performed.
8235 #. ``ugt``: interprets the operands as unsigned values and yields
8236 ``true`` if ``op1`` is greater than ``op2``.
8237 #. ``uge``: interprets the operands as unsigned values and yields
8238 ``true`` if ``op1`` is greater than or equal to ``op2``.
8239 #. ``ult``: interprets the operands as unsigned values and yields
8240 ``true`` if ``op1`` is less than ``op2``.
8241 #. ``ule``: interprets the operands as unsigned values and yields
8242 ``true`` if ``op1`` is less than or equal to ``op2``.
8243 #. ``sgt``: interprets the operands as signed values and yields ``true``
8244 if ``op1`` is greater than ``op2``.
8245 #. ``sge``: interprets the operands as signed values and yields ``true``
8246 if ``op1`` is greater than or equal to ``op2``.
8247 #. ``slt``: interprets the operands as signed values and yields ``true``
8248 if ``op1`` is less than ``op2``.
8249 #. ``sle``: interprets the operands as signed values and yields ``true``
8250 if ``op1`` is less than or equal to ``op2``.
8252 If the operands are :ref:`pointer <t_pointer>` typed, the pointer values
8253 are compared as if they were integers.
8255 If the operands are integer vectors, then they are compared element by
8256 element. The result is an ``i1`` vector with the same number of elements
8257 as the values being compared. Otherwise, the result is an ``i1``.
8262 .. code-block:: llvm
8264 <result> = icmp eq i32 4, 5 ; yields: result=false
8265 <result> = icmp ne float* %X, %X ; yields: result=false
8266 <result> = icmp ult i16 4, 5 ; yields: result=true
8267 <result> = icmp sgt i16 4, 5 ; yields: result=false
8268 <result> = icmp ule i16 -4, 5 ; yields: result=false
8269 <result> = icmp sge i16 4, 5 ; yields: result=false
8271 Note that the code generator does not yet support vector types with the
8272 ``icmp`` instruction.
8276 '``fcmp``' Instruction
8277 ^^^^^^^^^^^^^^^^^^^^^^
8284 <result> = fcmp [fast-math flags]* <cond> <ty> <op1>, <op2> ; yields i1 or <N x i1>:result
8289 The '``fcmp``' instruction returns a boolean value or vector of boolean
8290 values based on comparison of its operands.
8292 If the operands are floating point scalars, then the result type is a
8293 boolean (:ref:`i1 <t_integer>`).
8295 If the operands are floating point vectors, then the result type is a
8296 vector of boolean with the same number of elements as the operands being
8302 The '``fcmp``' instruction takes three operands. The first operand is
8303 the condition code indicating the kind of comparison to perform. It is
8304 not a value, just a keyword. The possible condition code are:
8306 #. ``false``: no comparison, always returns false
8307 #. ``oeq``: ordered and equal
8308 #. ``ogt``: ordered and greater than
8309 #. ``oge``: ordered and greater than or equal
8310 #. ``olt``: ordered and less than
8311 #. ``ole``: ordered and less than or equal
8312 #. ``one``: ordered and not equal
8313 #. ``ord``: ordered (no nans)
8314 #. ``ueq``: unordered or equal
8315 #. ``ugt``: unordered or greater than
8316 #. ``uge``: unordered or greater than or equal
8317 #. ``ult``: unordered or less than
8318 #. ``ule``: unordered or less than or equal
8319 #. ``une``: unordered or not equal
8320 #. ``uno``: unordered (either nans)
8321 #. ``true``: no comparison, always returns true
8323 *Ordered* means that neither operand is a QNAN while *unordered* means
8324 that either operand may be a QNAN.
8326 Each of ``val1`` and ``val2`` arguments must be either a :ref:`floating
8327 point <t_floating>` type or a :ref:`vector <t_vector>` of floating point
8328 type. They must have identical types.
8333 The '``fcmp``' instruction compares ``op1`` and ``op2`` according to the
8334 condition code given as ``cond``. If the operands are vectors, then the
8335 vectors are compared element by element. Each comparison performed
8336 always yields an :ref:`i1 <t_integer>` result, as follows:
8338 #. ``false``: always yields ``false``, regardless of operands.
8339 #. ``oeq``: yields ``true`` if both operands are not a QNAN and ``op1``
8340 is equal to ``op2``.
8341 #. ``ogt``: yields ``true`` if both operands are not a QNAN and ``op1``
8342 is greater than ``op2``.
8343 #. ``oge``: yields ``true`` if both operands are not a QNAN and ``op1``
8344 is greater than or equal to ``op2``.
8345 #. ``olt``: yields ``true`` if both operands are not a QNAN and ``op1``
8346 is less than ``op2``.
8347 #. ``ole``: yields ``true`` if both operands are not a QNAN and ``op1``
8348 is less than or equal to ``op2``.
8349 #. ``one``: yields ``true`` if both operands are not a QNAN and ``op1``
8350 is not equal to ``op2``.
8351 #. ``ord``: yields ``true`` if both operands are not a QNAN.
8352 #. ``ueq``: yields ``true`` if either operand is a QNAN or ``op1`` is
8354 #. ``ugt``: yields ``true`` if either operand is a QNAN or ``op1`` is
8355 greater than ``op2``.
8356 #. ``uge``: yields ``true`` if either operand is a QNAN or ``op1`` is
8357 greater than or equal to ``op2``.
8358 #. ``ult``: yields ``true`` if either operand is a QNAN or ``op1`` is
8360 #. ``ule``: yields ``true`` if either operand is a QNAN or ``op1`` is
8361 less than or equal to ``op2``.
8362 #. ``une``: yields ``true`` if either operand is a QNAN or ``op1`` is
8363 not equal to ``op2``.
8364 #. ``uno``: yields ``true`` if either operand is a QNAN.
8365 #. ``true``: always yields ``true``, regardless of operands.
8367 The ``fcmp`` instruction can also optionally take any number of
8368 :ref:`fast-math flags <fastmath>`, which are optimization hints to enable
8369 otherwise unsafe floating point optimizations.
8371 Any set of fast-math flags are legal on an ``fcmp`` instruction, but the
8372 only flags that have any effect on its semantics are those that allow
8373 assumptions to be made about the values of input arguments; namely
8374 ``nnan``, ``ninf``, and ``nsz``. See :ref:`fastmath` for more information.
8379 .. code-block:: llvm
8381 <result> = fcmp oeq float 4.0, 5.0 ; yields: result=false
8382 <result> = fcmp one float 4.0, 5.0 ; yields: result=true
8383 <result> = fcmp olt float 4.0, 5.0 ; yields: result=true
8384 <result> = fcmp ueq double 1.0, 2.0 ; yields: result=false
8386 Note that the code generator does not yet support vector types with the
8387 ``fcmp`` instruction.
8391 '``phi``' Instruction
8392 ^^^^^^^^^^^^^^^^^^^^^
8399 <result> = phi <ty> [ <val0>, <label0>], ...
8404 The '``phi``' instruction is used to implement the φ node in the SSA
8405 graph representing the function.
8410 The type of the incoming values is specified with the first type field.
8411 After this, the '``phi``' instruction takes a list of pairs as
8412 arguments, with one pair for each predecessor basic block of the current
8413 block. Only values of :ref:`first class <t_firstclass>` type may be used as
8414 the value arguments to the PHI node. Only labels may be used as the
8417 There must be no non-phi instructions between the start of a basic block
8418 and the PHI instructions: i.e. PHI instructions must be first in a basic
8421 For the purposes of the SSA form, the use of each incoming value is
8422 deemed to occur on the edge from the corresponding predecessor block to
8423 the current block (but after any definition of an '``invoke``'
8424 instruction's return value on the same edge).
8429 At runtime, the '``phi``' instruction logically takes on the value
8430 specified by the pair corresponding to the predecessor basic block that
8431 executed just prior to the current block.
8436 .. code-block:: llvm
8438 Loop: ; Infinite loop that counts from 0 on up...
8439 %indvar = phi i32 [ 0, %LoopHeader ], [ %nextindvar, %Loop ]
8440 %nextindvar = add i32 %indvar, 1
8445 '``select``' Instruction
8446 ^^^^^^^^^^^^^^^^^^^^^^^^
8453 <result> = select selty <cond>, <ty> <val1>, <ty> <val2> ; yields ty
8455 selty is either i1 or {<N x i1>}
8460 The '``select``' instruction is used to choose one value based on a
8461 condition, without IR-level branching.
8466 The '``select``' instruction requires an 'i1' value or a vector of 'i1'
8467 values indicating the condition, and two values of the same :ref:`first
8468 class <t_firstclass>` type.
8473 If the condition is an i1 and it evaluates to 1, the instruction returns
8474 the first value argument; otherwise, it returns the second value
8477 If the condition is a vector of i1, then the value arguments must be
8478 vectors of the same size, and the selection is done element by element.
8480 If the condition is an i1 and the value arguments are vectors of the
8481 same size, then an entire vector is selected.
8486 .. code-block:: llvm
8488 %X = select i1 true, i8 17, i8 42 ; yields i8:17
8492 '``call``' Instruction
8493 ^^^^^^^^^^^^^^^^^^^^^^
8500 <result> = [tail | musttail | notail ] call [cconv] [ret attrs] <ty> [<fnty>*] <fnptrval>(<function args>) [fn attrs]
8506 The '``call``' instruction represents a simple function call.
8511 This instruction requires several arguments:
8513 #. The optional ``tail`` and ``musttail`` markers indicate that the optimizers
8514 should perform tail call optimization. The ``tail`` marker is a hint that
8515 `can be ignored <CodeGenerator.html#sibcallopt>`_. The ``musttail`` marker
8516 means that the call must be tail call optimized in order for the program to
8517 be correct. The ``musttail`` marker provides these guarantees:
8519 #. The call will not cause unbounded stack growth if it is part of a
8520 recursive cycle in the call graph.
8521 #. Arguments with the :ref:`inalloca <attr_inalloca>` attribute are
8524 Both markers imply that the callee does not access allocas or varargs from
8525 the caller. Calls marked ``musttail`` must obey the following additional
8528 - The call must immediately precede a :ref:`ret <i_ret>` instruction,
8529 or a pointer bitcast followed by a ret instruction.
8530 - The ret instruction must return the (possibly bitcasted) value
8531 produced by the call or void.
8532 - The caller and callee prototypes must match. Pointer types of
8533 parameters or return types may differ in pointee type, but not
8535 - The calling conventions of the caller and callee must match.
8536 - All ABI-impacting function attributes, such as sret, byval, inreg,
8537 returned, and inalloca, must match.
8538 - The callee must be varargs iff the caller is varargs. Bitcasting a
8539 non-varargs function to the appropriate varargs type is legal so
8540 long as the non-varargs prefixes obey the other rules.
8542 Tail call optimization for calls marked ``tail`` is guaranteed to occur if
8543 the following conditions are met:
8545 - Caller and callee both have the calling convention ``fastcc``.
8546 - The call is in tail position (ret immediately follows call and ret
8547 uses value of call or is void).
8548 - Option ``-tailcallopt`` is enabled, or
8549 ``llvm::GuaranteedTailCallOpt`` is ``true``.
8550 - `Platform-specific constraints are
8551 met. <CodeGenerator.html#tailcallopt>`_
8553 #. The optional ``notail`` marker indicates that the optimizers should not add
8554 ``tail`` or ``musttail`` markers to the call. It is used to prevent tail
8555 call optimization from being performed on the call.
8557 #. The optional "cconv" marker indicates which :ref:`calling
8558 convention <callingconv>` the call should use. If none is
8559 specified, the call defaults to using C calling conventions. The
8560 calling convention of the call must match the calling convention of
8561 the target function, or else the behavior is undefined.
8562 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
8563 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
8565 #. '``ty``': the type of the call instruction itself which is also the
8566 type of the return value. Functions that return no value are marked
8568 #. '``fnty``': shall be the signature of the pointer to function value
8569 being invoked. The argument types must match the types implied by
8570 this signature. This type can be omitted if the function is not
8571 varargs and if the function type does not return a pointer to a
8573 #. '``fnptrval``': An LLVM value containing a pointer to a function to
8574 be invoked. In most cases, this is a direct function invocation, but
8575 indirect ``call``'s are just as possible, calling an arbitrary pointer
8577 #. '``function args``': argument list whose types match the function
8578 signature argument types and parameter attributes. All arguments must
8579 be of :ref:`first class <t_firstclass>` type. If the function signature
8580 indicates the function accepts a variable number of arguments, the
8581 extra arguments can be specified.
8582 #. The optional :ref:`function attributes <fnattrs>` list. Only
8583 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
8584 attributes are valid here.
8585 #. The optional :ref:`operand bundles <opbundles>` list.
8590 The '``call``' instruction is used to cause control flow to transfer to
8591 a specified function, with its incoming arguments bound to the specified
8592 values. Upon a '``ret``' instruction in the called function, control
8593 flow continues with the instruction after the function call, and the
8594 return value of the function is bound to the result argument.
8599 .. code-block:: llvm
8601 %retval = call i32 @test(i32 %argc)
8602 call i32 (i8*, ...)* @printf(i8* %msg, i32 12, i8 42) ; yields i32
8603 %X = tail call i32 @foo() ; yields i32
8604 %Y = tail call fastcc i32 @foo() ; yields i32
8605 call void %foo(i8 97 signext)
8607 %struct.A = type { i32, i8 }
8608 %r = call %struct.A @foo() ; yields { i32, i8 }
8609 %gr = extractvalue %struct.A %r, 0 ; yields i32
8610 %gr1 = extractvalue %struct.A %r, 1 ; yields i8
8611 %Z = call void @foo() noreturn ; indicates that %foo never returns normally
8612 %ZZ = call zeroext i32 @bar() ; Return value is %zero extended
8614 llvm treats calls to some functions with names and arguments that match
8615 the standard C99 library as being the C99 library functions, and may
8616 perform optimizations or generate code for them under that assumption.
8617 This is something we'd like to change in the future to provide better
8618 support for freestanding environments and non-C-based languages.
8622 '``va_arg``' Instruction
8623 ^^^^^^^^^^^^^^^^^^^^^^^^
8630 <resultval> = va_arg <va_list*> <arglist>, <argty>
8635 The '``va_arg``' instruction is used to access arguments passed through
8636 the "variable argument" area of a function call. It is used to implement
8637 the ``va_arg`` macro in C.
8642 This instruction takes a ``va_list*`` value and the type of the
8643 argument. It returns a value of the specified argument type and
8644 increments the ``va_list`` to point to the next argument. The actual
8645 type of ``va_list`` is target specific.
8650 The '``va_arg``' instruction loads an argument of the specified type
8651 from the specified ``va_list`` and causes the ``va_list`` to point to
8652 the next argument. For more information, see the variable argument
8653 handling :ref:`Intrinsic Functions <int_varargs>`.
