xref: /llvm-project-15.0.7/llvm/docs/LangRef.rst (revision cdfcbbd3)
1==============================
2LLVM Language Reference Manual
3==============================
4
5.. contents::
6   :local:
7   :depth: 4
8
9Abstract
10========
11
12This document is a reference manual for the LLVM assembly language. LLVM
13is a Static Single Assignment (SSA) based representation that provides
14type safety, low-level operations, flexibility, and the capability of
15representing 'all' high-level languages cleanly. It is the common code
16representation used throughout all phases of the LLVM compilation
17strategy.
18
19Introduction
20============
21
22The LLVM code representation is designed to be used in three different
23forms: 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
25readable assembly language representation. This allows LLVM to provide a
26powerful intermediate representation for efficient compiler
27transformations and analysis, while providing a natural means to debug
28and visualize the transformations. The three different forms of LLVM are
29all equivalent. This document describes the human readable
30representation and notation.
31
32The LLVM representation aims to be light-weight and low-level while
33being expressive, typed, and extensible at the same time. It aims to be
34a "universal IR" of sorts, by being at a low enough level that
35high-level ideas may be cleanly mapped to it (similar to how
36microprocessors are "universal IR's", allowing many source languages to
37be mapped to them). By providing type information, LLVM can be used as
38the target of optimizations: for example, through pointer analysis, it
39can be proven that a C automatic variable is never accessed outside of
40the current function, allowing it to be promoted to a simple SSA value
41instead of a memory location.
42
43.. _wellformed:
44
45Well-Formedness
46---------------
47
48It is important to note that this document describes 'well formed' LLVM
49assembly language. There is a difference between what the parser accepts
50and what is considered 'well formed'. For example, the following
51instruction is syntactically okay, but not well formed:
52
53.. code-block:: llvm
54
55    %x = add i32 1, %x
56
57because the definition of ``%x`` does not dominate all of its uses. The
58LLVM infrastructure provides a verification pass that may be used to
59verify that an LLVM module is well formed. This pass is automatically
60run by the parser after parsing input assembly and by the optimizer
61before it outputs bitcode. The violations pointed out by the verifier
62pass indicate bugs in transformation passes or input to the parser.
63
64.. _identifiers:
65
66Identifiers
67===========
68
69LLVM identifiers come in two basic types: global and local. Global
70identifiers (functions, global variables) begin with the ``'@'``
71character. Local identifiers (register names, types) begin with the
72``'%'`` character. Additionally, there are three different formats for
73identifiers, for different purposes:
74
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 values 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.
87
88LLVM requires that values start with a prefix for two reasons: Compilers
89don't need to worry about name clashes with reserved words, and the set
90of reserved words may be expanded in the future without penalty.
91Additionally, unnamed identifiers allow a compiler to quickly come up
92with a temporary variable without having to avoid symbol table
93conflicts.
94
95Reserved words in LLVM are very similar to reserved words in other
96languages. 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
99with variable names, because none of them start with a prefix character
100(``'%'`` or ``'@'``).
101
102Here is an example of LLVM code to multiply the integer variable
103'``%X``' by 8:
104
105The easy way:
106
107.. code-block:: llvm
108
109    %result = mul i32 %X, 8
110
111After strength reduction:
112
113.. code-block:: llvm
114
115    %result = shl i32 %X, 3
116
117And the hard way:
118
119.. code-block:: llvm
120
121    %0 = add i32 %X, %X           ; yields i32:%0
122    %1 = add i32 %0, %0           ; yields i32:%1
123    %result = add i32 %1, %1
124
125This last way of multiplying ``%X`` by 8 illustrates several important
126lexical features of LLVM:
127
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.
136
137It also shows a convention that we follow in this document. When
138demonstrating instructions, we will follow an instruction with a comment
139that defines the type and name of value produced.
140
141High Level Structure
142====================
143
144Module Structure
145----------------
146
147LLVM programs are composed of ``Module``'s, each of which is a
148translation unit of the input programs. Each module consists of
149functions, global variables, and symbol table entries. Modules may be
150combined together with the LLVM linker, which merges function (and
151global variable) definitions, resolves forward declarations, and merges
152symbol table entries. Here is an example of the "hello world" module:
153
154.. code-block:: llvm
155
156    ; Declare the string constant as a global constant.
157    @.str = private unnamed_addr constant [13 x i8] c"hello world\0A\00"
158
159    ; External declaration of the puts function
160    declare i32 @puts(i8* nocapture) nounwind
161
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
166
167      ; Call puts function to write out the string to stdout.
168      call i32 @puts(i8* %cast210)
169      ret i32 0
170    }
171
172    ; Named metadata
173    !0 = !{i32 42, null, !"string"}
174    !foo = !{!0}
175
176This 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``".
180
181In general, a module is made up of a list of global values (where both
182functions and global variables are global values). Global values are
183represented by a pointer to a memory location (in this case, a pointer
184to an array of char, and a pointer to a function), and have one of the
185following :ref:`linkage types <linkage>`.
186
187.. _linkage:
188
189Linkage Types
190-------------
191
192All Global Variables and Functions have one of the following types of
193linkage:
194
195``private``
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 a 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.
202``internal``
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 into
208    the object file corresponding to the LLVM module. From the linker's
209    perspective, an ``available_externally`` global is equivalent to
210    an external declaration. They exist to allow inlining and other
211    optimizations to take place given knowledge of the definition of the
212    global, which is known to be somewhere outside the module. Globals
213    with ``available_externally`` linkage are allowed to be discarded at
214    will, and allow inlining and other optimizations. This linkage type is
215    only allowed on definitions, not declarations.
216``linkonce``
217    Globals with "``linkonce``" linkage are merged with other globals of
218    the same name when linkage occurs. This can be used to implement
219    some forms of inline functions, templates, or other code which must
220    be generated in each translation unit that uses it, but where the
221    body may be overridden with a more definitive definition later.
222    Unreferenced ``linkonce`` globals are allowed to be discarded. Note
223    that ``linkonce`` linkage does not actually allow the optimizer to
224    inline the body of this function into callers because it doesn't
225    know if this definition of the function is the definitive definition
226    within the program or whether it will be overridden by a stronger
227    definition. To enable inlining and other optimizations, use
228    "``linkonce_odr``" linkage.
229``weak``
230    "``weak``" linkage has the same merging semantics as ``linkonce``
231    linkage, except that unreferenced globals with ``weak`` linkage may
232    not be discarded. This is used for globals that are declared "weak"
233    in C source code.
234``common``
235    "``common``" linkage is most similar to "``weak``" linkage, but they
236    are used for tentative definitions in C, such as "``int X;``" at
237    global scope. Symbols with "``common``" linkage are merged in the
238    same way as ``weak symbols``, and they may not be deleted if
239    unreferenced. ``common`` symbols may not have an explicit section,
240    must have a zero initializer, and may not be marked
241    ':ref:`constant <globalvars>`'. Functions and aliases may not have
242    common linkage.
243
244.. _linkage_appending:
245
246``appending``
247    "``appending``" linkage may only be applied to global variables of
248    pointer to array type. When two global variables with appending
249    linkage are linked together, the two global arrays are appended
250    together. This is the LLVM, typesafe, equivalent of having the
251    system linker append together "sections" with identical names when
252    .o files are linked.
253
254    Unfortunately this doesn't correspond to any feature in .o files, so it
255    can only be used for variables like ``llvm.global_ctors`` which llvm
256    interprets specially.
257
258``extern_weak``
259    The semantics of this linkage follow the ELF object file model: the
260    symbol is weak until linked, if not linked, the symbol becomes null
261    instead of being an undefined reference.
262``linkonce_odr``, ``weak_odr``
263    Some languages allow differing globals to be merged, such as two
264    functions with different semantics. Other languages, such as
265    ``C++``, ensure that only equivalent globals are ever merged (the
266    "one definition rule" --- "ODR"). Such languages can use the
267    ``linkonce_odr`` and ``weak_odr`` linkage types to indicate that the
268    global will only be merged with equivalent globals. These linkage
269    types are otherwise the same as their non-``odr`` versions.
270``external``
271    If none of the above identifiers are used, the global is externally
272    visible, meaning that it participates in linkage and can be used to
273    resolve external symbol references.
274
275It is illegal for a function *declaration* to have any linkage type
276other than ``external`` or ``extern_weak``.
277
278.. _callingconv:
279
280Calling Conventions
281-------------------
282
283LLVM :ref:`functions <functionstructure>`, :ref:`calls <i_call>` and
284:ref:`invokes <i_invoke>` can all have an optional calling convention
285specified for the call. The calling convention of any pair of dynamic
286caller/callee must match, or the behavior of the program is undefined.
287The following calling conventions are supported by LLVM, and more may be
288added in the future:
289
290"``ccc``" - The C calling convention
291    This calling convention (the default if no other calling convention
292    is specified) matches the target C calling conventions. This calling
293    convention supports varargs function calls and tolerates some
294    mismatch in the declared prototype and implemented declaration of
295    the function (as does normal C).
296"``fastcc``" - The fast calling convention
297    This calling convention attempts to make calls as fast as possible
298    (e.g. by passing things in registers). This calling convention
299    allows the target to use whatever tricks it wants to produce fast
300    code for the target, without having to conform to an externally
301    specified ABI (Application Binary Interface). `Tail calls can only
302    be optimized when this, the GHC or the HiPE convention is
303    used. <CodeGenerator.html#id80>`_ This calling convention does not
304    support varargs and requires the prototype of all callees to exactly
305    match the prototype of the function definition.
306"``coldcc``" - The cold calling convention
307    This calling convention attempts to make code in the caller as
308    efficient as possible under the assumption that the call is not
309    commonly executed. As such, these calls often preserve all registers
310    so that the call does not break any live ranges in the caller side.
311    This calling convention does not support varargs and requires the
312    prototype of all callees to exactly match the prototype of the
313    function definition. Furthermore the inliner doesn't consider such function
314    calls for inlining.
315"``cc 10``" - GHC convention
316    This calling convention has been implemented specifically for use by
317    the `Glasgow Haskell Compiler (GHC) <http://www.haskell.org/ghc>`_.
318    It passes everything in registers, going to extremes to achieve this
319    by disabling callee save registers. This calling convention should
320    not be used lightly but only for specific situations such as an
321    alternative to the *register pinning* performance technique often
322    used when implementing functional programming languages. At the
323    moment only X86 supports this convention and it has the following
324    limitations:
325
326    -  On *X86-32* only supports up to 4 bit type parameters. No
327       floating-point types are supported.
328    -  On *X86-64* only supports up to 10 bit type parameters and 6
329       floating-point parameters.
330
331    This calling convention supports `tail call
332    optimization <CodeGenerator.html#id80>`_ but requires both the
333    caller and callee are using it.
334"``cc 11``" - The HiPE calling convention
335    This calling convention has been implemented specifically for use by
336    the `High-Performance Erlang
337    (HiPE) <http://www.it.uu.se/research/group/hipe/>`_ compiler, *the*
338    native code compiler of the `Ericsson's Open Source Erlang/OTP
339    system <http://www.erlang.org/download.shtml>`_. It uses more
340    registers for argument passing than the ordinary C calling
341    convention and defines no callee-saved registers. The calling
342    convention properly supports `tail call
343    optimization <CodeGenerator.html#id80>`_ but requires that both the
344    caller and the callee use it. It uses a *register pinning*
345    mechanism, similar to GHC's convention, for keeping frequently
346    accessed runtime components pinned to specific hardware registers.
347    At the moment only X86 supports this convention (both 32 and 64
348    bit).
349"``webkit_jscc``" - WebKit's JavaScript calling convention
350    This calling convention has been implemented for `WebKit FTL JIT
351    <https://trac.webkit.org/wiki/FTLJIT>`_. It passes arguments on the
352    stack right to left (as cdecl does), and returns a value in the
353    platform's customary return register.
354"``anyregcc``" - Dynamic calling convention for code patching
355    This is a special convention that supports patching an arbitrary code
356    sequence in place of a call site. This convention forces the call
357    arguments into registers but allows them to be dynamically
358    allocated. This can currently only be used with calls to
359    llvm.experimental.patchpoint because only this intrinsic records
360    the location of its arguments in a side table. See :doc:`StackMaps`.
361"``preserve_mostcc``" - The `PreserveMost` calling convention
362    This calling convention attempts to make the code in the caller as
363    unintrusive as possible. This convention behaves identically to the `C`
364    calling convention on how arguments and return values are passed, but it
365    uses a different set of caller/callee-saved registers. This alleviates the
366    burden of saving and recovering a large register set before and after the
367    call in the caller. If the arguments are passed in callee-saved registers,
368    then they will be preserved by the callee across the call. This doesn't
369    apply for values returned in callee-saved registers.
370
371    - On X86-64 the callee preserves all general purpose registers, except for
372      R11. R11 can be used as a scratch register. Floating-point registers
373      (XMMs/YMMs) are not preserved and need to be saved by the caller.
374
375    The idea behind this convention is to support calls to runtime functions
376    that have a hot path and a cold path. The hot path is usually a small piece
377    of code that doesn't use many registers. The cold path might need to call out to
378    another function and therefore only needs to preserve the caller-saved
379    registers, which haven't already been saved by the caller. The
380    `PreserveMost` calling convention is very similar to the `cold` calling
381    convention in terms of caller/callee-saved registers, but they are used for
382    different types of function calls. `coldcc` is for function calls that are
383    rarely executed, whereas `preserve_mostcc` function calls are intended to be
384    on the hot path and definitely executed a lot. Furthermore `preserve_mostcc`
385    doesn't prevent the inliner from inlining the function call.
386
387    This calling convention will be used by a future version of the ObjectiveC
388    runtime and should therefore still be considered experimental at this time.
389    Although this convention was created to optimize certain runtime calls to
390    the ObjectiveC runtime, it is not limited to this runtime and might be used
391    by other runtimes in the future too. The current implementation only
392    supports X86-64, but the intention is to support more architectures in the
393    future.
394"``preserve_allcc``" - The `PreserveAll` calling convention
395    This calling convention attempts to make the code in the caller even less
396    intrusive than the `PreserveMost` calling convention. This calling
397    convention also behaves identical to the `C` calling convention on how
398    arguments and return values are passed, but it uses a different set of
399    caller/callee-saved registers. This removes the burden of saving and
400    recovering a large register set before and after the call in the caller. If
401    the arguments are passed in callee-saved registers, then they will be
402    preserved by the callee across the call. This doesn't apply for values
403    returned in callee-saved registers.
404
405    - On X86-64 the callee preserves all general purpose registers, except for
406      R11. R11 can be used as a scratch register. Furthermore it also preserves
407      all floating-point registers (XMMs/YMMs).
408
409    The idea behind this convention is to support calls to runtime functions
410    that don't need to call out to any other functions.
411
412    This calling convention, like the `PreserveMost` calling convention, will be
413    used by a future version of the ObjectiveC runtime and should be considered
414    experimental at this time.
415"``cxx_fast_tlscc``" - The `CXX_FAST_TLS` calling convention for access functions
416    Clang generates an access function to access C++-style TLS. The access
417    function generally has an entry block, an exit block and an initialization
418    block that is run at the first time. The entry and exit blocks can access
419    a few TLS IR variables, each access will be lowered to a platform-specific
420    sequence.
421
422    This calling convention aims to minimize overhead in the caller by
423    preserving as many registers as possible (all the registers that are
424    perserved on the fast path, composed of the entry and exit blocks).
425
426    This calling convention behaves identical to the `C` calling convention on
427    how arguments and return values are passed, but it uses a different set of
428    caller/callee-saved registers.
429
430    Given that each platform has its own lowering sequence, hence its own set
431    of preserved registers, we can't use the existing `PreserveMost`.
432
433    - On X86-64 the callee preserves all general purpose registers, except for
434      RDI and RAX.
435"``swiftcc``" - This calling convention is used for Swift language.
436    - On X86-64 RCX and R8 are available for additional integer returns, and
437      XMM2 and XMM3 are available for additional FP/vector returns.
438    - On iOS platforms, we use AAPCS-VFP calling convention.
439"``cc <n>``" - Numbered convention
440    Any calling convention may be specified by number, allowing
441    target-specific calling conventions to be used. Target specific
442    calling conventions start at 64.
443
444More calling conventions can be added/defined on an as-needed basis, to
445support Pascal conventions or any other well-known target-independent
446convention.
447
448.. _visibilitystyles:
449
450Visibility Styles
451-----------------
452
453All Global Variables and Functions have one of the following visibility
454styles:
455
456"``default``" - Default style
457    On targets that use the ELF object file format, default visibility
458    means that the declaration is visible to other modules and, in
459    shared libraries, means that the declared entity may be overridden.
460    On Darwin, default visibility means that the declaration is visible
461    to other modules. Default visibility corresponds to "external
462    linkage" in the language.
463"``hidden``" - Hidden style
464    Two declarations of an object with hidden visibility refer to the
465    same object if they are in the same shared object. Usually, hidden
466    visibility indicates that the symbol will not be placed into the
467    dynamic symbol table, so no other module (executable or shared
468    library) can reference it directly.
469"``protected``" - Protected style
470    On ELF, protected visibility indicates that the symbol will be
471    placed in the dynamic symbol table, but that references within the
472    defining module will bind to the local symbol. That is, the symbol
473    cannot be overridden by another module.
474
475A symbol with ``internal`` or ``private`` linkage must have ``default``
476visibility.
477
478.. _dllstorageclass:
479
480DLL Storage Classes
481-------------------
482
483All Global Variables, Functions and Aliases can have one of the following
484DLL storage class:
485
486``dllimport``
487    "``dllimport``" causes the compiler to reference a function or variable via
488    a global pointer to a pointer that is set up by the DLL exporting the
489    symbol. On Microsoft Windows targets, the pointer name is formed by
490    combining ``__imp_`` and the function or variable name.
491``dllexport``
492    "``dllexport``" causes the compiler to provide a global pointer to a pointer
493    in a DLL, so that it can be referenced with the ``dllimport`` attribute. On
494    Microsoft Windows targets, the pointer name is formed by combining
495    ``__imp_`` and the function or variable name. Since this storage class
496    exists for defining a dll interface, the compiler, assembler and linker know
497    it is externally referenced and must refrain from deleting the symbol.
498
499.. _tls_model:
500
501Thread Local Storage Models
502---------------------------
503
504A variable may be defined as ``thread_local``, which means that it will
505not be shared by threads (each thread will have a separated copy of the
506variable). Not all targets support thread-local variables. Optionally, a
507TLS model may be specified:
508
509``localdynamic``
510    For variables that are only used within the current shared library.
511``initialexec``
512    For variables in modules that will not be loaded dynamically.
513``localexec``
514    For variables defined in the executable and only used within it.
515
516If no explicit model is given, the "general dynamic" model is used.
517
518The models correspond to the ELF TLS models; see `ELF Handling For
519Thread-Local Storage <http://people.redhat.com/drepper/tls.pdf>`_ for
520more information on under which circumstances the different models may
521be used. The target may choose a different TLS model if the specified
522model is not supported, or if a better choice of model can be made.
523
524A model can also be specified in an alias, but then it only governs how
525the alias is accessed. It will not have any effect in the aliasee.
526
527For platforms without linker support of ELF TLS model, the -femulated-tls
528flag can be used to generate GCC compatible emulated TLS code.
529
530.. _runtime_preemption_model:
531
532Runtime Preemption Specifiers
533-----------------------------
534
535Global variables, functions and aliases may have an optional runtime preemption
536specifier. If a preemption specifier isn't given explicitly, then a
537symbol is assumed to be ``dso_preemptable``.
538
539``dso_preemptable``
540    Indicates that the function or variable may be replaced by a symbol from
541    outside the linkage unit at runtime.
542
543``dso_local``
544    The compiler may assume that a function or variable marked as ``dso_local``
545    will resolve to a symbol within the same linkage unit. Direct access will
546    be generated even if the definition is not within this compilation unit.
547
548.. _namedtypes:
549
550Structure Types
551---------------
552
553LLVM IR allows you to specify both "identified" and "literal" :ref:`structure
554types <t_struct>`. Literal types are uniqued structurally, but identified types
555are never uniqued. An :ref:`opaque structural type <t_opaque>` can also be used
556to forward declare a type that is not yet available.
557
558An example of an identified structure specification is:
559
560.. code-block:: llvm
561
562    %mytype = type { %mytype*, i32 }
563
564Prior to the LLVM 3.0 release, identified types were structurally uniqued. Only
565literal types are uniqued in recent versions of LLVM.
566
567.. _nointptrtype:
568
569Non-Integral Pointer Type
570-------------------------
571
572Note: non-integral pointer types are a work in progress, and they should be
573considered experimental at this time.
574
575LLVM IR optionally allows the frontend to denote pointers in certain address
576spaces as "non-integral" via the :ref:`datalayout string<langref_datalayout>`.
577Non-integral pointer types represent pointers that have an *unspecified* bitwise
578representation; that is, the integral representation may be target dependent or
579unstable (not backed by a fixed integer).
580
581``inttoptr`` instructions converting integers to non-integral pointer types are
582ill-typed, and so are ``ptrtoint`` instructions converting values of
583non-integral pointer types to integers.  Vector versions of said instructions
584are ill-typed as well.
585
586.. _globalvars:
587
588Global Variables
589----------------
590
591Global variables define regions of memory allocated at compilation time
592instead of run-time.
593
594Global variable definitions must be initialized.
595
596Global variables in other translation units can also be declared, in which
597case they don't have an initializer.
598
599Either global variable definitions or declarations may have an explicit section
600to be placed in and may have an optional explicit alignment specified. If there
601is a mismatch between the explicit or inferred section information for the
602variable declaration and its definition the resulting behavior is undefined.
603
604A variable may be defined as a global ``constant``, which indicates that
605the contents of the variable will **never** be modified (enabling better
606optimization, allowing the global data to be placed in the read-only
607section of an executable, etc). Note that variables that need runtime
608initialization cannot be marked ``constant`` as there is a store to the
609variable.
610
611LLVM explicitly allows *declarations* of global variables to be marked
612constant, even if the final definition of the global is not. This
613capability can be used to enable slightly better optimization of the
614program, but requires the language definition to guarantee that
615optimizations based on the 'constantness' are valid for the translation
616units that do not include the definition.
617
618As SSA values, global variables define pointer values that are in scope
619(i.e. they dominate) all basic blocks in the program. Global variables
620always define a pointer to their "content" type because they describe a
621region of memory, and all memory objects in LLVM are accessed through
622pointers.
623
624Global variables can be marked with ``unnamed_addr`` which indicates
625that the address is not significant, only the content. Constants marked
626like this can be merged with other constants if they have the same
627initializer. Note that a constant with significant address *can* be
628merged with a ``unnamed_addr`` constant, the result being a constant
629whose address is significant.
630
631If the ``local_unnamed_addr`` attribute is given, the address is known to
632not be significant within the module.
633
634A global variable may be declared to reside in a target-specific
635numbered address space. For targets that support them, address spaces
636may affect how optimizations are performed and/or what target
637instructions are used to access the variable. The default address space
638is zero. The address space qualifier must precede any other attributes.
639
640LLVM allows an explicit section to be specified for globals. If the
641target supports it, it will emit globals to the section specified.
642Additionally, the global can placed in a comdat if the target has the necessary
643support.
644
645External declarations may have an explicit section specified. Section
646information is retained in LLVM IR for targets that make use of this
647information. Attaching section information to an external declaration is an
648assertion that its definition is located in the specified section. If the
649definition is located in a different section, the behavior is undefined.
650
651By default, global initializers are optimized by assuming that global
652variables defined within the module are not modified from their
653initial values before the start of the global initializer. This is
654true even for variables potentially accessible from outside the
655module, including those with external linkage or appearing in
656``@llvm.used`` or dllexported variables. This assumption may be suppressed
657by marking the variable with ``externally_initialized``.
658
659An explicit alignment may be specified for a global, which must be a
660power of 2. If not present, or if the alignment is set to zero, the
661alignment of the global is set by the target to whatever it feels
662convenient. If an explicit alignment is specified, the global is forced
663to have exactly that alignment. Targets and optimizers are not allowed
664to over-align the global if the global has an assigned section. In this
665case, the extra alignment could be observable: for example, code could
666assume that the globals are densely packed in their section and try to
667iterate over them as an array, alignment padding would break this
668iteration. The maximum alignment is ``1 << 29``.
669
670Globals can also have a :ref:`DLL storage class <dllstorageclass>`,
671an optional :ref:`runtime preemption specifier <runtime_preemption_model>`,
672an optional :ref:`global attributes <glattrs>` and
673an optional list of attached :ref:`metadata <metadata>`.
674
675Variables and aliases can have a
676:ref:`Thread Local Storage Model <tls_model>`.
677
678Syntax::
679
680      @<GlobalVarName> = [Linkage] [PreemptionSpecifier] [Visibility]
681                         [DLLStorageClass] [ThreadLocal]
682                         [(unnamed_addr|local_unnamed_addr)] [AddrSpace]
683                         [ExternallyInitialized]
684                         <global | constant> <Type> [<InitializerConstant>]
685                         [, section "name"] [, comdat [($name)]]
686                         [, align <Alignment>] (, !name !N)*
687
688For example, the following defines a global in a numbered address space
689with an initializer, section, and alignment:
690
691.. code-block:: llvm
692
693    @G = addrspace(5) constant float 1.0, section "foo", align 4
694
695The following example just declares a global variable
696
697.. code-block:: llvm
698
699   @G = external global i32
700
701The following example defines a thread-local global with the
702``initialexec`` TLS model:
703
704.. code-block:: llvm
705
706    @G = thread_local(initialexec) global i32 0, align 4
707
708.. _functionstructure:
709
710Functions
711---------
712
713LLVM function definitions consist of the "``define``" keyword, an
714optional :ref:`linkage type <linkage>`, an optional :ref:`runtime preemption
715specifier <runtime_preemption_model>`,  an optional :ref:`visibility
716style <visibility>`, an optional :ref:`DLL storage class <dllstorageclass>`,
717an optional :ref:`calling convention <callingconv>`,
718an optional ``unnamed_addr`` attribute, a return type, an optional
719:ref:`parameter attribute <paramattrs>` for the return type, a function
720name, a (possibly empty) argument list (each with optional :ref:`parameter
721attributes <paramattrs>`), optional :ref:`function attributes <fnattrs>`,
722an optional address space, an optional section, an optional alignment,
723an optional :ref:`comdat <langref_comdats>`,
724an optional :ref:`garbage collector name <gc>`, an optional :ref:`prefix <prefixdata>`,
725an optional :ref:`prologue <prologuedata>`,
726an optional :ref:`personality <personalityfn>`,
727an optional list of attached :ref:`metadata <metadata>`,
728an opening curly brace, a list of basic blocks, and a closing curly brace.
729
730LLVM function declarations consist of the "``declare``" keyword, an
731optional :ref:`linkage type <linkage>`, an optional :ref:`visibility style
732<visibility>`, an optional :ref:`DLL storage class <dllstorageclass>`, an
733optional :ref:`calling convention <callingconv>`, an optional ``unnamed_addr``
734or ``local_unnamed_addr`` attribute, an optional address space, a return type,
735an optional :ref:`parameter attribute <paramattrs>` for the return type, a function name, a possibly
736empty list of arguments, an optional alignment, an optional :ref:`garbage
737collector name <gc>`, an optional :ref:`prefix <prefixdata>`, and an optional
738:ref:`prologue <prologuedata>`.
739
740A function definition contains a list of basic blocks, forming the CFG (Control
741Flow Graph) for the function. Each basic block may optionally start with a label
742(giving the basic block a symbol table entry), contains a list of instructions,
743and ends with a :ref:`terminator <terminators>` instruction (such as a branch or
744function return). If an explicit label name is not provided, a block is assigned
745an implicit numbered label, using the next value from the same counter as used
746for unnamed temporaries (:ref:`see above<identifiers>`). For example, if a
747function entry block does not have an explicit label, it will be assigned label
748"%0", then the first unnamed temporary in that block will be "%1", etc. If a
749numeric label is explicitly specified, it must match the numeric label that
750would be used implicitly.
751
752The first basic block in a function is special in two ways: it is
753immediately executed on entrance to the function, and it is not allowed
754to have predecessor basic blocks (i.e. there can not be any branches to
755the entry block of a function). Because the block can have no
756predecessors, it also cannot have any :ref:`PHI nodes <i_phi>`.
757
758LLVM allows an explicit section to be specified for functions. If the
759target supports it, it will emit functions to the section specified.
760Additionally, the function can be placed in a COMDAT.
761
762An explicit alignment may be specified for a function. If not present,
763or if the alignment is set to zero, the alignment of the function is set
764by the target to whatever it feels convenient. If an explicit alignment
765is specified, the function is forced to have at least that much
766alignment. All alignments must be a power of 2.
767
768If the ``unnamed_addr`` attribute is given, the address is known to not
769be significant and two identical functions can be merged.
770
771If the ``local_unnamed_addr`` attribute is given, the address is known to
772not be significant within the module.
773
774If an explicit address space is not given, it will default to the program
775address space from the :ref:`datalayout string<langref_datalayout>`.
776
777Syntax::
778
779    define [linkage] [PreemptionSpecifier] [visibility] [DLLStorageClass]
780           [cconv] [ret attrs]
781           <ResultType> @<FunctionName> ([argument list])
782           [(unnamed_addr|local_unnamed_addr)] [AddrSpace] [fn Attrs]
783           [section "name"] [comdat [($name)]] [align N] [gc] [prefix Constant]
784           [prologue Constant] [personality Constant] (!name !N)* { ... }
785
786The argument list is a comma separated sequence of arguments where each
787argument is of the following form:
788
789Syntax::
790
791   <type> [parameter Attrs] [name]
792
793
794.. _langref_aliases:
795
796Aliases
797-------
798
799Aliases, unlike function or variables, don't create any new data. They
800are just a new symbol and metadata for an existing position.
801
802Aliases have a name and an aliasee that is either a global value or a
803constant expression.
804
805Aliases may have an optional :ref:`linkage type <linkage>`, an optional
806:ref:`runtime preemption specifier <runtime_preemption_model>`, an optional
807:ref:`visibility style <visibility>`, an optional :ref:`DLL storage class
808<dllstorageclass>` and an optional :ref:`tls model <tls_model>`.
809
810Syntax::
811
812    @<Name> = [Linkage] [PreemptionSpecifier] [Visibility] [DLLStorageClass] [ThreadLocal] [(unnamed_addr|local_unnamed_addr)] alias <AliaseeTy>, <AliaseeTy>* @<Aliasee>
813
814The linkage must be one of ``private``, ``internal``, ``linkonce``, ``weak``,
815``linkonce_odr``, ``weak_odr``, ``external``. Note that some system linkers
816might not correctly handle dropping a weak symbol that is aliased.
817
818Aliases that are not ``unnamed_addr`` are guaranteed to have the same address as
819the aliasee expression. ``unnamed_addr`` ones are only guaranteed to point
820to the same content.
821
822If the ``local_unnamed_addr`` attribute is given, the address is known to
823not be significant within the module.
824
825Since aliases are only a second name, some restrictions apply, of which
826some can only be checked when producing an object file:
827
828* The expression defining the aliasee must be computable at assembly
829  time. Since it is just a name, no relocations can be used.
830
831* No alias in the expression can be weak as the possibility of the
832  intermediate alias being overridden cannot be represented in an
833  object file.
834
835* No global value in the expression can be a declaration, since that
836  would require a relocation, which is not possible.
837
838.. _langref_ifunc:
839
840IFuncs
841-------
842
843IFuncs, like as aliases, don't create any new data or func. They are just a new
844symbol that dynamic linker resolves at runtime by calling a resolver function.
845
846IFuncs have a name and a resolver that is a function called by dynamic linker
847that returns address of another function associated with the name.
848
849IFunc may have an optional :ref:`linkage type <linkage>` and an optional
850:ref:`visibility style <visibility>`.
851
852Syntax::
853
854    @<Name> = [Linkage] [Visibility] ifunc <IFuncTy>, <ResolverTy>* @<Resolver>
855
856
857.. _langref_comdats:
858
859Comdats
860-------
861
862Comdat IR provides access to COFF and ELF object file COMDAT functionality.
863
864Comdats have a name which represents the COMDAT key. All global objects that
865specify this key will only end up in the final object file if the linker chooses
866that key over some other key. Aliases are placed in the same COMDAT that their
867aliasee computes to, if any.
868
869Comdats have a selection kind to provide input on how the linker should
870choose between keys in two different object files.
871
872Syntax::
873
874    $<Name> = comdat SelectionKind
875
876The selection kind must be one of the following:
877
878``any``
879    The linker may choose any COMDAT key, the choice is arbitrary.
880``exactmatch``
881    The linker may choose any COMDAT key but the sections must contain the
882    same data.
883``largest``
884    The linker will choose the section containing the largest COMDAT key.
885``noduplicates``
886    The linker requires that only section with this COMDAT key exist.
887``samesize``
888    The linker may choose any COMDAT key but the sections must contain the
889    same amount of data.
890
891Note that the Mach-O platform doesn't support COMDATs, and ELF and WebAssembly
892only support ``any`` as a selection kind.
893
894Here is an example of a COMDAT group where a function will only be selected if
895the COMDAT key's section is the largest:
896
897.. code-block:: text
898
899   $foo = comdat largest
900   @foo = global i32 2, comdat($foo)
901
902   define void @bar() comdat($foo) {
903     ret void
904   }
905
906As a syntactic sugar the ``$name`` can be omitted if the name is the same as
907the global name:
908
909.. code-block:: text
910
911  $foo = comdat any
912  @foo = global i32 2, comdat
913
914
915In a COFF object file, this will create a COMDAT section with selection kind
916``IMAGE_COMDAT_SELECT_LARGEST`` containing the contents of the ``@foo`` symbol
917and another COMDAT section with selection kind
918``IMAGE_COMDAT_SELECT_ASSOCIATIVE`` which is associated with the first COMDAT
919section and contains the contents of the ``@bar`` symbol.
920
921There are some restrictions on the properties of the global object.
922It, or an alias to it, must have the same name as the COMDAT group when
923targeting COFF.
924The contents and size of this object may be used during link-time to determine
925which COMDAT groups get selected depending on the selection kind.
926Because the name of the object must match the name of the COMDAT group, the
927linkage of the global object must not be local; local symbols can get renamed
928if a collision occurs in the symbol table.
929
930The combined use of COMDATS and section attributes may yield surprising results.
931For example:
932
933.. code-block:: text
934
935   $foo = comdat any
936   $bar = comdat any
937   @g1 = global i32 42, section "sec", comdat($foo)
938   @g2 = global i32 42, section "sec", comdat($bar)
939
940From the object file perspective, this requires the creation of two sections
941with the same name. This is necessary because both globals belong to different
942COMDAT groups and COMDATs, at the object file level, are represented by
943sections.
944
945Note that certain IR constructs like global variables and functions may
946create COMDATs in the object file in addition to any which are specified using
947COMDAT IR. This arises when the code generator is configured to emit globals
948in individual sections (e.g. when `-data-sections` or `-function-sections`
949is supplied to `llc`).
950
951.. _namedmetadatastructure:
952
953Named Metadata
954--------------
955
956Named metadata is a collection of metadata. :ref:`Metadata
957nodes <metadata>` (but not metadata strings) are the only valid
958operands for a named metadata.
959
960#. Named metadata are represented as a string of characters with the
961   metadata prefix. The rules for metadata names are the same as for
962   identifiers, but quoted names are not allowed. ``"\xx"`` type escapes
963   are still valid, which allows any character to be part of a name.
964
965Syntax::
966
967    ; Some unnamed metadata nodes, which are referenced by the named metadata.
968    !0 = !{!"zero"}
969    !1 = !{!"one"}
970    !2 = !{!"two"}
971    ; A named metadata.
972    !name = !{!0, !1, !2}
973
974.. _paramattrs:
975
976Parameter Attributes
977--------------------
978
979The return type and each parameter of a function type may have a set of
980*parameter attributes* associated with them. Parameter attributes are
981used to communicate additional information about the result or
982parameters of a function. Parameter attributes are considered to be part
983of the function, not of the function type, so functions with different
984parameter attributes can have the same function type.
985
986Parameter attributes are simple keywords that follow the type specified.
987If multiple parameter attributes are needed, they are space separated.
988For example:
989
990.. code-block:: llvm
991
992    declare i32 @printf(i8* noalias nocapture, ...)
993    declare i32 @atoi(i8 zeroext)
994    declare signext i8 @returns_signed_char()
995
996Note that any attributes for the function result (``nounwind``,
997``readonly``) come immediately after the argument list.
998
999Currently, only the following parameter attributes are defined:
1000
1001``zeroext``
1002    This indicates to the code generator that the parameter or return
1003    value should be zero-extended to the extent required by the target's
1004    ABI by the caller (for a parameter) or the callee (for a return value).
1005``signext``
1006    This indicates to the code generator that the parameter or return
1007    value should be sign-extended to the extent required by the target's
1008    ABI (which is usually 32-bits) by the caller (for a parameter) or
1009    the callee (for a return value).
1010``inreg``
1011    This indicates that this parameter or return value should be treated
1012    in a special target-dependent fashion while emitting code for
1013    a function call or return (usually, by putting it in a register as
1014    opposed to memory, though some targets use it to distinguish between
1015    two different kinds of registers). Use of this attribute is
1016    target-specific.
1017``byval``
1018    This indicates that the pointer parameter should really be passed by
1019    value to the function. The attribute implies that a hidden copy of
1020    the pointee is made between the caller and the callee, so the callee
1021    is unable to modify the value in the caller. This attribute is only
1022    valid on LLVM pointer arguments. It is generally used to pass
1023    structs and arrays by value, but is also valid on pointers to
1024    scalars. The copy is considered to belong to the caller not the
1025    callee (for example, ``readonly`` functions should not write to
1026    ``byval`` parameters). This is not a valid attribute for return
1027    values.
1028
1029    The byval attribute also supports specifying an alignment with the
1030    align attribute. It indicates the alignment of the stack slot to
1031    form and the known alignment of the pointer specified to the call
1032    site. If the alignment is not specified, then the code generator
1033    makes a target-specific assumption.
1034
1035.. _attr_inalloca:
1036
1037``inalloca``
1038
1039    The ``inalloca`` argument attribute allows the caller to take the
1040    address of outgoing stack arguments. An ``inalloca`` argument must
1041    be a pointer to stack memory produced by an ``alloca`` instruction.
1042    The alloca, or argument allocation, must also be tagged with the
1043    inalloca keyword. Only the last argument may have the ``inalloca``
1044    attribute, and that argument is guaranteed to be passed in memory.
1045
1046    An argument allocation may be used by a call at most once because
1047    the call may deallocate it. The ``inalloca`` attribute cannot be
1048    used in conjunction with other attributes that affect argument
1049    storage, like ``inreg``, ``nest``, ``sret``, or ``byval``. The
1050    ``inalloca`` attribute also disables LLVM's implicit lowering of
1051    large aggregate return values, which means that frontend authors
1052    must lower them with ``sret`` pointers.
1053
1054    When the call site is reached, the argument allocation must have
1055    been the most recent stack allocation that is still live, or the
1056    behavior is undefined. It is possible to allocate additional stack
1057    space after an argument allocation and before its call site, but it
1058    must be cleared off with :ref:`llvm.stackrestore
1059    <int_stackrestore>`.
1060
1061    See :doc:`InAlloca` for more information on how to use this
1062    attribute.
1063
1064``sret``
1065    This indicates that the pointer parameter specifies the address of a
1066    structure that is the return value of the function in the source
1067    program. This pointer must be guaranteed by the caller to be valid:
1068    loads and stores to the structure may be assumed by the callee not
1069    to trap and to be properly aligned. This is not a valid attribute
1070    for return values.
1071
1072.. _attr_align:
1073
1074``align <n>``
1075    This indicates that the pointer value may be assumed by the optimizer to
1076    have the specified alignment.  If the pointer value does not have the
1077    specified alignment, behavior is undefined.
1078
1079    Note that this attribute has additional semantics when combined with the
1080    ``byval`` attribute, which are documented there.
1081
1082.. _noalias:
1083
1084``noalias``
1085    This indicates that objects accessed via pointer values
1086    :ref:`based <pointeraliasing>` on the argument or return value are not also
1087    accessed, during the execution of the function, via pointer values not
1088    *based* on the argument or return value. The attribute on a return value
1089    also has additional semantics described below. The caller shares the
1090    responsibility with the callee for ensuring that these requirements are met.
1091    For further details, please see the discussion of the NoAlias response in
1092    :ref:`alias analysis <Must, May, or No>`.
1093
1094    Note that this definition of ``noalias`` is intentionally similar
1095    to the definition of ``restrict`` in C99 for function arguments.
1096
1097    For function return values, C99's ``restrict`` is not meaningful,
1098    while LLVM's ``noalias`` is. Furthermore, the semantics of the ``noalias``
1099    attribute on return values are stronger than the semantics of the attribute
1100    when used on function arguments. On function return values, the ``noalias``
1101    attribute indicates that the function acts like a system memory allocation
1102    function, returning a pointer to allocated storage disjoint from the
1103    storage for any other object accessible to the caller.
1104
1105``nocapture``
1106    This indicates that the callee does not make any copies of the
1107    pointer that outlive the callee itself. This is not a valid
1108    attribute for return values.  Addresses used in volatile operations
1109    are considered to be captured.
1110
1111.. _nest:
1112
1113``nest``
1114    This indicates that the pointer parameter can be excised using the
1115    :ref:`trampoline intrinsics <int_trampoline>`. This is not a valid
1116    attribute for return values and can only be applied to one parameter.
1117
1118``returned``
1119    This indicates that the function always returns the argument as its return
1120    value. This is a hint to the optimizer and code generator used when
1121    generating the caller, allowing value propagation, tail call optimization,
1122    and omission of register saves and restores in some cases; it is not
1123    checked or enforced when generating the callee. The parameter and the
1124    function return type must be valid operands for the
1125    :ref:`bitcast instruction <i_bitcast>`. This is not a valid attribute for
1126    return values and can only be applied to one parameter.
1127
1128``nonnull``
1129    This indicates that the parameter or return pointer is not null. This
1130    attribute may only be applied to pointer typed parameters. This is not
1131    checked or enforced by LLVM; if the parameter or return pointer is null,
1132    the behavior is undefined.
1133
1134``dereferenceable(<n>)``
1135    This indicates that the parameter or return pointer is dereferenceable. This
1136    attribute may only be applied to pointer typed parameters. A pointer that
1137    is dereferenceable can be loaded from speculatively without a risk of
1138    trapping. The number of bytes known to be dereferenceable must be provided
1139    in parentheses. It is legal for the number of bytes to be less than the
1140    size of the pointee type. The ``nonnull`` attribute does not imply
1141    dereferenceability (consider a pointer to one element past the end of an
1142    array), however ``dereferenceable(<n>)`` does imply ``nonnull`` in
1143    ``addrspace(0)`` (which is the default address space).
1144
1145``dereferenceable_or_null(<n>)``
1146    This indicates that the parameter or return value isn't both
1147    non-null and non-dereferenceable (up to ``<n>`` bytes) at the same
1148    time. All non-null pointers tagged with
1149    ``dereferenceable_or_null(<n>)`` are ``dereferenceable(<n>)``.
1150    For address space 0 ``dereferenceable_or_null(<n>)`` implies that
1151    a pointer is exactly one of ``dereferenceable(<n>)`` or ``null``,
1152    and in other address spaces ``dereferenceable_or_null(<n>)``
1153    implies that a pointer is at least one of ``dereferenceable(<n>)``
1154    or ``null`` (i.e. it may be both ``null`` and
1155    ``dereferenceable(<n>)``). This attribute may only be applied to
1156    pointer typed parameters.
1157
1158``swiftself``
1159    This indicates that the parameter is the self/context parameter. This is not
1160    a valid attribute for return values and can only be applied to one
1161    parameter.
1162
1163``swifterror``
1164    This attribute is motivated to model and optimize Swift error handling. It
1165    can be applied to a parameter with pointer to pointer type or a
1166    pointer-sized alloca. At the call site, the actual argument that corresponds
1167    to a ``swifterror`` parameter has to come from a ``swifterror`` alloca or
1168    the ``swifterror`` parameter of the caller. A ``swifterror`` value (either
1169    the parameter or the alloca) can only be loaded and stored from, or used as
1170    a ``swifterror`` argument. This is not a valid attribute for return values
1171    and can only be applied to one parameter.
1172
1173    These constraints allow the calling convention to optimize access to
1174    ``swifterror`` variables by associating them with a specific register at
1175    call boundaries rather than placing them in memory. Since this does change
1176    the calling convention, a function which uses the ``swifterror`` attribute
1177    on a parameter is not ABI-compatible with one which does not.
1178
1179    These constraints also allow LLVM to assume that a ``swifterror`` argument
1180    does not alias any other memory visible within a function and that a
1181    ``swifterror`` alloca passed as an argument does not escape.
1182
1183``immarg``
1184    This indicates the parameter is required to be an immediate
1185    value. This must be a trivial immediate integer or floating-point
1186    constant. Undef or constant expressions are not valid. This is
1187    only valid on intrinsic declarations and cannot be applied to a
1188    call site or arbitrary function.
1189
1190.. _gc:
1191
1192Garbage Collector Strategy Names
1193--------------------------------
1194
1195Each function may specify a garbage collector strategy name, which is simply a
1196string:
1197
1198.. code-block:: llvm
1199
1200    define void @f() gc "name" { ... }
1201
1202The supported values of *name* includes those :ref:`built in to LLVM
1203<builtin-gc-strategies>` and any provided by loaded plugins. Specifying a GC
1204strategy will cause the compiler to alter its output in order to support the
1205named garbage collection algorithm. Note that LLVM itself does not contain a
1206garbage collector, this functionality is restricted to generating machine code
1207which can interoperate with a collector provided externally.
1208
1209.. _prefixdata:
1210
1211Prefix Data
1212-----------
1213
1214Prefix data is data associated with a function which the code
1215generator will emit immediately before the function's entrypoint.
1216The purpose of this feature is to allow frontends to associate
1217language-specific runtime metadata with specific functions and make it
1218available through the function pointer while still allowing the
1219function pointer to be called.
1220
1221To access the data for a given function, a program may bitcast the
1222function pointer to a pointer to the constant's type and dereference
1223index -1. This implies that the IR symbol points just past the end of
1224the prefix data. For instance, take the example of a function annotated
1225with a single ``i32``,
1226
1227.. code-block:: llvm
1228
1229    define void @f() prefix i32 123 { ... }
1230
1231The prefix data can be referenced as,
1232
1233.. code-block:: llvm
1234
1235    %0 = bitcast void* () @f to i32*
1236    %a = getelementptr inbounds i32, i32* %0, i32 -1
1237    %b = load i32, i32* %a
1238
1239Prefix data is laid out as if it were an initializer for a global variable
1240of the prefix data's type. The function will be placed such that the
1241beginning of the prefix data is aligned. This means that if the size
1242of the prefix data is not a multiple of the alignment size, the
1243function's entrypoint will not be aligned. If alignment of the
1244function's entrypoint is desired, padding must be added to the prefix
1245data.
1246
1247A function may have prefix data but no body. This has similar semantics
1248to the ``available_externally`` linkage in that the data may be used by the
1249optimizers but will not be emitted in the object file.
1250
1251.. _prologuedata:
1252
1253Prologue Data
1254-------------
1255
1256The ``prologue`` attribute allows arbitrary code (encoded as bytes) to
1257be inserted prior to the function body. This can be used for enabling
1258function hot-patching and instrumentation.
1259
1260To maintain the semantics of ordinary function calls, the prologue data must
1261have a particular format. Specifically, it must begin with a sequence of
1262bytes which decode to a sequence of machine instructions, valid for the
1263module's target, which transfer control to the point immediately succeeding
1264the prologue data, without performing any other visible action. This allows
1265the inliner and other passes to reason about the semantics of the function
1266definition without needing to reason about the prologue data. Obviously this
1267makes the format of the prologue data highly target dependent.
1268
1269A trivial example of valid prologue data for the x86 architecture is ``i8 144``,
1270which encodes the ``nop`` instruction:
1271
1272.. code-block:: text
1273
1274    define void @f() prologue i8 144 { ... }
1275
1276Generally prologue data can be formed by encoding a relative branch instruction
1277which skips the metadata, as in this example of valid prologue data for the
1278x86_64 architecture, where the first two bytes encode ``jmp .+10``:
1279
1280.. code-block:: text
1281
1282    %0 = type <{ i8, i8, i8* }>
1283
1284    define void @f() prologue %0 <{ i8 235, i8 8, i8* @md}> { ... }
1285
1286A function may have prologue data but no body. This has similar semantics
1287to the ``available_externally`` linkage in that the data may be used by the
1288optimizers but will not be emitted in the object file.
1289
1290.. _personalityfn:
1291
1292Personality Function
1293--------------------
1294
1295The ``personality`` attribute permits functions to specify what function
1296to use for exception handling.
1297
1298.. _attrgrp:
1299
1300Attribute Groups
1301----------------
1302
1303Attribute groups are groups of attributes that are referenced by objects within
1304the IR. They are important for keeping ``.ll`` files readable, because a lot of
1305functions will use the same set of attributes. In the degenerative case of a
1306``.ll`` file that corresponds to a single ``.c`` file, the single attribute
1307group will capture the important command line flags used to build that file.
1308
1309An attribute group is a module-level object. To use an attribute group, an
1310object references the attribute group's ID (e.g. ``#37``). An object may refer
1311to more than one attribute group. In that situation, the attributes from the
1312different groups are merged.
1313
1314Here is an example of attribute groups for a function that should always be
1315inlined, has a stack alignment of 4, and which shouldn't use SSE instructions:
1316
1317.. code-block:: llvm
1318
1319   ; Target-independent attributes:
1320   attributes #0 = { alwaysinline alignstack=4 }
1321
1322   ; Target-dependent attributes:
1323   attributes #1 = { "no-sse" }
1324
1325   ; Function @f has attributes: alwaysinline, alignstack=4, and "no-sse".
1326   define void @f() #0 #1 { ... }
1327
1328.. _fnattrs:
1329
1330Function Attributes
1331-------------------
1332
1333Function attributes are set to communicate additional information about
1334a function. Function attributes are considered to be part of the
1335function, not of the function type, so functions with different function
1336attributes can have the same function type.
1337
1338Function attributes are simple keywords that follow the type specified.
1339If multiple attributes are needed, they are space separated. For
1340example:
1341
1342.. code-block:: llvm
1343
1344    define void @f() noinline { ... }
1345    define void @f() alwaysinline { ... }
1346    define void @f() alwaysinline optsize { ... }
1347    define void @f() optsize { ... }
1348
1349``alignstack(<n>)``
1350    This attribute indicates that, when emitting the prologue and
1351    epilogue, the backend should forcibly align the stack pointer.
1352    Specify the desired alignment, which must be a power of two, in
1353    parentheses.
1354``allocsize(<EltSizeParam>[, <NumEltsParam>])``
1355    This attribute indicates that the annotated function will always return at
1356    least a given number of bytes (or null). Its arguments are zero-indexed
1357    parameter numbers; if one argument is provided, then it's assumed that at
1358    least ``CallSite.Args[EltSizeParam]`` bytes will be available at the
1359    returned pointer. If two are provided, then it's assumed that
1360    ``CallSite.Args[EltSizeParam] * CallSite.Args[NumEltsParam]`` bytes are
1361    available. The referenced parameters must be integer types. No assumptions
1362    are made about the contents of the returned block of memory.
1363``alwaysinline``
1364    This attribute indicates that the inliner should attempt to inline
1365    this function into callers whenever possible, ignoring any active
1366    inlining size threshold for this caller.
1367``builtin``
1368    This indicates that the callee function at a call site should be
1369    recognized as a built-in function, even though the function's declaration
1370    uses the ``nobuiltin`` attribute. This is only valid at call sites for
1371    direct calls to functions that are declared with the ``nobuiltin``
1372    attribute.
1373``cold``
1374    This attribute indicates that this function is rarely called. When
1375    computing edge weights, basic blocks post-dominated by a cold
1376    function call are also considered to be cold; and, thus, given low
1377    weight.
1378``convergent``
1379    In some parallel execution models, there exist operations that cannot be
1380    made control-dependent on any additional values.  We call such operations
1381    ``convergent``, and mark them with this attribute.
1382
1383    The ``convergent`` attribute may appear on functions or call/invoke
1384    instructions.  When it appears on a function, it indicates that calls to
1385    this function should not be made control-dependent on additional values.
1386    For example, the intrinsic ``llvm.nvvm.barrier0`` is ``convergent``, so
1387    calls to this intrinsic cannot be made control-dependent on additional
1388    values.
1389
1390    When it appears on a call/invoke, the ``convergent`` attribute indicates
1391    that we should treat the call as though we're calling a convergent
1392    function.  This is particularly useful on indirect calls; without this we
1393    may treat such calls as though the target is non-convergent.
1394
1395    The optimizer may remove the ``convergent`` attribute on functions when it
1396    can prove that the function does not execute any convergent operations.
1397    Similarly, the optimizer may remove ``convergent`` on calls/invokes when it
1398    can prove that the call/invoke cannot call a convergent function.
1399``inaccessiblememonly``
1400    This attribute indicates that the function may only access memory that
1401    is not accessible by the module being compiled. This is a weaker form
1402    of ``readnone``. If the function reads or writes other memory, the
1403    behavior is undefined.
1404``inaccessiblemem_or_argmemonly``
1405    This attribute indicates that the function may only access memory that is
1406    either not accessible by the module being compiled, or is pointed to
1407    by its pointer arguments. This is a weaker form of  ``argmemonly``. If the
1408    function reads or writes other memory, the behavior is undefined.
1409``inlinehint``
1410    This attribute indicates that the source code contained a hint that
1411    inlining this function is desirable (such as the "inline" keyword in
1412    C/C++). It is just a hint; it imposes no requirements on the
1413    inliner.
1414``jumptable``
1415    This attribute indicates that the function should be added to a
1416    jump-instruction table at code-generation time, and that all address-taken
1417    references to this function should be replaced with a reference to the
1418    appropriate jump-instruction-table function pointer. Note that this creates
1419    a new pointer for the original function, which means that code that depends
1420    on function-pointer identity can break. So, any function annotated with
1421    ``jumptable`` must also be ``unnamed_addr``.
1422``minsize``
1423    This attribute suggests that optimization passes and code generator
1424    passes make choices that keep the code size of this function as small
1425    as possible and perform optimizations that may sacrifice runtime
1426    performance in order to minimize the size of the generated code.
1427``naked``
1428    This attribute disables prologue / epilogue emission for the
1429    function. This can have very system-specific consequences.
1430``no-jump-tables``
1431    When this attribute is set to true, the jump tables and lookup tables that
1432    can be generated from a switch case lowering are disabled.
1433``nobuiltin``
1434    This indicates that the callee function at a call site is not recognized as
1435    a built-in function. LLVM will retain the original call and not replace it
1436    with equivalent code based on the semantics of the built-in function, unless
1437    the call site uses the ``builtin`` attribute. This is valid at call sites
1438    and on function declarations and definitions.
1439``noduplicate``
1440    This attribute indicates that calls to the function cannot be
1441    duplicated. A call to a ``noduplicate`` function may be moved
1442    within its parent function, but may not be duplicated within
1443    its parent function.
1444
1445    A function containing a ``noduplicate`` call may still
1446    be an inlining candidate, provided that the call is not
1447    duplicated by inlining. That implies that the function has
1448    internal linkage and only has one call site, so the original
1449    call is dead after inlining.
1450``noimplicitfloat``
1451    This attributes disables implicit floating-point instructions.
1452``noinline``
1453    This attribute indicates that the inliner should never inline this
1454    function in any situation. This attribute may not be used together
1455    with the ``alwaysinline`` attribute.
1456``nonlazybind``
1457    This attribute suppresses lazy symbol binding for the function. This
1458    may make calls to the function faster, at the cost of extra program
1459    startup time if the function is not called during program startup.
1460``noredzone``
1461    This attribute indicates that the code generator should not use a
1462    red zone, even if the target-specific ABI normally permits it.
1463``indirect-tls-seg-refs``
1464    This attribute indicates that the code generator should not use
1465    direct TLS access through segment registers, even if the
1466    target-specific ABI normally permits it.
1467``noreturn``
1468    This function attribute indicates that the function never returns
1469    normally. This produces undefined behavior at runtime if the
1470    function ever does dynamically return.
1471``norecurse``
1472    This function attribute indicates that the function does not call itself
1473    either directly or indirectly down any possible call path. This produces
1474    undefined behavior at runtime if the function ever does recurse.
1475``nounwind``
1476    This function attribute indicates that the function never raises an
1477    exception. If the function does raise an exception, its runtime
1478    behavior is undefined. However, functions marked nounwind may still
1479    trap or generate asynchronous exceptions. Exception handling schemes
1480    that are recognized by LLVM to handle asynchronous exceptions, such
1481    as SEH, will still provide their implementation defined semantics.
1482``"null-pointer-is-valid"``
1483   If ``"null-pointer-is-valid"`` is set to ``"true"``, then ``null`` address
1484   in address-space 0 is considered to be a valid address for memory loads and
1485   stores. Any analysis or optimization should not treat dereferencing a
1486   pointer to ``null`` as undefined behavior in this function.
1487   Note: Comparing address of a global variable to ``null`` may still
1488   evaluate to false because of a limitation in querying this attribute inside
1489   constant expressions.
1490``optforfuzzing``
1491    This attribute indicates that this function should be optimized
1492    for maximum fuzzing signal.
1493``optnone``
1494    This function attribute indicates that most optimization passes will skip
1495    this function, with the exception of interprocedural optimization passes.
1496    Code generation defaults to the "fast" instruction selector.
1497    This attribute cannot be used together with the ``alwaysinline``
1498    attribute; this attribute is also incompatible
1499    with the ``minsize`` attribute and the ``optsize`` attribute.
1500
1501    This attribute requires the ``noinline`` attribute to be specified on
1502    the function as well, so the function is never inlined into any caller.
1503    Only functions with the ``alwaysinline`` attribute are valid
1504    candidates for inlining into the body of this function.
1505``optsize``
1506    This attribute suggests that optimization passes and code generator
1507    passes make choices that keep the code size of this function low,
1508    and otherwise do optimizations specifically to reduce code size as
1509    long as they do not significantly impact runtime performance.
1510``"patchable-function"``
1511    This attribute tells the code generator that the code
1512    generated for this function needs to follow certain conventions that
1513    make it possible for a runtime function to patch over it later.
1514    The exact effect of this attribute depends on its string value,
1515    for which there currently is one legal possibility:
1516
1517     * ``"prologue-short-redirect"`` - This style of patchable
1518       function is intended to support patching a function prologue to
1519       redirect control away from the function in a thread safe
1520       manner.  It guarantees that the first instruction of the
1521       function will be large enough to accommodate a short jump
1522       instruction, and will be sufficiently aligned to allow being
1523       fully changed via an atomic compare-and-swap instruction.