8655 It is legal for this instruction to be called in a function which does
8656 not take a variable number of arguments, for example, the ``vfprintf``
8659 ``va_arg`` is an LLVM instruction instead of an :ref:`intrinsic
8660 function <intrinsics>` because it takes a type as an argument.
8665 See the :ref:`variable argument processing <int_varargs>` section.
8667 Note that the code generator does not yet fully support va\_arg on many
8668 targets. Also, it does not currently support va\_arg with aggregate
8669 types on any target.
8673 '``landingpad``' Instruction
8674 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8681 <resultval> = landingpad <resultty> <clause>+
8682 <resultval> = landingpad <resultty> cleanup <clause>*
8684 <clause> := catch <type> <value>
8685 <clause> := filter <array constant type> <array constant>
8690 The '``landingpad``' instruction is used by `LLVM's exception handling
8691 system <ExceptionHandling.html#overview>`_ to specify that a basic block
8692 is a landing pad --- one where the exception lands, and corresponds to the
8693 code found in the ``catch`` portion of a ``try``/``catch`` sequence. It
8694 defines values supplied by the :ref:`personality function <personalityfn>` upon
8695 re-entry to the function. The ``resultval`` has the type ``resultty``.
8701 ``cleanup`` flag indicates that the landing pad block is a cleanup.
8703 A ``clause`` begins with the clause type --- ``catch`` or ``filter`` --- and
8704 contains the global variable representing the "type" that may be caught
8705 or filtered respectively. Unlike the ``catch`` clause, the ``filter``
8706 clause takes an array constant as its argument. Use
8707 "``[0 x i8**] undef``" for a filter which cannot throw. The
8708 '``landingpad``' instruction must contain *at least* one ``clause`` or
8709 the ``cleanup`` flag.
8714 The '``landingpad``' instruction defines the values which are set by the
8715 :ref:`personality function <personalityfn>` upon re-entry to the function, and
8716 therefore the "result type" of the ``landingpad`` instruction. As with
8717 calling conventions, how the personality function results are
8718 represented in LLVM IR is target specific.
8720 The clauses are applied in order from top to bottom. If two
8721 ``landingpad`` instructions are merged together through inlining, the
8722 clauses from the calling function are appended to the list of clauses.
8723 When the call stack is being unwound due to an exception being thrown,
8724 the exception is compared against each ``clause`` in turn. If it doesn't
8725 match any of the clauses, and the ``cleanup`` flag is not set, then
8726 unwinding continues further up the call stack.
8728 The ``landingpad`` instruction has several restrictions:
8730 - A landing pad block is a basic block which is the unwind destination
8731 of an '``invoke``' instruction.
8732 - A landing pad block must have a '``landingpad``' instruction as its
8733 first non-PHI instruction.
8734 - There can be only one '``landingpad``' instruction within the landing
8736 - A basic block that is not a landing pad block may not include a
8737 '``landingpad``' instruction.
8742 .. code-block:: llvm
8744 ;; A landing pad which can catch an integer.
8745 %res = landingpad { i8*, i32 }
8747 ;; A landing pad that is a cleanup.
8748 %res = landingpad { i8*, i32 }
8750 ;; A landing pad which can catch an integer and can only throw a double.
8751 %res = landingpad { i8*, i32 }
8753 filter [1 x i8**] [@_ZTId]
8757 '``cleanuppad``' Instruction
8758 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8765 <resultval> = cleanuppad [<args>*]
8770 The '``cleanuppad``' instruction is used by `LLVM's exception handling
8771 system <ExceptionHandling.html#overview>`_ to specify that a basic block
8772 is a cleanup block --- one where a personality routine attempts to
8773 transfer control to run cleanup actions.
8774 The ``args`` correspond to whatever additional
8775 information the :ref:`personality function <personalityfn>` requires to
8776 execute the cleanup.
8777 The ``resultval`` has the type :ref:`token <t_token>` and is used to
8778 match the ``cleanuppad`` to corresponding :ref:`cleanuprets <i_cleanupret>`
8779 and :ref:`cleanupendpads <i_cleanupendpad>`.
8784 The instruction takes a list of arbitrary values which are interpreted
8785 by the :ref:`personality function <personalityfn>`.
8790 When the call stack is being unwound due to an exception being thrown,
8791 the :ref:`personality function <personalityfn>` transfers control to the
8792 ``cleanuppad`` with the aid of the personality-specific arguments.
8793 As with calling conventions, how the personality function results are
8794 represented in LLVM IR is target specific.
8796 The ``cleanuppad`` instruction has several restrictions:
8798 - A cleanup block is a basic block which is the unwind destination of
8799 an exceptional instruction.
8800 - A cleanup block must have a '``cleanuppad``' instruction as its
8801 first non-PHI instruction.
8802 - There can be only one '``cleanuppad``' instruction within the
8804 - A basic block that is not a cleanup block may not include a
8805 '``cleanuppad``' instruction.
8806 - All '``cleanupret``'s and '``cleanupendpad``'s which consume a ``cleanuppad``
8807 must have the same exceptional successor.
8808 - It is undefined behavior for control to transfer from a ``cleanuppad`` to a
8809 ``ret`` without first executing a ``cleanupret`` or ``cleanupendpad`` that
8810 consumes the ``cleanuppad``.
8811 - It is undefined behavior for control to transfer from a ``cleanuppad`` to
8812 itself without first executing a ``cleanupret`` or ``cleanupendpad`` that
8813 consumes the ``cleanuppad``.
8818 .. code-block:: llvm
8820 %tok = cleanuppad []
8827 LLVM supports the notion of an "intrinsic function". These functions
8828 have well known names and semantics and are required to follow certain
8829 restrictions. Overall, these intrinsics represent an extension mechanism
8830 for the LLVM language that does not require changing all of the
8831 transformations in LLVM when adding to the language (or the bitcode
8832 reader/writer, the parser, etc...).
8834 Intrinsic function names must all start with an "``llvm.``" prefix. This
8835 prefix is reserved in LLVM for intrinsic names; thus, function names may
8836 not begin with this prefix. Intrinsic functions must always be external
8837 functions: you cannot define the body of intrinsic functions. Intrinsic
8838 functions may only be used in call or invoke instructions: it is illegal
8839 to take the address of an intrinsic function. Additionally, because
8840 intrinsic functions are part of the LLVM language, it is required if any
8841 are added that they be documented here.
8843 Some intrinsic functions can be overloaded, i.e., the intrinsic
8844 represents a family of functions that perform the same operation but on
8845 different data types. Because LLVM can represent over 8 million
8846 different integer types, overloading is used commonly to allow an
8847 intrinsic function to operate on any integer type. One or more of the
8848 argument types or the result type can be overloaded to accept any
8849 integer type. Argument types may also be defined as exactly matching a
8850 previous argument's type or the result type. This allows an intrinsic
8851 function which accepts multiple arguments, but needs all of them to be
8852 of the same type, to only be overloaded with respect to a single
8853 argument or the result.
8855 Overloaded intrinsics will have the names of its overloaded argument
8856 types encoded into its function name, each preceded by a period. Only
8857 those types which are overloaded result in a name suffix. Arguments
8858 whose type is matched against another type do not. For example, the
8859 ``llvm.ctpop`` function can take an integer of any width and returns an
8860 integer of exactly the same integer width. This leads to a family of
8861 functions such as ``i8 @llvm.ctpop.i8(i8 %val)`` and
8862 ``i29 @llvm.ctpop.i29(i29 %val)``. Only one type, the return type, is
8863 overloaded, and only one type suffix is required. Because the argument's
8864 type is matched against the return type, it does not require its own
8867 To learn how to add an intrinsic function, please see the `Extending
8868 LLVM Guide <ExtendingLLVM.html>`_.
8872 Variable Argument Handling Intrinsics
8873 -------------------------------------
8875 Variable argument support is defined in LLVM with the
8876 :ref:`va_arg <i_va_arg>` instruction and these three intrinsic
8877 functions. These functions are related to the similarly named macros
8878 defined in the ``<stdarg.h>`` header file.
8880 All of these functions operate on arguments that use a target-specific
8881 value type "``va_list``". The LLVM assembly language reference manual
8882 does not define what this type is, so all transformations should be
8883 prepared to handle these functions regardless of the type used.
8885 This example shows how the :ref:`va_arg <i_va_arg>` instruction and the
8886 variable argument handling intrinsic functions are used.
8888 .. code-block:: llvm
8890 ; This struct is different for every platform. For most platforms,
8891 ; it is merely an i8*.
8892 %struct.va_list = type { i8* }
8894 ; For Unix x86_64 platforms, va_list is the following struct:
8895 ; %struct.va_list = type { i32, i32, i8*, i8* }
8897 define i32 @test(i32 %X, ...) {
8898 ; Initialize variable argument processing
8899 %ap = alloca %struct.va_list
8900 %ap2 = bitcast %struct.va_list* %ap to i8*
8901 call void @llvm.va_start(i8* %ap2)
8903 ; Read a single integer argument
8904 %tmp = va_arg i8* %ap2, i32
8906 ; Demonstrate usage of llvm.va_copy and llvm.va_end
8908 %aq2 = bitcast i8** %aq to i8*
8909 call void @llvm.va_copy(i8* %aq2, i8* %ap2)
8910 call void @llvm.va_end(i8* %aq2)
8912 ; Stop processing of arguments.
8913 call void @llvm.va_end(i8* %ap2)
8917 declare void @llvm.va_start(i8*)
8918 declare void @llvm.va_copy(i8*, i8*)
8919 declare void @llvm.va_end(i8*)
8923 '``llvm.va_start``' Intrinsic
8924 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8931 declare void @llvm.va_start(i8* <arglist>)
8936 The '``llvm.va_start``' intrinsic initializes ``*<arglist>`` for
8937 subsequent use by ``va_arg``.
8942 The argument is a pointer to a ``va_list`` element to initialize.
8947 The '``llvm.va_start``' intrinsic works just like the ``va_start`` macro
8948 available in C. In a target-dependent way, it initializes the
8949 ``va_list`` element to which the argument points, so that the next call
8950 to ``va_arg`` will produce the first variable argument passed to the
8951 function. Unlike the C ``va_start`` macro, this intrinsic does not need
8952 to know the last argument of the function as the compiler can figure
8955 '``llvm.va_end``' Intrinsic
8956 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8963 declare void @llvm.va_end(i8* <arglist>)
8968 The '``llvm.va_end``' intrinsic destroys ``*<arglist>``, which has been
8969 initialized previously with ``llvm.va_start`` or ``llvm.va_copy``.
8974 The argument is a pointer to a ``va_list`` to destroy.
8979 The '``llvm.va_end``' intrinsic works just like the ``va_end`` macro
8980 available in C. In a target-dependent way, it destroys the ``va_list``
8981 element to which the argument points. Calls to
8982 :ref:`llvm.va_start <int_va_start>` and
8983 :ref:`llvm.va_copy <int_va_copy>` must be matched exactly with calls to
8988 '``llvm.va_copy``' Intrinsic
8989 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8996 declare void @llvm.va_copy(i8* <destarglist>, i8* <srcarglist>)
9001 The '``llvm.va_copy``' intrinsic copies the current argument position
9002 from the source argument list to the destination argument list.
9007 The first argument is a pointer to a ``va_list`` element to initialize.
9008 The second argument is a pointer to a ``va_list`` element to copy from.
9013 The '``llvm.va_copy``' intrinsic works just like the ``va_copy`` macro
9014 available in C. In a target-dependent way, it copies the source
9015 ``va_list`` element into the destination ``va_list`` element. This
9016 intrinsic is necessary because the `` llvm.va_start`` intrinsic may be
9017 arbitrarily complex and require, for example, memory allocation.
9019 Accurate Garbage Collection Intrinsics
9020 --------------------------------------
9022 LLVM's support for `Accurate Garbage Collection <GarbageCollection.html>`_
9023 (GC) requires the frontend to generate code containing appropriate intrinsic
9024 calls and select an appropriate GC strategy which knows how to lower these
9025 intrinsics in a manner which is appropriate for the target collector.
9027 These intrinsics allow identification of :ref:`GC roots on the
9028 stack <int_gcroot>`, as well as garbage collector implementations that
9029 require :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers.
9030 Frontends for type-safe garbage collected languages should generate
9031 these intrinsics to make use of the LLVM garbage collectors. For more
9032 details, see `Garbage Collection with LLVM <GarbageCollection.html>`_.
9034 Experimental Statepoint Intrinsics
9035 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9037 LLVM provides an second experimental set of intrinsics for describing garbage
9038 collection safepoints in compiled code. These intrinsics are an alternative
9039 to the ``llvm.gcroot`` intrinsics, but are compatible with the ones for
9040 :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers. The
9041 differences in approach are covered in the `Garbage Collection with LLVM
9042 <GarbageCollection.html>`_ documentation. The intrinsics themselves are
9043 described in :doc:`Statepoints`.
9047 '``llvm.gcroot``' Intrinsic
9048 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9055 declare void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
9060 The '``llvm.gcroot``' intrinsic declares the existence of a GC root to
9061 the code generator, and allows some metadata to be associated with it.
9066 The first argument specifies the address of a stack object that contains
9067 the root pointer. The second pointer (which must be either a constant or
9068 a global value address) contains the meta-data to be associated with the
9074 At runtime, a call to this intrinsic stores a null pointer into the
9075 "ptrloc" location. At compile-time, the code generator generates
9076 information to allow the runtime to find the pointer at GC safe points.
9077 The '``llvm.gcroot``' intrinsic may only be used in a function which
9078 :ref:`specifies a GC algorithm <gc>`.
9082 '``llvm.gcread``' Intrinsic
9083 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9090 declare i8* @llvm.gcread(i8* %ObjPtr, i8** %Ptr)
9095 The '``llvm.gcread``' intrinsic identifies reads of references from heap
9096 locations, allowing garbage collector implementations that require read
9102 The second argument is the address to read from, which should be an
9103 address allocated from the garbage collector. The first object is a
9104 pointer to the start of the referenced object, if needed by the language
9105 runtime (otherwise null).
9110 The '``llvm.gcread``' intrinsic has the same semantics as a load
9111 instruction, but may be replaced with substantially more complex code by
9112 the garbage collector runtime, as needed. The '``llvm.gcread``'
9113 intrinsic may only be used in a function which :ref:`specifies a GC
9118 '``llvm.gcwrite``' Intrinsic
9119 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9126 declare void @llvm.gcwrite(i8* %P1, i8* %Obj, i8** %P2)
9131 The '``llvm.gcwrite``' intrinsic identifies writes of references to heap
9132 locations, allowing garbage collector implementations that require write
9133 barriers (such as generational or reference counting collectors).
9138 The first argument is the reference to store, the second is the start of
9139 the object to store it to, and the third is the address of the field of
9140 Obj to store to. If the runtime does not require a pointer to the
9141 object, Obj may be null.
9146 The '``llvm.gcwrite``' intrinsic has the same semantics as a store
9147 instruction, but may be replaced with substantially more complex code by
9148 the garbage collector runtime, as needed. The '``llvm.gcwrite``'
9149 intrinsic may only be used in a function which :ref:`specifies a GC
9152 Code Generator Intrinsics
9153 -------------------------
9155 These intrinsics are provided by LLVM to expose special features that
9156 may only be implemented with code generator support.
9158 '``llvm.returnaddress``' Intrinsic
9159 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9166 declare i8 *@llvm.returnaddress(i32 <level>)
9171 The '``llvm.returnaddress``' intrinsic attempts to compute a
9172 target-specific value indicating the return address of the current
9173 function or one of its callers.
9178 The argument to this intrinsic indicates which function to return the
9179 address for. Zero indicates the calling function, one indicates its
9180 caller, etc. The argument is **required** to be a constant integer
9186 The '``llvm.returnaddress``' intrinsic either returns a pointer
9187 indicating the return address of the specified call frame, or zero if it
9188 cannot be identified. The value returned by this intrinsic is likely to
9189 be incorrect or 0 for arguments other than zero, so it should only be
9190 used for debugging purposes.
9192 Note that calling this intrinsic does not prevent function inlining or
9193 other aggressive transformations, so the value returned may not be that
9194 of the obvious source-language caller.
9196 '``llvm.frameaddress``' Intrinsic
9197 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9204 declare i8* @llvm.frameaddress(i32 <level>)
9209 The '``llvm.frameaddress``' intrinsic attempts to return the
9210 target-specific frame pointer value for the specified stack frame.
9215 The argument to this intrinsic indicates which function to return the
9216 frame pointer for. Zero indicates the calling function, one indicates
9217 its caller, etc. The argument is **required** to be a constant integer
9223 The '``llvm.frameaddress``' intrinsic either returns a pointer
9224 indicating the frame address of the specified call frame, or zero if it
9225 cannot be identified. The value returned by this intrinsic is likely to
9226 be incorrect or 0 for arguments other than zero, so it should only be
9227 used for debugging purposes.
9229 Note that calling this intrinsic does not prevent function inlining or
9230 other aggressive transformations, so the value returned may not be that
9231 of the obvious source-language caller.
9233 '``llvm.localescape``' and '``llvm.localrecover``' Intrinsics
9234 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9241 declare void @llvm.localescape(...)
9242 declare i8* @llvm.localrecover(i8* %func, i8* %fp, i32 %idx)
9247 The '``llvm.localescape``' intrinsic escapes offsets of a collection of static
9248 allocas, and the '``llvm.localrecover``' intrinsic applies those offsets to a
9249 live frame pointer to recover the address of the allocation. The offset is
9250 computed during frame layout of the caller of ``llvm.localescape``.
9255 All arguments to '``llvm.localescape``' must be pointers to static allocas or
9256 casts of static allocas. Each function can only call '``llvm.localescape``'
9257 once, and it can only do so from the entry block.
9259 The ``func`` argument to '``llvm.localrecover``' must be a constant
9260 bitcasted pointer to a function defined in the current module. The code
9261 generator cannot determine the frame allocation offset of functions defined in
9264 The ``fp`` argument to '``llvm.localrecover``' must be a frame pointer of a
9265 call frame that is currently live. The return value of '``llvm.localaddress``'
9266 is one way to produce such a value, but various runtimes also expose a suitable
9267 pointer in platform-specific ways.
9269 The ``idx`` argument to '``llvm.localrecover``' indicates which alloca passed to
9270 '``llvm.localescape``' to recover. It is zero-indexed.
9275 These intrinsics allow a group of functions to share access to a set of local
9276 stack allocations of a one parent function. The parent function may call the
9277 '``llvm.localescape``' intrinsic once from the function entry block, and the
9278 child functions can use '``llvm.localrecover``' to access the escaped allocas.
9279 The '``llvm.localescape``' intrinsic blocks inlining, as inlining changes where
9280 the escaped allocas are allocated, which would break attempts to use
9281 '``llvm.localrecover``'.
9283 .. _int_read_register:
9284 .. _int_write_register:
9286 '``llvm.read_register``' and '``llvm.write_register``' Intrinsics
9287 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9294 declare i32 @llvm.read_register.i32(metadata)
9295 declare i64 @llvm.read_register.i64(metadata)
9296 declare void @llvm.write_register.i32(metadata, i32 @value)
9297 declare void @llvm.write_register.i64(metadata, i64 @value)
9303 The '``llvm.read_register``' and '``llvm.write_register``' intrinsics
9304 provides access to the named register. The register must be valid on
9305 the architecture being compiled to. The type needs to be compatible
9306 with the register being read.
9311 The '``llvm.read_register``' intrinsic returns the current value of the
9312 register, where possible. The '``llvm.write_register``' intrinsic sets
9313 the current value of the register, where possible.
9315 This is useful to implement named register global variables that need
9316 to always be mapped to a specific register, as is common practice on
9317 bare-metal programs including OS kernels.
9319 The compiler doesn't check for register availability or use of the used
9320 register in surrounding code, including inline assembly. Because of that,
9321 allocatable registers are not supported.
9323 Warning: So far it only works with the stack pointer on selected
9324 architectures (ARM, AArch64, PowerPC and x86_64). Significant amount of
9325 work is needed to support other registers and even more so, allocatable
9330 '``llvm.stacksave``' Intrinsic
9331 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9338 declare i8* @llvm.stacksave()
9343 The '``llvm.stacksave``' intrinsic is used to remember the current state
9344 of the function stack, for use with
9345 :ref:`llvm.stackrestore <int_stackrestore>`. This is useful for
9346 implementing language features like scoped automatic variable sized
9352 This intrinsic returns a opaque pointer value that can be passed to
9353 :ref:`llvm.stackrestore <int_stackrestore>`. When an
9354 ``llvm.stackrestore`` intrinsic is executed with a value saved from
9355 ``llvm.stacksave``, it effectively restores the state of the stack to
9356 the state it was in when the ``llvm.stacksave`` intrinsic executed. In
9357 practice, this pops any :ref:`alloca <i_alloca>` blocks from the stack that
9358 were allocated after the ``llvm.stacksave`` was executed.
9360 .. _int_stackrestore:
9362 '``llvm.stackrestore``' Intrinsic
9363 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9370 declare void @llvm.stackrestore(i8* %ptr)
9375 The '``llvm.stackrestore``' intrinsic is used to restore the state of
9376 the function stack to the state it was in when the corresponding
9377 :ref:`llvm.stacksave <int_stacksave>` intrinsic executed. This is
9378 useful for implementing language features like scoped automatic variable
9379 sized arrays in C99.
9384 See the description for :ref:`llvm.stacksave <int_stacksave>`.
9386 .. _int_get_dynamic_area_offset:
9388 '``llvm.get.dynamic.area.offset``' Intrinsic
9389 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9396 declare i32 @llvm.get.dynamic.area.offset.i32()
9397 declare i64 @llvm.get.dynamic.area.offset.i64()
9402 The '``llvm.get.dynamic.area.offset.*``' intrinsic family is used to
9403 get the offset from native stack pointer to the address of the most
9404 recent dynamic alloca on the caller's stack. These intrinsics are
9405 intendend for use in combination with
9406 :ref:`llvm.stacksave <int_stacksave>` to get a
9407 pointer to the most recent dynamic alloca. This is useful, for example,
9408 for AddressSanitizer's stack unpoisoning routines.
9413 These intrinsics return a non-negative integer value that can be used to
9414 get the address of the most recent dynamic alloca, allocated by :ref:`alloca <i_alloca>`
9415 on the caller's stack. In particular, for targets where stack grows downwards,
9416 adding this offset to the native stack pointer would get the address of the most
9417 recent dynamic alloca. For targets where stack grows upwards, the situation is a bit more
9418 complicated, because substracting this value from stack pointer would get the address
9419 one past the end of the most recent dynamic alloca.
9421 Although for most targets `llvm.get.dynamic.area.offset <int_get_dynamic_area_offset>`
9422 returns just a zero, for others, such as PowerPC and PowerPC64, it returns a
9423 compile-time-known constant value.
9425 The return value type of :ref:`llvm.get.dynamic.area.offset <int_get_dynamic_area_offset>`
9426 must match the target's generic address space's (address space 0) pointer type.
9428 '``llvm.prefetch``' Intrinsic
9429 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9436 declare void @llvm.prefetch(i8* <address>, i32 <rw>, i32 <locality>, i32 <cache type>)
9441 The '``llvm.prefetch``' intrinsic is a hint to the code generator to
9442 insert a prefetch instruction if supported; otherwise, it is a noop.
9443 Prefetches have no effect on the behavior of the program but can change
9444 its performance characteristics.
9449 ``address`` is the address to be prefetched, ``rw`` is the specifier
9450 determining if the fetch should be for a read (0) or write (1), and
9451 ``locality`` is a temporal locality specifier ranging from (0) - no
9452 locality, to (3) - extremely local keep in cache. The ``cache type``
9453 specifies whether the prefetch is performed on the data (1) or
9454 instruction (0) cache. The ``rw``, ``locality`` and ``cache type``
9455 arguments must be constant integers.
9460 This intrinsic does not modify the behavior of the program. In
9461 particular, prefetches cannot trap and do not produce a value. On
9462 targets that support this intrinsic, the prefetch can provide hints to
9463 the processor cache for better performance.
9465 '``llvm.pcmarker``' Intrinsic
9466 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9473 declare void @llvm.pcmarker(i32 <id>)
9478 The '``llvm.pcmarker``' intrinsic is a method to export a Program
9479 Counter (PC) in a region of code to simulators and other tools. The
9480 method is target specific, but it is expected that the marker will use
9481 exported symbols to transmit the PC of the marker. The marker makes no
9482 guarantees that it will remain with any specific instruction after
9483 optimizations. It is possible that the presence of a marker will inhibit
9484 optimizations. The intended use is to be inserted after optimizations to
9485 allow correlations of simulation runs.
9490 ``id`` is a numerical id identifying the marker.
9495 This intrinsic does not modify the behavior of the program. Backends
9496 that do not support this intrinsic may ignore it.
9498 '``llvm.readcyclecounter``' Intrinsic
9499 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9506 declare i64 @llvm.readcyclecounter()
9511 The '``llvm.readcyclecounter``' intrinsic provides access to the cycle
9512 counter register (or similar low latency, high accuracy clocks) on those
9513 targets that support it. On X86, it should map to RDTSC. On Alpha, it
9514 should map to RPCC. As the backing counters overflow quickly (on the
9515 order of 9 seconds on alpha), this should only be used for small
9521 When directly supported, reading the cycle counter should not modify any
9522 memory. Implementations are allowed to either return a application
9523 specific value or a system wide value. On backends without support, this
9524 is lowered to a constant 0.
9526 Note that runtime support may be conditional on the privilege-level code is
9527 running at and the host platform.
9529 '``llvm.clear_cache``' Intrinsic
9530 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9537 declare void @llvm.clear_cache(i8*, i8*)
9542 The '``llvm.clear_cache``' intrinsic ensures visibility of modifications
9543 in the specified range to the execution unit of the processor. On
9544 targets with non-unified instruction and data cache, the implementation
9545 flushes the instruction cache.
9550 On platforms with coherent instruction and data caches (e.g. x86), this
9551 intrinsic is a nop. On platforms with non-coherent instruction and data
9552 cache (e.g. ARM, MIPS), the intrinsic is lowered either to appropriate
9553 instructions or a system call, if cache flushing requires special
9556 The default behavior is to emit a call to ``__clear_cache`` from the run
9559 This instrinsic does *not* empty the instruction pipeline. Modifications
9560 of the current function are outside the scope of the intrinsic.
9562 '``llvm.instrprof_increment``' Intrinsic
9563 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9570 declare void @llvm.instrprof_increment(i8* <name>, i64 <hash>,
9571 i32 <num-counters>, i32 <index>)
9576 The '``llvm.instrprof_increment``' intrinsic can be emitted by a
9577 frontend for use with instrumentation based profiling. These will be
9578 lowered by the ``-instrprof`` pass to generate execution counts of a
9584 The first argument is a pointer to a global variable containing the
9585 name of the entity being instrumented. This should generally be the
9586 (mangled) function name for a set of counters.
9588 The second argument is a hash value that can be used by the consumer
9589 of the profile data to detect changes to the instrumented source, and
9590 the third is the number of counters associated with ``name``. It is an
9591 error if ``hash`` or ``num-counters`` differ between two instances of
9592 ``instrprof_increment`` that refer to the same name.
9594 The last argument refers to which of the counters for ``name`` should
9595 be incremented. It should be a value between 0 and ``num-counters``.
9600 This intrinsic represents an increment of a profiling counter. It will
9601 cause the ``-instrprof`` pass to generate the appropriate data
9602 structures and the code to increment the appropriate value, in a
9603 format that can be written out by a compiler runtime and consumed via
9604 the ``llvm-profdata`` tool.
9606 '``llvm.instrprof_value_profile``' Intrinsic
9607 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9614 declare void @llvm.instrprof_value_profile(i8* <name>, i64 <hash>,
9615 i64 <value>, i32 <value_kind>,
9621 The '``llvm.instrprof_value_profile``' intrinsic can be emitted by a
9622 frontend for use with instrumentation based profiling. This will be
9623 lowered by the ``-instrprof`` pass to find out the target values,
9624 instrumented expressions take in a program at runtime.
9629 The first argument is a pointer to a global variable containing the
9630 name of the entity being instrumented. ``name`` should generally be the
9631 (mangled) function name for a set of counters.
9633 The second argument is a hash value that can be used by the consumer
9634 of the profile data to detect changes to the instrumented source. It
9635 is an error if ``hash`` differs between two instances of
9636 ``llvm.instrprof_*`` that refer to the same name.
9638 The third argument is the value of the expression being profiled. The profiled
9639 expression's value should be representable as an unsigned 64-bit value. The
9640 fourth argument represents the kind of value profiling that is being done. The
9641 supported value profiling kinds are enumerated through the
9642 ``InstrProfValueKind`` type declared in the
9643 ``<include/llvm/ProfileData/InstrProf.h>`` header file. The last argument is the
9644 index of the instrumented expression within ``name``. It should be >= 0.
9649 This intrinsic represents the point where a call to a runtime routine
9650 should be inserted for value profiling of target expressions. ``-instrprof``
9651 pass will generate the appropriate data structures and replace the
9652 ``llvm.instrprof_value_profile`` intrinsic with the call to the profile
9653 runtime library with proper arguments.
9655 Standard C Library Intrinsics
9656 -----------------------------
9658 LLVM provides intrinsics for a few important standard C library
9659 functions. These intrinsics allow source-language front-ends to pass
9660 information about the alignment of the pointer arguments to the code
9661 generator, providing opportunity for more efficient code generation.
9665 '``llvm.memcpy``' Intrinsic
9666 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9671 This is an overloaded intrinsic. You can use ``llvm.memcpy`` on any
9672 integer bit width and for different address spaces. Not all targets
9673 support all bit widths however.