1524       While the first requirement can be satisfied by inserting large
1525       enough NOP, LLVM can and will try to re-purpose an existing
1526       instruction (i.e. one that would have to be emitted anyway) as
1527       the patchable instruction larger than a short jump.
1528
1529       ``"prologue-short-redirect"`` is currently only supported on
1530       x86-64.
1531
1532    This attribute by itself does not imply restrictions on
1533    inter-procedural optimizations.  All of the semantic effects the
1534    patching may have to be separately conveyed via the linkage type.
1535``"probe-stack"``
1536    This attribute indicates that the function will trigger a guard region
1537    in the end of the stack. It ensures that accesses to the stack must be
1538    no further apart than the size of the guard region to a previous
1539    access of the stack. It takes one required string value, the name of
1540    the stack probing function that will be called.
1541
1542    If a function that has a ``"probe-stack"`` attribute is inlined into
1543    a function with another ``"probe-stack"`` attribute, the resulting
1544    function has the ``"probe-stack"`` attribute of the caller. If a
1545    function that has a ``"probe-stack"`` attribute is inlined into a
1546    function that has no ``"probe-stack"`` attribute at all, the resulting
1547    function has the ``"probe-stack"`` attribute of the callee.
1548``readnone``
1549    On a function, this attribute indicates that the function computes its
1550    result (or decides to unwind an exception) based strictly on its arguments,
1551    without dereferencing any pointer arguments or otherwise accessing
1552    any mutable state (e.g. memory, control registers, etc) visible to
1553    caller functions. It does not write through any pointer arguments
1554    (including ``byval`` arguments) and never changes any state visible
1555    to callers. This means while it cannot unwind exceptions by calling
1556    the ``C++`` exception throwing methods (since they write to memory), there may
1557    be non-``C++`` mechanisms that throw exceptions without writing to LLVM
1558    visible memory.
1559
1560    On an argument, this attribute indicates that the function does not
1561    dereference that pointer argument, even though it may read or write the
1562    memory that the pointer points to if accessed through other pointers.
1563
1564    If a readnone function reads or writes memory visible to the program, or
1565    has other side-effects, the behavior is undefined. If a function reads from
1566    or writes to a readnone pointer argument, the behavior is undefined.
1567``readonly``
1568    On a function, this attribute indicates that the function does not write
1569    through any pointer arguments (including ``byval`` arguments) or otherwise
1570    modify any state (e.g. memory, control registers, etc) visible to
1571    caller functions. It may dereference pointer arguments and read
1572    state that may be set in the caller. A readonly function always
1573    returns the same value (or unwinds an exception identically) when
1574    called with the same set of arguments and global state.  This means while it
1575    cannot unwind exceptions by calling the ``C++`` exception throwing methods
1576    (since they write to memory), there may be non-``C++`` mechanisms that throw
1577    exceptions without writing to LLVM visible memory.
1578
1579    On an argument, this attribute indicates that the function does not write
1580    through this pointer argument, even though it may write to the memory that
1581    the pointer points to.
1582
1583    If a readonly function writes memory visible to the program, or
1584    has other side-effects, the behavior is undefined. If a function writes to
1585    a readonly pointer argument, the behavior is undefined.
1586``"stack-probe-size"``
1587    This attribute controls the behavior of stack probes: either
1588    the ``"probe-stack"`` attribute, or ABI-required stack probes, if any.
1589    It defines the size of the guard region. It ensures that if the function
1590    may use more stack space than the size of the guard region, stack probing
1591    sequence will be emitted. It takes one required integer value, which
1592    is 4096 by default.
1593
1594    If a function that has a ``"stack-probe-size"`` attribute is inlined into
1595    a function with another ``"stack-probe-size"`` attribute, the resulting
1596    function has the ``"stack-probe-size"`` attribute that has the lower
1597    numeric value. If a function that has a ``"stack-probe-size"`` attribute is
1598    inlined into a function that has no ``"stack-probe-size"`` attribute
1599    at all, the resulting function has the ``"stack-probe-size"`` attribute
1600    of the callee.
1601``"no-stack-arg-probe"``
1602    This attribute disables ABI-required stack probes, if any.
1603``writeonly``
1604    On a function, this attribute indicates that the function may write to but
1605    does not read from memory.
1606
1607    On an argument, this attribute indicates that the function may write to but
1608    does not read through this pointer argument (even though it may read from
1609    the memory that the pointer points to).
1610
1611    If a writeonly function reads memory visible to the program, or
1612    has other side-effects, the behavior is undefined. If a function reads
1613    from a writeonly pointer argument, the behavior is undefined.
1614``argmemonly``
1615    This attribute indicates that the only memory accesses inside function are
1616    loads and stores from objects pointed to by its pointer-typed arguments,
1617    with arbitrary offsets. Or in other words, all memory operations in the
1618    function can refer to memory only using pointers based on its function
1619    arguments.
1620
1621    Note that ``argmemonly`` can be used together with ``readonly`` attribute
1622    in order to specify that function reads only from its arguments.
1623
1624    If an argmemonly function reads or writes memory other than the pointer
1625    arguments, or has other side-effects, the behavior is undefined.
1626``returns_twice``
1627    This attribute indicates that this function can return twice. The C
1628    ``setjmp`` is an example of such a function. The compiler disables
1629    some optimizations (like tail calls) in the caller of these
1630    functions.
1631``safestack``
1632    This attribute indicates that
1633    `SafeStack <http://clang.llvm.org/docs/SafeStack.html>`_
1634    protection is enabled for this function.
1635
1636    If a function that has a ``safestack`` attribute is inlined into a
1637    function that doesn't have a ``safestack`` attribute or which has an
1638    ``ssp``, ``sspstrong`` or ``sspreq`` attribute, then the resulting
1639    function will have a ``safestack`` attribute.
1640``sanitize_address``
1641    This attribute indicates that AddressSanitizer checks
1642    (dynamic address safety analysis) are enabled for this function.
1643``sanitize_memory``
1644    This attribute indicates that MemorySanitizer checks (dynamic detection
1645    of accesses to uninitialized memory) are enabled for this function.
1646``sanitize_thread``
1647    This attribute indicates that ThreadSanitizer checks
1648    (dynamic thread safety analysis) are enabled for this function.
1649``sanitize_hwaddress``
1650    This attribute indicates that HWAddressSanitizer checks
1651    (dynamic address safety analysis based on tagged pointers) are enabled for
1652    this function.
1653``speculative_load_hardening``
1654    This attribute indicates that
1655    `Speculative Load Hardening <https://llvm.org/docs/SpeculativeLoadHardening.html>`_
1656    should be enabled for the function body.
1657
1658    Speculative Load Hardening is a best-effort mitigation against
1659    information leak attacks that make use of control flow
1660    miss-speculation - specifically miss-speculation of whether a branch
1661    is taken or not. Typically vulnerabilities enabling such attacks are
1662    classified as "Spectre variant #1". Notably, this does not attempt to
1663    mitigate against miss-speculation of branch target, classified as
1664    "Spectre variant #2" vulnerabilities.
1665
1666    When inlining, the attribute is sticky. Inlining a function that carries
1667    this attribute will cause the caller to gain the attribute. This is intended
1668    to provide a maximally conservative model where the code in a function
1669    annotated with this attribute will always (even after inlining) end up
1670    hardened.
1671``speculatable``
1672    This function attribute indicates that the function does not have any
1673    effects besides calculating its result and does not have undefined behavior.
1674    Note that ``speculatable`` is not enough to conclude that along any
1675    particular execution path the number of calls to this function will not be
1676    externally observable. This attribute is only valid on functions
1677    and declarations, not on individual call sites. If a function is
1678    incorrectly marked as speculatable and really does exhibit
1679    undefined behavior, the undefined behavior may be observed even
1680    if the call site is dead code.
1681
1682``ssp``
1683    This attribute indicates that the function should emit a stack
1684    smashing protector. It is in the form of a "canary" --- a random value
1685    placed on the stack before the local variables that's checked upon
1686    return from the function to see if it has been overwritten. A
1687    heuristic is used to determine if a function needs stack protectors
1688    or not. The heuristic used will enable protectors for functions with:
1689
1690    - Character arrays larger than ``ssp-buffer-size`` (default 8).
1691    - Aggregates containing character arrays larger than ``ssp-buffer-size``.
1692    - Calls to alloca() with variable sizes or constant sizes greater than
1693      ``ssp-buffer-size``.
1694
1695    Variables that are identified as requiring a protector will be arranged
1696    on the stack such that they are adjacent to the stack protector guard.
1697
1698    If a function that has an ``ssp`` attribute is inlined into a
1699    function that doesn't have an ``ssp`` attribute, then the resulting
1700    function will have an ``ssp`` attribute.
1701``sspreq``
1702    This attribute indicates that the function should *always* emit a
1703    stack smashing protector. This overrides the ``ssp`` function
1704    attribute.
1705
1706    Variables that are identified as requiring a protector will be arranged
1707    on the stack such that they are adjacent to the stack protector guard.
1708    The specific layout rules are:
1709
1710    #. Large arrays and structures containing large arrays
1711       (``>= ssp-buffer-size``) are closest to the stack protector.
1712    #. Small arrays and structures containing small arrays
1713       (``< ssp-buffer-size``) are 2nd closest to the protector.
1714    #. Variables that have had their address taken are 3rd closest to the
1715       protector.
1716
1717    If a function that has an ``sspreq`` attribute is inlined into a
1718    function that doesn't have an ``sspreq`` attribute or which has an
1719    ``ssp`` or ``sspstrong`` attribute, then the resulting function will have
1720    an ``sspreq`` attribute.
1721``sspstrong``
1722    This attribute indicates that the function should emit a stack smashing
1723    protector. This attribute causes a strong heuristic to be used when
1724    determining if a function needs stack protectors. The strong heuristic
1725    will enable protectors for functions with:
1726
1727    - Arrays of any size and type
1728    - Aggregates containing an array of any size and type.
1729    - Calls to alloca().
1730    - Local variables that have had their address taken.
1731
1732    Variables that are identified as requiring a protector will be arranged
1733    on the stack such that they are adjacent to the stack protector guard.
1734    The specific layout rules are:
1735
1736    #. Large arrays and structures containing large arrays
1737       (``>= ssp-buffer-size``) are closest to the stack protector.
1738    #. Small arrays and structures containing small arrays
1739       (``< ssp-buffer-size``) are 2nd closest to the protector.
1740    #. Variables that have had their address taken are 3rd closest to the
1741       protector.
1742
1743    This overrides the ``ssp`` function attribute.
1744
1745    If a function that has an ``sspstrong`` attribute is inlined into a
1746    function that doesn't have an ``sspstrong`` attribute, then the
1747    resulting function will have an ``sspstrong`` attribute.
1748``strictfp``
1749    This attribute indicates that the function was called from a scope that
1750    requires strict floating-point semantics.  LLVM will not attempt any
1751    optimizations that require assumptions about the floating-point rounding
1752    mode or that might alter the state of floating-point status flags that
1753    might otherwise be set or cleared by calling this function.
1754``"thunk"``
1755    This attribute indicates that the function will delegate to some other
1756    function with a tail call. The prototype of a thunk should not be used for
1757    optimization purposes. The caller is expected to cast the thunk prototype to
1758    match the thunk target prototype.
1759``uwtable``
1760    This attribute indicates that the ABI being targeted requires that
1761    an unwind table entry be produced for this function even if we can
1762    show that no exceptions passes by it. This is normally the case for
1763    the ELF x86-64 abi, but it can be disabled for some compilation
1764    units.
1765``nocf_check``
1766    This attribute indicates that no control-flow check will be performed on
1767    the attributed entity. It disables -fcf-protection=<> for a specific
1768    entity to fine grain the HW control flow protection mechanism. The flag
1769    is target independent and currently appertains to a function or function
1770    pointer.
1771``shadowcallstack``
1772    This attribute indicates that the ShadowCallStack checks are enabled for
1773    the function. The instrumentation checks that the return address for the
1774    function has not changed between the function prolog and eiplog. It is
1775    currently x86_64-specific.
1776
1777.. _glattrs:
1778
1779Global Attributes
1780-----------------
1781
1782Attributes may be set to communicate additional information about a global variable.
1783Unlike :ref:`function attributes <fnattrs>`, attributes on a global variable
1784are grouped into a single :ref:`attribute group <attrgrp>`.
1785
1786.. _opbundles:
1787
1788Operand Bundles
1789---------------
1790
1791Operand bundles are tagged sets of SSA values that can be associated
1792with certain LLVM instructions (currently only ``call`` s and
1793``invoke`` s).  In a way they are like metadata, but dropping them is
1794incorrect and will change program semantics.
1795
1796Syntax::
1797
1798    operand bundle set ::= '[' operand bundle (, operand bundle )* ']'
1799    operand bundle ::= tag '(' [ bundle operand ] (, bundle operand )* ')'
1800    bundle operand ::= SSA value
1801    tag ::= string constant
1802
1803Operand bundles are **not** part of a function's signature, and a
1804given function may be called from multiple places with different kinds
1805of operand bundles.  This reflects the fact that the operand bundles
1806are conceptually a part of the ``call`` (or ``invoke``), not the
1807callee being dispatched to.
1808
1809Operand bundles are a generic mechanism intended to support
1810runtime-introspection-like functionality for managed languages.  While
1811the exact semantics of an operand bundle depend on the bundle tag,
1812there are certain limitations to how much the presence of an operand
1813bundle can influence the semantics of a program.  These restrictions
1814are described as the semantics of an "unknown" operand bundle.  As
1815long as the behavior of an operand bundle is describable within these
1816restrictions, LLVM does not need to have special knowledge of the
1817operand bundle to not miscompile programs containing it.
1818
1819- The bundle operands for an unknown operand bundle escape in unknown
1820  ways before control is transferred to the callee or invokee.
1821- Calls and invokes with operand bundles have unknown read / write
1822  effect on the heap on entry and exit (even if the call target is
1823  ``readnone`` or ``readonly``), unless they're overridden with
1824  callsite specific attributes.
1825- An operand bundle at a call site cannot change the implementation
1826  of the called function.  Inter-procedural optimizations work as
1827  usual as long as they take into account the first two properties.
1828
1829More specific types of operand bundles are described below.
1830
1831.. _deopt_opbundles:
1832
1833Deoptimization Operand Bundles
1834^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1835
1836Deoptimization operand bundles are characterized by the ``"deopt"``
1837operand bundle tag.  These operand bundles represent an alternate
1838"safe" continuation for the call site they're attached to, and can be
1839used by a suitable runtime to deoptimize the compiled frame at the
1840specified call site.  There can be at most one ``"deopt"`` operand
1841bundle attached to a call site.  Exact details of deoptimization is
1842out of scope for the language reference, but it usually involves
1843rewriting a compiled frame into a set of interpreted frames.
1844
1845From the compiler's perspective, deoptimization operand bundles make
1846the call sites they're attached to at least ``readonly``.  They read
1847through all of their pointer typed operands (even if they're not
1848otherwise escaped) and the entire visible heap.  Deoptimization
1849operand bundles do not capture their operands except during
1850deoptimization, in which case control will not be returned to the
1851compiled frame.
1852
1853The inliner knows how to inline through calls that have deoptimization
1854operand bundles.  Just like inlining through a normal call site
1855involves composing the normal and exceptional continuations, inlining
1856through a call site with a deoptimization operand bundle needs to
1857appropriately compose the "safe" deoptimization continuation.  The
1858inliner does this by prepending the parent's deoptimization
1859continuation to every deoptimization continuation in the inlined body.
1860E.g. inlining ``@f`` into ``@g`` in the following example
1861
1862.. code-block:: llvm
1863
1864    define void @f() {
1865      call void @x()  ;; no deopt state
1866      call void @y() [ "deopt"(i32 10) ]
1867      call void @y() [ "deopt"(i32 10), "unknown"(i8* null) ]
1868      ret void
1869    }
1870
1871    define void @g() {
1872      call void @f() [ "deopt"(i32 20) ]
1873      ret void
1874    }
1875
1876will result in
1877
1878.. code-block:: llvm
1879
1880    define void @g() {
1881      call void @x()  ;; still no deopt state
1882      call void @y() [ "deopt"(i32 20, i32 10) ]
1883      call void @y() [ "deopt"(i32 20, i32 10), "unknown"(i8* null) ]
1884      ret void
1885    }
1886
1887It is the frontend's responsibility to structure or encode the
1888deoptimization state in a way that syntactically prepending the
1889caller's deoptimization state to the callee's deoptimization state is
1890semantically equivalent to composing the caller's deoptimization
1891continuation after the callee's deoptimization continuation.
1892
1893.. _ob_funclet:
1894
1895Funclet Operand Bundles
1896^^^^^^^^^^^^^^^^^^^^^^^
1897
1898Funclet operand bundles are characterized by the ``"funclet"``
1899operand bundle tag.  These operand bundles indicate that a call site
1900is within a particular funclet.  There can be at most one
1901``"funclet"`` operand bundle attached to a call site and it must have
1902exactly one bundle operand.
1903
1904If any funclet EH pads have been "entered" but not "exited" (per the
1905`description in the EH doc\ <ExceptionHandling.html#wineh-constraints>`_),
1906it is undefined behavior to execute a ``call`` or ``invoke`` which:
1907
1908* does not have a ``"funclet"`` bundle and is not a ``call`` to a nounwind
1909  intrinsic, or
1910* has a ``"funclet"`` bundle whose operand is not the most-recently-entered
1911  not-yet-exited funclet EH pad.
1912
1913Similarly, if no funclet EH pads have been entered-but-not-yet-exited,
1914executing a ``call`` or ``invoke`` with a ``"funclet"`` bundle is undefined behavior.
1915
1916GC Transition Operand Bundles
1917^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1918
1919GC transition operand bundles are characterized by the
1920``"gc-transition"`` operand bundle tag. These operand bundles mark a
1921call as a transition between a function with one GC strategy to a
1922function with a different GC strategy. If coordinating the transition
1923between GC strategies requires additional code generation at the call
1924site, these bundles may contain any values that are needed by the
1925generated code.  For more details, see :ref:`GC Transitions
1926<gc_transition_args>`.
1927
1928.. _moduleasm:
1929
1930Module-Level Inline Assembly
1931----------------------------
1932
1933Modules may contain "module-level inline asm" blocks, which corresponds
1934to the GCC "file scope inline asm" blocks. These blocks are internally
1935concatenated by LLVM and treated as a single unit, but may be separated
1936in the ``.ll`` file if desired. The syntax is very simple:
1937
1938.. code-block:: llvm
1939
1940    module asm "inline asm code goes here"
1941    module asm "more can go here"
1942
1943The strings can contain any character by escaping non-printable
1944characters. The escape sequence used is simply "\\xx" where "xx" is the
1945two digit hex code for the number.
1946
1947Note that the assembly string *must* be parseable by LLVM's integrated assembler
1948(unless it is disabled), even when emitting a ``.s`` file.
1949
1950.. _langref_datalayout:
1951
1952Data Layout
1953-----------
1954
1955A module may specify a target specific data layout string that specifies
1956how data is to be laid out in memory. The syntax for the data layout is
1957simply:
1958
1959.. code-block:: llvm
1960
1961    target datalayout = "layout specification"
1962
1963The *layout specification* consists of a list of specifications
1964separated by the minus sign character ('-'). Each specification starts
1965with a letter and may include other information after the letter to
1966define some aspect of the data layout. The specifications accepted are
1967as follows:
1968
1969``E``
1970    Specifies that the target lays out data in big-endian form. That is,
1971    the bits with the most significance have the lowest address
1972    location.
1973``e``
1974    Specifies that the target lays out data in little-endian form. That
1975    is, the bits with the least significance have the lowest address
1976    location.
1977``S<size>``
1978    Specifies the natural alignment of the stack in bits. Alignment
1979    promotion of stack variables is limited to the natural stack
1980    alignment to avoid dynamic stack realignment. The stack alignment
1981    must be a multiple of 8-bits. If omitted, the natural stack
1982    alignment defaults to "unspecified", which does not prevent any
1983    alignment promotions.
1984``P<address space>``
1985    Specifies the address space that corresponds to program memory.
1986    Harvard architectures can use this to specify what space LLVM
1987    should place things such as functions into. If omitted, the
1988    program memory space defaults to the default address space of 0,
1989    which corresponds to a Von Neumann architecture that has code
1990    and data in the same space.
1991``A<address space>``
1992    Specifies the address space of objects created by '``alloca``'.
1993    Defaults to the default address space of 0.
1994``p[n]:<size>:<abi>:<pref>:<idx>``
1995    This specifies the *size* of a pointer and its ``<abi>`` and
1996    ``<pref>``\erred alignments for address space ``n``. The fourth parameter
1997    ``<idx>`` is a size of index that used for address calculation. If not
1998    specified, the default index size is equal to the pointer size. All sizes
1999    are in bits. The address space, ``n``, is optional, and if not specified,
2000    denotes the default address space 0. The value of ``n`` must be
2001    in the range [1,2^23).
2002``i<size>:<abi>:<pref>``
2003    This specifies the alignment for an integer type of a given bit
2004    ``<size>``. The value of ``<size>`` must be in the range [1,2^23).
2005``v<size>:<abi>:<pref>``
2006    This specifies the alignment for a vector type of a given bit
2007    ``<size>``.
2008``f<size>:<abi>:<pref>``
2009    This specifies the alignment for a floating-point type of a given bit
2010    ``<size>``. Only values of ``<size>`` that are supported by the target
2011    will work. 32 (float) and 64 (double) are supported on all targets; 80
2012    or 128 (different flavors of long double) are also supported on some
2013    targets.
2014``a:<abi>:<pref>``
2015    This specifies the alignment for an object of aggregate type.
2016``F<type><abi>``
2017    This specifies the alignment for function pointers.
2018    The options for ``<type>`` are:
2019
2020    * ``i``: The alignment of function pointers is independent of the alignment
2021      of functions, and is a multiple of ``<abi>``.
2022    * ``n``: The alignment of function pointers is a multiple of the explicit
2023      alignment specified on the function, and is a multiple of ``<abi>``.
2024``m:<mangling>``
2025    If present, specifies that llvm names are mangled in the output. Symbols
2026    prefixed with the mangling escape character ``\01`` are passed through
2027    directly to the assembler without the escape character. The mangling style
2028    options are
2029
2030    * ``e``: ELF mangling: Private symbols get a ``.L`` prefix.
2031    * ``m``: Mips mangling: Private symbols get a ``$`` prefix.
2032    * ``o``: Mach-O mangling: Private symbols get ``L`` prefix. Other
2033      symbols get a ``_`` prefix.
2034    * ``x``: Windows x86 COFF mangling: Private symbols get the usual prefix.
2035      Regular C symbols get a ``_`` prefix. Functions with ``__stdcall``,
2036      ``__fastcall``, and ``__vectorcall`` have custom mangling that appends
2037      ``@N`` where N is the number of bytes used to pass parameters. C++ symbols
2038      starting with ``?`` are not mangled in any way.
2039    * ``w``: Windows COFF mangling: Similar to ``x``, except that normal C
2040      symbols do not receive a ``_`` prefix.
2041``n<size1>:<size2>:<size3>...``
2042    This specifies a set of native integer widths for the target CPU in
2043    bits. For example, it might contain ``n32`` for 32-bit PowerPC,
2044    ``n32:64`` for PowerPC 64, or ``n8:16:32:64`` for X86-64. Elements of
2045    this set are considered to support most general arithmetic operations
2046    efficiently.
2047``ni:<address space0>:<address space1>:<address space2>...``
2048    This specifies pointer types with the specified address spaces
2049    as :ref:`Non-Integral Pointer Type <nointptrtype>` s.  The ``0``
2050    address space cannot be specified as non-integral.
2051
2052On every specification that takes a ``<abi>:<pref>``, specifying the
2053``<pref>`` alignment is optional. If omitted, the preceding ``:``
2054should be omitted too and ``<pref>`` will be equal to ``<abi>``.
2055
2056When constructing the data layout for a given target, LLVM starts with a
2057default set of specifications which are then (possibly) overridden by
2058the specifications in the ``datalayout`` keyword. The default
2059specifications are given in this list:
2060
2061-  ``E`` - big endian
2062-  ``p:64:64:64`` - 64-bit pointers with 64-bit alignment.
2063-  ``p[n]:64:64:64`` - Other address spaces are assumed to be the
2064   same as the default address space.
2065-  ``S0`` - natural stack alignment is unspecified
2066-  ``i1:8:8`` - i1 is 8-bit (byte) aligned
2067-  ``i8:8:8`` - i8 is 8-bit (byte) aligned
2068-  ``i16:16:16`` - i16 is 16-bit aligned
2069-  ``i32:32:32`` - i32 is 32-bit aligned
2070-  ``i64:32:64`` - i64 has ABI alignment of 32-bits but preferred
2071   alignment of 64-bits
2072-  ``f16:16:16`` - half is 16-bit aligned
2073-  ``f32:32:32`` - float is 32-bit aligned
2074-  ``f64:64:64`` - double is 64-bit aligned
2075-  ``f128:128:128`` - quad is 128-bit aligned
2076-  ``v64:64:64`` - 64-bit vector is 64-bit aligned
2077-  ``v128:128:128`` - 128-bit vector is 128-bit aligned
2078-  ``a:0:64`` - aggregates are 64-bit aligned
2079
2080When LLVM is determining the alignment for a given type, it uses the
2081following rules:
2082
2083#. If the type sought is an exact match for one of the specifications,
2084   that specification is used.
2085#. If no match is found, and the type sought is an integer type, then
2086   the smallest integer type that is larger than the bitwidth of the
2087   sought type is used. If none of the specifications are larger than
2088   the bitwidth then the largest integer type is used. For example,
2089   given the default specifications above, the i7 type will use the
2090   alignment of i8 (next largest) while both i65 and i256 will use the
2091   alignment of i64 (largest specified).
2092#. If no match is found, and the type sought is a vector type, then the
2093   largest vector type that is smaller than the sought vector type will
2094   be used as a fall back. This happens because <128 x double> can be
2095   implemented in terms of 64 <2 x double>, for example.
2096
2097The function of the data layout string may not be what you expect.
2098Notably, this is not a specification from the frontend of what alignment
2099the code generator should use.
2100
2101Instead, if specified, the target data layout is required to match what
2102the ultimate *code generator* expects. This string is used by the
2103mid-level optimizers to improve code, and this only works if it matches
2104what the ultimate code generator uses. There is no way to generate IR
2105that does not embed this target-specific detail into the IR. If you
2106don't specify the string, the default specifications will be used to
2107generate a Data Layout and the optimization phases will operate
2108accordingly and introduce target specificity into the IR with respect to
2109these default specifications.
2110
2111.. _langref_triple:
2112
2113Target Triple
2114-------------
2115
2116A module may specify a target triple string that describes the target
2117host. The syntax for the target triple is simply:
2118
2119.. code-block:: llvm
2120
2121    target triple = "x86_64-apple-macosx10.7.0"
2122
2123The *target triple* string consists of a series of identifiers delimited
2124by the minus sign character ('-'). The canonical forms are:
2125
2126::
2127
2128    ARCHITECTURE-VENDOR-OPERATING_SYSTEM
2129    ARCHITECTURE-VENDOR-OPERATING_SYSTEM-ENVIRONMENT
2130
2131This information is passed along to the backend so that it generates
2132code for the proper architecture. It's possible to override this on the
2133command line with the ``-mtriple`` command line option.
2134
2135.. _pointeraliasing:
2136
2137Pointer Aliasing Rules
2138----------------------
2139
2140Any memory access must be done through a pointer value associated with
2141an address range of the memory access, otherwise the behavior is
2142undefined. Pointer values are associated with address ranges according
2143to the following rules:
2144
2145-  A pointer value is associated with the addresses associated with any
2146   value it is *based* on.
2147-  An address of a global variable is associated with the address range
2148   of the variable's storage.
2149-  The result value of an allocation instruction is associated with the
2150   address range of the allocated storage.
2151-  A null pointer in the default address-space is associated with no
2152   address.
2153-  An integer constant other than zero or a pointer value returned from
2154   a function not defined within LLVM may be associated with address
2155   ranges allocated through mechanisms other than those provided by
2156   LLVM. Such ranges shall not overlap with any ranges of addresses
2157   allocated by mechanisms provided by LLVM.
2158
2159A pointer value is *based* on another pointer value according to the
2160following rules:
2161
2162-  A pointer value formed from a scalar ``getelementptr`` operation is *based* on
2163   the pointer-typed operand of the ``getelementptr``.
2164-  The pointer in lane *l* of the result of a vector ``getelementptr`` operation
2165   is *based* on the pointer in lane *l* of the vector-of-pointers-typed operand
2166   of the ``getelementptr``.
2167-  The result value of a ``bitcast`` is *based* on the operand of the
2168   ``bitcast``.
2169-  A pointer value formed by an ``inttoptr`` is *based* on all pointer
2170   values that contribute (directly or indirectly) to the computation of
2171   the pointer's value.
2172-  The "*based* on" relationship is transitive.
2173
2174Note that this definition of *"based"* is intentionally similar to the
2175definition of *"based"* in C99, though it is slightly weaker.
2176
2177LLVM IR does not associate types with memory. The result type of a
2178``load`` merely indicates the size and alignment of the memory from
2179which to load, as well as the interpretation of the value. The first
2180operand type of a ``store`` similarly only indicates the size and
2181alignment of the store.
2182
2183Consequently, type-based alias analysis, aka TBAA, aka
2184``-fstrict-aliasing``, is not applicable to general unadorned LLVM IR.