9677 declare void @llvm.memcpy.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
9678 i32 <len>, i32 <align>, i1 <isvolatile>)
9679 declare void @llvm.memcpy.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
9680 i64 <len>, i32 <align>, i1 <isvolatile>)
9685 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
9686 source location to the destination location.
9688 Note that, unlike the standard libc function, the ``llvm.memcpy.*``
9689 intrinsics do not return a value, takes extra alignment/isvolatile
9690 arguments and the pointers can be in specified address spaces.
9695 The first argument is a pointer to the destination, the second is a
9696 pointer to the source. The third argument is an integer argument
9697 specifying the number of bytes to copy, the fourth argument is the
9698 alignment of the source and destination locations, and the fifth is a
9699 boolean indicating a volatile access.
9701 If the call to this intrinsic has an alignment value that is not 0 or 1,
9702 then the caller guarantees that both the source and destination pointers
9703 are aligned to that boundary.
9705 If the ``isvolatile`` parameter is ``true``, the ``llvm.memcpy`` call is
9706 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
9707 very cleanly specified and it is unwise to depend on it.
9712 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
9713 source location to the destination location, which are not allowed to
9714 overlap. It copies "len" bytes of memory over. If the argument is known
9715 to be aligned to some boundary, this can be specified as the fourth
9716 argument, otherwise it should be set to 0 or 1 (both meaning no alignment).
9718 '``llvm.memmove``' Intrinsic
9719 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9724 This is an overloaded intrinsic. You can use llvm.memmove on any integer
9725 bit width and for different address space. Not all targets support all
9730 declare void @llvm.memmove.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
9731 i32 <len>, i32 <align>, i1 <isvolatile>)
9732 declare void @llvm.memmove.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
9733 i64 <len>, i32 <align>, i1 <isvolatile>)
9738 The '``llvm.memmove.*``' intrinsics move a block of memory from the
9739 source location to the destination location. It is similar to the
9740 '``llvm.memcpy``' intrinsic but allows the two memory locations to
9743 Note that, unlike the standard libc function, the ``llvm.memmove.*``
9744 intrinsics do not return a value, takes extra alignment/isvolatile
9745 arguments and the pointers can be in specified address spaces.
9750 The first argument is a pointer to the destination, the second is a
9751 pointer to the source. The third argument is an integer argument
9752 specifying the number of bytes to copy, the fourth argument is the
9753 alignment of the source and destination locations, and the fifth is a
9754 boolean indicating a volatile access.
9756 If the call to this intrinsic has an alignment value that is not 0 or 1,
9757 then the caller guarantees that the source and destination pointers are
9758 aligned to that boundary.
9760 If the ``isvolatile`` parameter is ``true``, the ``llvm.memmove`` call
9761 is a :ref:`volatile operation <volatile>`. The detailed access behavior is
9762 not very cleanly specified and it is unwise to depend on it.
9767 The '``llvm.memmove.*``' intrinsics copy a block of memory from the
9768 source location to the destination location, which may overlap. It
9769 copies "len" bytes of memory over. If the argument is known to be
9770 aligned to some boundary, this can be specified as the fourth argument,
9771 otherwise it should be set to 0 or 1 (both meaning no alignment).
9773 '``llvm.memset.*``' Intrinsics
9774 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9779 This is an overloaded intrinsic. You can use llvm.memset on any integer
9780 bit width and for different address spaces. However, not all targets
9781 support all bit widths.
9785 declare void @llvm.memset.p0i8.i32(i8* <dest>, i8 <val>,
9786 i32 <len>, i32 <align>, i1 <isvolatile>)
9787 declare void @llvm.memset.p0i8.i64(i8* <dest>, i8 <val>,
9788 i64 <len>, i32 <align>, i1 <isvolatile>)
9793 The '``llvm.memset.*``' intrinsics fill a block of memory with a
9794 particular byte value.
9796 Note that, unlike the standard libc function, the ``llvm.memset``
9797 intrinsic does not return a value and takes extra alignment/volatile
9798 arguments. Also, the destination can be in an arbitrary address space.
9803 The first argument is a pointer to the destination to fill, the second
9804 is the byte value with which to fill it, the third argument is an
9805 integer argument specifying the number of bytes to fill, and the fourth
9806 argument is the known alignment of the destination location.
9808 If the call to this intrinsic has an alignment value that is not 0 or 1,
9809 then the caller guarantees that the destination pointer is aligned to
9812 If the ``isvolatile`` parameter is ``true``, the ``llvm.memset`` call is
9813 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
9814 very cleanly specified and it is unwise to depend on it.
9819 The '``llvm.memset.*``' intrinsics fill "len" bytes of memory starting
9820 at the destination location. If the argument is known to be aligned to
9821 some boundary, this can be specified as the fourth argument, otherwise
9822 it should be set to 0 or 1 (both meaning no alignment).
9824 '``llvm.sqrt.*``' Intrinsic
9825 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9830 This is an overloaded intrinsic. You can use ``llvm.sqrt`` on any
9831 floating point or vector of floating point type. Not all targets support
9836 declare float @llvm.sqrt.f32(float %Val)
9837 declare double @llvm.sqrt.f64(double %Val)
9838 declare x86_fp80 @llvm.sqrt.f80(x86_fp80 %Val)
9839 declare fp128 @llvm.sqrt.f128(fp128 %Val)
9840 declare ppc_fp128 @llvm.sqrt.ppcf128(ppc_fp128 %Val)
9845 The '``llvm.sqrt``' intrinsics return the sqrt of the specified operand,
9846 returning the same value as the libm '``sqrt``' functions would. Unlike
9847 ``sqrt`` in libm, however, ``llvm.sqrt`` has undefined behavior for
9848 negative numbers other than -0.0 (which allows for better optimization,
9849 because there is no need to worry about errno being set).
9850 ``llvm.sqrt(-0.0)`` is defined to return -0.0 like IEEE sqrt.
9855 The argument and return value are floating point numbers of the same
9861 This function returns the sqrt of the specified operand if it is a
9862 nonnegative floating point number.
9864 '``llvm.powi.*``' Intrinsic
9865 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9870 This is an overloaded intrinsic. You can use ``llvm.powi`` on any
9871 floating point or vector of floating point type. Not all targets support
9876 declare float @llvm.powi.f32(float %Val, i32 %power)
9877 declare double @llvm.powi.f64(double %Val, i32 %power)
9878 declare x86_fp80 @llvm.powi.f80(x86_fp80 %Val, i32 %power)
9879 declare fp128 @llvm.powi.f128(fp128 %Val, i32 %power)
9880 declare ppc_fp128 @llvm.powi.ppcf128(ppc_fp128 %Val, i32 %power)
9885 The '``llvm.powi.*``' intrinsics return the first operand raised to the
9886 specified (positive or negative) power. The order of evaluation of
9887 multiplications is not defined. When a vector of floating point type is
9888 used, the second argument remains a scalar integer value.
9893 The second argument is an integer power, and the first is a value to
9894 raise to that power.
9899 This function returns the first value raised to the second power with an
9900 unspecified sequence of rounding operations.
9902 '``llvm.sin.*``' Intrinsic
9903 ^^^^^^^^^^^^^^^^^^^^^^^^^^
9908 This is an overloaded intrinsic. You can use ``llvm.sin`` on any
9909 floating point or vector of floating point type. Not all targets support
9914 declare float @llvm.sin.f32(float %Val)
9915 declare double @llvm.sin.f64(double %Val)
9916 declare x86_fp80 @llvm.sin.f80(x86_fp80 %Val)
9917 declare fp128 @llvm.sin.f128(fp128 %Val)
9918 declare ppc_fp128 @llvm.sin.ppcf128(ppc_fp128 %Val)
9923 The '``llvm.sin.*``' intrinsics return the sine of the operand.
9928 The argument and return value are floating point numbers of the same
9934 This function returns the sine of the specified operand, returning the
9935 same values as the libm ``sin`` functions would, and handles error
9936 conditions in the same way.
9938 '``llvm.cos.*``' Intrinsic
9939 ^^^^^^^^^^^^^^^^^^^^^^^^^^
9944 This is an overloaded intrinsic. You can use ``llvm.cos`` on any
9945 floating point or vector of floating point type. Not all targets support
9950 declare float @llvm.cos.f32(float %Val)
9951 declare double @llvm.cos.f64(double %Val)
9952 declare x86_fp80 @llvm.cos.f80(x86_fp80 %Val)
9953 declare fp128 @llvm.cos.f128(fp128 %Val)
9954 declare ppc_fp128 @llvm.cos.ppcf128(ppc_fp128 %Val)
9959 The '``llvm.cos.*``' intrinsics return the cosine of the operand.
9964 The argument and return value are floating point numbers of the same
9970 This function returns the cosine of the specified operand, returning the
9971 same values as the libm ``cos`` functions would, and handles error
9972 conditions in the same way.
9974 '``llvm.pow.*``' Intrinsic
9975 ^^^^^^^^^^^^^^^^^^^^^^^^^^
9980 This is an overloaded intrinsic. You can use ``llvm.pow`` on any
9981 floating point or vector of floating point type. Not all targets support
9986 declare float @llvm.pow.f32(float %Val, float %Power)
9987 declare double @llvm.pow.f64(double %Val, double %Power)
9988 declare x86_fp80 @llvm.pow.f80(x86_fp80 %Val, x86_fp80 %Power)
9989 declare fp128 @llvm.pow.f128(fp128 %Val, fp128 %Power)
9990 declare ppc_fp128 @llvm.pow.ppcf128(ppc_fp128 %Val, ppc_fp128 Power)
9995 The '``llvm.pow.*``' intrinsics return the first operand raised to the
9996 specified (positive or negative) power.
10001 The second argument is a floating point power, and the first is a value
10002 to raise to that power.
10007 This function returns the first value raised to the second power,
10008 returning the same values as the libm ``pow`` functions would, and
10009 handles error conditions in the same way.
10011 '``llvm.exp.*``' Intrinsic
10012 ^^^^^^^^^^^^^^^^^^^^^^^^^^
10017 This is an overloaded intrinsic. You can use ``llvm.exp`` on any
10018 floating point or vector of floating point type. Not all targets support
10023 declare float @llvm.exp.f32(float %Val)
10024 declare double @llvm.exp.f64(double %Val)
10025 declare x86_fp80 @llvm.exp.f80(x86_fp80 %Val)
10026 declare fp128 @llvm.exp.f128(fp128 %Val)
10027 declare ppc_fp128 @llvm.exp.ppcf128(ppc_fp128 %Val)
10032 The '``llvm.exp.*``' intrinsics perform the exp function.
10037 The argument and return value are floating point numbers of the same
10043 This function returns the same values as the libm ``exp`` functions
10044 would, and handles error conditions in the same way.
10046 '``llvm.exp2.*``' Intrinsic
10047 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10052 This is an overloaded intrinsic. You can use ``llvm.exp2`` on any
10053 floating point or vector of floating point type. Not all targets support
10058 declare float @llvm.exp2.f32(float %Val)
10059 declare double @llvm.exp2.f64(double %Val)
10060 declare x86_fp80 @llvm.exp2.f80(x86_fp80 %Val)
10061 declare fp128 @llvm.exp2.f128(fp128 %Val)
10062 declare ppc_fp128 @llvm.exp2.ppcf128(ppc_fp128 %Val)
10067 The '``llvm.exp2.*``' intrinsics perform the exp2 function.
10072 The argument and return value are floating point numbers of the same
10078 This function returns the same values as the libm ``exp2`` functions
10079 would, and handles error conditions in the same way.
10081 '``llvm.log.*``' Intrinsic
10082 ^^^^^^^^^^^^^^^^^^^^^^^^^^
10087 This is an overloaded intrinsic. You can use ``llvm.log`` on any
10088 floating point or vector of floating point type. Not all targets support
10093 declare float @llvm.log.f32(float %Val)
10094 declare double @llvm.log.f64(double %Val)
10095 declare x86_fp80 @llvm.log.f80(x86_fp80 %Val)
10096 declare fp128 @llvm.log.f128(fp128 %Val)
10097 declare ppc_fp128 @llvm.log.ppcf128(ppc_fp128 %Val)
10102 The '``llvm.log.*``' intrinsics perform the log function.
10107 The argument and return value are floating point numbers of the same
10113 This function returns the same values as the libm ``log`` functions
10114 would, and handles error conditions in the same way.
10116 '``llvm.log10.*``' Intrinsic
10117 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10122 This is an overloaded intrinsic. You can use ``llvm.log10`` on any
10123 floating point or vector of floating point type. Not all targets support
10128 declare float @llvm.log10.f32(float %Val)
10129 declare double @llvm.log10.f64(double %Val)
10130 declare x86_fp80 @llvm.log10.f80(x86_fp80 %Val)
10131 declare fp128 @llvm.log10.f128(fp128 %Val)
10132 declare ppc_fp128 @llvm.log10.ppcf128(ppc_fp128 %Val)
10137 The '``llvm.log10.*``' intrinsics perform the log10 function.
10142 The argument and return value are floating point numbers of the same
10148 This function returns the same values as the libm ``log10`` functions
10149 would, and handles error conditions in the same way.
10151 '``llvm.log2.*``' Intrinsic
10152 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10157 This is an overloaded intrinsic. You can use ``llvm.log2`` on any
10158 floating point or vector of floating point type. Not all targets support
10163 declare float @llvm.log2.f32(float %Val)
10164 declare double @llvm.log2.f64(double %Val)
10165 declare x86_fp80 @llvm.log2.f80(x86_fp80 %Val)
10166 declare fp128 @llvm.log2.f128(fp128 %Val)
10167 declare ppc_fp128 @llvm.log2.ppcf128(ppc_fp128 %Val)
10172 The '``llvm.log2.*``' intrinsics perform the log2 function.
10177 The argument and return value are floating point numbers of the same
10183 This function returns the same values as the libm ``log2`` functions
10184 would, and handles error conditions in the same way.
10186 '``llvm.fma.*``' Intrinsic
10187 ^^^^^^^^^^^^^^^^^^^^^^^^^^
10192 This is an overloaded intrinsic. You can use ``llvm.fma`` on any
10193 floating point or vector of floating point type. Not all targets support
10198 declare float @llvm.fma.f32(float %a, float %b, float %c)
10199 declare double @llvm.fma.f64(double %a, double %b, double %c)
10200 declare x86_fp80 @llvm.fma.f80(x86_fp80 %a, x86_fp80 %b, x86_fp80 %c)
10201 declare fp128 @llvm.fma.f128(fp128 %a, fp128 %b, fp128 %c)
10202 declare ppc_fp128 @llvm.fma.ppcf128(ppc_fp128 %a, ppc_fp128 %b, ppc_fp128 %c)
10207 The '``llvm.fma.*``' intrinsics perform the fused multiply-add
10213 The argument and return value are floating point numbers of the same
10219 This function returns the same values as the libm ``fma`` functions
10220 would, and does not set errno.
10222 '``llvm.fabs.*``' Intrinsic
10223 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10228 This is an overloaded intrinsic. You can use ``llvm.fabs`` on any
10229 floating point or vector of floating point type. Not all targets support
10234 declare float @llvm.fabs.f32(float %Val)
10235 declare double @llvm.fabs.f64(double %Val)
10236 declare x86_fp80 @llvm.fabs.f80(x86_fp80 %Val)
10237 declare fp128 @llvm.fabs.f128(fp128 %Val)
10238 declare ppc_fp128 @llvm.fabs.ppcf128(ppc_fp128 %Val)
10243 The '``llvm.fabs.*``' intrinsics return the absolute value of the
10249 The argument and return value are floating point numbers of the same
10255 This function returns the same values as the libm ``fabs`` functions
10256 would, and handles error conditions in the same way.
10258 '``llvm.minnum.*``' Intrinsic
10259 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10264 This is an overloaded intrinsic. You can use ``llvm.minnum`` on any
10265 floating point or vector of floating point type. Not all targets support
10270 declare float @llvm.minnum.f32(float %Val0, float %Val1)
10271 declare double @llvm.minnum.f64(double %Val0, double %Val1)
10272 declare x86_fp80 @llvm.minnum.f80(x86_fp80 %Val0, x86_fp80 %Val1)
10273 declare fp128 @llvm.minnum.f128(fp128 %Val0, fp128 %Val1)
10274 declare ppc_fp128 @llvm.minnum.ppcf128(ppc_fp128 %Val0, ppc_fp128 %Val1)
10279 The '``llvm.minnum.*``' intrinsics return the minimum of the two
10286 The arguments and return value are floating point numbers of the same
10292 Follows the IEEE-754 semantics for minNum, which also match for libm's
10295 If either operand is a NaN, returns the other non-NaN operand. Returns
10296 NaN only if both operands are NaN. If the operands compare equal,
10297 returns a value that compares equal to both operands. This means that
10298 fmin(+/-0.0, +/-0.0) could return either -0.0 or 0.0.
10300 '``llvm.maxnum.*``' Intrinsic
10301 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10306 This is an overloaded intrinsic. You can use ``llvm.maxnum`` on any
10307 floating point or vector of floating point type. Not all targets support
10312 declare float @llvm.maxnum.f32(float %Val0, float %Val1l)
10313 declare double @llvm.maxnum.f64(double %Val0, double %Val1)
10314 declare x86_fp80 @llvm.maxnum.f80(x86_fp80 %Val0, x86_fp80 %Val1)
10315 declare fp128 @llvm.maxnum.f128(fp128 %Val0, fp128 %Val1)
10316 declare ppc_fp128 @llvm.maxnum.ppcf128(ppc_fp128 %Val0, ppc_fp128 %Val1)
10321 The '``llvm.maxnum.*``' intrinsics return the maximum of the two
10328 The arguments and return value are floating point numbers of the same
10333 Follows the IEEE-754 semantics for maxNum, which also match for libm's
10336 If either operand is a NaN, returns the other non-NaN operand. Returns
10337 NaN only if both operands are NaN. If the operands compare equal,
10338 returns a value that compares equal to both operands. This means that
10339 fmax(+/-0.0, +/-0.0) could return either -0.0 or 0.0.
10341 '``llvm.copysign.*``' Intrinsic
10342 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10347 This is an overloaded intrinsic. You can use ``llvm.copysign`` on any
10348 floating point or vector of floating point type. Not all targets support
10353 declare float @llvm.copysign.f32(float %Mag, float %Sgn)
10354 declare double @llvm.copysign.f64(double %Mag, double %Sgn)
10355 declare x86_fp80 @llvm.copysign.f80(x86_fp80 %Mag, x86_fp80 %Sgn)
10356 declare fp128 @llvm.copysign.f128(fp128 %Mag, fp128 %Sgn)
10357 declare ppc_fp128 @llvm.copysign.ppcf128(ppc_fp128 %Mag, ppc_fp128 %Sgn)
10362 The '``llvm.copysign.*``' intrinsics return a value with the magnitude of the
10363 first operand and the sign of the second operand.
10368 The arguments and return value are floating point numbers of the same
10374 This function returns the same values as the libm ``copysign``
10375 functions would, and handles error conditions in the same way.
10377 '``llvm.floor.*``' Intrinsic
10378 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10383 This is an overloaded intrinsic. You can use ``llvm.floor`` on any
10384 floating point or vector of floating point type. Not all targets support
10389 declare float @llvm.floor.f32(float %Val)
10390 declare double @llvm.floor.f64(double %Val)
10391 declare x86_fp80 @llvm.floor.f80(x86_fp80 %Val)
10392 declare fp128 @llvm.floor.f128(fp128 %Val)
10393 declare ppc_fp128 @llvm.floor.ppcf128(ppc_fp128 %Val)
10398 The '``llvm.floor.*``' intrinsics return the floor of the operand.
10403 The argument and return value are floating point numbers of the same
10409 This function returns the same values as the libm ``floor`` functions
10410 would, and handles error conditions in the same way.
10412 '``llvm.ceil.*``' Intrinsic
10413 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10418 This is an overloaded intrinsic. You can use ``llvm.ceil`` on any
10419 floating point or vector of floating point type. Not all targets support
10424 declare float @llvm.ceil.f32(float %Val)
10425 declare double @llvm.ceil.f64(double %Val)
10426 declare x86_fp80 @llvm.ceil.f80(x86_fp80 %Val)
10427 declare fp128 @llvm.ceil.f128(fp128 %Val)
10428 declare ppc_fp128 @llvm.ceil.ppcf128(ppc_fp128 %Val)
10433 The '``llvm.ceil.*``' intrinsics return the ceiling of the operand.
10438 The argument and return value are floating point numbers of the same
10444 This function returns the same values as the libm ``ceil`` functions
10445 would, and handles error conditions in the same way.
10447 '``llvm.trunc.*``' Intrinsic
10448 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10453 This is an overloaded intrinsic. You can use ``llvm.trunc`` on any
10454 floating point or vector of floating point type. Not all targets support
10459 declare float @llvm.trunc.f32(float %Val)
10460 declare double @llvm.trunc.f64(double %Val)
10461 declare x86_fp80 @llvm.trunc.f80(x86_fp80 %Val)
10462 declare fp128 @llvm.trunc.f128(fp128 %Val)
10463 declare ppc_fp128 @llvm.trunc.ppcf128(ppc_fp128 %Val)
10468 The '``llvm.trunc.*``' intrinsics returns the operand rounded to the
10469 nearest integer not larger in magnitude than the operand.
10474 The argument and return value are floating point numbers of the same
10480 This function returns the same values as the libm ``trunc`` functions
10481 would, and handles error conditions in the same way.
10483 '``llvm.rint.*``' Intrinsic
10484 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10489 This is an overloaded intrinsic. You can use ``llvm.rint`` on any
10490 floating point or vector of floating point type. Not all targets support
10495 declare float @llvm.rint.f32(float %Val)
10496 declare double @llvm.rint.f64(double %Val)
10497 declare x86_fp80 @llvm.rint.f80(x86_fp80 %Val)
10498 declare fp128 @llvm.rint.f128(fp128 %Val)
10499 declare ppc_fp128 @llvm.rint.ppcf128(ppc_fp128 %Val)
10504 The '``llvm.rint.*``' intrinsics returns the operand rounded to the
10505 nearest integer. It may raise an inexact floating-point exception if the
10506 operand isn't an integer.
10511 The argument and return value are floating point numbers of the same
10517 This function returns the same values as the libm ``rint`` functions
10518 would, and handles error conditions in the same way.
10520 '``llvm.nearbyint.*``' Intrinsic
10521 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10526 This is an overloaded intrinsic. You can use ``llvm.nearbyint`` on any
10527 floating point or vector of floating point type. Not all targets support
10532 declare float @llvm.nearbyint.f32(float %Val)
10533 declare double @llvm.nearbyint.f64(double %Val)
10534 declare x86_fp80 @llvm.nearbyint.f80(x86_fp80 %Val)
10535 declare fp128 @llvm.nearbyint.f128(fp128 %Val)
10536 declare ppc_fp128 @llvm.nearbyint.ppcf128(ppc_fp128 %Val)
10541 The '``llvm.nearbyint.*``' intrinsics returns the operand rounded to the
10547 The argument and return value are floating point numbers of the same
10553 This function returns the same values as the libm ``nearbyint``
10554 functions would, and handles error conditions in the same way.
10556 '``llvm.round.*``' Intrinsic
10557 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10562 This is an overloaded intrinsic. You can use ``llvm.round`` on any
10563 floating point or vector of floating point type. Not all targets support
10568 declare float @llvm.round.f32(float %Val)
10569 declare double @llvm.round.f64(double %Val)
10570 declare x86_fp80 @llvm.round.f80(x86_fp80 %Val)
10571 declare fp128 @llvm.round.f128(fp128 %Val)
10572 declare ppc_fp128 @llvm.round.ppcf128(ppc_fp128 %Val)
10577 The '``llvm.round.*``' intrinsics returns the operand rounded to the
10583 The argument and return value are floating point numbers of the same
10589 This function returns the same values as the libm ``round``
10590 functions would, and handles error conditions in the same way.
10592 Bit Manipulation Intrinsics
10593 ---------------------------
10595 LLVM provides intrinsics for a few important bit manipulation
10596 operations. These allow efficient code generation for some algorithms.
10598 '``llvm.bitreverse.*``' Intrinsics
10599 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10604 This is an overloaded intrinsic function. You can use bitreverse on any
10609 declare i16 @llvm.bitreverse.i16(i16 <id>)
10610 declare i32 @llvm.bitreverse.i32(i32 <id>)
10611 declare i64 @llvm.bitreverse.i64(i64 <id>)
10616 The '``llvm.bitreverse``' family of intrinsics is used to reverse the
10617 bitpattern of an integer value; for example ``0b1234567`` becomes
10623 The ``llvm.bitreverse.iN`` intrinsic returns an i16 value that has bit
10624 ``M`` in the input moved to bit ``N-M`` in the output.
10626 '``llvm.bswap.*``' Intrinsics
10627 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10632 This is an overloaded intrinsic function. You can use bswap on any
10633 integer type that is an even number of bytes (i.e. BitWidth % 16 == 0).
10637 declare i16 @llvm.bswap.i16(i16 <id>)
10638 declare i32 @llvm.bswap.i32(i32 <id>)
10639 declare i64 @llvm.bswap.i64(i64 <id>)
10644 The '``llvm.bswap``' family of intrinsics is used to byte swap integer
10645 values with an even number of bytes (positive multiple of 16 bits).
10646 These are useful for performing operations on data that is not in the
10647 target's native byte order.
10652 The ``llvm.bswap.i16`` intrinsic returns an i16 value that has the high
10653 and low byte of the input i16 swapped. Similarly, the ``llvm.bswap.i32``
10654 intrinsic returns an i32 value that has the four bytes of the input i32
10655 swapped, so that if the input bytes are numbered 0, 1, 2, 3 then the
10656 returned i32 will have its bytes in 3, 2, 1, 0 order. The
10657 ``llvm.bswap.i48``, ``llvm.bswap.i64`` and other intrinsics extend this
10658 concept to additional even-byte lengths (6 bytes, 8 bytes and more,
10661 '``llvm.ctpop.*``' Intrinsic
10662 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10667 This is an overloaded intrinsic. You can use llvm.ctpop on any integer
10668 bit width, or on any vector with integer elements. Not all targets
10669 support all bit widths or vector types, however.
10673 declare i8 @llvm.ctpop.i8(i8 <src>)
10674 declare i16 @llvm.ctpop.i16(i16 <src>)
10675 declare i32 @llvm.ctpop.i32(i32 <src>)
10676 declare i64 @llvm.ctpop.i64(i64 <src>)
10677 declare i256 @llvm.ctpop.i256(i256 <src>)
10678 declare <2 x i32> @llvm.ctpop.v2i32(<2 x i32> <src>)
10683 The '``llvm.ctpop``' family of intrinsics counts the number of bits set
10689 The only argument is the value to be counted. The argument may be of any
10690 integer type, or a vector with integer elements. The return type must
10691 match the argument type.
10696 The '``llvm.ctpop``' intrinsic counts the 1's in a variable, or within
10697 each element of a vector.
10699 '``llvm.ctlz.*``' Intrinsic
10700 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10705 This is an overloaded intrinsic. You can use ``llvm.ctlz`` on any
10706 integer bit width, or any vector whose elements are integers. Not all
10707 targets support all bit widths or vector types, however.
10711 declare i8 @llvm.ctlz.i8 (i8 <src>, i1 <is_zero_undef>)
10712 declare i16 @llvm.ctlz.i16 (i16 <src>, i1 <is_zero_undef>)
10713 declare i32 @llvm.ctlz.i32 (i32 <src>, i1 <is_zero_undef>)
10714 declare i64 @llvm.ctlz.i64 (i64 <src>, i1 <is_zero_undef>)
10715 declare i256 @llvm.ctlz.i256(i256 <src>, i1 <is_zero_undef>)
10716 declase <2 x i32> @llvm.ctlz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
10721 The '``llvm.ctlz``' family of intrinsic functions counts the number of
10722 leading zeros in a variable.
10727 The first argument is the value to be counted. This argument may be of
10728 any integer type, or a vector with integer element type. The return
10729 type must match the first argument type.
10731 The second argument must be a constant and is a flag to indicate whether
10732 the intrinsic should ensure that a zero as the first argument produces a
10733 defined result. Historically some architectures did not provide a
10734 defined result for zero values as efficiently, and many algorithms are
10735 now predicated on avoiding zero-value inputs.
10740 The '``llvm.ctlz``' intrinsic counts the leading (most significant)
10741 zeros in a variable, or within each element of the vector. If
10742 ``src == 0`` then the result is the size in bits of the type of ``src``
10743 if ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
10744 ``llvm.ctlz(i32 2) = 30``.
10746 '``llvm.cttz.*``' Intrinsic
10747 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10752 This is an overloaded intrinsic. You can use ``llvm.cttz`` on any
10753 integer bit width, or any vector of integer elements. Not all targets
10754 support all bit widths or vector types, however.
10758 declare i8 @llvm.cttz.i8 (i8 <src>, i1 <is_zero_undef>)
10759 declare i16 @llvm.cttz.i16 (i16 <src>, i1 <is_zero_undef>)
10760 declare i32 @llvm.cttz.i32 (i32 <src>, i1 <is_zero_undef>)
10761 declare i64 @llvm.cttz.i64 (i64 <src>, i1 <is_zero_undef>)
10762 declare i256 @llvm.cttz.i256(i256 <src>, i1 <is_zero_undef>)
10763 declase <2 x i32> @llvm.cttz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
10768 The '``llvm.cttz``' family of intrinsic functions counts the number of
10774 The first argument is the value to be counted. This argument may be of
10775 any integer type, or a vector with integer element type. The return
10776 type must match the first argument type.