2185:ref:`Metadata <metadata>` may be used to encode additional information
2186which specialized optimization passes may use to implement type-based
2187alias analysis.
2188
2189.. _volatile:
2190
2191Volatile Memory Accesses
2192------------------------
2193
2194Certain memory accesses, such as :ref:`load <i_load>`'s,
2195:ref:`store <i_store>`'s, and :ref:`llvm.memcpy <int_memcpy>`'s may be
2196marked ``volatile``. The optimizers must not change the number of
2197volatile operations or change their order of execution relative to other
2198volatile operations. The optimizers *may* change the order of volatile
2199operations relative to non-volatile operations. This is not Java's
2200"volatile" and has no cross-thread synchronization behavior.
2201
2202A volatile load or store may have additional target-specific semantics.
2203Any volatile operation can have side effects, and any volatile operation
2204can read and/or modify state which is not accessible via a regular load
2205or store in this module. Volatile operations may use addresses which do
2206not point to memory (like MMIO registers). This means the compiler may
2207not use a volatile operation to prove a non-volatile access to that
2208address has defined behavior.
2209
2210The allowed side-effects for volatile accesses are limited.  If a
2211non-volatile store to a given address would be legal, a volatile
2212operation may modify the memory at that address. A volatile operation
2213may not modify any other memory accessible by the module being compiled.
2214A volatile operation may not call any code in the current module.
2215
2216The compiler may assume execution will continue after a volatile operation,
2217so operations which modify memory or may have undefined behavior can be
2218hoisted past a volatile operation.
2219
2220IR-level volatile loads and stores cannot safely be optimized into
2221llvm.memcpy or llvm.memmove intrinsics even when those intrinsics are
2222flagged volatile. Likewise, the backend should never split or merge
2223target-legal volatile load/store instructions.
2224
2225.. admonition:: Rationale
2226
2227 Platforms may rely on volatile loads and stores of natively supported
2228 data width to be executed as single instruction. For example, in C
2229 this holds for an l-value of volatile primitive type with native
2230 hardware support, but not necessarily for aggregate types. The
2231 frontend upholds these expectations, which are intentionally
2232 unspecified in the IR. The rules above ensure that IR transformations
2233 do not violate the frontend's contract with the language.
2234
2235.. _memmodel:
2236
2237Memory Model for Concurrent Operations
2238--------------------------------------
2239
2240The LLVM IR does not define any way to start parallel threads of
2241execution or to register signal handlers. Nonetheless, there are
2242platform-specific ways to create them, and we define LLVM IR's behavior
2243in their presence. This model is inspired by the C++0x memory model.
2244
2245For a more informal introduction to this model, see the :doc:`Atomics`.
2246
2247We define a *happens-before* partial order as the least partial order
2248that
2249
2250-  Is a superset of single-thread program order, and
2251-  When a *synchronizes-with* ``b``, includes an edge from ``a`` to
2252   ``b``. *Synchronizes-with* pairs are introduced by platform-specific
2253   techniques, like pthread locks, thread creation, thread joining,
2254   etc., and by atomic instructions. (See also :ref:`Atomic Memory Ordering
2255   Constraints <ordering>`).
2256
2257Note that program order does not introduce *happens-before* edges
2258between a thread and signals executing inside that thread.
2259
2260Every (defined) read operation (load instructions, memcpy, atomic
2261loads/read-modify-writes, etc.) R reads a series of bytes written by
2262(defined) write operations (store instructions, atomic
2263stores/read-modify-writes, memcpy, etc.). For the purposes of this
2264section, initialized globals are considered to have a write of the
2265initializer which is atomic and happens before any other read or write
2266of the memory in question. For each byte of a read R, R\ :sub:`byte`
2267may see any write to the same byte, except:
2268
2269-  If write\ :sub:`1`  happens before write\ :sub:`2`, and
2270   write\ :sub:`2` happens before R\ :sub:`byte`, then
2271   R\ :sub:`byte` does not see write\ :sub:`1`.
2272-  If R\ :sub:`byte` happens before write\ :sub:`3`, then
2273   R\ :sub:`byte` does not see write\ :sub:`3`.
2274
2275Given that definition, R\ :sub:`byte` is defined as follows:
2276
2277-  If R is volatile, the result is target-dependent. (Volatile is
2278   supposed to give guarantees which can support ``sig_atomic_t`` in
2279   C/C++, and may be used for accesses to addresses that do not behave
2280   like normal memory. It does not generally provide cross-thread
2281   synchronization.)
2282-  Otherwise, if there is no write to the same byte that happens before
2283   R\ :sub:`byte`, R\ :sub:`byte` returns ``undef`` for that byte.
2284-  Otherwise, if R\ :sub:`byte` may see exactly one write,
2285   R\ :sub:`byte` returns the value written by that write.
2286-  Otherwise, if R is atomic, and all the writes R\ :sub:`byte` may
2287   see are atomic, it chooses one of the values written. See the :ref:`Atomic
2288   Memory Ordering Constraints <ordering>` section for additional
2289   constraints on how the choice is made.
2290-  Otherwise R\ :sub:`byte` returns ``undef``.
2291
2292R returns the value composed of the series of bytes it read. This
2293implies that some bytes within the value may be ``undef`` **without**
2294the entire value being ``undef``. Note that this only defines the
2295semantics of the operation; it doesn't mean that targets will emit more
2296than one instruction to read the series of bytes.
2297
2298Note that in cases where none of the atomic intrinsics are used, this
2299model places only one restriction on IR transformations on top of what
2300is required for single-threaded execution: introducing a store to a byte
2301which might not otherwise be stored is not allowed in general.
2302(Specifically, in the case where another thread might write to and read
2303from an address, introducing a store can change a load that may see
2304exactly one write into a load that may see multiple writes.)
2305
2306.. _ordering:
2307
2308Atomic Memory Ordering Constraints
2309----------------------------------
2310
2311Atomic instructions (:ref:`cmpxchg <i_cmpxchg>`,
2312:ref:`atomicrmw <i_atomicrmw>`, :ref:`fence <i_fence>`,
2313:ref:`atomic load <i_load>`, and :ref:`atomic store <i_store>`) take
2314ordering parameters that determine which other atomic instructions on
2315the same address they *synchronize with*. These semantics are borrowed
2316from Java and C++0x, but are somewhat more colloquial. If these
2317descriptions aren't precise enough, check those specs (see spec
2318references in the :doc:`atomics guide <Atomics>`).
2319:ref:`fence <i_fence>` instructions treat these orderings somewhat
2320differently since they don't take an address. See that instruction's
2321documentation for details.
2322
2323For a simpler introduction to the ordering constraints, see the
2324:doc:`Atomics`.
2325
2326``unordered``
2327    The set of values that can be read is governed by the happens-before
2328    partial order. A value cannot be read unless some operation wrote
2329    it. This is intended to provide a guarantee strong enough to model
2330    Java's non-volatile shared variables. This ordering cannot be
2331    specified for read-modify-write operations; it is not strong enough
2332    to make them atomic in any interesting way.
2333``monotonic``
2334    In addition to the guarantees of ``unordered``, there is a single
2335    total order for modifications by ``monotonic`` operations on each
2336    address. All modification orders must be compatible with the
2337    happens-before order. There is no guarantee that the modification
2338    orders can be combined to a global total order for the whole program
2339    (and this often will not be possible). The read in an atomic
2340    read-modify-write operation (:ref:`cmpxchg <i_cmpxchg>` and
2341    :ref:`atomicrmw <i_atomicrmw>`) reads the value in the modification
2342    order immediately before the value it writes. If one atomic read
2343    happens before another atomic read of the same address, the later
2344    read must see the same value or a later value in the address's
2345    modification order. This disallows reordering of ``monotonic`` (or
2346    stronger) operations on the same address. If an address is written
2347    ``monotonic``-ally by one thread, and other threads ``monotonic``-ally
2348    read that address repeatedly, the other threads must eventually see
2349    the write. This corresponds to the C++0x/C1x
2350    ``memory_order_relaxed``.
2351``acquire``
2352    In addition to the guarantees of ``monotonic``, a
2353    *synchronizes-with* edge may be formed with a ``release`` operation.
2354    This is intended to model C++'s ``memory_order_acquire``.
2355``release``
2356    In addition to the guarantees of ``monotonic``, if this operation
2357    writes a value which is subsequently read by an ``acquire``
2358    operation, it *synchronizes-with* that operation. (This isn't a
2359    complete description; see the C++0x definition of a release
2360    sequence.) This corresponds to the C++0x/C1x
2361    ``memory_order_release``.
2362``acq_rel`` (acquire+release)
2363    Acts as both an ``acquire`` and ``release`` operation on its
2364    address. This corresponds to the C++0x/C1x ``memory_order_acq_rel``.
2365``seq_cst`` (sequentially consistent)
2366    In addition to the guarantees of ``acq_rel`` (``acquire`` for an
2367    operation that only reads, ``release`` for an operation that only
2368    writes), there is a global total order on all
2369    sequentially-consistent operations on all addresses, which is
2370    consistent with the *happens-before* partial order and with the
2371    modification orders of all the affected addresses. Each
2372    sequentially-consistent read sees the last preceding write to the
2373    same address in this global order. This corresponds to the C++0x/C1x
2374    ``memory_order_seq_cst`` and Java volatile.
2375
2376.. _syncscope:
2377
2378If an atomic operation is marked ``syncscope("singlethread")``, it only
2379*synchronizes with* and only participates in the seq\_cst total orderings of
2380other operations running in the same thread (for example, in signal handlers).
2381
2382If an atomic operation is marked ``syncscope("<target-scope>")``, where
2383``<target-scope>`` is a target specific synchronization scope, then it is target
2384dependent if it *synchronizes with* and participates in the seq\_cst total
2385orderings of other operations.
2386
2387Otherwise, an atomic operation that is not marked ``syncscope("singlethread")``
2388or ``syncscope("<target-scope>")`` *synchronizes with* and participates in the
2389seq\_cst total orderings of other operations that are not marked
2390``syncscope("singlethread")`` or ``syncscope("<target-scope>")``.
2391
2392.. _floatenv:
2393
2394Floating-Point Environment
2395--------------------------
2396
2397The default LLVM floating-point environment assumes that floating-point
2398instructions do not have side effects. Results assume the round-to-nearest
2399rounding mode. No floating-point exception state is maintained in this
2400environment. Therefore, there is no attempt to create or preserve invalid
2401operation (SNaN) or division-by-zero exceptions.
2402
2403The benefit of this exception-free assumption is that floating-point
2404operations may be speculated freely without any other fast-math relaxations
2405to the floating-point model.
2406
2407Code that requires different behavior than this should use the
2408:ref:`Constrained Floating-Point Intrinsics <constrainedfp>`.
2409
2410.. _fastmath:
2411
2412Fast-Math Flags
2413---------------
2414
2415LLVM IR floating-point operations (:ref:`fadd <i_fadd>`,
2416:ref:`fsub <i_fsub>`, :ref:`fmul <i_fmul>`, :ref:`fdiv <i_fdiv>`,
2417:ref:`frem <i_frem>`, :ref:`fcmp <i_fcmp>`) and :ref:`call <i_call>`
2418may use the following flags to enable otherwise unsafe
2419floating-point transformations.
2420
2421``nnan``
2422   No NaNs - Allow optimizations to assume the arguments and result are not
2423   NaN. If an argument is a nan, or the result would be a nan, it produces
2424   a :ref:`poison value <poisonvalues>` instead.
2425
2426``ninf``
2427   No Infs - Allow optimizations to assume the arguments and result are not
2428   +/-Inf. If an argument is +/-Inf, or the result would be +/-Inf, it
2429   produces a :ref:`poison value <poisonvalues>` instead.
2430
2431``nsz``
2432   No Signed Zeros - Allow optimizations to treat the sign of a zero
2433   argument or result as insignificant.
2434
2435``arcp``
2436   Allow Reciprocal - Allow optimizations to use the reciprocal of an
2437   argument rather than perform division.
2438
2439``contract``
2440   Allow floating-point contraction (e.g. fusing a multiply followed by an
2441   addition into a fused multiply-and-add).
2442
2443``afn``
2444   Approximate functions - Allow substitution of approximate calculations for
2445   functions (sin, log, sqrt, etc). See floating-point intrinsic definitions
2446   for places where this can apply to LLVM's intrinsic math functions.
2447
2448``reassoc``
2449   Allow reassociation transformations for floating-point instructions.
2450   This may dramatically change results in floating-point.
2451
2452``fast``
2453   This flag implies all of the others.
2454
2455.. _uselistorder:
2456
2457Use-list Order Directives
2458-------------------------
2459
2460Use-list directives encode the in-memory order of each use-list, allowing the
2461order to be recreated. ``<order-indexes>`` is a comma-separated list of
2462indexes that are assigned to the referenced value's uses. The referenced
2463value's use-list is immediately sorted by these indexes.
2464
2465Use-list directives may appear at function scope or global scope. They are not
2466instructions, and have no effect on the semantics of the IR. When they're at
2467function scope, they must appear after the terminator of the final basic block.
2468
2469If basic blocks have their address taken via ``blockaddress()`` expressions,
2470``uselistorder_bb`` can be used to reorder their use-lists from outside their
2471function's scope.
2472
2473:Syntax:
2474
2475::
2476
2477    uselistorder <ty> <value>, { <order-indexes> }
2478    uselistorder_bb @function, %block { <order-indexes> }
2479
2480:Examples:
2481
2482::
2483
2484    define void @foo(i32 %arg1, i32 %arg2) {
2485    entry:
2486      ; ... instructions ...
2487    bb:
2488      ; ... instructions ...
2489
2490      ; At function scope.
2491      uselistorder i32 %arg1, { 1, 0, 2 }
2492      uselistorder label %bb, { 1, 0 }
2493    }
2494
2495    ; At global scope.
2496    uselistorder i32* @global, { 1, 2, 0 }
2497    uselistorder i32 7, { 1, 0 }
2498    uselistorder i32 (i32) @bar, { 1, 0 }
2499    uselistorder_bb @foo, %bb, { 5, 1, 3, 2, 0, 4 }
2500
2501.. _source_filename:
2502
2503Source Filename
2504---------------
2505
2506The *source filename* string is set to the original module identifier,
2507which will be the name of the compiled source file when compiling from
2508source through the clang front end, for example. It is then preserved through
2509the IR and bitcode.
2510
2511This is currently necessary to generate a consistent unique global
2512identifier for local functions used in profile data, which prepends the
2513source file name to the local function name.
2514
2515The syntax for the source file name is simply:
2516
2517.. code-block:: text
2518
2519    source_filename = "/path/to/source.c"
2520
2521.. _typesystem:
2522
2523Type System
2524===========
2525
2526The LLVM type system is one of the most important features of the
2527intermediate representation. Being typed enables a number of
2528optimizations to be performed on the intermediate representation
2529directly, without having to do extra analyses on the side before the
2530transformation. A strong type system makes it easier to read the
2531generated code and enables novel analyses and transformations that are
2532not feasible to perform on normal three address code representations.
2533
2534.. _t_void:
2535
2536Void Type
2537---------
2538
2539:Overview:
2540
2541
2542The void type does not represent any value and has no size.
2543
2544:Syntax:
2545
2546
2547::
2548
2549      void
2550
2551
2552.. _t_function:
2553
2554Function Type
2555-------------
2556
2557:Overview:
2558
2559
2560The function type can be thought of as a function signature. It consists of a
2561return type and a list of formal parameter types. The return type of a function
2562type is a void type or first class type --- except for :ref:`label <t_label>`
2563and :ref:`metadata <t_metadata>` types.
2564
2565:Syntax:
2566
2567::
2568
2569      <returntype> (<parameter list>)
2570
2571...where '``<parameter list>``' is a comma-separated list of type
2572specifiers. Optionally, the parameter list may include a type ``...``, which
2573indicates that the function takes a variable number of arguments. Variable
2574argument functions can access their arguments with the :ref:`variable argument
2575handling intrinsic <int_varargs>` functions. '``<returntype>``' is any type
2576except :ref:`label <t_label>` and :ref:`metadata <t_metadata>`.
2577
2578:Examples:
2579
2580+---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2581| ``i32 (i32)``                   | function taking an ``i32``, returning an ``i32``                                                                                                                    |
2582+---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2583| ``float (i16, i32 *) *``        | :ref:`Pointer <t_pointer>` to a function that takes an ``i16`` and a :ref:`pointer <t_pointer>` to ``i32``, returning ``float``.                                    |
2584+---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2585| ``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. |
2586+---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2587| ``{i32, i32} (i32)``            | A function taking an ``i32``, returning a :ref:`structure <t_struct>` containing two ``i32`` values                                                                 |
2588+---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2589
2590.. _t_firstclass:
2591
2592First Class Types
2593-----------------
2594
2595The :ref:`first class <t_firstclass>` types are perhaps the most important.
2596Values of these types are the only ones which can be produced by
2597instructions.
2598
2599.. _t_single_value:
2600
2601Single Value Types
2602^^^^^^^^^^^^^^^^^^
2603
2604These are the types that are valid in registers from CodeGen's perspective.
2605
2606.. _t_integer:
2607
2608Integer Type
2609""""""""""""
2610
2611:Overview:
2612
2613The integer type is a very simple type that simply specifies an
2614arbitrary bit width for the integer type desired. Any bit width from 1
2615bit to 2\ :sup:`23`\ -1 (about 8 million) can be specified.
2616
2617:Syntax:
2618
2619::
2620
2621      iN
2622
2623The number of bits the integer will occupy is specified by the ``N``
2624value.
2625
2626Examples:
2627*********
2628
2629+----------------+------------------------------------------------+
2630| ``i1``         | a single-bit integer.                          |
2631+----------------+------------------------------------------------+
2632| ``i32``        | a 32-bit integer.                              |
2633+----------------+------------------------------------------------+
2634| ``i1942652``   | a really big integer of over 1 million bits.   |
2635+----------------+------------------------------------------------+
2636
2637.. _t_floating:
2638
2639Floating-Point Types
2640""""""""""""""""""""
2641
2642.. list-table::
2643   :header-rows: 1
2644
2645   * - Type
2646     - Description
2647
2648   * - ``half``
2649     - 16-bit floating-point value
2650
2651   * - ``float``
2652     - 32-bit floating-point value
2653
2654   * - ``double``
2655     - 64-bit floating-point value
2656
2657   * - ``fp128``
2658     - 128-bit floating-point value (112-bit mantissa)
2659
2660   * - ``x86_fp80``
2661     -  80-bit floating-point value (X87)
2662
2663   * - ``ppc_fp128``
2664     - 128-bit floating-point value (two 64-bits)
2665
2666The binary format of half, float, double, and fp128 correspond to the
2667IEEE-754-2008 specifications for binary16, binary32, binary64, and binary128
2668respectively.
2669
2670X86_mmx Type
2671""""""""""""
2672
2673:Overview:
2674
2675The x86_mmx type represents a value held in an MMX register on an x86
2676machine. The operations allowed on it are quite limited: parameters and
2677return values, load and store, and bitcast. User-specified MMX
2678instructions are represented as intrinsic or asm calls with arguments
2679and/or results of this type. There are no arrays, vectors or constants
2680of this type.
2681
2682:Syntax:
2683
2684::
2685
2686      x86_mmx
2687
2688
2689.. _t_pointer:
2690
2691Pointer Type
2692""""""""""""
2693
2694:Overview:
2695
2696The pointer type is used to specify memory locations. Pointers are
2697commonly used to reference objects in memory.
2698
2699Pointer types may have an optional address space attribute defining the
2700numbered address space where the pointed-to object resides. The default
2701address space is number zero. The semantics of non-zero address spaces
2702are target-specific.
2703
2704Note that LLVM does not permit pointers to void (``void*``) nor does it
2705permit pointers to labels (``label*``). Use ``i8*`` instead.
2706
2707:Syntax:
2708
2709::
2710
2711      <type> *
2712
2713:Examples:
2714
2715+-------------------------+--------------------------------------------------------------------------------------------------------------+
2716| ``[4 x i32]*``          | A :ref:`pointer <t_pointer>` to :ref:`array <t_array>` of four ``i32`` values.                               |
2717+-------------------------+--------------------------------------------------------------------------------------------------------------+
2718| ``i32 (i32*) *``        | A :ref:`pointer <t_pointer>` to a :ref:`function <t_function>` that takes an ``i32*``, returning an ``i32``. |
2719+-------------------------+--------------------------------------------------------------------------------------------------------------+
2720| ``i32 addrspace(5)*``   | A :ref:`pointer <t_pointer>` to an ``i32`` value that resides in address space #5.                           |
2721+-------------------------+--------------------------------------------------------------------------------------------------------------+
2722
2723.. _t_vector:
2724
2725Vector Type
2726"""""""""""
2727
2728:Overview:
2729
2730A vector type is a simple derived type that represents a vector of
2731elements. Vector types are used when multiple primitive data are
2732operated in parallel using a single instruction (SIMD). A vector type
2733requires a size (number of elements) and an underlying primitive data
2734type. Vector types are considered :ref:`first class <t_firstclass>`.
2735
2736:Syntax:
2737
2738::
2739
2740      < <# elements> x <elementtype> >
2741
2742The number of elements is a constant integer value larger than 0;
2743elementtype may be any integer, floating-point or pointer type. Vectors
2744of size zero are not allowed.
2745
2746:Examples:
2747
2748+-------------------+--------------------------------------------------+
2749| ``<4 x i32>``     | Vector of 4 32-bit integer values.               |
2750+-------------------+--------------------------------------------------+
2751| ``<8 x float>``   | Vector of 8 32-bit floating-point values.        |
2752+-------------------+--------------------------------------------------+
2753| ``<2 x i64>``     | Vector of 2 64-bit integer values.               |
2754+-------------------+--------------------------------------------------+
2755| ``<4 x i64*>``    | Vector of 4 pointers to 64-bit integer values.   |
2756+-------------------+--------------------------------------------------+
2757
2758.. _t_label:
2759
2760Label Type
2761^^^^^^^^^^
2762
2763:Overview:
2764
2765The label type represents code labels.
2766
2767:Syntax:
2768
2769::
2770
2771      label
2772
2773.. _t_token:
2774
2775Token Type
2776^^^^^^^^^^
2777
2778:Overview:
2779
2780The token type is used when a value is associated with an instruction
2781but all uses of the value must not attempt to introspect or obscure it.
2782As such, it is not appropriate to have a :ref:`phi <i_phi>` or
2783:ref:`select <i_select>` of type token.
2784
2785:Syntax:
2786
2787::
2788
2789      token
2790
2791
2792
2793.. _t_metadata:
2794
2795Metadata Type
2796^^^^^^^^^^^^^
2797
2798:Overview:
2799
2800The metadata type represents embedded metadata. No derived types may be
2801created from metadata except for :ref:`function <t_function>` arguments.
2802
2803:Syntax:
2804
2805::
2806
2807      metadata
2808
2809.. _t_aggregate:
2810
2811Aggregate Types
2812^^^^^^^^^^^^^^^
2813
2814Aggregate Types are a subset of derived types that can contain multiple
2815member types. :ref:`Arrays <t_array>` and :ref:`structs <t_struct>` are
2816aggregate types. :ref:`Vectors <t_vector>` are not considered to be
2817aggregate types.
2818
2819.. _t_array:
2820
2821Array Type
2822""""""""""
2823
2824:Overview:
2825
2826The array type is a very simple derived type that arranges elements
2827sequentially in memory. The array type requires a size (number of
2828elements) and an underlying data type.
2829
2830:Syntax:
2831
2832::
2833
2834      [<# elements> x <elementtype>]
2835
2836The number of elements is a constant integer value; ``elementtype`` may
2837be any type with a size.
2838
2839:Examples:
2840
2841+------------------+--------------------------------------+
2842| ``[40 x i32]``   | Array of 40 32-bit integer values.   |
2843+------------------+--------------------------------------+
2844| ``[41 x i32]``   | Array of 41 32-bit integer values.   |
2845+------------------+--------------------------------------+
2846| ``[4 x i8]``     | Array of 4 8-bit integer values.     |
2847+------------------+--------------------------------------+
2848
2849Here are some examples of multidimensional arrays:
2850
2851+-----------------------------+----------------------------------------------------------+
2852| ``[3 x [4 x i32]]``         | 3x4 array of 32-bit integer values.                      |
2853+-----------------------------+----------------------------------------------------------+
2854| ``[12 x [10 x float]]``     | 12x10 array of single precision floating-point values.   |
2855+-----------------------------+----------------------------------------------------------+
2856| ``[2 x [3 x [4 x i16]]]``   | 2x3x4 array of 16-bit integer values.                    |
2857+-----------------------------+----------------------------------------------------------+
2858
2859There is no restriction on indexing beyond the end of the array implied
2860by a static type (though there are restrictions on indexing beyond the
2861bounds of an allocated object in some cases). This means that
2862single-dimension 'variable sized array' addressing can be implemented in
2863LLVM with a zero length array type. An implementation of 'pascal style
2864arrays' in LLVM could use the type "``{ i32, [0 x float]}``", for
2865example.
2866
2867.. _t_struct:
2868
2869Structure Type
2870""""""""""""""
2871
2872:Overview:
2873
2874The structure type is used to represent a collection of data members
2875together in memory. The elements of a structure may be any type that has
2876a size.
2877
2878Structures in memory are accessed using '``load``' and '``store``' by
2879getting a pointer to a field with the '``getelementptr``' instruction.
2880Structures in registers are accessed using the '``extractvalue``' and
2881'``insertvalue``' instructions.
2882
2883Structures may optionally be "packed" structures, which indicate that
2884the alignment of the struct is one byte, and that there is no padding
2885between the elements. In non-packed structs, padding between field types
2886is inserted as defined by the DataLayout string in the module, which is
2887required to match what the underlying code generator expects.
2888
2889Structures can either be "literal" or "identified". A literal structure
2890is defined inline with other types (e.g. ``{i32, i32}*``) whereas
2891identified types are always defined at the top level with a name.
2892Literal types are uniqued by their contents and can never be recursive
2893or opaque since there is no way to write one. Identified types can be
2894recursive, can be opaqued, and are never uniqued.
2895
2896:Syntax:
2897
2898::
2899
2900      %T1 = type { <type list> }     ; Identified normal struct type
2901      %T2 = type <{ <type list> }>   ; Identified packed struct type
2902
2903:Examples:
2904
2905+------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2906| ``{ i32, i32, i32 }``        | A triple of three ``i32`` values                                                                                                                                                      |
2907+------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2908| ``{ 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``.  |
2909+------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2910| ``<{ i8, i32 }>``            | A packed struct known to be 5 bytes in size.                                                                                                                                          |
2911+------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2912
2913.. _t_opaque:
2914
2915Opaque Structure Types
2916""""""""""""""""""""""
2917
2918:Overview:
2919
2920Opaque structure types are used to represent named structure types that
2921do not have a body specified. This corresponds (for example) to the C
2922notion of a forward declared structure.
2923
2924:Syntax:
2925
2926::
2927
2928      %X = type opaque
2929      %52 = type opaque
2930
2931:Examples:
2932
2933+--------------+-------------------+
2934| ``opaque``   | An opaque type.   |
2935+--------------+-------------------+
2936
2937.. _constants:
2938
2939Constants
2940=========
2941
2942LLVM has several different basic types of constants. This section
2943describes them all and their syntax.
2944
2945Simple Constants
2946----------------
2947
2948**Boolean constants**
2949    The two strings '``true``' and '``false``' are both valid constants
2950    of the ``i1`` type.
2951**Integer constants**
2952    Standard integers (such as '4') are constants of the
2953    :ref:`integer <t_integer>` type. Negative numbers may be used with
2954    integer types.
2955**Floating-point constants**
2956    Floating-point constants use standard decimal notation (e.g.
2957    123.421), exponential notation (e.g. 1.23421e+2), or a more precise
2958    hexadecimal notation (see below). The assembler requires the exact
2959    decimal value of a floating-point constant. For example, the
2960    assembler accepts 1.25 but rejects 1.3 because 1.3 is a repeating
2961    decimal in binary. Floating-point constants must have a
2962    :ref:`floating-point <t_floating>` type.
2963**Null pointer constants**
2964    The identifier '``null``' is recognized as a null pointer constant
2965    and must be of :ref:`pointer type <t_pointer>`.
2966**Token constants**
2967    The identifier '``none``' is recognized as an empty token constant
2968    and must be of :ref:`token type <t_token>`.
2969
2970The one non-intuitive notation for constants is the hexadecimal form of
2971floating-point constants. For example, the form
2972'``double    0x432ff973cafa8000``' is equivalent to (but harder to read
2973than) '``double 4.5e+15``'. The only time hexadecimal floating-point
2974constants are required (and the only time that they are generated by the
2975disassembler) is when a floating-point constant must be emitted but it
2976cannot be represented as a decimal floating-point number in a reasonable
2977number of digits. For example, NaN's, infinities, and other special
2978values are represented in their IEEE hexadecimal format so that assembly
2979and disassembly do not cause any bits to change in the constants.
2980
2981When using the hexadecimal form, constants of types half, float, and
2982double are represented using the 16-digit form shown above (which
2983matches the IEEE754 representation for double); half and float values
2984must, however, be exactly representable as IEEE 754 half and single
2985precision, respectively. Hexadecimal format is always used for long
2986double, and there are three forms of long double. The 80-bit format used
2987by x86 is represented as ``0xK`` followed by 20 hexadecimal digits. The
2988128-bit format used by PowerPC (two adjacent doubles) is represented by
2989``0xM`` followed by 32 hexadecimal digits. The IEEE 128-bit format is
2990represented by ``0xL`` followed by 32 hexadecimal digits. Long doubles
2991will only work if they match the long double format on your target.