10778 The second argument must be a constant and is a flag to indicate whether
10779 the intrinsic should ensure that a zero as the first argument produces a
10780 defined result. Historically some architectures did not provide a
10781 defined result for zero values as efficiently, and many algorithms are
10782 now predicated on avoiding zero-value inputs.
10787 The '``llvm.cttz``' intrinsic counts the trailing (least significant)
10788 zeros in a variable, or within each element of a vector. If ``src == 0``
10789 then the result is the size in bits of the type of ``src`` if
10790 ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
10791 ``llvm.cttz(2) = 1``.
10795 Arithmetic with Overflow Intrinsics
10796 -----------------------------------
10798 LLVM provides intrinsics for some arithmetic with overflow operations.
10800 '``llvm.sadd.with.overflow.*``' Intrinsics
10801 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10806 This is an overloaded intrinsic. You can use ``llvm.sadd.with.overflow``
10807 on any integer bit width.
10811 declare {i16, i1} @llvm.sadd.with.overflow.i16(i16 %a, i16 %b)
10812 declare {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
10813 declare {i64, i1} @llvm.sadd.with.overflow.i64(i64 %a, i64 %b)
10818 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
10819 a signed addition of the two arguments, and indicate whether an overflow
10820 occurred during the signed summation.
10825 The arguments (%a and %b) and the first element of the result structure
10826 may be of integer types of any bit width, but they must have the same
10827 bit width. The second element of the result structure must be of type
10828 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
10834 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
10835 a signed addition of the two variables. They return a structure --- the
10836 first element of which is the signed summation, and the second element
10837 of which is a bit specifying if the signed summation resulted in an
10843 .. code-block:: llvm
10845 %res = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
10846 %sum = extractvalue {i32, i1} %res, 0
10847 %obit = extractvalue {i32, i1} %res, 1
10848 br i1 %obit, label %overflow, label %normal
10850 '``llvm.uadd.with.overflow.*``' Intrinsics
10851 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10856 This is an overloaded intrinsic. You can use ``llvm.uadd.with.overflow``
10857 on any integer bit width.
10861 declare {i16, i1} @llvm.uadd.with.overflow.i16(i16 %a, i16 %b)
10862 declare {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
10863 declare {i64, i1} @llvm.uadd.with.overflow.i64(i64 %a, i64 %b)
10868 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
10869 an unsigned addition of the two arguments, and indicate whether a carry
10870 occurred during the unsigned summation.
10875 The arguments (%a and %b) and the first element of the result structure
10876 may be of integer types of any bit width, but they must have the same
10877 bit width. The second element of the result structure must be of type
10878 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
10884 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
10885 an unsigned addition of the two arguments. They return a structure --- the
10886 first element of which is the sum, and the second element of which is a
10887 bit specifying if the unsigned summation resulted in a carry.
10892 .. code-block:: llvm
10894 %res = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
10895 %sum = extractvalue {i32, i1} %res, 0
10896 %obit = extractvalue {i32, i1} %res, 1
10897 br i1 %obit, label %carry, label %normal
10899 '``llvm.ssub.with.overflow.*``' Intrinsics
10900 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10905 This is an overloaded intrinsic. You can use ``llvm.ssub.with.overflow``
10906 on any integer bit width.
10910 declare {i16, i1} @llvm.ssub.with.overflow.i16(i16 %a, i16 %b)
10911 declare {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
10912 declare {i64, i1} @llvm.ssub.with.overflow.i64(i64 %a, i64 %b)
10917 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
10918 a signed subtraction of the two arguments, and indicate whether an
10919 overflow occurred during the signed subtraction.
10924 The arguments (%a and %b) and the first element of the result structure
10925 may be of integer types of any bit width, but they must have the same
10926 bit width. The second element of the result structure must be of type
10927 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
10933 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
10934 a signed subtraction of the two arguments. They return a structure --- the
10935 first element of which is the subtraction, and the second element of
10936 which is a bit specifying if the signed subtraction resulted in an
10942 .. code-block:: llvm
10944 %res = call {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
10945 %sum = extractvalue {i32, i1} %res, 0
10946 %obit = extractvalue {i32, i1} %res, 1
10947 br i1 %obit, label %overflow, label %normal
10949 '``llvm.usub.with.overflow.*``' Intrinsics
10950 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10955 This is an overloaded intrinsic. You can use ``llvm.usub.with.overflow``
10956 on any integer bit width.
10960 declare {i16, i1} @llvm.usub.with.overflow.i16(i16 %a, i16 %b)
10961 declare {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
10962 declare {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
10967 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
10968 an unsigned subtraction of the two arguments, and indicate whether an
10969 overflow occurred during the unsigned subtraction.
10974 The arguments (%a and %b) and the first element of the result structure
10975 may be of integer types of any bit width, but they must have the same
10976 bit width. The second element of the result structure must be of type
10977 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
10983 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
10984 an unsigned subtraction of the two arguments. They return a structure ---
10985 the first element of which is the subtraction, and the second element of
10986 which is a bit specifying if the unsigned subtraction resulted in an
10992 .. code-block:: llvm
10994 %res = call {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
10995 %sum = extractvalue {i32, i1} %res, 0
10996 %obit = extractvalue {i32, i1} %res, 1
10997 br i1 %obit, label %overflow, label %normal
10999 '``llvm.smul.with.overflow.*``' Intrinsics
11000 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11005 This is an overloaded intrinsic. You can use ``llvm.smul.with.overflow``
11006 on any integer bit width.
11010 declare {i16, i1} @llvm.smul.with.overflow.i16(i16 %a, i16 %b)
11011 declare {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
11012 declare {i64, i1} @llvm.smul.with.overflow.i64(i64 %a, i64 %b)
11017 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
11018 a signed multiplication of the two arguments, and indicate whether an
11019 overflow occurred during the signed multiplication.
11024 The arguments (%a and %b) and the first element of the result structure
11025 may be of integer types of any bit width, but they must have the same
11026 bit width. The second element of the result structure must be of type
11027 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
11033 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
11034 a signed multiplication of the two arguments. They return a structure ---
11035 the first element of which is the multiplication, and the second element
11036 of which is a bit specifying if the signed multiplication resulted in an
11042 .. code-block:: llvm
11044 %res = call {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
11045 %sum = extractvalue {i32, i1} %res, 0
11046 %obit = extractvalue {i32, i1} %res, 1
11047 br i1 %obit, label %overflow, label %normal
11049 '``llvm.umul.with.overflow.*``' Intrinsics
11050 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11055 This is an overloaded intrinsic. You can use ``llvm.umul.with.overflow``
11056 on any integer bit width.
11060 declare {i16, i1} @llvm.umul.with.overflow.i16(i16 %a, i16 %b)
11061 declare {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
11062 declare {i64, i1} @llvm.umul.with.overflow.i64(i64 %a, i64 %b)
11067 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
11068 a unsigned multiplication of the two arguments, and indicate whether an
11069 overflow occurred during the unsigned multiplication.
11074 The arguments (%a and %b) and the first element of the result structure
11075 may be of integer types of any bit width, but they must have the same
11076 bit width. The second element of the result structure must be of type
11077 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
11083 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
11084 an unsigned multiplication of the two arguments. They return a structure ---
11085 the first element of which is the multiplication, and the second
11086 element of which is a bit specifying if the unsigned multiplication
11087 resulted in an overflow.
11092 .. code-block:: llvm
11094 %res = call {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
11095 %sum = extractvalue {i32, i1} %res, 0
11096 %obit = extractvalue {i32, i1} %res, 1
11097 br i1 %obit, label %overflow, label %normal
11099 Specialised Arithmetic Intrinsics
11100 ---------------------------------
11102 '``llvm.canonicalize.*``' Intrinsic
11103 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11110 declare float @llvm.canonicalize.f32(float %a)
11111 declare double @llvm.canonicalize.f64(double %b)
11116 The '``llvm.canonicalize.*``' intrinsic returns the platform specific canonical
11117 encoding of a floating point number. This canonicalization is useful for
11118 implementing certain numeric primitives such as frexp. The canonical encoding is
11119 defined by IEEE-754-2008 to be:
11123 2.1.8 canonical encoding: The preferred encoding of a floating-point
11124 representation in a format. Applied to declets, significands of finite
11125 numbers, infinities, and NaNs, especially in decimal formats.
11127 This operation can also be considered equivalent to the IEEE-754-2008
11128 conversion of a floating-point value to the same format. NaNs are handled
11129 according to section 6.2.
11131 Examples of non-canonical encodings:
11133 - x87 pseudo denormals, pseudo NaNs, pseudo Infinity, Unnormals. These are
11134 converted to a canonical representation per hardware-specific protocol.
11135 - Many normal decimal floating point numbers have non-canonical alternative
11137 - Some machines, like GPUs or ARMv7 NEON, do not support subnormal values.
11138 These are treated as non-canonical encodings of zero and with be flushed to
11139 a zero of the same sign by this operation.
11141 Note that per IEEE-754-2008 6.2, systems that support signaling NaNs with
11142 default exception handling must signal an invalid exception, and produce a
11145 This function should always be implementable as multiplication by 1.0, provided
11146 that the compiler does not constant fold the operation. Likewise, division by
11147 1.0 and ``llvm.minnum(x, x)`` are possible implementations. Addition with
11148 -0.0 is also sufficient provided that the rounding mode is not -Infinity.
11150 ``@llvm.canonicalize`` must preserve the equality relation. That is:
11152 - ``(@llvm.canonicalize(x) == x)`` is equivalent to ``(x == x)``
11153 - ``(@llvm.canonicalize(x) == @llvm.canonicalize(y))`` is equivalent to
11156 Additionally, the sign of zero must be conserved:
11157 ``@llvm.canonicalize(-0.0) = -0.0`` and ``@llvm.canonicalize(+0.0) = +0.0``
11159 The payload bits of a NaN must be conserved, with two exceptions.
11160 First, environments which use only a single canonical representation of NaN
11161 must perform said canonicalization. Second, SNaNs must be quieted per the
11164 The canonicalization operation may be optimized away if:
11166 - The input is known to be canonical. For example, it was produced by a
11167 floating-point operation that is required by the standard to be canonical.
11168 - The result is consumed only by (or fused with) other floating-point
11169 operations. That is, the bits of the floating point value are not examined.
11171 '``llvm.fmuladd.*``' Intrinsic
11172 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11179 declare float @llvm.fmuladd.f32(float %a, float %b, float %c)
11180 declare double @llvm.fmuladd.f64(double %a, double %b, double %c)
11185 The '``llvm.fmuladd.*``' intrinsic functions represent multiply-add
11186 expressions that can be fused if the code generator determines that (a) the
11187 target instruction set has support for a fused operation, and (b) that the
11188 fused operation is more efficient than the equivalent, separate pair of mul
11189 and add instructions.
11194 The '``llvm.fmuladd.*``' intrinsics each take three arguments: two
11195 multiplicands, a and b, and an addend c.
11204 %0 = call float @llvm.fmuladd.f32(%a, %b, %c)
11206 is equivalent to the expression a \* b + c, except that rounding will
11207 not be performed between the multiplication and addition steps if the
11208 code generator fuses the operations. Fusion is not guaranteed, even if
11209 the target platform supports it. If a fused multiply-add is required the
11210 corresponding llvm.fma.\* intrinsic function should be used
11211 instead. This never sets errno, just as '``llvm.fma.*``'.
11216 .. code-block:: llvm
11218 %r2 = call float @llvm.fmuladd.f32(float %a, float %b, float %c) ; yields float:r2 = (a * b) + c
11221 '``llvm.uabsdiff.*``' and '``llvm.sabsdiff.*``' Intrinsics
11222 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11226 This is an overloaded intrinsic. The loaded data is a vector of any integer bit width.
11228 .. code-block:: llvm
11230 declare <4 x integer> @llvm.uabsdiff.v4i32(<4 x integer> %a, <4 x integer> %b)
11236 The ``llvm.uabsdiff`` intrinsic returns a vector result of the absolute difference
11237 of the two operands, treating them both as unsigned integers. The intermediate
11238 calculations are computed using infinitely precise unsigned arithmetic. The final
11239 result will be truncated to the given type.
11241 The ``llvm.sabsdiff`` intrinsic returns a vector result of the absolute difference of
11242 the two operands, treating them both as signed integers. If the result overflows, the
11243 behavior is undefined.
11247 These intrinsics are primarily used during the code generation stage of compilation.
11248 They are generated by compiler passes such as the Loop and SLP vectorizers. It is not
11249 recommended for users to create them manually.
11254 Both intrinsics take two integer of the same bitwidth.
11261 call <4 x i32> @llvm.uabsdiff.v4i32(<4 x i32> %a, <4 x i32> %b)
11265 %1 = zext <4 x i32> %a to <4 x i64>
11266 %2 = zext <4 x i32> %b to <4 x i64>
11267 %sub = sub <4 x i64> %1, %2
11268 %trunc = trunc <4 x i64> to <4 x i32>
11270 and the expression::
11272 call <4 x i32> @llvm.sabsdiff.v4i32(<4 x i32> %a, <4 x i32> %b)
11276 %sub = sub nsw <4 x i32> %a, %b
11277 %ispos = icmp sge <4 x i32> %sub, zeroinitializer
11278 %neg = sub nsw <4 x i32> zeroinitializer, %sub
11279 %1 = select <4 x i1> %ispos, <4 x i32> %sub, <4 x i32> %neg
11282 Half Precision Floating Point Intrinsics
11283 ----------------------------------------
11285 For most target platforms, half precision floating point is a
11286 storage-only format. This means that it is a dense encoding (in memory)
11287 but does not support computation in the format.
11289 This means that code must first load the half-precision floating point
11290 value as an i16, then convert it to float with
11291 :ref:`llvm.convert.from.fp16 <int_convert_from_fp16>`. Computation can
11292 then be performed on the float value (including extending to double
11293 etc). To store the value back to memory, it is first converted to float
11294 if needed, then converted to i16 with
11295 :ref:`llvm.convert.to.fp16 <int_convert_to_fp16>`, then storing as an
11298 .. _int_convert_to_fp16:
11300 '``llvm.convert.to.fp16``' Intrinsic
11301 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11308 declare i16 @llvm.convert.to.fp16.f32(float %a)
11309 declare i16 @llvm.convert.to.fp16.f64(double %a)
11314 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion from a
11315 conventional floating point type to half precision floating point format.
11320 The intrinsic function contains single argument - the value to be
11326 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion from a
11327 conventional floating point format to half precision floating point format. The
11328 return value is an ``i16`` which contains the converted number.
11333 .. code-block:: llvm
11335 %res = call i16 @llvm.convert.to.fp16.f32(float %a)
11336 store i16 %res, i16* @x, align 2
11338 .. _int_convert_from_fp16:
11340 '``llvm.convert.from.fp16``' Intrinsic
11341 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11348 declare float @llvm.convert.from.fp16.f32(i16 %a)
11349 declare double @llvm.convert.from.fp16.f64(i16 %a)
11354 The '``llvm.convert.from.fp16``' intrinsic function performs a
11355 conversion from half precision floating point format to single precision
11356 floating point format.