2992The IEEE 16-bit format (half precision) is represented by ``0xH``
2993followed by 4 hexadecimal digits. All hexadecimal formats are big-endian
2994(sign bit at the left).
2995
2996There are no constants of type x86_mmx.
2997
2998.. _complexconstants:
2999
3000Complex Constants
3001-----------------
3002
3003Complex constants are a (potentially recursive) combination of simple
3004constants and smaller complex constants.
3005
3006**Structure constants**
3007    Structure constants are represented with notation similar to
3008    structure type definitions (a comma separated list of elements,
3009    surrounded by braces (``{}``)). For example:
3010    "``{ i32 4, float 17.0, i32* @G }``", where "``@G``" is declared as
3011    "``@G = external global i32``". Structure constants must have
3012    :ref:`structure type <t_struct>`, and the number and types of elements
3013    must match those specified by the type.
3014**Array constants**
3015    Array constants are represented with notation similar to array type
3016    definitions (a comma separated list of elements, surrounded by
3017    square brackets (``[]``)). For example:
3018    "``[ i32 42, i32 11, i32 74 ]``". Array constants must have
3019    :ref:`array type <t_array>`, and the number and types of elements must
3020    match those specified by the type. As a special case, character array
3021    constants may also be represented as a double-quoted string using the ``c``
3022    prefix. For example: "``c"Hello World\0A\00"``".
3023**Vector constants**
3024    Vector constants are represented with notation similar to vector
3025    type definitions (a comma separated list of elements, surrounded by
3026    less-than/greater-than's (``<>``)). For example:
3027    "``< i32 42, i32 11, i32 74, i32 100 >``". Vector constants
3028    must have :ref:`vector type <t_vector>`, and the number and types of
3029    elements must match those specified by the type.
3030**Zero initialization**
3031    The string '``zeroinitializer``' can be used to zero initialize a
3032    value to zero of *any* type, including scalar and
3033    :ref:`aggregate <t_aggregate>` types. This is often used to avoid
3034    having to print large zero initializers (e.g. for large arrays) and
3035    is always exactly equivalent to using explicit zero initializers.
3036**Metadata node**
3037    A metadata node is a constant tuple without types. For example:
3038    "``!{!0, !{!2, !0}, !"test"}``". Metadata can reference constant values,
3039    for example: "``!{!0, i32 0, i8* @global, i64 (i64)* @function, !"str"}``".
3040    Unlike other typed constants that are meant to be interpreted as part of
3041    the instruction stream, metadata is a place to attach additional
3042    information such as debug info.
3043
3044Global Variable and Function Addresses
3045--------------------------------------
3046
3047The addresses of :ref:`global variables <globalvars>` and
3048:ref:`functions <functionstructure>` are always implicitly valid
3049(link-time) constants. These constants are explicitly referenced when
3050the :ref:`identifier for the global <identifiers>` is used and always have
3051:ref:`pointer <t_pointer>` type. For example, the following is a legal LLVM
3052file:
3053
3054.. code-block:: llvm
3055
3056    @X = global i32 17
3057    @Y = global i32 42
3058    @Z = global [2 x i32*] [ i32* @X, i32* @Y ]
3059
3060.. _undefvalues:
3061
3062Undefined Values
3063----------------
3064
3065The string '``undef``' can be used anywhere a constant is expected, and
3066indicates that the user of the value may receive an unspecified
3067bit-pattern. Undefined values may be of any type (other than '``label``'
3068or '``void``') and be used anywhere a constant is permitted.
3069
3070Undefined values are useful because they indicate to the compiler that
3071the program is well defined no matter what value is used. This gives the
3072compiler more freedom to optimize. Here are some examples of
3073(potentially surprising) transformations that are valid (in pseudo IR):
3074
3075.. code-block:: llvm
3076
3077      %A = add %X, undef
3078      %B = sub %X, undef
3079      %C = xor %X, undef
3080    Safe:
3081      %A = undef
3082      %B = undef
3083      %C = undef
3084
3085This is safe because all of the output bits are affected by the undef
3086bits. Any output bit can have a zero or one depending on the input bits.
3087
3088.. code-block:: llvm
3089
3090      %A = or %X, undef
3091      %B = and %X, undef
3092    Safe:
3093      %A = -1
3094      %B = 0
3095    Safe:
3096      %A = %X  ;; By choosing undef as 0
3097      %B = %X  ;; By choosing undef as -1
3098    Unsafe:
3099      %A = undef
3100      %B = undef
3101
3102These logical operations have bits that are not always affected by the
3103input. For example, if ``%X`` has a zero bit, then the output of the
3104'``and``' operation will always be a zero for that bit, no matter what
3105the corresponding bit from the '``undef``' is. As such, it is unsafe to
3106optimize or assume that the result of the '``and``' is '``undef``'.
3107However, it is safe to assume that all bits of the '``undef``' could be
31080, and optimize the '``and``' to 0. Likewise, it is safe to assume that
3109all the bits of the '``undef``' operand to the '``or``' could be set,
3110allowing the '``or``' to be folded to -1.
3111
3112.. code-block:: llvm
3113
3114      %A = select undef, %X, %Y
3115      %B = select undef, 42, %Y
3116      %C = select %X, %Y, undef
3117    Safe:
3118      %A = %X     (or %Y)
3119      %B = 42     (or %Y)
3120      %C = %Y
3121    Unsafe:
3122      %A = undef
3123      %B = undef
3124      %C = undef
3125
3126This set of examples shows that undefined '``select``' (and conditional
3127branch) conditions can go *either way*, but they have to come from one
3128of the two operands. In the ``%A`` example, if ``%X`` and ``%Y`` were
3129both known to have a clear low bit, then ``%A`` would have to have a
3130cleared low bit. However, in the ``%C`` example, the optimizer is
3131allowed to assume that the '``undef``' operand could be the same as
3132``%Y``, allowing the whole '``select``' to be eliminated.
3133
3134.. code-block:: text
3135
3136      %A = xor undef, undef
3137
3138      %B = undef
3139      %C = xor %B, %B
3140
3141      %D = undef
3142      %E = icmp slt %D, 4
3143      %F = icmp gte %D, 4
3144
3145    Safe:
3146      %A = undef
3147      %B = undef
3148      %C = undef
3149      %D = undef
3150      %E = undef
3151      %F = undef
3152
3153This example points out that two '``undef``' operands are not
3154necessarily the same. This can be surprising to people (and also matches
3155C semantics) where they assume that "``X^X``" is always zero, even if
3156``X`` is undefined. This isn't true for a number of reasons, but the
3157short answer is that an '``undef``' "variable" can arbitrarily change
3158its value over its "live range". This is true because the variable
3159doesn't actually *have a live range*. Instead, the value is logically
3160read from arbitrary registers that happen to be around when needed, so
3161the value is not necessarily consistent over time. In fact, ``%A`` and
3162``%C`` need to have the same semantics or the core LLVM "replace all
3163uses with" concept would not hold.
3164
3165.. code-block:: llvm
3166
3167      %A = sdiv undef, %X
3168      %B = sdiv %X, undef
3169    Safe:
3170      %A = 0
3171    b: unreachable
3172
3173These examples show the crucial difference between an *undefined value*
3174and *undefined behavior*. An undefined value (like '``undef``') is
3175allowed to have an arbitrary bit-pattern. This means that the ``%A``
3176operation can be constant folded to '``0``', because the '``undef``'
3177could be zero, and zero divided by any value is zero.
3178However, in the second example, we can make a more aggressive
3179assumption: because the ``undef`` is allowed to be an arbitrary value,
3180we are allowed to assume that it could be zero. Since a divide by zero
3181has *undefined behavior*, we are allowed to assume that the operation
3182does not execute at all. This allows us to delete the divide and all
3183code after it. Because the undefined operation "can't happen", the
3184optimizer can assume that it occurs in dead code.
3185
3186.. code-block:: text
3187
3188    a:  store undef -> %X
3189    b:  store %X -> undef
3190    Safe:
3191    a: <deleted>
3192    b: unreachable
3193
3194A store *of* an undefined value can be assumed to not have any effect;
3195we can assume that the value is overwritten with bits that happen to
3196match what was already there. However, a store *to* an undefined
3197location could clobber arbitrary memory, therefore, it has undefined
3198behavior.
3199
3200.. _poisonvalues:
3201
3202Poison Values
3203-------------
3204
3205Poison values are similar to :ref:`undef values <undefvalues>`, however
3206they also represent the fact that an instruction or constant expression
3207that cannot evoke side effects has nevertheless detected a condition
3208that results in undefined behavior.
3209
3210There is currently no way of representing a poison value in the IR; they
3211only exist when produced by operations such as :ref:`add <i_add>` with
3212the ``nsw`` flag.
3213
3214Poison value behavior is defined in terms of value *dependence*:
3215
3216-  Values other than :ref:`phi <i_phi>` nodes depend on their operands.
3217-  :ref:`Phi <i_phi>` nodes depend on the operand corresponding to
3218   their dynamic predecessor basic block.
3219-  Function arguments depend on the corresponding actual argument values
3220   in the dynamic callers of their functions.
3221-  :ref:`Call <i_call>` instructions depend on the :ref:`ret <i_ret>`
3222   instructions that dynamically transfer control back to them.
3223-  :ref:`Invoke <i_invoke>` instructions depend on the
3224   :ref:`ret <i_ret>`, :ref:`resume <i_resume>`, or exception-throwing
3225   call instructions that dynamically transfer control back to them.
3226-  Non-volatile loads and stores depend on the most recent stores to all
3227   of the referenced memory addresses, following the order in the IR
3228   (including loads and stores implied by intrinsics such as
3229   :ref:`@llvm.memcpy <int_memcpy>`.)
3230-  An instruction with externally visible side effects depends on the
3231   most recent preceding instruction with externally visible side
3232   effects, following the order in the IR. (This includes :ref:`volatile
3233   operations <volatile>`.)
3234-  An instruction *control-depends* on a :ref:`terminator
3235   instruction <terminators>` if the terminator instruction has
3236   multiple successors and the instruction is always executed when
3237   control transfers to one of the successors, and may not be executed
3238   when control is transferred to another.
3239-  Additionally, an instruction also *control-depends* on a terminator
3240   instruction if the set of instructions it otherwise depends on would
3241   be different if the terminator had transferred control to a different
3242   successor.
3243-  Dependence is transitive.
3244
3245Poison values have the same behavior as :ref:`undef values <undefvalues>`,
3246with the additional effect that any instruction that has a *dependence*
3247on a poison value has undefined behavior.
3248
3249Here are some examples:
3250
3251.. code-block:: llvm
3252
3253    entry:
3254      %poison = sub nuw i32 0, 1           ; Results in a poison value.
3255      %still_poison = and i32 %poison, 0   ; 0, but also poison.
3256      %poison_yet_again = getelementptr i32, i32* @h, i32 %still_poison
3257      store i32 0, i32* %poison_yet_again  ; memory at @h[0] is poisoned
3258
3259      store i32 %poison, i32* @g           ; Poison value stored to memory.
3260      %poison2 = load i32, i32* @g         ; Poison value loaded back from memory.
3261
3262      store volatile i32 %poison, i32* @g  ; External observation; undefined behavior.
3263
3264      %narrowaddr = bitcast i32* @g to i16*
3265      %wideaddr = bitcast i32* @g to i64*
3266      %poison3 = load i16, i16* %narrowaddr ; Returns a poison value.
3267      %poison4 = load i64, i64* %wideaddr  ; Returns a poison value.
3268
3269      %cmp = icmp slt i32 %poison, 0       ; Returns a poison value.
3270      br i1 %cmp, label %true, label %end  ; Branch to either destination.
3271
3272    true:
3273      store volatile i32 0, i32* @g        ; This is control-dependent on %cmp, so
3274                                           ; it has undefined behavior.
3275      br label %end
3276
3277    end:
3278      %p = phi i32 [ 0, %entry ], [ 1, %true ]
3279                                           ; Both edges into this PHI are
3280                                           ; control-dependent on %cmp, so this
3281                                           ; always results in a poison value.
3282
3283      store volatile i32 0, i32* @g        ; This would depend on the store in %true
3284                                           ; if %cmp is true, or the store in %entry
3285                                           ; otherwise, so this is undefined behavior.
3286
3287      br i1 %cmp, label %second_true, label %second_end
3288                                           ; The same branch again, but this time the
3289                                           ; true block doesn't have side effects.
3290
3291    second_true:
3292      ; No side effects!
3293      ret void
3294
3295    second_end:
3296      store volatile i32 0, i32* @g        ; This time, the instruction always depends
3297                                           ; on the store in %end. Also, it is
3298                                           ; control-equivalent to %end, so this is
3299                                           ; well-defined (ignoring earlier undefined
3300                                           ; behavior in this example).
3301
3302.. _blockaddress:
3303
3304Addresses of Basic Blocks
3305-------------------------
3306
3307``blockaddress(@function, %block)``
3308
3309The '``blockaddress``' constant computes the address of the specified
3310basic block in the specified function, and always has an ``i8*`` type.
3311Taking the address of the entry block is illegal.
3312
3313This value only has defined behavior when used as an operand to the
3314':ref:`indirectbr <i_indirectbr>`' or ':ref:`callbr <i_callbr>`'instruction, or
3315for comparisons against null. Pointer equality tests between labels addresses
3316results in undefined behavior --- though, again, comparison against null is ok,
3317and no label is equal to the null pointer. This may be passed around as an
3318opaque pointer sized value as long as the bits are not inspected. This
3319allows ``ptrtoint`` and arithmetic to be performed on these values so
3320long as the original value is reconstituted before the ``indirectbr`` or
3321``callbr`` instruction.
3322
3323Finally, some targets may provide defined semantics when using the value
3324as the operand to an inline assembly, but that is target specific.
3325
3326.. _constantexprs:
3327
3328Constant Expressions
3329--------------------
3330
3331Constant expressions are used to allow expressions involving other
3332constants to be used as constants. Constant expressions may be of any
3333:ref:`first class <t_firstclass>` type and may involve any LLVM operation
3334that does not have side effects (e.g. load and call are not supported).
3335The following is the syntax for constant expressions:
3336
3337``trunc (CST to TYPE)``
3338    Perform the :ref:`trunc operation <i_trunc>` on constants.
3339``zext (CST to TYPE)``
3340    Perform the :ref:`zext operation <i_zext>` on constants.
3341``sext (CST to TYPE)``
3342    Perform the :ref:`sext operation <i_sext>` on constants.
3343``fptrunc (CST to TYPE)``
3344    Truncate a floating-point constant to another floating-point type.
3345    The size of CST must be larger than the size of TYPE. Both types
3346    must be floating-point.
3347``fpext (CST to TYPE)``
3348    Floating-point extend a constant to another type. The size of CST
3349    must be smaller or equal to the size of TYPE. Both types must be
3350    floating-point.
3351``fptoui (CST to TYPE)``
3352    Convert a floating-point constant to the corresponding unsigned
3353    integer constant. TYPE must be a scalar or vector integer type. CST
3354    must be of scalar or vector floating-point type. Both CST and TYPE
3355    must be scalars, or vectors of the same number of elements. If the
3356    value won't fit in the integer type, the result is a
3357    :ref:`poison value <poisonvalues>`.
3358``fptosi (CST to TYPE)``
3359    Convert a floating-point constant to the corresponding signed
3360    integer constant. TYPE must be a scalar or vector integer type. CST
3361    must be of scalar or vector floating-point type. Both CST and TYPE
3362    must be scalars, or vectors of the same number of elements. If the
3363    value won't fit in the integer type, the result is a
3364    :ref:`poison value <poisonvalues>`.
3365``uitofp (CST to TYPE)``
3366    Convert an unsigned integer constant to the corresponding
3367    floating-point constant. TYPE must be a scalar or vector floating-point
3368    type.  CST must be of scalar or vector integer type. Both CST and TYPE must
3369    be scalars, or vectors of the same number of elements.
3370``sitofp (CST to TYPE)``
3371    Convert a signed integer constant to the corresponding floating-point
3372    constant. TYPE must be a scalar or vector floating-point type.
3373    CST must be of scalar or vector integer type. Both CST and TYPE must
3374    be scalars, or vectors of the same number of elements.
3375``ptrtoint (CST to TYPE)``
3376    Perform the :ref:`ptrtoint operation <i_ptrtoint>` on constants.
3377``inttoptr (CST to TYPE)``
3378    Perform the :ref:`inttoptr operation <i_inttoptr>` on constants.
3379    This one is *really* dangerous!
3380``bitcast (CST to TYPE)``
3381    Convert a constant, CST, to another TYPE.
3382    The constraints of the operands are the same as those for the
3383    :ref:`bitcast instruction <i_bitcast>`.
3384``addrspacecast (CST to TYPE)``
3385    Convert a constant pointer or constant vector of pointer, CST, to another
3386    TYPE in a different address space. The constraints of the operands are the
3387    same as those for the :ref:`addrspacecast instruction <i_addrspacecast>`.
3388``getelementptr (TY, CSTPTR, IDX0, IDX1, ...)``, ``getelementptr inbounds (TY, CSTPTR, IDX0, IDX1, ...)``
3389    Perform the :ref:`getelementptr operation <i_getelementptr>` on
3390    constants. As with the :ref:`getelementptr <i_getelementptr>`
3391    instruction, the index list may have one or more indexes, which are
3392    required to make sense for the type of "pointer to TY".
3393``select (COND, VAL1, VAL2)``
3394    Perform the :ref:`select operation <i_select>` on constants.
3395``icmp COND (VAL1, VAL2)``
3396    Perform the :ref:`icmp operation <i_icmp>` on constants.
3397``fcmp COND (VAL1, VAL2)``
3398    Perform the :ref:`fcmp operation <i_fcmp>` on constants.
3399``extractelement (VAL, IDX)``
3400    Perform the :ref:`extractelement operation <i_extractelement>` on
3401    constants.
3402``insertelement (VAL, ELT, IDX)``
3403    Perform the :ref:`insertelement operation <i_insertelement>` on
3404    constants.
3405``shufflevector (VEC1, VEC2, IDXMASK)``
3406    Perform the :ref:`shufflevector operation <i_shufflevector>` on
3407    constants.
3408``extractvalue (VAL, IDX0, IDX1, ...)``
3409    Perform the :ref:`extractvalue operation <i_extractvalue>` on
3410    constants. The index list is interpreted in a similar manner as
3411    indices in a ':ref:`getelementptr <i_getelementptr>`' operation. At
3412    least one index value must be specified.
3413``insertvalue (VAL, ELT, IDX0, IDX1, ...)``
3414    Perform the :ref:`insertvalue operation <i_insertvalue>` on constants.
3415    The index list is interpreted in a similar manner as indices in a
3416    ':ref:`getelementptr <i_getelementptr>`' operation. At least one index
3417    value must be specified.
3418``OPCODE (LHS, RHS)``
3419    Perform the specified operation of the LHS and RHS constants. OPCODE
3420    may be any of the :ref:`binary <binaryops>` or :ref:`bitwise
3421    binary <bitwiseops>` operations. The constraints on operands are
3422    the same as those for the corresponding instruction (e.g. no bitwise
3423    operations on floating-point values are allowed).
3424
3425Other Values
3426============
3427
3428.. _inlineasmexprs:
3429
3430Inline Assembler Expressions
3431----------------------------
3432
3433LLVM supports inline assembler expressions (as opposed to :ref:`Module-Level
3434Inline Assembly <moduleasm>`) through the use of a special value. This value
3435represents the inline assembler as a template string (containing the
3436instructions to emit), a list of operand constraints (stored as a string), a
3437flag that indicates whether or not the inline asm expression has side effects,
3438and a flag indicating whether the function containing the asm needs to align its
3439stack conservatively.
3440
3441The template string supports argument substitution of the operands using "``$``"
3442followed by a number, to indicate substitution of the given register/memory
3443location, as specified by the constraint string. "``${NUM:MODIFIER}``" may also
3444be used, where ``MODIFIER`` is a target-specific annotation for how to print the
3445operand (See :ref:`inline-asm-modifiers`).
3446
3447A literal "``$``" may be included by using "``$$``" in the template. To include
3448other special characters into the output, the usual "``\XX``" escapes may be
3449used, just as in other strings. Note that after template substitution, the
3450resulting assembly string is parsed by LLVM's integrated assembler unless it is
3451disabled -- even when emitting a ``.s`` file -- and thus must contain assembly
3452syntax known to LLVM.
3453
3454LLVM also supports a few more substitions useful for writing inline assembly:
3455
3456- ``${:uid}``: Expands to a decimal integer unique to this inline assembly blob.
3457  This substitution is useful when declaring a local label. Many standard
3458  compiler optimizations, such as inlining, may duplicate an inline asm blob.
3459  Adding a blob-unique identifier ensures that the two labels will not conflict
3460  during assembly. This is used to implement `GCC's %= special format
3461  string <https://gcc.gnu.org/onlinedocs/gcc/Extended-Asm.html>`_.
3462- ``${:comment}``: Expands to the comment character of the current target's
3463  assembly dialect. This is usually ``#``, but many targets use other strings,
3464  such as ``;``, ``//``, or ``!``.
3465- ``${:private}``: Expands to the assembler private label prefix. Labels with
3466  this prefix will not appear in the symbol table of the assembled object.
3467  Typically the prefix is ``L``, but targets may use other strings. ``.L`` is
3468  relatively popular.
3469
3470LLVM's support for inline asm is modeled closely on the requirements of Clang's
3471GCC-compatible inline-asm support. Thus, the feature-set and the constraint and
3472modifier codes listed here are similar or identical to those in GCC's inline asm
3473support. However, to be clear, the syntax of the template and constraint strings
3474described here is *not* the same as the syntax accepted by GCC and Clang, and,
3475while most constraint letters are passed through as-is by Clang, some get
3476translated to other codes when converting from the C source to the LLVM
3477assembly.
3478
3479An example inline assembler expression is:
3480
3481.. code-block:: llvm
3482
3483    i32 (i32) asm "bswap $0", "=r,r"
3484
3485Inline assembler expressions may **only** be used as the callee operand
3486of a :ref:`call <i_call>` or an :ref:`invoke <i_invoke>` instruction.
3487Thus, typically we have:
3488
3489.. code-block:: llvm
3490
3491    %X = call i32 asm "bswap $0", "=r,r"(i32 %Y)
3492
3493Inline asms with side effects not visible in the constraint list must be
3494marked as having side effects. This is done through the use of the
3495'``sideeffect``' keyword, like so:
3496
3497.. code-block:: llvm
3498
3499    call void asm sideeffect "eieio", ""()
3500
3501In some cases inline asms will contain code that will not work unless
3502the stack is aligned in some way, such as calls or SSE instructions on
3503x86, yet will not contain code that does that alignment within the asm.
3504The compiler should make conservative assumptions about what the asm
3505might contain and should generate its usual stack alignment code in the
3506prologue if the '``alignstack``' keyword is present:
3507
3508.. code-block:: llvm
3509
3510    call void asm alignstack "eieio", ""()
3511
3512Inline asms also support using non-standard assembly dialects. The
3513assumed dialect is ATT. When the '``inteldialect``' keyword is present,
3514the inline asm is using the Intel dialect. Currently, ATT and Intel are
3515the only supported dialects. An example is:
3516
3517.. code-block:: llvm
3518
3519    call void asm inteldialect "eieio", ""()
3520
3521If multiple keywords appear the '``sideeffect``' keyword must come
3522first, the '``alignstack``' keyword second and the '``inteldialect``'
3523keyword last.
3524
3525Inline Asm Constraint String
3526^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3527
3528The constraint list is a comma-separated string, each element containing one or
3529more constraint codes.
3530
3531For each element in the constraint list an appropriate register or memory
3532operand will be chosen, and it will be made available to assembly template
3533string expansion as ``$0`` for the first constraint in the list, ``$1`` for the
3534second, etc.
3535
3536There are three different types of constraints, which are distinguished by a
3537prefix symbol in front of the constraint code: Output, Input, and Clobber. The
3538constraints must always be given in that order: outputs first, then inputs, then
3539clobbers. They cannot be intermingled.
3540
3541There are also three different categories of constraint codes:
3542
3543- Register constraint. This is either a register class, or a fixed physical
3544  register. This kind of constraint will allocate a register, and if necessary,
3545  bitcast the argument or result to the appropriate type.
3546- Memory constraint. This kind of constraint is for use with an instruction
3547  taking a memory operand. Different constraints allow for different addressing
3548  modes used by the target.
3549- Immediate value constraint. This kind of constraint is for an integer or other
3550  immediate value which can be rendered directly into an instruction. The
3551  various target-specific constraints allow the selection of a value in the
3552  proper range for the instruction you wish to use it with.
3553
3554Output constraints
3555""""""""""""""""""
3556
3557Output constraints are specified by an "``=``" prefix (e.g. "``=r``"). This
3558indicates that the assembly will write to this operand, and the operand will
3559then be made available as a return value of the ``asm`` expression. Output
3560constraints do not consume an argument from the call instruction. (Except, see
3561below about indirect outputs).
3562
3563Normally, it is expected that no output locations are written to by the assembly
3564expression until *all* of the inputs have been read. As such, LLVM may assign
3565the same register to an output and an input. If this is not safe (e.g. if the
3566assembly contains two instructions, where the first writes to one output, and
3567the second reads an input and writes to a second output), then the "``&``"
3568modifier must be used (e.g. "``=&r``") to specify that the output is an
3569"early-clobber" output. Marking an output as "early-clobber" ensures that LLVM
3570will not use the same register for any inputs (other than an input tied to this
3571output).
3572
3573Input constraints
3574"""""""""""""""""
3575
3576Input constraints do not have a prefix -- just the constraint codes. Each input
3577constraint will consume one argument from the call instruction. It is not
3578permitted for the asm to write to any input register or memory location (unless
3579that input is tied to an output). Note also that multiple inputs may all be
3580assigned to the same register, if LLVM can determine that they necessarily all
3581contain the same value.
3582
3583Instead of providing a Constraint Code, input constraints may also "tie"
3584themselves to an output constraint, by providing an integer as the constraint
3585string. Tied inputs still consume an argument from the call instruction, and
3586take up a position in the asm template numbering as is usual -- they will simply
3587be constrained to always use the same register as the output they've been tied
3588to. For example, a constraint string of "``=r,0``" says to assign a register for
3589output, and use that register as an input as well (it being the 0'th
3590constraint).
3591
3592It is permitted to tie an input to an "early-clobber" output. In that case, no
3593*other* input may share the same register as the input tied to the early-clobber
3594(even when the other input has the same value).
3595
3596You may only tie an input to an output which has a register constraint, not a
3597memory constraint. Only a single input may be tied to an output.
3598
3599There is also an "interesting" feature which deserves a bit of explanation: if a
3600register class constraint allocates a register which is too small for the value
3601type operand provided as input, the input value will be split into multiple
3602registers, and all of them passed to the inline asm.
3603
3604However, this feature is often not as useful as you might think.
3605
3606Firstly, the registers are *not* guaranteed to be consecutive. So, on those
3607architectures that have instructions which operate on multiple consecutive
3608instructions, this is not an appropriate way to support them. (e.g. the 32-bit
3609SparcV8 has a 64-bit load, which instruction takes a single 32-bit register. The
3610hardware then loads into both the named register, and the next register. This
3611feature of inline asm would not be useful to support that.)
3612
3613A few of the targets provide a template string modifier allowing explicit access
3614to the second register of a two-register operand (e.g. MIPS ``L``, ``M``, and
3615``D``). On such an architecture, you can actually access the second allocated
3616register (yet, still, not any subsequent ones). But, in that case, you're still
3617probably better off simply splitting the value into two separate operands, for
3618clarity. (e.g. see the description of the ``A`` constraint on X86, which,
3619despite existing only for use with this feature, is not really a good idea to
3620use)
3621
3622Indirect inputs and outputs
3623"""""""""""""""""""""""""""
3624
3625Indirect output or input constraints can be specified by the "``*``" modifier
3626(which goes after the "``=``" in case of an output). This indicates that the asm
3627will write to or read from the contents of an *address* provided as an input
3628argument. (Note that in this way, indirect outputs act more like an *input* than
3629an output: just like an input, they consume an argument of the call expression,
3630rather than producing a return value. An indirect output constraint is an
3631"output" only in that the asm is expected to write to the contents of the input
3632memory location, instead of just read from it).
3633
3634This is most typically used for memory constraint, e.g. "``=*m``", to pass the
3635address of a variable as a value.
3636
3637It is also possible to use an indirect *register* constraint, but only on output
3638(e.g. "``=*r``"). This will cause LLVM to allocate a register for an output
3639value normally, and then, separately emit a store to the address provided as
3640input, after the provided inline asm. (It's not clear what value this
3641functionality provides, compared to writing the store explicitly after the asm
3642statement, and it can only produce worse code, since it bypasses many
3643optimization passes. I would recommend not using it.)
3644
3645
3646Clobber constraints
3647"""""""""""""""""""
3648
3649A clobber constraint is indicated by a "``~``" prefix. A clobber does not
3650consume an input operand, nor generate an output. Clobbers cannot use any of the
3651general constraint code letters -- they may use only explicit register
3652constraints, e.g. "``~{eax}``". The one exception is that a clobber string of
3653"``~{memory}``" indicates that the assembly writes to arbitrary undeclared
3654memory locations -- not only the memory pointed to by a declared indirect
3655output.
3656
3657Note that clobbering named registers that are also present in output
3658constraints is not legal.
3659
3660
3661Constraint Codes
3662""""""""""""""""
3663After a potential prefix comes constraint code, or codes.