11361 The intrinsic function contains single argument - the value to be
11367 The '``llvm.convert.from.fp16``' intrinsic function performs a
11368 conversion from half single precision floating point format to single
11369 precision floating point format. The input half-float value is
11370 represented by an ``i16`` value.
11375 .. code-block:: llvm
11377 %a = load i16, i16* @x, align 2
11378 %res = call float @llvm.convert.from.fp16(i16 %a)
11380 .. _dbg_intrinsics:
11382 Debugger Intrinsics
11383 -------------------
11385 The LLVM debugger intrinsics (which all start with ``llvm.dbg.``
11386 prefix), are described in the `LLVM Source Level
11387 Debugging <SourceLevelDebugging.html#format_common_intrinsics>`_
11390 Exception Handling Intrinsics
11391 -----------------------------
11393 The LLVM exception handling intrinsics (which all start with
11394 ``llvm.eh.`` prefix), are described in the `LLVM Exception
11395 Handling <ExceptionHandling.html#format_common_intrinsics>`_ document.
11397 .. _int_trampoline:
11399 Trampoline Intrinsics
11400 ---------------------
11402 These intrinsics make it possible to excise one parameter, marked with
11403 the :ref:`nest <nest>` attribute, from a function. The result is a
11404 callable function pointer lacking the nest parameter - the caller does
11405 not need to provide a value for it. Instead, the value to use is stored
11406 in advance in a "trampoline", a block of memory usually allocated on the
11407 stack, which also contains code to splice the nest value into the
11408 argument list. This is used to implement the GCC nested function address
11411 For example, if the function is ``i32 f(i8* nest %c, i32 %x, i32 %y)``
11412 then the resulting function pointer has signature ``i32 (i32, i32)*``.
11413 It can be created as follows:
11415 .. code-block:: llvm
11417 %tramp = alloca [10 x i8], align 4 ; size and alignment only correct for X86
11418 %tramp1 = getelementptr [10 x i8], [10 x i8]* %tramp, i32 0, i32 0
11419 call i8* @llvm.init.trampoline(i8* %tramp1, i8* bitcast (i32 (i8*, i32, i32)* @f to i8*), i8* %nval)
11420 %p = call i8* @llvm.adjust.trampoline(i8* %tramp1)
11421 %fp = bitcast i8* %p to i32 (i32, i32)*
11423 The call ``%val = call i32 %fp(i32 %x, i32 %y)`` is then equivalent to
11424 ``%val = call i32 %f(i8* %nval, i32 %x, i32 %y)``.
11428 '``llvm.init.trampoline``' Intrinsic
11429 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11436 declare void @llvm.init.trampoline(i8* <tramp>, i8* <func>, i8* <nval>)
11441 This fills the memory pointed to by ``tramp`` with executable code,
11442 turning it into a trampoline.
11447 The ``llvm.init.trampoline`` intrinsic takes three arguments, all
11448 pointers. The ``tramp`` argument must point to a sufficiently large and
11449 sufficiently aligned block of memory; this memory is written to by the
11450 intrinsic. Note that the size and the alignment are target-specific -
11451 LLVM currently provides no portable way of determining them, so a
11452 front-end that generates this intrinsic needs to have some
11453 target-specific knowledge. The ``func`` argument must hold a function
11454 bitcast to an ``i8*``.
11459 The block of memory pointed to by ``tramp`` is filled with target
11460 dependent code, turning it into a function. Then ``tramp`` needs to be
11461 passed to :ref:`llvm.adjust.trampoline <int_at>` to get a pointer which can
11462 be :ref:`bitcast (to a new function) and called <int_trampoline>`. The new
11463 function's signature is the same as that of ``func`` with any arguments
11464 marked with the ``nest`` attribute removed. At most one such ``nest``
11465 argument is allowed, and it must be of pointer type. Calling the new
11466 function is equivalent to calling ``func`` with the same argument list,
11467 but with ``nval`` used for the missing ``nest`` argument. If, after
11468 calling ``llvm.init.trampoline``, the memory pointed to by ``tramp`` is
11469 modified, then the effect of any later call to the returned function
11470 pointer is undefined.
11474 '``llvm.adjust.trampoline``' Intrinsic
11475 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11482 declare i8* @llvm.adjust.trampoline(i8* <tramp>)
11487 This performs any required machine-specific adjustment to the address of
11488 a trampoline (passed as ``tramp``).
11493 ``tramp`` must point to a block of memory which already has trampoline
11494 code filled in by a previous call to
11495 :ref:`llvm.init.trampoline <int_it>`.
11500 On some architectures the address of the code to be executed needs to be
11501 different than the address where the trampoline is actually stored. This
11502 intrinsic returns the executable address corresponding to ``tramp``
11503 after performing the required machine specific adjustments. The pointer
11504 returned can then be :ref:`bitcast and executed <int_trampoline>`.
11506 .. _int_mload_mstore:
11508 Masked Vector Load and Store Intrinsics
11509 ---------------------------------------
11511 LLVM provides intrinsics for predicated vector load and store operations. The predicate is specified by a mask operand, which holds one bit per vector element, switching the associated vector lane on or off. The memory addresses corresponding to the "off" lanes are not accessed. When all bits of the mask are on, the intrinsic is identical to a regular vector load or store. When all bits are off, no memory is accessed.
11515 '``llvm.masked.load.*``' Intrinsics
11516 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11520 This is an overloaded intrinsic. The loaded data is a vector of any integer, floating point or pointer data type.
11524 declare <16 x float> @llvm.masked.load.v16f32 (<16 x float>* <ptr>, i32 <alignment>, <16 x i1> <mask>, <16 x float> <passthru>)
11525 declare <2 x double> @llvm.masked.load.v2f64 (<2 x double>* <ptr>, i32 <alignment>, <2 x i1> <mask>, <2 x double> <passthru>)
11526 ;; The data is a vector of pointers to double
11527 declare <8 x double*> @llvm.masked.load.v8p0f64 (<8 x double*>* <ptr>, i32 <alignment>, <8 x i1> <mask>, <8 x double*> <passthru>)
11528 ;; The data is a vector of function pointers
11529 declare <8 x i32 ()*> @llvm.masked.load.v8p0f_i32f (<8 x i32 ()*>* <ptr>, i32 <alignment>, <8 x i1> <mask>, <8 x i32 ()*> <passthru>)
11534 Reads a vector from memory according to the provided mask. The mask holds a bit for each vector lane, and is used to prevent memory accesses to the masked-off lanes. The masked-off lanes in the result vector are taken from the corresponding lanes of the '``passthru``' operand.
11540 The first operand is the base pointer for the load. The second operand is the alignment of the source location. It must be a constant integer value. The third operand, mask, is a vector of boolean values with the same number of elements as the return type. The fourth is a pass-through value that is used to fill the masked-off lanes of the result. The return type, underlying type of the base pointer and the type of the '``passthru``' operand are the same vector types.
11546 The '``llvm.masked.load``' intrinsic is designed for conditional reading of selected vector elements in a single IR operation. It is useful for targets that support vector masked loads and allows vectorizing predicated basic blocks on these targets. Other targets may support this intrinsic differently, for example by lowering it into a sequence of branches that guard scalar load operations.
11547 The result of this operation is equivalent to a regular vector load instruction followed by a 'select' between the loaded and the passthru values, predicated on the same mask. However, using this intrinsic prevents exceptions on memory access to masked-off lanes.
11552 %res = call <16 x float> @llvm.masked.load.v16f32 (<16 x float>* %ptr, i32 4, <16 x i1>%mask, <16 x float> %passthru)
11554 ;; The result of the two following instructions is identical aside from potential memory access exception
11555 %loadlal = load <16 x float>, <16 x float>* %ptr, align 4
11556 %res = select <16 x i1> %mask, <16 x float> %loadlal, <16 x float> %passthru
11560 '``llvm.masked.store.*``' Intrinsics
11561 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11565 This is an overloaded intrinsic. The data stored in memory is a vector of any integer, floating point or pointer data type.
11569 declare void @llvm.masked.store.v8i32 (<8 x i32> <value>, <8 x i32>* <ptr>, i32 <alignment>, <8 x i1> <mask>)
11570 declare void @llvm.masked.store.v16f32 (<16 x float> <value>, <16 x float>* <ptr>, i32 <alignment>, <16 x i1> <mask>)
11571 ;; The data is a vector of pointers to double
11572 declare void @llvm.masked.store.v8p0f64 (<8 x double*> <value>, <8 x double*>* <ptr>, i32 <alignment>, <8 x i1> <mask>)
11573 ;; The data is a vector of function pointers
11574 declare void @llvm.masked.store.v4p0f_i32f (<4 x i32 ()*> <value>, <4 x i32 ()*>* <ptr>, i32 <alignment>, <4 x i1> <mask>)
11579 Writes a vector to memory according to the provided mask. The mask holds a bit for each vector lane, and is used to prevent memory accesses to the masked-off lanes.
11584 The first operand is the vector value to be written to memory. The second operand is the base pointer for the store, it has the same underlying type as the value operand. The third operand is the alignment of the destination location. The fourth operand, mask, is a vector of boolean values. The types of the mask and the value operand must have the same number of vector elements.
11590 The '``llvm.masked.store``' intrinsics is designed for conditional writing of selected vector elements in a single IR operation. It is useful for targets that support vector masked store and allows vectorizing predicated basic blocks on these targets. Other targets may support this intrinsic differently, for example by lowering it into a sequence of branches that guard scalar store operations.
11591 The result of this operation is equivalent to a load-modify-store sequence. However, using this intrinsic prevents exceptions and data races on memory access to masked-off lanes.
11595 call void @llvm.masked.store.v16f32(<16 x float> %value, <16 x float>* %ptr, i32 4, <16 x i1> %mask)
11597 ;; The result of the following instructions is identical aside from potential data races and memory access exceptions
11598 %oldval = load <16 x float>, <16 x float>* %ptr, align 4
11599 %res = select <16 x i1> %mask, <16 x float> %value, <16 x float> %oldval
11600 store <16 x float> %res, <16 x float>* %ptr, align 4
11603 Masked Vector Gather and Scatter Intrinsics
11604 -------------------------------------------
11606 LLVM provides intrinsics for vector gather and scatter operations. They are similar to :ref:`Masked Vector Load and Store <int_mload_mstore>`, except they are designed for arbitrary memory accesses, rather than sequential memory accesses. Gather and scatter also employ a mask operand, which holds one bit per vector element, switching the associated vector lane on or off. The memory addresses corresponding to the "off" lanes are not accessed. When all bits are off, no memory is accessed.
11610 '``llvm.masked.gather.*``' Intrinsics
11611 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11615 This is an overloaded intrinsic. The loaded data are multiple scalar values of any integer, floating point or pointer data type gathered together into one vector.
11619 declare <16 x float> @llvm.masked.gather.v16f32 (<16 x float*> <ptrs>, i32 <alignment>, <16 x i1> <mask>, <16 x float> <passthru>)
11620 declare <2 x double> @llvm.masked.gather.v2f64 (<2 x double*> <ptrs>, i32 <alignment>, <2 x i1> <mask>, <2 x double> <passthru>)
11621 declare <8 x float*> @llvm.masked.gather.v8p0f32 (<8 x float**> <ptrs>, i32 <alignment>, <8 x i1> <mask>, <8 x float*> <passthru>)
11626 Reads scalar values from arbitrary memory locations and gathers them into one vector. The memory locations are provided in the vector of pointers '``ptrs``'. The memory is accessed according to the provided mask. The mask holds a bit for each vector lane, and is used to prevent memory accesses to the masked-off lanes. The masked-off lanes in the result vector are taken from the corresponding lanes of the '``passthru``' operand.
11632 The first operand is a vector of pointers which holds all memory addresses to read. The second operand is an alignment of the source addresses. It must be a constant integer value. The third operand, mask, is a vector of boolean values with the same number of elements as the return type. The fourth is a pass-through value that is used to fill the masked-off lanes of the result. The return type, underlying type of the vector of pointers and the type of the '``passthru``' operand are the same vector types.
11638 The '``llvm.masked.gather``' intrinsic is designed for conditional reading of multiple scalar values from arbitrary memory locations in a single IR operation. It is useful for targets that support vector masked gathers and allows vectorizing basic blocks with data and control divergence. Other targets may support this intrinsic differently, for example by lowering it into a sequence of scalar load operations.
11639 The semantics of this operation are equivalent to a sequence of conditional scalar loads with subsequent gathering all loaded values into a single vector. The mask restricts memory access to certain lanes and facilitates vectorization of predicated basic blocks.
11644 %res = call <4 x double> @llvm.masked.gather.v4f64 (<4 x double*> %ptrs, i32 8, <4 x i1>%mask, <4 x double> <true, true, true, true>)
11646 ;; The gather with all-true mask is equivalent to the following instruction sequence
11647 %ptr0 = extractelement <4 x double*> %ptrs, i32 0
11648 %ptr1 = extractelement <4 x double*> %ptrs, i32 1
11649 %ptr2 = extractelement <4 x double*> %ptrs, i32 2
11650 %ptr3 = extractelement <4 x double*> %ptrs, i32 3
11652 %val0 = load double, double* %ptr0, align 8
11653 %val1 = load double, double* %ptr1, align 8
11654 %val2 = load double, double* %ptr2, align 8
11655 %val3 = load double, double* %ptr3, align 8
11657 %vec0 = insertelement <4 x double>undef, %val0, 0
11658 %vec01 = insertelement <4 x double>%vec0, %val1, 1
11659 %vec012 = insertelement <4 x double>%vec01, %val2, 2
11660 %vec0123 = insertelement <4 x double>%vec012, %val3, 3
11664 '``llvm.masked.scatter.*``' Intrinsics
11665 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11669 This is an overloaded intrinsic. The data stored in memory is a vector of any integer, floating point or pointer data type. Each vector element is stored in an arbitrary memory address. Scatter with overlapping addresses is guaranteed to be ordered from least-significant to most-significant element.
11673 declare void @llvm.masked.scatter.v8i32 (<8 x i32> <value>, <8 x i32*> <ptrs>, i32 <alignment>, <8 x i1> <mask>)
11674 declare void @llvm.masked.scatter.v16f32 (<16 x float> <value>, <16 x float*> <ptrs>, i32 <alignment>, <16 x i1> <mask>)
11675 declare void @llvm.masked.scatter.v4p0f64 (<4 x double*> <value>, <4 x double**> <ptrs>, i32 <alignment>, <4 x i1> <mask>)
11680 Writes each element from the value vector to the corresponding memory address. The memory addresses are represented as a vector of pointers. Writing is done according to the provided mask. The mask holds a bit for each vector lane, and is used to prevent memory accesses to the masked-off lanes.