3664
3665A Constraint Code is either a single letter (e.g. "``r``"), a "``^``" character
3666followed by two letters (e.g. "``^wc``"), or "``{``" register-name "``}``"
3667(e.g. "``{eax}``").
3668
3669The one and two letter constraint codes are typically chosen to be the same as
3670GCC's constraint codes.
3671
3672A single constraint may include one or more than constraint code in it, leaving
3673it up to LLVM to choose which one to use. This is included mainly for
3674compatibility with the translation of GCC inline asm coming from clang.
3675
3676There are two ways to specify alternatives, and either or both may be used in an
3677inline asm constraint list:
3678
36791) Append the codes to each other, making a constraint code set. E.g. "``im``"
3680   or "``{eax}m``". This means "choose any of the options in the set". The
3681   choice of constraint is made independently for each constraint in the
3682   constraint list.
3683
36842) Use "``|``" between constraint code sets, creating alternatives. Every
3685   constraint in the constraint list must have the same number of alternative
3686   sets. With this syntax, the same alternative in *all* of the items in the
3687   constraint list will be chosen together.
3688
3689Putting those together, you might have a two operand constraint string like
3690``"rm|r,ri|rm"``. This indicates that if operand 0 is ``r`` or ``m``, then
3691operand 1 may be one of ``r`` or ``i``. If operand 0 is ``r``, then operand 1
3692may be one of ``r`` or ``m``. But, operand 0 and 1 cannot both be of type m.
3693
3694However, the use of either of the alternatives features is *NOT* recommended, as
3695LLVM is not able to make an intelligent choice about which one to use. (At the
3696point it currently needs to choose, not enough information is available to do so
3697in a smart way.) Thus, it simply tries to make a choice that's most likely to
3698compile, not one that will be optimal performance. (e.g., given "``rm``", it'll
3699always choose to use memory, not registers). And, if given multiple registers,
3700or multiple register classes, it will simply choose the first one. (In fact, it
3701doesn't currently even ensure explicitly specified physical registers are
3702unique, so specifying multiple physical registers as alternatives, like
3703``{r11}{r12},{r11}{r12}``, will assign r11 to both operands, not at all what was
3704intended.)
3705
3706Supported Constraint Code List
3707""""""""""""""""""""""""""""""
3708
3709The constraint codes are, in general, expected to behave the same way they do in
3710GCC. LLVM's support is often implemented on an 'as-needed' basis, to support C
3711inline asm code which was supported by GCC. A mismatch in behavior between LLVM
3712and GCC likely indicates a bug in LLVM.
3713
3714Some constraint codes are typically supported by all targets:
3715
3716- ``r``: A register in the target's general purpose register class.
3717- ``m``: A memory address operand. It is target-specific what addressing modes
3718  are supported, typical examples are register, or register + register offset,
3719  or register + immediate offset (of some target-specific size).
3720- ``i``: An integer constant (of target-specific width). Allows either a simple
3721  immediate, or a relocatable value.
3722- ``n``: An integer constant -- *not* including relocatable values.
3723- ``s``: An integer constant, but allowing *only* relocatable values.
3724- ``X``: Allows an operand of any kind, no constraint whatsoever. Typically
3725  useful to pass a label for an asm branch or call.
3726
3727  .. FIXME: but that surely isn't actually okay to jump out of an asm
3728     block without telling llvm about the control transfer???)
3729
3730- ``{register-name}``: Requires exactly the named physical register.
3731
3732Other constraints are target-specific:
3733
3734AArch64:
3735
3736- ``z``: An immediate integer 0. Outputs ``WZR`` or ``XZR``, as appropriate.
3737- ``I``: An immediate integer valid for an ``ADD`` or ``SUB`` instruction,
3738  i.e. 0 to 4095 with optional shift by 12.
3739- ``J``: An immediate integer that, when negated, is valid for an ``ADD`` or
3740  ``SUB`` instruction, i.e. -1 to -4095 with optional left shift by 12.
3741- ``K``: An immediate integer that is valid for the 'bitmask immediate 32' of a
3742  logical instruction like ``AND``, ``EOR``, or ``ORR`` with a 32-bit register.
3743- ``L``: An immediate integer that is valid for the 'bitmask immediate 64' of a
3744  logical instruction like ``AND``, ``EOR``, or ``ORR`` with a 64-bit register.
3745- ``M``: An immediate integer for use with the ``MOV`` assembly alias on a
3746  32-bit register. This is a superset of ``K``: in addition to the bitmask
3747  immediate, also allows immediate integers which can be loaded with a single
3748  ``MOVZ`` or ``MOVL`` instruction.
3749- ``N``: An immediate integer for use with the ``MOV`` assembly alias on a
3750  64-bit register. This is a superset of ``L``.
3751- ``Q``: Memory address operand must be in a single register (no
3752  offsets). (However, LLVM currently does this for the ``m`` constraint as
3753  well.)
3754- ``r``: A 32 or 64-bit integer register (W* or X*).
3755- ``w``: A 32, 64, or 128-bit floating-point/SIMD register.
3756- ``x``: A lower 128-bit floating-point/SIMD register (``V0`` to ``V15``).
3757
3758AMDGPU:
3759
3760- ``r``: A 32 or 64-bit integer register.
3761- ``[0-9]v``: The 32-bit VGPR register, number 0-9.
3762- ``[0-9]s``: The 32-bit SGPR register, number 0-9.
3763
3764
3765All ARM modes:
3766
3767- ``Q``, ``Um``, ``Un``, ``Uq``, ``Us``, ``Ut``, ``Uv``, ``Uy``: Memory address
3768  operand. Treated the same as operand ``m``, at the moment.
3769
3770ARM and ARM's Thumb2 mode:
3771
3772- ``j``: An immediate integer between 0 and 65535 (valid for ``MOVW``)
3773- ``I``: An immediate integer valid for a data-processing instruction.
3774- ``J``: An immediate integer between -4095 and 4095.
3775- ``K``: An immediate integer whose bitwise inverse is valid for a
3776  data-processing instruction. (Can be used with template modifier "``B``" to
3777  print the inverted value).
3778- ``L``: An immediate integer whose negation is valid for a data-processing
3779  instruction. (Can be used with template modifier "``n``" to print the negated
3780  value).
3781- ``M``: A power of two or a integer between 0 and 32.
3782- ``N``: Invalid immediate constraint.
3783- ``O``: Invalid immediate constraint.
3784- ``r``: A general-purpose 32-bit integer register (``r0-r15``).
3785- ``l``: In Thumb2 mode, low 32-bit GPR registers (``r0-r7``). In ARM mode, same
3786  as ``r``.
3787- ``h``: In Thumb2 mode, a high 32-bit GPR register (``r8-r15``). In ARM mode,
3788  invalid.
3789- ``w``: A 32, 64, or 128-bit floating-point/SIMD register: ``s0-s31``,
3790  ``d0-d31``, or ``q0-q15``.
3791- ``x``: A 32, 64, or 128-bit floating-point/SIMD register: ``s0-s15``,
3792  ``d0-d7``, or ``q0-q3``.
3793- ``t``: A low floating-point/SIMD register: ``s0-s31``, ``d0-d16``, or
3794  ``q0-q8``.
3795
3796ARM's Thumb1 mode:
3797
3798- ``I``: An immediate integer between 0 and 255.
3799- ``J``: An immediate integer between -255 and -1.
3800- ``K``: An immediate integer between 0 and 255, with optional left-shift by
3801  some amount.
3802- ``L``: An immediate integer between -7 and 7.
3803- ``M``: An immediate integer which is a multiple of 4 between 0 and 1020.
3804- ``N``: An immediate integer between 0 and 31.
3805- ``O``: An immediate integer which is a multiple of 4 between -508 and 508.
3806- ``r``: A low 32-bit GPR register (``r0-r7``).
3807- ``l``: A low 32-bit GPR register (``r0-r7``).
3808- ``h``: A high GPR register (``r0-r7``).
3809- ``w``: A 32, 64, or 128-bit floating-point/SIMD register: ``s0-s31``,
3810  ``d0-d31``, or ``q0-q15``.
3811- ``x``: A 32, 64, or 128-bit floating-point/SIMD register: ``s0-s15``,
3812  ``d0-d7``, or ``q0-q3``.
3813- ``t``: A low floating-point/SIMD register: ``s0-s31``, ``d0-d16``, or
3814  ``q0-q8``.
3815
3816
3817Hexagon:
3818
3819- ``o``, ``v``: A memory address operand, treated the same as constraint ``m``,
3820  at the moment.
3821- ``r``: A 32 or 64-bit register.
3822
3823MSP430:
3824
3825- ``r``: An 8 or 16-bit register.
3826
3827MIPS:
3828
3829- ``I``: An immediate signed 16-bit integer.
3830- ``J``: An immediate integer zero.
3831- ``K``: An immediate unsigned 16-bit integer.
3832- ``L``: An immediate 32-bit integer, where the lower 16 bits are 0.
3833- ``N``: An immediate integer between -65535 and -1.
3834- ``O``: An immediate signed 15-bit integer.
3835- ``P``: An immediate integer between 1 and 65535.
3836- ``m``: A memory address operand. In MIPS-SE mode, allows a base address
3837  register plus 16-bit immediate offset. In MIPS mode, just a base register.
3838- ``R``: A memory address operand. In MIPS-SE mode, allows a base address
3839  register plus a 9-bit signed offset. In MIPS mode, the same as constraint
3840  ``m``.
3841- ``ZC``: A memory address operand, suitable for use in a ``pref``, ``ll``, or
3842  ``sc`` instruction on the given subtarget (details vary).
3843- ``r``, ``d``,  ``y``: A 32 or 64-bit GPR register.
3844- ``f``: A 32 or 64-bit FPU register (``F0-F31``), or a 128-bit MSA register
3845  (``W0-W31``). In the case of MSA registers, it is recommended to use the ``w``
3846  argument modifier for compatibility with GCC.
3847- ``c``: A 32-bit or 64-bit GPR register suitable for indirect jump (always
3848  ``25``).
3849- ``l``: The ``lo`` register, 32 or 64-bit.
3850- ``x``: Invalid.
3851
3852NVPTX:
3853
3854- ``b``: A 1-bit integer register.
3855- ``c`` or ``h``: A 16-bit integer register.
3856- ``r``: A 32-bit integer register.
3857- ``l`` or ``N``: A 64-bit integer register.
3858- ``f``: A 32-bit float register.
3859- ``d``: A 64-bit float register.
3860
3861
3862PowerPC:
3863
3864- ``I``: An immediate signed 16-bit integer.
3865- ``J``: An immediate unsigned 16-bit integer, shifted left 16 bits.
3866- ``K``: An immediate unsigned 16-bit integer.
3867- ``L``: An immediate signed 16-bit integer, shifted left 16 bits.
3868- ``M``: An immediate integer greater than 31.
3869- ``N``: An immediate integer that is an exact power of 2.
3870- ``O``: The immediate integer constant 0.
3871- ``P``: An immediate integer constant whose negation is a signed 16-bit
3872  constant.
3873- ``es``, ``o``, ``Q``, ``Z``, ``Zy``: A memory address operand, currently
3874  treated the same as ``m``.
3875- ``r``: A 32 or 64-bit integer register.
3876- ``b``: A 32 or 64-bit integer register, excluding ``R0`` (that is:
3877  ``R1-R31``).
3878- ``f``: A 32 or 64-bit float register (``F0-F31``), or when QPX is enabled, a
3879  128 or 256-bit QPX register (``Q0-Q31``; aliases the ``F`` registers).
3880- ``v``: For ``4 x f32`` or ``4 x f64`` types, when QPX is enabled, a
3881  128 or 256-bit QPX register (``Q0-Q31``), otherwise a 128-bit
3882  altivec vector register (``V0-V31``).
3883
3884  .. FIXME: is this a bug that v accepts QPX registers? I think this
3885     is supposed to only use the altivec vector registers?
3886
3887- ``y``: Condition register (``CR0-CR7``).
3888- ``wc``: An individual CR bit in a CR register.
3889- ``wa``, ``wd``, ``wf``: Any 128-bit VSX vector register, from the full VSX
3890  register set (overlapping both the floating-point and vector register files).
3891- ``ws``: A 32 or 64-bit floating-point register, from the full VSX register
3892  set.
3893
3894Sparc:
3895
3896- ``I``: An immediate 13-bit signed integer.
3897- ``r``: A 32-bit integer register.
3898- ``f``: Any floating-point register on SparcV8, or a floating-point
3899  register in the "low" half of the registers on SparcV9.
3900- ``e``: Any floating-point register. (Same as ``f`` on SparcV8.)
3901
3902SystemZ:
3903
3904- ``I``: An immediate unsigned 8-bit integer.
3905- ``J``: An immediate unsigned 12-bit integer.
3906- ``K``: An immediate signed 16-bit integer.
3907- ``L``: An immediate signed 20-bit integer.
3908- ``M``: An immediate integer 0x7fffffff.
3909- ``Q``: A memory address operand with a base address and a 12-bit immediate
3910  unsigned displacement.
3911- ``R``: A memory address operand with a base address, a 12-bit immediate
3912  unsigned displacement, and an index register.
3913- ``S``: A memory address operand with a base address and a 20-bit immediate
3914  signed displacement.
3915- ``T``: A memory address operand with a base address, a 20-bit immediate
3916  signed displacement, and an index register.
3917- ``r`` or ``d``: A 32, 64, or 128-bit integer register.
3918- ``a``: A 32, 64, or 128-bit integer address register (excludes R0, which in an
3919  address context evaluates as zero).
3920- ``h``: A 32-bit value in the high part of a 64bit data register
3921  (LLVM-specific)
3922- ``f``: A 32, 64, or 128-bit floating-point register.
3923
3924X86:
3925
3926- ``I``: An immediate integer between 0 and 31.
3927- ``J``: An immediate integer between 0 and 64.
3928- ``K``: An immediate signed 8-bit integer.
3929- ``L``: An immediate integer, 0xff or 0xffff or (in 64-bit mode only)
3930  0xffffffff.
3931- ``M``: An immediate integer between 0 and 3.
3932- ``N``: An immediate unsigned 8-bit integer.
3933- ``O``: An immediate integer between 0 and 127.
3934- ``e``: An immediate 32-bit signed integer.
3935- ``Z``: An immediate 32-bit unsigned integer.
3936- ``o``, ``v``: Treated the same as ``m``, at the moment.
3937- ``q``: An 8, 16, 32, or 64-bit register which can be accessed as an 8-bit
3938  ``l`` integer register. On X86-32, this is the ``a``, ``b``, ``c``, and ``d``
3939  registers, and on X86-64, it is all of the integer registers.
3940- ``Q``: An 8, 16, 32, or 64-bit register which can be accessed as an 8-bit
3941  ``h`` integer register. This is the ``a``, ``b``, ``c``, and ``d`` registers.
3942- ``r`` or ``l``: An 8, 16, 32, or 64-bit integer register.
3943- ``R``: An 8, 16, 32, or 64-bit "legacy" integer register -- one which has
3944  existed since i386, and can be accessed without the REX prefix.
3945- ``f``: A 32, 64, or 80-bit '387 FPU stack pseudo-register.
3946- ``y``: A 64-bit MMX register, if MMX is enabled.
3947- ``x``: If SSE is enabled: a 32 or 64-bit scalar operand, or 128-bit vector
3948  operand in a SSE register. If AVX is also enabled, can also be a 256-bit
3949  vector operand in an AVX register. If AVX-512 is also enabled, can also be a
3950  512-bit vector operand in an AVX512 register, Otherwise, an error.
3951- ``Y``: The same as ``x``, if *SSE2* is enabled, otherwise an error.
3952- ``A``: Special case: allocates EAX first, then EDX, for a single operand (in
3953  32-bit mode, a 64-bit integer operand will get split into two registers). It
3954  is not recommended to use this constraint, as in 64-bit mode, the 64-bit
3955  operand will get allocated only to RAX -- if two 32-bit operands are needed,
3956  you're better off splitting it yourself, before passing it to the asm
3957  statement.
3958
3959XCore:
3960
3961- ``r``: A 32-bit integer register.
3962
3963
3964.. _inline-asm-modifiers:
3965
3966Asm template argument modifiers
3967^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3968
3969In the asm template string, modifiers can be used on the operand reference, like
3970"``${0:n}``".
3971
3972The modifiers are, in general, expected to behave the same way they do in
3973GCC. LLVM's support is often implemented on an 'as-needed' basis, to support C
3974inline asm code which was supported by GCC. A mismatch in behavior between LLVM
3975and GCC likely indicates a bug in LLVM.
3976
3977Target-independent:
3978
3979- ``c``: Print an immediate integer constant unadorned, without
3980  the target-specific immediate punctuation (e.g. no ``$`` prefix).
3981- ``n``: Negate and print immediate integer constant unadorned, without the
3982  target-specific immediate punctuation (e.g. no ``$`` prefix).
3983- ``l``: Print as an unadorned label, without the target-specific label
3984  punctuation (e.g. no ``$`` prefix).
3985
3986AArch64:
3987
3988- ``w``: Print a GPR register with a ``w*`` name instead of ``x*`` name. E.g.,
3989  instead of ``x30``, print ``w30``.
3990- ``x``: Print a GPR register with a ``x*`` name. (this is the default, anyhow).
3991- ``b``, ``h``, ``s``, ``d``, ``q``: Print a floating-point/SIMD register with a
3992  ``b*``, ``h*``, ``s*``, ``d*``, or ``q*`` name, rather than the default of
3993  ``v*``.
3994
3995AMDGPU:
3996
3997- ``r``: No effect.
3998
3999ARM:
4000
4001- ``a``: Print an operand as an address (with ``[`` and ``]`` surrounding a
4002  register).
4003- ``P``: No effect.
4004- ``q``: No effect.
4005- ``y``: Print a VFP single-precision register as an indexed double (e.g. print
4006  as ``d4[1]`` instead of ``s9``)
4007- ``B``: Bitwise invert and print an immediate integer constant without ``#``
4008  prefix.
4009- ``L``: Print the low 16-bits of an immediate integer constant.
4010- ``M``: Print as a register set suitable for ldm/stm. Also prints *all*
4011  register operands subsequent to the specified one (!), so use carefully.
4012- ``Q``: Print the low-order register of a register-pair, or the low-order
4013  register of a two-register operand.
4014- ``R``: Print the high-order register of a register-pair, or the high-order
4015  register of a two-register operand.
4016- ``H``: Print the second register of a register-pair. (On a big-endian system,
4017  ``H`` is equivalent to ``Q``, and on little-endian system, ``H`` is equivalent
4018  to ``R``.)
4019
4020  .. FIXME: H doesn't currently support printing the second register
4021     of a two-register operand.
4022
4023- ``e``: Print the low doubleword register of a NEON quad register.
4024- ``f``: Print the high doubleword register of a NEON quad register.
4025- ``m``: Print the base register of a memory operand without the ``[`` and ``]``
4026  adornment.
4027
4028Hexagon:
4029
4030- ``L``: Print the second register of a two-register operand. Requires that it
4031  has been allocated consecutively to the first.
4032
4033  .. FIXME: why is it restricted to consecutive ones? And there's
4034     nothing that ensures that happens, is there?
4035
4036- ``I``: Print the letter 'i' if the operand is an integer constant, otherwise
4037  nothing. Used to print 'addi' vs 'add' instructions.
4038
4039MSP430:
4040
4041No additional modifiers.
4042
4043MIPS:
4044
4045- ``X``: Print an immediate integer as hexadecimal
4046- ``x``: Print the low 16 bits of an immediate integer as hexadecimal.
4047- ``d``: Print an immediate integer as decimal.
4048- ``m``: Subtract one and print an immediate integer as decimal.
4049- ``z``: Print $0 if an immediate zero, otherwise print normally.
4050- ``L``: Print the low-order register of a two-register operand, or prints the
4051  address of the low-order word of a double-word memory operand.
4052
4053  .. FIXME: L seems to be missing memory operand support.
4054
4055- ``M``: Print the high-order register of a two-register operand, or prints the
4056  address of the high-order word of a double-word memory operand.
4057
4058  .. FIXME: M seems to be missing memory operand support.
4059
4060- ``D``: Print the second register of a two-register operand, or prints the
4061  second word of a double-word memory operand. (On a big-endian system, ``D`` is
4062  equivalent to ``L``, and on little-endian system, ``D`` is equivalent to
4063  ``M``.)
4064- ``w``: No effect. Provided for compatibility with GCC which requires this
4065  modifier in order to print MSA registers (``W0-W31``) with the ``f``
4066  constraint.
4067
4068NVPTX:
4069
4070- ``r``: No effect.
4071
4072PowerPC:
4073
4074- ``L``: Print the second register of a two-register operand. Requires that it
4075  has been allocated consecutively to the first.
4076
4077  .. FIXME: why is it restricted to consecutive ones? And there's
4078     nothing that ensures that happens, is there?
4079
4080- ``I``: Print the letter 'i' if the operand is an integer constant, otherwise
4081  nothing. Used to print 'addi' vs 'add' instructions.
4082- ``y``: For a memory operand, prints formatter for a two-register X-form
4083  instruction. (Currently always prints ``r0,OPERAND``).
4084- ``U``: Prints 'u' if the memory operand is an update form, and nothing
4085  otherwise. (NOTE: LLVM does not support update form, so this will currently
4086  always print nothing)
4087- ``X``: Prints 'x' if the memory operand is an indexed form. (NOTE: LLVM does
4088  not support indexed form, so this will currently always print nothing)
4089
4090Sparc:
4091
4092- ``r``: No effect.
4093
4094SystemZ:
4095
4096SystemZ implements only ``n``, and does *not* support any of the other
4097target-independent modifiers.
4098
4099X86:
4100
4101- ``c``: Print an unadorned integer or symbol name. (The latter is
4102  target-specific behavior for this typically target-independent modifier).
4103- ``A``: Print a register name with a '``*``' before it.
4104- ``b``: Print an 8-bit register name (e.g. ``al``); do nothing on a memory
4105  operand.
4106- ``h``: Print the upper 8-bit register name (e.g. ``ah``); do nothing on a
4107  memory operand.
4108- ``w``: Print the 16-bit register name (e.g. ``ax``); do nothing on a memory
4109  operand.
4110- ``k``: Print the 32-bit register name (e.g. ``eax``); do nothing on a memory
4111  operand.
4112- ``q``: Print the 64-bit register name (e.g. ``rax``), if 64-bit registers are
4113  available, otherwise the 32-bit register name; do nothing on a memory operand.
4114- ``n``: Negate and print an unadorned integer, or, for operands other than an
4115  immediate integer (e.g. a relocatable symbol expression), print a '-' before
4116  the operand. (The behavior for relocatable symbol expressions is a
4117  target-specific behavior for this typically target-independent modifier)
4118- ``H``: Print a memory reference with additional offset +8.
4119- ``P``: Print a memory reference or operand for use as the argument of a call
4120  instruction. (E.g. omit ``(rip)``, even though it's PC-relative.)
4121
4122XCore:
4123
4124No additional modifiers.
4125
4126
4127Inline Asm Metadata
4128^^^^^^^^^^^^^^^^^^^
4129
4130The call instructions that wrap inline asm nodes may have a
4131"``!srcloc``" MDNode attached to it that contains a list of constant
4132integers. If present, the code generator will use the integer as the
4133location cookie value when report errors through the ``LLVMContext``
4134error reporting mechanisms. This allows a front-end to correlate backend
4135errors that occur with inline asm back to the source code that produced
4136it. For example:
4137
4138.. code-block:: llvm
4139
4140    call void asm sideeffect "something bad", ""(), !srcloc !42
4141    ...
4142    !42 = !{ i32 1234567 }
4143
4144It is up to the front-end to make sense of the magic numbers it places
4145in the IR. If the MDNode contains multiple constants, the code generator
4146will use the one that corresponds to the line of the asm that the error
4147occurs on.
4148
4149.. _metadata:
4150
4151Metadata
4152========
4153
4154LLVM IR allows metadata to be attached to instructions in the program
4155that can convey extra information about the code to the optimizers and
4156code generator. One example application of metadata is source-level
4157debug information. There are two metadata primitives: strings and nodes.
4158
4159Metadata does not have a type, and is not a value. If referenced from a
4160``call`` instruction, it uses the ``metadata`` type.
4161
4162All metadata are identified in syntax by a exclamation point ('``!``').
4163
4164.. _metadata-string:
4165
4166Metadata Nodes and Metadata Strings
4167-----------------------------------
4168
4169A metadata string is a string surrounded by double quotes. It can
4170contain any character by escaping non-printable characters with
4171"``\xx``" where "``xx``" is the two digit hex code. For example:
4172"``!"test\00"``".
4173
4174Metadata nodes are represented with notation similar to structure
4175constants (a comma separated list of elements, surrounded by braces and
4176preceded by an exclamation point). Metadata nodes can have any values as
4177their operand. For example:
4178
4179.. code-block:: llvm
4180
4181    !{ !"test\00", i32 10}
4182
4183Metadata nodes that aren't uniqued use the ``distinct`` keyword. For example:
4184
4185.. code-block:: text
4186
4187    !0 = distinct !{!"test\00", i32 10}
4188
4189``distinct`` nodes are useful when nodes shouldn't be merged based on their
4190content. They can also occur when transformations cause uniquing collisions
4191when metadata operands change.
4192
4193A :ref:`named metadata <namedmetadatastructure>` is a collection of
4194metadata nodes, which can be looked up in the module symbol table. For
4195example:
4196
4197.. code-block:: llvm
4198
4199    !foo = !{!4, !3}
4200
4201Metadata can be used as function arguments. Here the ``llvm.dbg.value``
4202intrinsic is using three metadata arguments:
4203
4204.. code-block:: llvm
4205
4206    call void @llvm.dbg.value(metadata !24, metadata !25, metadata !26)
4207
4208Metadata can be attached to an instruction. Here metadata ``!21`` is attached
4209to the ``add`` instruction using the ``!dbg`` identifier:
4210
4211.. code-block:: llvm
4212
4213    %indvar.next = add i64 %indvar, 1, !dbg !21
4214
4215Metadata can also be attached to a function or a global variable. Here metadata
4216``!22`` is attached to the ``f1`` and ``f2 functions, and the globals ``g1``
4217and ``g2`` using the ``!dbg`` identifier:
4218
4219.. code-block:: llvm
4220
4221    declare !dbg !22 void @f1()
4222    define void @f2() !dbg !22 {
4223      ret void
4224    }
4225
4226    @g1 = global i32 0, !dbg !22
4227    @g2 = external global i32, !dbg !22
4228
4229A transformation is required to drop any metadata attachment that it does not
4230know or know it can't preserve. Currently there is an exception for metadata
4231attachment to globals for ``!type`` and ``!absolute_symbol`` which can't be
4232unconditionally dropped unless the global is itself deleted.
4233
4234Metadata attached to a module using named metadata may not be dropped, with
4235the exception of debug metadata (named metadata with the name ``!llvm.dbg.*``).
4236
4237More information about specific metadata nodes recognized by the
4238optimizers and code generator is found below.
4239
4240.. _specialized-metadata:
4241
4242Specialized Metadata Nodes
4243^^^^^^^^^^^^^^^^^^^^^^^^^^
4244
4245Specialized metadata nodes are custom data structures in metadata (as opposed
4246to generic tuples). Their fields are labelled, and can be specified in any
4247order.
4248
4249These aren't inherently debug info centric, but currently all the specialized
4250metadata nodes are related to debug info.
4251
4252.. _DICompileUnit:
4253
4254DICompileUnit
4255"""""""""""""
4256
4257``DICompileUnit`` nodes represent a compile unit. The ``enums:``,
4258``retainedTypes:``, ``globals:``, ``imports:`` and ``macros:`` fields are tuples
4259containing the debug info to be emitted along with the compile unit, regardless
4260of code optimizations (some nodes are only emitted if there are references to
4261them from instructions). The ``debugInfoForProfiling:`` field is a boolean
4262indicating whether or not line-table discriminators are updated to provide
4263more-accurate debug info for profiling results.
4264
4265.. code-block:: text
4266
4267    !0 = !DICompileUnit(language: DW_LANG_C99, file: !1, producer: "clang",
4268                        isOptimized: true, flags: "-O2", runtimeVersion: 2,
4269                        splitDebugFilename: "abc.debug", emissionKind: FullDebug,
4270                        enums: !2, retainedTypes: !3, globals: !4, imports: !5,
4271                        macros: !6, dwoId: 0x0abcd)
4272
4273Compile unit descriptors provide the root scope for objects declared in a
4274specific compilation unit. File descriptors are defined using this scope.  These
4275descriptors are collected by a named metadata node ``!llvm.dbg.cu``. They keep
4276track of global variables, type information, and imported entities (declarations
4277and namespaces).
4278
4279.. _DIFile:
4280
4281DIFile
4282""""""
4283
4284``DIFile`` nodes represent files. The ``filename:`` can include slashes.
4285
4286.. code-block:: none
4287
4288    !0 = !DIFile(filename: "path/to/file", directory: "/path/to/dir",
4289                 checksumkind: CSK_MD5,
4290                 checksum: "000102030405060708090a0b0c0d0e0f")
4291
4292Files are sometimes used in ``scope:`` fields, and are the only valid target
4293for ``file:`` fields.
4294Valid values for ``checksumkind:`` field are: {CSK_None, CSK_MD5, CSK_SHA1}
4295
4296.. _DIBasicType:
4297
4298DIBasicType
4299"""""""""""
4300
4301``DIBasicType`` nodes represent primitive types, such as ``int``, ``bool`` and
4302``float``. ``tag:`` defaults to ``DW_TAG_base_type``.