11685 The first operand is a vector value to be written to memory. The second operand is a vector of pointers, pointing to where the value elements should be stored. It has the same underlying type as the value operand. The third operand is an alignment of the destination addresses. The fourth operand, mask, is a vector of boolean values. The types of the mask and the value operand must have the same number of vector elements.
11691 The '``llvm.masked.scatter``' intrinsics is designed for writing selected vector elements to arbitrary memory addresses in a single IR operation. The operation may be conditional, when not all bits in the mask are switched on. It is useful for targets that support vector masked scatter and allows vectorizing basic blocks with data and control divergence. Other targets may support this intrinsic differently, for example by lowering it into a sequence of branches that guard scalar store operations.
11695 ;; This instruction unconditionaly stores data vector in multiple addresses
11696 call @llvm.masked.scatter.v8i32 (<8 x i32> %value, <8 x i32*> %ptrs, i32 4, <8 x i1> <true, true, .. true>)
11698 ;; It is equivalent to a list of scalar stores
11699 %val0 = extractelement <8 x i32> %value, i32 0
11700 %val1 = extractelement <8 x i32> %value, i32 1
11702 %val7 = extractelement <8 x i32> %value, i32 7
11703 %ptr0 = extractelement <8 x i32*> %ptrs, i32 0
11704 %ptr1 = extractelement <8 x i32*> %ptrs, i32 1
11706 %ptr7 = extractelement <8 x i32*> %ptrs, i32 7
11707 ;; Note: the order of the following stores is important when they overlap:
11708 store i32 %val0, i32* %ptr0, align 4
11709 store i32 %val1, i32* %ptr1, align 4
11711 store i32 %val7, i32* %ptr7, align 4
11717 This class of intrinsics provides information about the lifetime of
11718 memory objects and ranges where variables are immutable.
11722 '``llvm.lifetime.start``' Intrinsic
11723 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11730 declare void @llvm.lifetime.start(i64 <size>, i8* nocapture <ptr>)
11735 The '``llvm.lifetime.start``' intrinsic specifies the start of a memory
11741 The first argument is a constant integer representing the size of the
11742 object, or -1 if it is variable sized. The second argument is a pointer
11748 This intrinsic indicates that before this point in the code, the value
11749 of the memory pointed to by ``ptr`` is dead. This means that it is known
11750 to never be used and has an undefined value. A load from the pointer
11751 that precedes this intrinsic can be replaced with ``'undef'``.
11755 '``llvm.lifetime.end``' Intrinsic
11756 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11763 declare void @llvm.lifetime.end(i64 <size>, i8* nocapture <ptr>)
11768 The '``llvm.lifetime.end``' intrinsic specifies the end of a memory
11774 The first argument is a constant integer representing the size of the
11775 object, or -1 if it is variable sized. The second argument is a pointer
11781 This intrinsic indicates that after this point in the code, the value of
11782 the memory pointed to by ``ptr`` is dead. This means that it is known to
11783 never be used and has an undefined value. Any stores into the memory
11784 object following this intrinsic may be removed as dead.
11786 '``llvm.invariant.start``' Intrinsic
11787 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11794 declare {}* @llvm.invariant.start(i64 <size>, i8* nocapture <ptr>)
11799 The '``llvm.invariant.start``' intrinsic specifies that the contents of
11800 a memory object will not change.
11805 The first argument is a constant integer representing the size of the
11806 object, or -1 if it is variable sized. The second argument is a pointer
11812 This intrinsic indicates that until an ``llvm.invariant.end`` that uses
11813 the return value, the referenced memory location is constant and
11816 '``llvm.invariant.end``' Intrinsic
11817 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11824 declare void @llvm.invariant.end({}* <start>, i64 <size>, i8* nocapture <ptr>)
11829 The '``llvm.invariant.end``' intrinsic specifies that the contents of a
11830 memory object are mutable.
11835 The first argument is the matching ``llvm.invariant.start`` intrinsic.
11836 The second argument is a constant integer representing the size of the
11837 object, or -1 if it is variable sized and the third argument is a
11838 pointer to the object.
11843 This intrinsic indicates that the memory is mutable again.
11845 '``llvm.invariant.group.barrier``' Intrinsic
11846 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11853 declare i8* @llvm.invariant.group.barrier(i8* <ptr>)
11858 The '``llvm.invariant.group.barrier``' intrinsic can be used when an invariant
11859 established by invariant.group metadata no longer holds, to obtain a new pointer
11860 value that does not carry the invariant information.
11866 The ``llvm.invariant.group.barrier`` takes only one argument, which is
11867 the pointer to the memory for which the ``invariant.group`` no longer holds.
11872 Returns another pointer that aliases its argument but which is considered different
11873 for the purposes of ``load``/``store`` ``invariant.group`` metadata.
11878 This class of intrinsics is designed to be generic and has no specific
11881 '``llvm.var.annotation``' Intrinsic
11882 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11889 declare void @llvm.var.annotation(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
11894 The '``llvm.var.annotation``' intrinsic.
11899 The first argument is a pointer to a value, the second is a pointer to a
11900 global string, the third is a pointer to a global string which is the
11901 source file name, and the last argument is the line number.
11906 This intrinsic allows annotation of local variables with arbitrary
11907 strings. This can be useful for special purpose optimizations that want
11908 to look for these annotations. These have no other defined use; they are
11909 ignored by code generation and optimization.
11911 '``llvm.ptr.annotation.*``' Intrinsic
11912 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11917 This is an overloaded intrinsic. You can use '``llvm.ptr.annotation``' on a
11918 pointer to an integer of any width. *NOTE* you must specify an address space for
11919 the pointer. The identifier for the default address space is the integer
11924 declare i8* @llvm.ptr.annotation.p<address space>i8(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
11925 declare i16* @llvm.ptr.annotation.p<address space>i16(i16* <val>, i8* <str>, i8* <str>, i32 <int>)
11926 declare i32* @llvm.ptr.annotation.p<address space>i32(i32* <val>, i8* <str>, i8* <str>, i32 <int>)
11927 declare i64* @llvm.ptr.annotation.p<address space>i64(i64* <val>, i8* <str>, i8* <str>, i32 <int>)
11928 declare i256* @llvm.ptr.annotation.p<address space>i256(i256* <val>, i8* <str>, i8* <str>, i32 <int>)
11933 The '``llvm.ptr.annotation``' intrinsic.
11938 The first argument is a pointer to an integer value of arbitrary bitwidth
11939 (result of some expression), the second is a pointer to a global string, the
11940 third is a pointer to a global string which is the source file name, and the
11941 last argument is the line number. It returns the value of the first argument.
11946 This intrinsic allows annotation of a pointer to an integer with arbitrary
11947 strings. This can be useful for special purpose optimizations that want to look
11948 for these annotations. These have no other defined use; they are ignored by code
11949 generation and optimization.
11951 '``llvm.annotation.*``' Intrinsic
11952 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11957 This is an overloaded intrinsic. You can use '``llvm.annotation``' on
11958 any integer bit width.
11962 declare i8 @llvm.annotation.i8(i8 <val>, i8* <str>, i8* <str>, i32 <int>)
11963 declare i16 @llvm.annotation.i16(i16 <val>, i8* <str>, i8* <str>, i32 <int>)
11964 declare i32 @llvm.annotation.i32(i32 <val>, i8* <str>, i8* <str>, i32 <int>)
11965 declare i64 @llvm.annotation.i64(i64 <val>, i8* <str>, i8* <str>, i32 <int>)
11966 declare i256 @llvm.annotation.i256(i256 <val>, i8* <str>, i8* <str>, i32 <int>)
11971 The '``llvm.annotation``' intrinsic.
11976 The first argument is an integer value (result of some expression), the
11977 second is a pointer to a global string, the third is a pointer to a
11978 global string which is the source file name, and the last argument is
11979 the line number. It returns the value of the first argument.
11984 This intrinsic allows annotations to be put on arbitrary expressions
11985 with arbitrary strings. This can be useful for special purpose
11986 optimizations that want to look for these annotations. These have no
11987 other defined use; they are ignored by code generation and optimization.
11989 '``llvm.trap``' Intrinsic
11990 ^^^^^^^^^^^^^^^^^^^^^^^^^
11997 declare void @llvm.trap() noreturn nounwind
12002 The '``llvm.trap``' intrinsic.
12012 This intrinsic is lowered to the target dependent trap instruction. If
12013 the target does not have a trap instruction, this intrinsic will be
12014 lowered to a call of the ``abort()`` function.
12016 '``llvm.debugtrap``' Intrinsic
12017 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12024 declare void @llvm.debugtrap() nounwind
12029 The '``llvm.debugtrap``' intrinsic.
12039 This intrinsic is lowered to code which is intended to cause an
12040 execution trap with the intention of requesting the attention of a
12043 '``llvm.stackprotector``' Intrinsic
12044 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12051 declare void @llvm.stackprotector(i8* <guard>, i8** <slot>)
12056 The ``llvm.stackprotector`` intrinsic takes the ``guard`` and stores it
12057 onto the stack at ``slot``. The stack slot is adjusted to ensure that it
12058 is placed on the stack before local variables.
12063 The ``llvm.stackprotector`` intrinsic requires two pointer arguments.
12064 The first argument is the value loaded from the stack guard
12065 ``@__stack_chk_guard``. The second variable is an ``alloca`` that has
12066 enough space to hold the value of the guard.
12071 This intrinsic causes the prologue/epilogue inserter to force the position of
12072 the ``AllocaInst`` stack slot to be before local variables on the stack. This is
12073 to ensure that if a local variable on the stack is overwritten, it will destroy
12074 the value of the guard. When the function exits, the guard on the stack is
12075 checked against the original guard by ``llvm.stackprotectorcheck``. If they are
12076 different, then ``llvm.stackprotectorcheck`` causes the program to abort by
12077 calling the ``__stack_chk_fail()`` function.
12079 '``llvm.stackprotectorcheck``' Intrinsic
12080 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12087 declare void @llvm.stackprotectorcheck(i8** <guard>)
12092 The ``llvm.stackprotectorcheck`` intrinsic compares ``guard`` against an already
12093 created stack protector and if they are not equal calls the
12094 ``__stack_chk_fail()`` function.
12099 The ``llvm.stackprotectorcheck`` intrinsic requires one pointer argument, the
12100 the variable ``@__stack_chk_guard``.
12105 This intrinsic is provided to perform the stack protector check by comparing
12106 ``guard`` with the stack slot created by ``llvm.stackprotector`` and if the
12107 values do not match call the ``__stack_chk_fail()`` function.
12109 The reason to provide this as an IR level intrinsic instead of implementing it
12110 via other IR operations is that in order to perform this operation at the IR
12111 level without an intrinsic, one would need to create additional basic blocks to
12112 handle the success/failure cases. This makes it difficult to stop the stack
12113 protector check from disrupting sibling tail calls in Codegen. With this
12114 intrinsic, we are able to generate the stack protector basic blocks late in
12115 codegen after the tail call decision has occurred.
12117 '``llvm.objectsize``' Intrinsic
12118 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12125 declare i32 @llvm.objectsize.i32(i8* <object>, i1 <min>)
12126 declare i64 @llvm.objectsize.i64(i8* <object>, i1 <min>)
12131 The ``llvm.objectsize`` intrinsic is designed to provide information to
12132 the optimizers to determine at compile time whether a) an operation
12133 (like memcpy) will overflow a buffer that corresponds to an object, or
12134 b) that a runtime check for overflow isn't necessary. An object in this
12135 context means an allocation of a specific class, structure, array, or
12141 The ``llvm.objectsize`` intrinsic takes two arguments. The first
12142 argument is a pointer to or into the ``object``. The second argument is
12143 a boolean and determines whether ``llvm.objectsize`` returns 0 (if true)
12144 or -1 (if false) when the object size is unknown. The second argument
12145 only accepts constants.
12150 The ``llvm.objectsize`` intrinsic is lowered to a constant representing
12151 the size of the object concerned. If the size cannot be determined at
12152 compile time, ``llvm.objectsize`` returns ``i32/i64 -1 or 0`` (depending
12153 on the ``min`` argument).
12155 '``llvm.expect``' Intrinsic
12156 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
12161 This is an overloaded intrinsic. You can use ``llvm.expect`` on any
12166 declare i1 @llvm.expect.i1(i1 <val>, i1 <expected_val>)
12167 declare i32 @llvm.expect.i32(i32 <val>, i32 <expected_val>)
12168 declare i64 @llvm.expect.i64(i64 <val>, i64 <expected_val>)
12173 The ``llvm.expect`` intrinsic provides information about expected (the
12174 most probable) value of ``val``, which can be used by optimizers.
12179 The ``llvm.expect`` intrinsic takes two arguments. The first argument is
12180 a value. The second argument is an expected value, this needs to be a
12181 constant value, variables are not allowed.
12186 This intrinsic is lowered to the ``val``.
12190 '``llvm.assume``' Intrinsic
12191 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12198 declare void @llvm.assume(i1 %cond)
12203 The ``llvm.assume`` allows the optimizer to assume that the provided
12204 condition is true. This information can then be used in simplifying other parts
12210 The condition which the optimizer may assume is always true.
12215 The intrinsic allows the optimizer to assume that the provided condition is
12216 always true whenever the control flow reaches the intrinsic call. No code is
12217 generated for this intrinsic, and instructions that contribute only to the
12218 provided condition are not used for code generation. If the condition is
12219 violated during execution, the behavior is undefined.
12221 Note that the optimizer might limit the transformations performed on values
12222 used by the ``llvm.assume`` intrinsic in order to preserve the instructions
12223 only used to form the intrinsic's input argument. This might prove undesirable
12224 if the extra information provided by the ``llvm.assume`` intrinsic does not cause
12225 sufficient overall improvement in code quality. For this reason,
12226 ``llvm.assume`` should not be used to document basic mathematical invariants
12227 that the optimizer can otherwise deduce or facts that are of little use to the
12232 '``llvm.bitset.test``' Intrinsic
12233 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12240 declare i1 @llvm.bitset.test(i8* %ptr, metadata %bitset) nounwind readnone
12246 The first argument is a pointer to be tested. The second argument is a
12247 metadata object representing an identifier for a :doc:`bitset <BitSets>`.
12252 The ``llvm.bitset.test`` intrinsic tests whether the given pointer is a
12253 member of the given bitset.
12255 '``llvm.donothing``' Intrinsic
12256 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12263 declare void @llvm.donothing() nounwind readnone
12268 The ``llvm.donothing`` intrinsic doesn't perform any operation. It's one of only
12269 two intrinsics (besides ``llvm.experimental.patchpoint``) that can be called
12270 with an invoke instruction.
12280 This intrinsic does nothing, and it's removed by optimizers and ignored
12283 Stack Map Intrinsics
12284 --------------------
12286 LLVM provides experimental intrinsics to support runtime patching
12287 mechanisms commonly desired in dynamic language JITs. These intrinsics
12288 are described in :doc:`StackMaps`.