4303
4304.. code-block:: text
4305
4306    !0 = !DIBasicType(name: "unsigned char", size: 8, align: 8,
4307                      encoding: DW_ATE_unsigned_char)
4308    !1 = !DIBasicType(tag: DW_TAG_unspecified_type, name: "decltype(nullptr)")
4309
4310The ``encoding:`` describes the details of the type. Usually it's one of the
4311following:
4312
4313.. code-block:: text
4314
4315  DW_ATE_address       = 1
4316  DW_ATE_boolean       = 2
4317  DW_ATE_float         = 4
4318  DW_ATE_signed        = 5
4319  DW_ATE_signed_char   = 6
4320  DW_ATE_unsigned      = 7
4321  DW_ATE_unsigned_char = 8
4322
4323.. _DISubroutineType:
4324
4325DISubroutineType
4326""""""""""""""""
4327
4328``DISubroutineType`` nodes represent subroutine types. Their ``types:`` field
4329refers to a tuple; the first operand is the return type, while the rest are the
4330types of the formal arguments in order. If the first operand is ``null``, that
4331represents a function with no return value (such as ``void foo() {}`` in C++).
4332
4333.. code-block:: text
4334
4335    !0 = !BasicType(name: "int", size: 32, align: 32, DW_ATE_signed)
4336    !1 = !BasicType(name: "char", size: 8, align: 8, DW_ATE_signed_char)
4337    !2 = !DISubroutineType(types: !{null, !0, !1}) ; void (int, char)
4338
4339.. _DIDerivedType:
4340
4341DIDerivedType
4342"""""""""""""
4343
4344``DIDerivedType`` nodes represent types derived from other types, such as
4345qualified types.
4346
4347.. code-block:: text
4348
4349    !0 = !DIBasicType(name: "unsigned char", size: 8, align: 8,
4350                      encoding: DW_ATE_unsigned_char)
4351    !1 = !DIDerivedType(tag: DW_TAG_pointer_type, baseType: !0, size: 32,
4352                        align: 32)
4353
4354The following ``tag:`` values are valid:
4355
4356.. code-block:: text
4357
4358  DW_TAG_member             = 13
4359  DW_TAG_pointer_type       = 15
4360  DW_TAG_reference_type     = 16
4361  DW_TAG_typedef            = 22
4362  DW_TAG_inheritance        = 28
4363  DW_TAG_ptr_to_member_type = 31
4364  DW_TAG_const_type         = 38
4365  DW_TAG_friend             = 42
4366  DW_TAG_volatile_type      = 53
4367  DW_TAG_restrict_type      = 55
4368  DW_TAG_atomic_type        = 71
4369
4370.. _DIDerivedTypeMember:
4371
4372``DW_TAG_member`` is used to define a member of a :ref:`composite type
4373<DICompositeType>`. The type of the member is the ``baseType:``. The
4374``offset:`` is the member's bit offset.  If the composite type has an ODR
4375``identifier:`` and does not set ``flags: DIFwdDecl``, then the member is
4376uniqued based only on its ``name:`` and ``scope:``.
4377
4378``DW_TAG_inheritance`` and ``DW_TAG_friend`` are used in the ``elements:``
4379field of :ref:`composite types <DICompositeType>` to describe parents and
4380friends.
4381
4382``DW_TAG_typedef`` is used to provide a name for the ``baseType:``.
4383
4384``DW_TAG_pointer_type``, ``DW_TAG_reference_type``, ``DW_TAG_const_type``,
4385``DW_TAG_volatile_type``, ``DW_TAG_restrict_type`` and ``DW_TAG_atomic_type``
4386are used to qualify the ``baseType:``.
4387
4388Note that the ``void *`` type is expressed as a type derived from NULL.
4389
4390.. _DICompositeType:
4391
4392DICompositeType
4393"""""""""""""""
4394
4395``DICompositeType`` nodes represent types composed of other types, like
4396structures and unions. ``elements:`` points to a tuple of the composed types.
4397
4398If the source language supports ODR, the ``identifier:`` field gives the unique
4399identifier used for type merging between modules.  When specified,
4400:ref:`subprogram declarations <DISubprogramDeclaration>` and :ref:`member
4401derived types <DIDerivedTypeMember>` that reference the ODR-type in their
4402``scope:`` change uniquing rules.
4403
4404For a given ``identifier:``, there should only be a single composite type that
4405does not have  ``flags: DIFlagFwdDecl`` set.  LLVM tools that link modules
4406together will unique such definitions at parse time via the ``identifier:``
4407field, even if the nodes are ``distinct``.
4408
4409.. code-block:: text
4410
4411    !0 = !DIEnumerator(name: "SixKind", value: 7)
4412    !1 = !DIEnumerator(name: "SevenKind", value: 7)
4413    !2 = !DIEnumerator(name: "NegEightKind", value: -8)
4414    !3 = !DICompositeType(tag: DW_TAG_enumeration_type, name: "Enum", file: !12,
4415                          line: 2, size: 32, align: 32, identifier: "_M4Enum",
4416                          elements: !{!0, !1, !2})
4417
4418The following ``tag:`` values are valid:
4419
4420.. code-block:: text
4421
4422  DW_TAG_array_type       = 1
4423  DW_TAG_class_type       = 2
4424  DW_TAG_enumeration_type = 4
4425  DW_TAG_structure_type   = 19
4426  DW_TAG_union_type       = 23
4427
4428For ``DW_TAG_array_type``, the ``elements:`` should be :ref:`subrange
4429descriptors <DISubrange>`, each representing the range of subscripts at that
4430level of indexing. The ``DIFlagVector`` flag to ``flags:`` indicates that an
4431array type is a native packed vector.
4432
4433For ``DW_TAG_enumeration_type``, the ``elements:`` should be :ref:`enumerator
4434descriptors <DIEnumerator>`, each representing the definition of an enumeration
4435value for the set. All enumeration type descriptors are collected in the
4436``enums:`` field of the :ref:`compile unit <DICompileUnit>`.
4437
4438For ``DW_TAG_structure_type``, ``DW_TAG_class_type``, and
4439``DW_TAG_union_type``, the ``elements:`` should be :ref:`derived types
4440<DIDerivedType>` with ``tag: DW_TAG_member``, ``tag: DW_TAG_inheritance``, or
4441``tag: DW_TAG_friend``; or :ref:`subprograms <DISubprogram>` with
4442``isDefinition: false``.
4443
4444.. _DISubrange:
4445
4446DISubrange
4447""""""""""
4448
4449``DISubrange`` nodes are the elements for ``DW_TAG_array_type`` variants of
4450:ref:`DICompositeType`.
4451
4452- ``count: -1`` indicates an empty array.
4453- ``count: !9`` describes the count with a :ref:`DILocalVariable`.
4454- ``count: !11`` describes the count with a :ref:`DIGlobalVariable`.
4455
4456.. code-block:: text
4457
4458    !0 = !DISubrange(count: 5, lowerBound: 0) ; array counting from 0
4459    !1 = !DISubrange(count: 5, lowerBound: 1) ; array counting from 1
4460    !2 = !DISubrange(count: -1) ; empty array.
4461
4462    ; Scopes used in rest of example
4463    !6 = !DIFile(filename: "vla.c", directory: "/path/to/file")
4464    !7 = distinct !DICompileUnit(language: DW_LANG_C99, file: !6)
4465    !8 = distinct !DISubprogram(name: "foo", scope: !7, file: !6, line: 5)
4466
4467    ; Use of local variable as count value
4468    !9 = !DIBasicType(name: "int", size: 32, encoding: DW_ATE_signed)
4469    !10 = !DILocalVariable(name: "count", scope: !8, file: !6, line: 42, type: !9)
4470    !11 = !DISubrange(count: !10, lowerBound: 0)
4471
4472    ; Use of global variable as count value
4473    !12 = !DIGlobalVariable(name: "count", scope: !8, file: !6, line: 22, type: !9)
4474    !13 = !DISubrange(count: !12, lowerBound: 0)
4475
4476.. _DIEnumerator:
4477
4478DIEnumerator
4479""""""""""""
4480
4481``DIEnumerator`` nodes are the elements for ``DW_TAG_enumeration_type``
4482variants of :ref:`DICompositeType`.
4483
4484.. code-block:: text
4485
4486    !0 = !DIEnumerator(name: "SixKind", value: 7)
4487    !1 = !DIEnumerator(name: "SevenKind", value: 7)
4488    !2 = !DIEnumerator(name: "NegEightKind", value: -8)
4489
4490DITemplateTypeParameter
4491"""""""""""""""""""""""
4492
4493``DITemplateTypeParameter`` nodes represent type parameters to generic source
4494language constructs. They are used (optionally) in :ref:`DICompositeType` and
4495:ref:`DISubprogram` ``templateParams:`` fields.
4496
4497.. code-block:: text
4498
4499    !0 = !DITemplateTypeParameter(name: "Ty", type: !1)
4500
4501DITemplateValueParameter
4502""""""""""""""""""""""""
4503
4504``DITemplateValueParameter`` nodes represent value parameters to generic source
4505language constructs. ``tag:`` defaults to ``DW_TAG_template_value_parameter``,
4506but if specified can also be set to ``DW_TAG_GNU_template_template_param`` or
4507``DW_TAG_GNU_template_param_pack``. They are used (optionally) in
4508:ref:`DICompositeType` and :ref:`DISubprogram` ``templateParams:`` fields.
4509
4510.. code-block:: text
4511
4512    !0 = !DITemplateValueParameter(name: "Ty", type: !1, value: i32 7)
4513
4514DINamespace
4515"""""""""""
4516
4517``DINamespace`` nodes represent namespaces in the source language.
4518
4519.. code-block:: text
4520
4521    !0 = !DINamespace(name: "myawesomeproject", scope: !1, file: !2, line: 7)
4522
4523.. _DIGlobalVariable:
4524
4525DIGlobalVariable
4526""""""""""""""""
4527
4528``DIGlobalVariable`` nodes represent global variables in the source language.
4529
4530.. code-block:: text
4531
4532    @foo = global i32, !dbg !0
4533    !0 = !DIGlobalVariableExpression(var: !1, expr: !DIExpression())
4534    !1 = !DIGlobalVariable(name: "foo", linkageName: "foo", scope: !2,
4535                           file: !3, line: 7, type: !4, isLocal: true,
4536                           isDefinition: false, declaration: !5)
4537
4538
4539DIGlobalVariableExpression
4540""""""""""""""""""""""""""
4541
4542``DIGlobalVariableExpression`` nodes tie a :ref:`DIGlobalVariable` together
4543with a :ref:`DIExpression`.
4544
4545.. code-block:: text
4546
4547    @lower = global i32, !dbg !0
4548    @upper = global i32, !dbg !1
4549    !0 = !DIGlobalVariableExpression(
4550             var: !2,
4551             expr: !DIExpression(DW_OP_LLVM_fragment, 0, 32)
4552             )
4553    !1 = !DIGlobalVariableExpression(
4554             var: !2,
4555             expr: !DIExpression(DW_OP_LLVM_fragment, 32, 32)
4556             )
4557    !2 = !DIGlobalVariable(name: "split64", linkageName: "split64", scope: !3,
4558                           file: !4, line: 8, type: !5, declaration: !6)
4559
4560All global variable expressions should be referenced by the `globals:` field of
4561a :ref:`compile unit <DICompileUnit>`.
4562
4563.. _DISubprogram:
4564
4565DISubprogram
4566""""""""""""
4567
4568``DISubprogram`` nodes represent functions from the source language. A
4569``DISubprogram`` may be attached to a function definition using ``!dbg``
4570metadata. The ``variables:`` field points at :ref:`variables <DILocalVariable>`
4571that must be retained, even if their IR counterparts are optimized out of
4572the IR. The ``type:`` field must point at an :ref:`DISubroutineType`.
4573
4574.. _DISubprogramDeclaration:
4575
4576When ``isDefinition: false``, subprograms describe a declaration in the type
4577tree as opposed to a definition of a function.  If the scope is a composite
4578type with an ODR ``identifier:`` and that does not set ``flags: DIFwdDecl``,
4579then the subprogram declaration is uniqued based only on its ``linkageName:``
4580and ``scope:``.
4581
4582.. code-block:: text
4583
4584    define void @_Z3foov() !dbg !0 {
4585      ...
4586    }
4587
4588    !0 = distinct !DISubprogram(name: "foo", linkageName: "_Zfoov", scope: !1,
4589                                file: !2, line: 7, type: !3, isLocal: true,
4590                                isDefinition: true, scopeLine: 8,
4591                                containingType: !4,
4592                                virtuality: DW_VIRTUALITY_pure_virtual,
4593                                virtualIndex: 10, flags: DIFlagPrototyped,
4594                                isOptimized: true, unit: !5, templateParams: !6,
4595                                declaration: !7, variables: !8, thrownTypes: !9)
4596
4597.. _DILexicalBlock:
4598
4599DILexicalBlock
4600""""""""""""""
4601
4602``DILexicalBlock`` nodes describe nested blocks within a :ref:`subprogram
4603<DISubprogram>`. The line number and column numbers are used to distinguish
4604two lexical blocks at same depth. They are valid targets for ``scope:``
4605fields.
4606
4607.. code-block:: text
4608
4609    !0 = distinct !DILexicalBlock(scope: !1, file: !2, line: 7, column: 35)
4610
4611Usually lexical blocks are ``distinct`` to prevent node merging based on
4612operands.
4613
4614.. _DILexicalBlockFile:
4615
4616DILexicalBlockFile
4617""""""""""""""""""
4618
4619``DILexicalBlockFile`` nodes are used to discriminate between sections of a
4620:ref:`lexical block <DILexicalBlock>`. The ``file:`` field can be changed to
4621indicate textual inclusion, or the ``discriminator:`` field can be used to
4622discriminate between control flow within a single block in the source language.
4623
4624.. code-block:: text
4625
4626    !0 = !DILexicalBlock(scope: !3, file: !4, line: 7, column: 35)
4627    !1 = !DILexicalBlockFile(scope: !0, file: !4, discriminator: 0)
4628    !2 = !DILexicalBlockFile(scope: !0, file: !4, discriminator: 1)
4629
4630.. _DILocation:
4631
4632DILocation
4633""""""""""
4634
4635``DILocation`` nodes represent source debug locations. The ``scope:`` field is
4636mandatory, and points at an :ref:`DILexicalBlockFile`, an
4637:ref:`DILexicalBlock`, or an :ref:`DISubprogram`.
4638
4639.. code-block:: text
4640
4641    !0 = !DILocation(line: 2900, column: 42, scope: !1, inlinedAt: !2)
4642
4643.. _DILocalVariable:
4644
4645DILocalVariable
4646"""""""""""""""
4647
4648``DILocalVariable`` nodes represent local variables in the source language. If
4649the ``arg:`` field is set to non-zero, then this variable is a subprogram
4650parameter, and it will be included in the ``variables:`` field of its
4651:ref:`DISubprogram`.
4652
4653.. code-block:: text
4654
4655    !0 = !DILocalVariable(name: "this", arg: 1, scope: !3, file: !2, line: 7,
4656                          type: !3, flags: DIFlagArtificial)
4657    !1 = !DILocalVariable(name: "x", arg: 2, scope: !4, file: !2, line: 7,
4658                          type: !3)
4659    !2 = !DILocalVariable(name: "y", scope: !5, file: !2, line: 7, type: !3)
4660
4661.. _DIExpression:
4662
4663DIExpression
4664""""""""""""
4665
4666``DIExpression`` nodes represent expressions that are inspired by the DWARF
4667expression language. They are used in :ref:`debug intrinsics<dbg_intrinsics>`
4668(such as ``llvm.dbg.declare`` and ``llvm.dbg.value``) to describe how the
4669referenced LLVM variable relates to the source language variable. Debug
4670intrinsics are interpreted left-to-right: start by pushing the value/address
4671operand of the intrinsic onto a stack, then repeatedly push and evaluate
4672opcodes from the DIExpression until the final variable description is produced.
4673
4674The current supported opcode vocabulary is limited:
4675
4676- ``DW_OP_deref`` dereferences the top of the expression stack.
4677- ``DW_OP_plus`` pops the last two entries from the expression stack, adds
4678  them together and appends the result to the expression stack.
4679- ``DW_OP_minus`` pops the last two entries from the expression stack, subtracts
4680  the last entry from the second last entry and appends the result to the
4681  expression stack.
4682- ``DW_OP_plus_uconst, 93`` adds ``93`` to the working expression.
4683- ``DW_OP_LLVM_fragment, 16, 8`` specifies the offset and size (``16`` and ``8``
4684  here, respectively) of the variable fragment from the working expression. Note
4685  that contrary to DW_OP_bit_piece, the offset is describing the location
4686  within the described source variable.
4687- ``DW_OP_LLVM_convert, 16, DW_ATE_signed`` specifies a bit size and encoding
4688  (``16`` and ``DW_ATE_signed`` here, respectively) to which the top of the
4689  expression stack is to be converted. Maps into a ``DW_OP_convert`` operation
4690  that references a base type constructed from the supplied values.
4691- ``DW_OP_swap`` swaps top two stack entries.
4692- ``DW_OP_xderef`` provides extended dereference mechanism. The entry at the top
4693  of the stack is treated as an address. The second stack entry is treated as an
4694  address space identifier.
4695- ``DW_OP_stack_value`` marks a constant value.
4696
4697DWARF specifies three kinds of simple location descriptions: Register, memory,
4698and implicit location descriptions.  Note that a location description is
4699defined over certain ranges of a program, i.e the location of a variable may
4700change over the course of the program. Register and memory location
4701descriptions describe the *concrete location* of a source variable (in the
4702sense that a debugger might modify its value), whereas *implicit locations*
4703describe merely the actual *value* of a source variable which might not exist
4704in registers or in memory (see ``DW_OP_stack_value``).
4705
4706A ``llvm.dbg.addr`` or ``llvm.dbg.declare`` intrinsic describes an indirect
4707value (the address) of a source variable. The first operand of the intrinsic
4708must be an address of some kind. A DIExpression attached to the intrinsic
4709refines this address to produce a concrete location for the source variable.
4710
4711A ``llvm.dbg.value`` intrinsic describes the direct value of a source variable.
4712The first operand of the intrinsic may be a direct or indirect value. A
4713DIExpresion attached to the intrinsic refines the first operand to produce a
4714direct value. For example, if the first operand is an indirect value, it may be
4715necessary to insert ``DW_OP_deref`` into the DIExpresion in order to produce a
4716valid debug intrinsic.
4717
4718.. note::
4719
4720   A DIExpression is interpreted in the same way regardless of which kind of
4721   debug intrinsic it's attached to.
4722
4723.. code-block:: text
4724
4725    !0 = !DIExpression(DW_OP_deref)
4726    !1 = !DIExpression(DW_OP_plus_uconst, 3)
4727    !1 = !DIExpression(DW_OP_constu, 3, DW_OP_plus)
4728    !2 = !DIExpression(DW_OP_bit_piece, 3, 7)
4729    !3 = !DIExpression(DW_OP_deref, DW_OP_constu, 3, DW_OP_plus, DW_OP_LLVM_fragment, 3, 7)
4730    !4 = !DIExpression(DW_OP_constu, 2, DW_OP_swap, DW_OP_xderef)
4731    !5 = !DIExpression(DW_OP_constu, 42, DW_OP_stack_value)
4732
4733DIObjCProperty
4734""""""""""""""
4735
4736``DIObjCProperty`` nodes represent Objective-C property nodes.
4737
4738.. code-block:: text
4739
4740    !3 = !DIObjCProperty(name: "foo", file: !1, line: 7, setter: "setFoo",
4741                         getter: "getFoo", attributes: 7, type: !2)
4742
4743DIImportedEntity
4744""""""""""""""""
4745
4746``DIImportedEntity`` nodes represent entities (such as modules) imported into a
4747compile unit.
4748
4749.. code-block:: text
4750
4751   !2 = !DIImportedEntity(tag: DW_TAG_imported_module, name: "foo", scope: !0,
4752                          entity: !1, line: 7)
4753
4754DIMacro
4755"""""""
4756
4757``DIMacro`` nodes represent definition or undefinition of a macro identifiers.
4758The ``name:`` field is the macro identifier, followed by macro parameters when
4759defining a function-like macro, and the ``value`` field is the token-string
4760used to expand the macro identifier.
4761
4762.. code-block:: text
4763
4764   !2 = !DIMacro(macinfo: DW_MACINFO_define, line: 7, name: "foo(x)",
4765                 value: "((x) + 1)")
4766   !3 = !DIMacro(macinfo: DW_MACINFO_undef, line: 30, name: "foo")
4767
4768DIMacroFile
4769"""""""""""
4770
4771``DIMacroFile`` nodes represent inclusion of source files.
4772The ``nodes:`` field is a list of ``DIMacro`` and ``DIMacroFile`` nodes that
4773appear in the included source file.
4774
4775.. code-block:: text
4776
4777   !2 = !DIMacroFile(macinfo: DW_MACINFO_start_file, line: 7, file: !2,
4778                     nodes: !3)
4779
4780'``tbaa``' Metadata
4781^^^^^^^^^^^^^^^^^^^
4782
4783In LLVM IR, memory does not have types, so LLVM's own type system is not
4784suitable for doing type based alias analysis (TBAA). Instead, metadata is
4785added to the IR to describe a type system of a higher level language. This
4786can be used to implement C/C++ strict type aliasing rules, but it can also
4787be used to implement custom alias analysis behavior for other languages.
4788
4789This description of LLVM's TBAA system is broken into two parts:
4790:ref:`Semantics<tbaa_node_semantics>` talks about high level issues, and
4791:ref:`Representation<tbaa_node_representation>` talks about the metadata
4792encoding of various entities.
4793
4794It is always possible to trace any TBAA node to a "root" TBAA node (details
4795in the :ref:`Representation<tbaa_node_representation>` section).  TBAA
4796nodes with different roots have an unknown aliasing relationship, and LLVM
4797conservatively infers ``MayAlias`` between them.  The rules mentioned in
4798this section only pertain to TBAA nodes living under the same root.
4799
4800.. _tbaa_node_semantics:
4801
4802Semantics
4803"""""""""
4804
4805The TBAA metadata system, referred to as "struct path TBAA" (not to be
4806confused with ``tbaa.struct``), consists of the following high level
4807concepts: *Type Descriptors*, further subdivided into scalar type
4808descriptors and struct type descriptors; and *Access Tags*.
4809
4810**Type descriptors** describe the type system of the higher level language
4811being compiled.  **Scalar type descriptors** describe types that do not
4812contain other types.  Each scalar type has a parent type, which must also
4813be a scalar type or the TBAA root.  Via this parent relation, scalar types
4814within a TBAA root form a tree.  **Struct type descriptors** denote types
4815that contain a sequence of other type descriptors, at known offsets.  These
4816contained type descriptors can either be struct type descriptors themselves
4817or scalar type descriptors.
4818
4819**Access tags** are metadata nodes attached to load and store instructions.
4820Access tags use type descriptors to describe the *location* being accessed
4821in terms of the type system of the higher level language.  Access tags are
4822tuples consisting of a base type, an access type and an offset.  The base
4823type is a scalar type descriptor or a struct type descriptor, the access
4824type is a scalar type descriptor, and the offset is a constant integer.
4825
4826The access tag ``(BaseTy, AccessTy, Offset)`` can describe one of two
4827things:
4828
4829 * If ``BaseTy`` is a struct type, the tag describes a memory access (load
4830   or store) of a value of type ``AccessTy`` contained in the struct type
4831   ``BaseTy`` at offset ``Offset``.
4832
4833 * If ``BaseTy`` is a scalar type, ``Offset`` must be 0 and ``BaseTy`` and
4834   ``AccessTy`` must be the same; and the access tag describes a scalar
4835   access with scalar type ``AccessTy``.
4836
4837We first define an ``ImmediateParent`` relation on ``(BaseTy, Offset)``
4838tuples this way:
4839
4840 * If ``BaseTy`` is a scalar type then ``ImmediateParent(BaseTy, 0)`` is
4841   ``(ParentTy, 0)`` where ``ParentTy`` is the parent of the scalar type as
4842   described in the TBAA metadata.  ``ImmediateParent(BaseTy, Offset)`` is
4843   undefined if ``Offset`` is non-zero.
4844
4845 * If ``BaseTy`` is a struct type then ``ImmediateParent(BaseTy, Offset)``
4846   is ``(NewTy, NewOffset)`` where ``NewTy`` is the type contained in
4847   ``BaseTy`` at offset ``Offset`` and ``NewOffset`` is ``Offset`` adjusted
4848   to be relative within that inner type.
4849
4850A memory access with an access tag ``(BaseTy1, AccessTy1, Offset1)``
4851aliases a memory access with an access tag ``(BaseTy2, AccessTy2,
4852Offset2)`` if either ``(BaseTy1, Offset1)`` is reachable from ``(Base2,
4853Offset2)`` via the ``Parent`` relation or vice versa.
4854
4855As a concrete example, the type descriptor graph for the following program
4856
4857.. code-block:: c
4858
4859    struct Inner {
4860      int i;    // offset 0
4861      float f;  // offset 4
4862    };
4863
4864    struct Outer {
4865      float f;  // offset 0
4866      double d; // offset 4
4867      struct Inner inner_a;  // offset 12
4868    };
4869
4870    void f(struct Outer* outer, struct Inner* inner, float* f, int* i, char* c) {
4871      outer->f = 0;            // tag0: (OuterStructTy, FloatScalarTy, 0)
4872      outer->inner_a.i = 0;    // tag1: (OuterStructTy, IntScalarTy, 12)
4873      outer->inner_a.f = 0.0;  // tag2: (OuterStructTy, FloatScalarTy, 16)
4874      *f = 0.0;                // tag3: (FloatScalarTy, FloatScalarTy, 0)
4875    }
4876
4877is (note that in C and C++, ``char`` can be used to access any arbitrary
4878type):
4879
4880.. code-block:: text
4881
4882    Root = "TBAA Root"
4883    CharScalarTy = ("char", Root, 0)
4884    FloatScalarTy = ("float", CharScalarTy, 0)
4885    DoubleScalarTy = ("double", CharScalarTy, 0)
4886    IntScalarTy = ("int", CharScalarTy, 0)
4887    InnerStructTy = {"Inner" (IntScalarTy, 0), (FloatScalarTy, 4)}
4888    OuterStructTy = {"Outer", (FloatScalarTy, 0), (DoubleScalarTy, 4),
4889                     (InnerStructTy, 12)}
4890
4891
4892with (e.g.) ``ImmediateParent(OuterStructTy, 12)`` = ``(InnerStructTy,
48930)``, ``ImmediateParent(InnerStructTy, 0)`` = ``(IntScalarTy, 0)``, and
4894``ImmediateParent(IntScalarTy, 0)`` = ``(CharScalarTy, 0)``.
4895
4896.. _tbaa_node_representation:
4897
4898Representation
4899""""""""""""""
4900
4901The root node of a TBAA type hierarchy is an ``MDNode`` with 0 operands or
4902with exactly one ``MDString`` operand.
4903
4904Scalar type descriptors are represented as an ``MDNode`` s with two
4905operands.  The first operand is an ``MDString`` denoting the name of the
4906struct type.  LLVM does not assign meaning to the value of this operand, it
4907only cares about it being an ``MDString``.  The second operand is an
4908``MDNode`` which points to the parent for said scalar type descriptor,
4909which is either another scalar type descriptor or the TBAA root.  Scalar
4910type descriptors can have an optional third argument, but that must be the
4911constant integer zero.
4912
4913Struct type descriptors are represented as ``MDNode`` s with an odd number
4914of operands greater than 1.  The first operand is an ``MDString`` denoting
4915the name of the struct type.  Like in scalar type descriptors the actual
4916value of this name operand is irrelevant to LLVM.  After the name operand,
4917the struct type descriptors have a sequence of alternating ``MDNode`` and
4918``ConstantInt`` operands.  With N starting from 1, the 2N - 1 th operand,
4919an ``MDNode``, denotes a contained field, and the 2N th operand, a
4920``ConstantInt``, is the offset of the said contained field.  The offsets
4921must be in non-decreasing order.
4922
4923Access tags are represented as ``MDNode`` s with either 3 or 4 operands.
4924The first operand is an ``MDNode`` pointing to the node representing the
4925base type.  The second operand is an ``MDNode`` pointing to the node
4926representing the access type.  The third operand is a ``ConstantInt`` that
4927states the offset of the access.  If a fourth field is present, it must be
4928a ``ConstantInt`` valued at 0 or 1.  If it is 1 then the access tag states
4929that the location being accessed is "constant" (meaning
4930``pointsToConstantMemory`` should return true; see `other useful
4931AliasAnalysis methods <AliasAnalysis.html#OtherItfs>`_).  The TBAA root of
4932the access type and the base type of an access tag must be the same, and
4933that is the TBAA root of the access tag.
4934
4935'``tbaa.struct``' Metadata
4936^^^^^^^^^^^^^^^^^^^^^^^^^^
4937
4938The :ref:`llvm.memcpy <int_memcpy>` is often used to implement
4939aggregate assignment operations in C and similar languages, however it
4940is defined to copy a contiguous region of memory, which is more than
4941strictly necessary for aggregate types which contain holes due to
4942padding. Also, it doesn't contain any TBAA information about the fields
4943of the aggregate.
4944
4945``!tbaa.struct`` metadata can describe which memory subregions in a
4946memcpy are padding and what the TBAA tags of the struct are.
4947
4948The current metadata format is very simple. ``!tbaa.struct`` metadata
4949nodes are a list of operands which are in conceptual groups of three.
4950For each group of three, the first operand gives the byte offset of a
4951field in bytes, the second gives its size in bytes, and the third gives
4952its tbaa tag. e.g.:
4953
4954.. code-block:: llvm
4955
4956    !4 = !{ i64 0, i64 4, !1, i64 8, i64 4, !2 }
4957
4958This describes a struct with two fields. The first is at offset 0 bytes
4959with size 4 bytes, and has tbaa tag !1. The second is at offset 8 bytes
4960and has size 4 bytes and has tbaa tag !2.
4961
4962Note that the fields need not be contiguous. In this example, there is a
49634 byte gap between the two fields. This gap represents padding which
4964does not carry useful data and need not be preserved.
4965
4966'``noalias``' and '``alias.scope``' Metadata
4967^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4968
4969``noalias`` and ``alias.scope`` metadata provide the ability to specify generic
4970noalias memory-access sets. This means that some collection of memory access
4971instructions (loads, stores, memory-accessing calls, etc.) that carry
4972``noalias`` metadata can specifically be specified not to alias with some other
4973collection of memory access instructions that carry ``alias.scope`` metadata.
4974Each type of metadata specifies a list of scopes where each scope has an id and
4975a domain.
4976
4977When evaluating an aliasing query, if for some domain, the set
4978of scopes with that domain in one instruction's ``alias.scope`` list is a
4979subset of (or equal to) the set of scopes for that domain in another
4980instruction's ``noalias`` list, then the two memory accesses are assumed not to
4981alias.
4982
4983Because scopes in one domain don't affect scopes in other domains, separate
4984domains can be used to compose multiple independent noalias sets.  This is
4985used for example during inlining.  As the noalias function parameters are
4986turned into noalias scope metadata, a new domain is used every time the
4987function is inlined.
4988
4989The metadata identifying each domain is itself a list containing one or two
4990entries. The first entry is the name of the domain. Note that if the name is a
4991string then it can be combined across functions and translation units. A
4992self-reference can be used to create globally unique domain names. A
4993descriptive string may optionally be provided as a second list entry.
4994
4995The metadata identifying each scope is also itself a list containing two or
4996three entries. The first entry is the name of the scope. Note that if the name
4997is a string then it can be combined across functions and translation units. A
4998self-reference can be used to create globally unique scope names. A metadata
4999reference to the scope's domain is the second entry. A descriptive string may
5000optionally be provided as a third list entry.
5001
5002For example,
5003
5004.. code-block:: llvm
5005
5006    ; Two scope domains:
5007    !0 = !{!0}
5008    !1 = !{!1}
5009
5010    ; Some scopes in these domains:
5011    !2 = !{!2, !0}
5012    !3 = !{!3, !0}
5013    !4 = !{!4, !1}
5014
5015    ; Some scope lists:
5016    !5 = !{!4} ; A list containing only scope !4
5017    !6 = !{!4, !3, !2}
5018    !7 = !{!3}
5019
5020    ; These two instructions don't alias:
5021    %0 = load float, float* %c, align 4, !alias.scope !5
5022    store float %0, float* %arrayidx.i, align 4, !noalias !5
5023
5024    ; These two instructions also don't alias (for domain !1, the set of scopes
5025    ; in the !alias.scope equals that in the !noalias list):
5026    %2 = load float, float* %c, align 4, !alias.scope !5
5027    store float %2, float* %arrayidx.i2, align 4, !noalias !6
5028
5029    ; These two instructions may alias (for domain !0, the set of scopes in
5030    ; the !noalias list is not a superset of, or equal to, the scopes in the
5031    ; !alias.scope list):
5032    %2 = load float, float* %c, align 4, !alias.scope !6
5033    store float %0, float* %arrayidx.i, align 4, !noalias !7
5034
5035'``fpmath``' Metadata
5036^^^^^^^^^^^^^^^^^^^^^
5037
5038``fpmath`` metadata may be attached to any instruction of floating-point
5039type. It can be used to express the maximum acceptable error in the
5040result of that instruction, in ULPs, thus potentially allowing the
5041compiler to use a more efficient but less accurate method of computing
5042it. ULP is defined as follows:
5043
5044    If ``x`` is a real number that lies between two finite consecutive
5045    floating-point numbers ``a`` and ``b``, without being equal to one
5046    of them, then ``ulp(x) = |b - a|``, otherwise ``ulp(x)`` is the
5047    distance between the two non-equal finite floating-point numbers
5048    nearest ``x``. Moreover, ``ulp(NaN)`` is ``NaN``.
5049
5050The metadata node shall consist of a single positive float type number
5051representing the maximum relative error, for example:
5052
5053.. code-block:: llvm
5054
5055    !0 = !{ float 2.5 } ; maximum acceptable inaccuracy is 2.5 ULPs
5056
5057.. _range-metadata:
5058
5059'``range``' Metadata
5060^^^^^^^^^^^^^^^^^^^^
5061
5062``range`` metadata may be attached only to ``load``, ``call`` and ``invoke`` of
5063integer types. It expresses the possible ranges the loaded value or the value
5064returned by the called function at this call site is in. If the loaded or
5065returned value is not in the specified range, the behavior is undefined. The
5066ranges are represented with a flattened list of integers. The loaded value or
5067the value returned is known to be in the union of the ranges defined by each
5068consecutive pair. Each pair has the following properties:
5069
5070-  The type must match the type loaded by the instruction.
5071-  The pair ``a,b`` represents the range ``[a,b)``.
5072-  Both ``a`` and ``b`` are constants.
5073-  The range is allowed to wrap.
5074-  The range should not represent the full or empty set. That is,
5075   ``a!=b``.
5076
5077In addition, the pairs must be in signed order of the lower bound and
5078they must be non-contiguous.
5079
5080Examples:
5081
5082.. code-block:: llvm
5083
5084      %a = load i8, i8* %x, align 1, !range !0 ; Can only be 0 or 1
5085      %b = load i8, i8* %y, align 1, !range !1 ; Can only be 255 (-1), 0 or 1
5086      %c = call i8 @foo(),       !range !2 ; Can only be 0, 1, 3, 4 or 5
5087      %d = invoke i8 @bar() to label %cont
5088             unwind label %lpad, !range !3 ; Can only be -2, -1, 3, 4 or 5
5089    ...
5090    !0 = !{ i8 0, i8 2 }
5091    !1 = !{ i8 255, i8 2 }
5092    !2 = !{ i8 0, i8 2, i8 3, i8 6 }
5093    !3 = !{ i8 -2, i8 0, i8 3, i8 6 }
5094
5095'``absolute_symbol``' Metadata
5096^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5097
5098``absolute_symbol`` metadata may be attached to a global variable
5099declaration. It marks the declaration as a reference to an absolute symbol,
5100which causes the backend to use absolute relocations for the symbol even
5101in position independent code, and expresses the possible ranges that the
5102global variable's *address* (not its value) is in, in the same format as
5103``range`` metadata, with the extension that the pair ``all-ones,all-ones``
5104may be used to represent the full set.
5105
5106Example (assuming 64-bit pointers):
5107
5108.. code-block:: llvm
5109
5110      @a = external global i8, !absolute_symbol !0 ; Absolute symbol in range [0,256)
5111      @b = external global i8, !absolute_symbol !1 ; Absolute symbol in range [0,2^64)
5112
5113    ...
5114    !0 = !{ i64 0, i64 256 }
5115    !1 = !{ i64 -1, i64 -1 }
5116
5117'``callees``' Metadata
5118^^^^^^^^^^^^^^^^^^^^^^
5119
5120``callees`` metadata may be attached to indirect call sites. If ``callees``
5121metadata is attached to a call site, and any callee is not among the set of
5122functions provided by the metadata, the behavior is undefined. The intent of
5123this metadata is to facilitate optimizations such as indirect-call promotion.
5124For example, in the code below, the call instruction may only target the
5125``add`` or ``sub`` functions:
5126
5127.. code-block:: llvm
5128
5129    %result = call i64 %binop(i64 %x, i64 %y), !callees !0
5130
5131    ...
5132    !0 = !{i64 (i64, i64)* @add, i64 (i64, i64)* @sub}
5133
5134'``callback``' Metadata
5135^^^^^^^^^^^^^^^^^^^^^^^
5136
5137``callback`` metadata may be attached to a function declaration, or definition.
5138(Call sites are excluded only due to the lack of a use case.) For ease of
5139exposition, we'll refer to the function annotated w/ metadata as a broker
5140function. The metadata describes how the arguments of a call to the broker are
5141in turn passed to the callback function specified by the metadata. Thus, the
5142``callback`` metadata provides a partial description of a call site inside the
5143broker function with regards to the arguments of a call to the broker. The only
5144semantic restriction on the broker function itself is that it is not allowed to
5145inspect or modify arguments referenced in the ``callback`` metadata as
5146pass-through to the callback function.
5147
5148The broker is not required to actually invoke the callback function at runtime.
5149However, the assumptions about not inspecting or modifying arguments that would
5150be passed to the specified callback function still hold, even if the callback
5151function is not dynamically invoked. The broker is allowed to invoke the
5152callback function more than once per invocation of the broker. The broker is
5153also allowed to invoke (directly or indirectly) the function passed as a
5154callback through another use. Finally, the broker is also allowed to relay the
5155callback callee invocation to a different thread.
5156
5157The metadata is structured as follows: At the outer level, ``callback``
5158metadata is a list of ``callback`` encodings. Each encoding starts with a
5159constant ``i64`` which describes the argument position of the callback function
5160in the call to the broker. The following elements, except the last, describe
5161what arguments are passed to the callback function. Each element is again an
5162``i64`` constant identifying the argument of the broker that is passed through,
5163or ``i64 -1`` to indicate an unknown or inspected argument. The order in which
5164they are listed has to be the same in which they are passed to the callback
5165callee. The last element of the encoding is a boolean which specifies how
5166variadic arguments of the broker are handled. If it is true, all variadic
5167arguments of the broker are passed through to the callback function *after* the
5168arguments encoded explicitly before.
5169
5170In the code below, the ``pthread_create`` function is marked as a broker
5171through the ``!callback !1`` metadata. In the example, there is only one
5172callback encoding, namely ``!2``, associated with the broker. This encoding
5173identifies the callback function as the second argument of the broker (``i64
51742``) and the sole argument of the callback function as the third one of the
5175broker function (``i64 3``).
5176
5177.. FIXME why does the llvm-sphinx-docs builder give a highlighting
5178   error if the below is set to highlight as 'llvm', despite that we
5179   have misc.highlighting_failure set?
5180
5181.. code-block:: text
5182
5183    declare !callback !1 dso_local i32 @pthread_create(i64*, %union.pthread_attr_t*, i8* (i8*)*, i8*)
5184
5185    ...
5186    !2 = !{i64 2, i64 3, i1 false}
5187    !1 = !{!2}
5188
5189Another example is shown below. The callback callee is the second argument of
5190the ``__kmpc_fork_call`` function (``i64 2``). The callee is given two unknown
5191values (each identified by a ``i64 -1``) and afterwards all
5192variadic arguments that are passed to the ``__kmpc_fork_call`` call (due to the
5193final ``i1 true``).
5194
5195.. FIXME why does the llvm-sphinx-docs builder give a highlighting
5196   error if the below is set to highlight as 'llvm', despite that we
5197   have misc.highlighting_failure set?
5198
5199.. code-block:: text
5200
5201    declare !callback !0 dso_local void @__kmpc_fork_call(%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...)
5202
5203    ...
5204    !1 = !{i64 2, i64 -1, i64 -1, i1 true}
5205    !0 = !{!1}
5206
5207
5208'``unpredictable``' Metadata
5209^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5210
5211``unpredictable`` metadata may be attached to any branch or switch
5212instruction. It can be used to express the unpredictability of control
5213flow. Similar to the llvm.expect intrinsic, it may be used to alter
5214optimizations related to compare and branch instructions. The metadata
5215is treated as a boolean value; if it exists, it signals that the branch
5216or switch that it is attached to is completely unpredictable.
5217
5218.. _llvm.loop:
5219
5220'``llvm.loop``'
5221^^^^^^^^^^^^^^^
5222
5223It is sometimes useful to attach information to loop constructs. Currently,
5224loop metadata is implemented as metadata attached to the branch instruction
5225in the loop latch block. This type of metadata refer to a metadata node that is
5226guaranteed to be separate for each loop. The loop identifier metadata is
5227specified with the name ``llvm.loop``.
5228
5229The loop identifier metadata is implemented using a metadata that refers to
5230itself to avoid merging it with any other identifier metadata, e.g.,
5231during module linkage or function inlining. That is, each loop should refer
5232to their own identification metadata even if they reside in separate functions.
5233The following example contains loop identifier metadata for two separate loop
5234constructs:
5235
5236.. code-block:: llvm
5237
5238    !0 = !{!0}
5239    !1 = !{!1}
5240
5241The loop identifier metadata can be used to specify additional
5242per-loop metadata. Any operands after the first operand can be treated
5243as user-defined metadata. For example the ``llvm.loop.unroll.count``
5244suggests an unroll factor to the loop unroller:
5245
5246.. code-block:: llvm
5247
5248      br i1 %exitcond, label %._crit_edge, label %.lr.ph, !llvm.loop !0
5249    ...
5250    !0 = !{!0, !1}
5251    !1 = !{!"llvm.loop.unroll.count", i32 4}
5252
5253'``llvm.loop.disable_nonforced``'
5254^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5255
5256This metadata disables all optional loop transformations unless
5257explicitly instructed using other transformation metdata such as
5258``llvm.loop.unroll.enable``. That is, no heuristic will try to determine
5259whether a transformation is profitable. The purpose is to avoid that the
5260loop is transformed to a different loop before an explicitly requested
5261(forced) transformation is applied. For instance, loop fusion can make
5262other transformations impossible. Mandatory loop canonicalizations such
5263as loop rotation are still applied.
5264
5265It is recommended to use this metadata in addition to any llvm.loop.*
5266transformation directive. Also, any loop should have at most one
5267directive applied to it (and a sequence of transformations built using
5268followup-attributes). Otherwise, which transformation will be applied
5269depends on implementation details such as the pass pipeline order.
5270
5271See :ref:`transformation-metadata` for details.
5272
5273'``llvm.loop.vectorize``' and '``llvm.loop.interleave``'
5274^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5275
5276Metadata prefixed with ``llvm.loop.vectorize`` or ``llvm.loop.interleave`` are
5277used to control per-loop vectorization and interleaving parameters such as
5278vectorization width and interleave count. These metadata should be used in
5279conjunction with ``llvm.loop`` loop identification metadata. The
5280``llvm.loop.vectorize`` and ``llvm.loop.interleave`` metadata are only
5281optimization hints and the optimizer will only interleave and vectorize loops if
5282it believes it is safe to do so. The ``llvm.loop.parallel_accesses`` metadata
5283which contains information about loop-carried memory dependencies can be helpful
5284in determining the safety of these transformations.
5285
5286'``llvm.loop.interleave.count``' Metadata
5287^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5288
5289This metadata suggests an interleave count to the loop interleaver.
5290The first operand is the string ``llvm.loop.interleave.count`` and the
5291second operand is an integer specifying the interleave count. For
5292example:
5293
5294.. code-block:: llvm
5295
5296   !0 = !{!"llvm.loop.interleave.count", i32 4}
5297
5298Note that setting ``llvm.loop.interleave.count`` to 1 disables interleaving
5299multiple iterations of the loop. If ``llvm.loop.interleave.count`` is set to 0
5300then the interleave count will be determined automatically.
5301
5302'``llvm.loop.vectorize.enable``' Metadata
5303^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5304
5305This metadata selectively enables or disables vectorization for the loop. The
5306first operand is the string ``llvm.loop.vectorize.enable`` and the second operand
5307is a bit. If the bit operand value is 1 vectorization is enabled. A value of
53080 disables vectorization:
5309
5310.. code-block:: llvm
5311
5312   !0 = !{!"llvm.loop.vectorize.enable", i1 0}
5313   !1 = !{!"llvm.loop.vectorize.enable", i1 1}
5314
5315'``llvm.loop.vectorize.width``' Metadata
5316^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5317
5318This metadata sets the target width of the vectorizer. The first
5319operand is the string ``llvm.loop.vectorize.width`` and the second
5320operand is an integer specifying the width. For example:
5321
5322.. code-block:: llvm
5323
5324   !0 = !{!"llvm.loop.vectorize.width", i32 4}
5325
5326Note that setting ``llvm.loop.vectorize.width`` to 1 disables
5327vectorization of the loop. If ``llvm.loop.vectorize.width`` is set to
53280 or if the loop does not have this metadata the width will be
5329determined automatically.
5330
5331'``llvm.loop.vectorize.followup_vectorized``' Metadata
5332^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5333
5334This metadata defines which loop attributes the vectorized loop will
5335have. See :ref:`transformation-metadata` for details.
5336
5337'``llvm.loop.vectorize.followup_epilogue``' Metadata
5338^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5339
5340This metadata defines which loop attributes the epilogue will have. The
5341epilogue is not vectorized and is executed when either the vectorized
5342loop is not known to preserve semantics (because e.g., it processes two
5343arrays that are found to alias by a runtime check) or for the last
5344iterations that do not fill a complete set of vector lanes. See
5345:ref:`Transformation Metadata <transformation-metadata>` for details.
5346
5347'``llvm.loop.vectorize.followup_all``' Metadata
5348^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5349
5350Attributes in the metadata will be added to both the vectorized and
5351epilogue loop.
5352See :ref:`Transformation Metadata <transformation-metadata>` for details.
5353
5354'``llvm.loop.unroll``'
5355^^^^^^^^^^^^^^^^^^^^^^
5356
5357Metadata prefixed with ``llvm.loop.unroll`` are loop unrolling
5358optimization hints such as the unroll factor. ``llvm.loop.unroll``
5359metadata should be used in conjunction with ``llvm.loop`` loop
5360identification metadata. The ``llvm.loop.unroll`` metadata are only
5361optimization hints and the unrolling will only be performed if the
5362optimizer believes it is safe to do so.
5363
5364'``llvm.loop.unroll.count``' Metadata
5365^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5366
5367This metadata suggests an unroll factor to the loop unroller. The
5368first operand is the string ``llvm.loop.unroll.count`` and the second
5369operand is a positive integer specifying the unroll factor. For
5370example:
5371
5372.. code-block:: llvm
5373
5374   !0 = !{!"llvm.loop.unroll.count", i32 4}
5375
5376If the trip count of the loop is less than the unroll count the loop
5377will be partially unrolled.
5378
5379'``llvm.loop.unroll.disable``' Metadata
5380^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5381
5382This metadata disables loop unrolling. The metadata has a single operand
5383which is the string ``llvm.loop.unroll.disable``. For example:
5384
5385.. code-block:: llvm
5386
5387   !0 = !{!"llvm.loop.unroll.disable"}
5388
5389'``llvm.loop.unroll.runtime.disable``' Metadata
5390^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5391
5392This metadata disables runtime loop unrolling. The metadata has a single
5393operand which is the string ``llvm.loop.unroll.runtime.disable``. For example:
5394
5395.. code-block:: llvm
5396
5397   !0 = !{!"llvm.loop.unroll.runtime.disable"}
5398
5399'``llvm.loop.unroll.enable``' Metadata
5400^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5401
5402This metadata suggests that the loop should be fully unrolled if the trip count
5403is known at compile time and partially unrolled if the trip count is not known
5404at compile time. The metadata has a single operand which is the string
5405``llvm.loop.unroll.enable``.  For example:
5406
5407.. code-block:: llvm
5408
5409   !0 = !{!"llvm.loop.unroll.enable"}
5410
5411'``llvm.loop.unroll.full``' Metadata
5412^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5413
5414This metadata suggests that the loop should be unrolled fully. The
5415metadata has a single operand which is the string ``llvm.loop.unroll.full``.
5416For example:
5417
5418.. code-block:: llvm
5419
5420   !0 = !{!"llvm.loop.unroll.full"}
5421
5422'``llvm.loop.unroll.followup``' Metadata
5423^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5424
5425This metadata defines which loop attributes the unrolled loop will have.
5426See :ref:`Transformation Metadata <transformation-metadata>` for details.
5427
5428'``llvm.loop.unroll.followup_remainder``' Metadata
5429^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5430
5431This metadata defines which loop attributes the remainder loop after
5432partial/runtime unrolling will have. See
5433:ref:`Transformation Metadata <transformation-metadata>` for details.
5434
5435'``llvm.loop.unroll_and_jam``'
5436^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5437
5438This metadata is treated very similarly to the ``llvm.loop.unroll`` metadata
5439above, but affect the unroll and jam pass. In addition any loop with
5440``llvm.loop.unroll`` metadata but no ``llvm.loop.unroll_and_jam`` metadata will
5441disable unroll and jam (so ``llvm.loop.unroll`` metadata will be left to the
5442unroller, plus ``llvm.loop.unroll.disable`` metadata will disable unroll and jam
5443too.)
5444
5445The metadata for unroll and jam otherwise is the same as for ``unroll``.
5446``llvm.loop.unroll_and_jam.enable``, ``llvm.loop.unroll_and_jam.disable`` and
5447``llvm.loop.unroll_and_jam.count`` do the same as for unroll.
5448``llvm.loop.unroll_and_jam.full`` is not supported. Again these are only hints
5449and the normal safety checks will still be performed.
5450
5451'``llvm.loop.unroll_and_jam.count``' Metadata
5452^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5453
5454This metadata suggests an unroll and jam factor to use, similarly to
5455``llvm.loop.unroll.count``. The first operand is the string
5456``llvm.loop.unroll_and_jam.count`` and the second operand is a positive integer
5457specifying the unroll factor. For example:
5458
5459.. code-block:: llvm
5460
5461   !0 = !{!"llvm.loop.unroll_and_jam.count", i32 4}
5462
5463If the trip count of the loop is less than the unroll count the loop
5464will be partially unroll and jammed.
5465
5466'``llvm.loop.unroll_and_jam.disable``' Metadata
5467^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5468
5469This metadata disables loop unroll and jamming. The metadata has a single
5470operand which is the string ``llvm.loop.unroll_and_jam.disable``. For example:
5471
5472.. code-block:: llvm
5473
5474   !0 = !{!"llvm.loop.unroll_and_jam.disable"}
5475
5476'``llvm.loop.unroll_and_jam.enable``' Metadata
5477^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5478
5479This metadata suggests that the loop should be fully unroll and jammed if the
5480trip count is known at compile time and partially unrolled if the trip count is
5481not known at compile time. The metadata has a single operand which is the
5482string ``llvm.loop.unroll_and_jam.enable``.  For example:
5483
5484.. code-block:: llvm
5485
5486   !0 = !{!"llvm.loop.unroll_and_jam.enable"}
5487
5488'``llvm.loop.unroll_and_jam.followup_outer``' Metadata
5489^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5490
5491This metadata defines which loop attributes the outer unrolled loop will
5492have. See :ref:`Transformation Metadata <transformation-metadata>` for
5493details.
5494
5495'``llvm.loop.unroll_and_jam.followup_inner``' Metadata
5496^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5497
5498This metadata defines which loop attributes the inner jammed loop will
5499have. See :ref:`Transformation Metadata <transformation-metadata>` for
5500details.
5501
5502'``llvm.loop.unroll_and_jam.followup_remainder_outer``' Metadata
5503^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5504
5505This metadata defines which attributes the epilogue of the outer loop
5506will have. This loop is usually unrolled, meaning there is no such
5507loop. This attribute will be ignored in this case. See
5508:ref:`Transformation Metadata <transformation-metadata>` for details.
5509
5510'``llvm.loop.unroll_and_jam.followup_remainder_inner``' Metadata
5511^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5512
5513This metadata defines which attributes the inner loop of the epilogue
5514will have. The outer epilogue will usually be unrolled, meaning there
5515can be multiple inner remainder loops. See
5516:ref:`Transformation Metadata <transformation-metadata>` for details.
5517
5518'``llvm.loop.unroll_and_jam.followup_all``' Metadata
5519^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5520
5521Attributes specified in the metadata is added to all
5522``llvm.loop.unroll_and_jam.*`` loops. See
5523:ref:`Transformation Metadata <transformation-metadata>` for details.
5524
5525'``llvm.loop.licm_versioning.disable``' Metadata
5526^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5527
5528This metadata indicates that the loop should not be versioned for the purpose
5529of enabling loop-invariant code motion (LICM). The metadata has a single operand
5530which is the string ``llvm.loop.licm_versioning.disable``. For example:
5531
5532.. code-block:: llvm
5533
5534   !0 = !{!"llvm.loop.licm_versioning.disable"}
5535
5536'``llvm.loop.distribute.enable``' Metadata
5537^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5538
5539Loop distribution allows splitting a loop into multiple loops.  Currently,
5540this is only performed if the entire loop cannot be vectorized due to unsafe
5541memory dependencies.  The transformation will attempt to isolate the unsafe
5542dependencies into their own loop.
5543
5544This metadata can be used to selectively enable or disable distribution of the
5545loop.  The first operand is the string ``llvm.loop.distribute.enable`` and the
5546second operand is a bit. If the bit operand value is 1 distribution is
5547enabled. A value of 0 disables distribution:
5548
5549.. code-block:: llvm
5550
5551   !0 = !{!"llvm.loop.distribute.enable", i1 0}
5552   !1 = !{!"llvm.loop.distribute.enable", i1 1}
5553
5554This metadata should be used in conjunction with ``llvm.loop`` loop
5555identification metadata.
5556
5557'``llvm.loop.distribute.followup_coincident``' Metadata
5558^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5559
5560This metadata defines which attributes extracted loops with no cyclic
5561dependencies will have (i.e. can be vectorized). See
5562:ref:`Transformation Metadata <transformation-metadata>` for details.
5563
5564'``llvm.loop.distribute.followup_sequential``' Metadata
5565^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5566
5567This metadata defines which attributes the isolated loops with unsafe
5568memory dependencies will have. See
5569:ref:`Transformation Metadata <transformation-metadata>` for details.
5570
5571'``llvm.loop.distribute.followup_fallback``' Metadata
5572^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5573
5574If loop versioning is necessary, this metadata defined the attributes
5575the non-distributed fallback version will have. See
5576:ref:`Transformation Metadata <transformation-metadata>` for details.
5577
5578'``llvm.loop.distribute.followup_all``' Metadata
5579^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5580
5581Thes attributes in this metdata is added to all followup loops of the
5582loop distribution pass. See
5583:ref:`Transformation Metadata <transformation-metadata>` for details.
5584
5585'``llvm.access.group``' Metadata
5586^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5587
5588``llvm.access.group`` metadata can be attached to any instruction that
5589potentially accesses memory. It can point to a single distinct metadata
5590node, which we call access group. This node represents all memory access
5591instructions referring to it via ``llvm.access.group``. When an
5592instruction belongs to multiple access groups, it can also point to a
5593list of accesses groups, illustrated by the following example.
5594
5595.. code-block:: llvm
5596
5597   %val = load i32, i32* %arrayidx, !llvm.access.group !0
5598   ...
5599   !0 = !{!1, !2}
5600   !1 = distinct !{}
5601   !2 = distinct !{}
5602
5603It is illegal for the list node to be empty since it might be confused
5604with an access group.
5605
5606The access group metadata node must be 'distinct' to avoid collapsing
5607multiple access groups by content. A access group metadata node must
5608always be empty which can be used to distinguish an access group
5609metadata node from a list of access groups. Being empty avoids the
5610situation that the content must be updated which, because metadata is
5611immutable by design, would required finding and updating all references
5612to the access group node.
5613
5614The access group can be used to refer to a memory access instruction
5615without pointing to it directly (which is not possible in global
5616metadata). Currently, the only metadata making use of it is
5617``llvm.loop.parallel_accesses``.
5618
5619'``llvm.loop.parallel_accesses``' Metadata
5620^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5621
5622The ``llvm.loop.parallel_accesses`` metadata refers to one or more
5623access group metadata nodes (see ``llvm.access.group``). It denotes that
5624no loop-carried memory dependence exist between it and other instructions
5625in the loop with this metadata.
5626
5627Let ``m1`` and ``m2`` be two instructions that both have the
5628``llvm.access.group`` metadata to the access group ``g1``, respectively
5629``g2`` (which might be identical). If a loop contains both access groups
5630in its ``llvm.loop.parallel_accesses`` metadata, then the compiler can
5631assume that there is no dependency between ``m1`` and ``m2`` carried by
5632this loop. Instructions that belong to multiple access groups are
5633considered having this property if at least one of the access groups
5634matches the ``llvm.loop.parallel_accesses`` list.
5635
5636If all memory-accessing instructions in a loop have
5637``llvm.loop.parallel_accesses`` metadata that refers to that loop, then the
5638loop has no loop carried memory dependences and is considered to be a
5639parallel loop.
5640
5641Note that if not all memory access instructions belong to an access
5642group referred to by ``llvm.loop.parallel_accesses``, then the loop must
5643not be considered trivially parallel. Additional
5644memory dependence analysis is required to make that determination. As a fail
5645safe mechanism, this causes loops that were originally parallel to be considered
5646sequential (if optimization passes that are unaware of the parallel semantics
5647insert new memory instructions into the loop body).
5648
5649Example of a loop that is considered parallel due to its correct use of
5650both ``llvm.access.group`` and ``llvm.loop.parallel_accesses``
5651metadata types.
5652
5653.. code-block:: llvm
5654
5655   for.body:
5656     ...
5657     %val0 = load i32, i32* %arrayidx, !llvm.access.group !1
5658     ...
5659     store i32 %val0, i32* %arrayidx1, !llvm.access.group !1
5660     ...
5661     br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !0
5662
5663   for.end:
5664   ...
5665   !0 = distinct !{!0, !{!"llvm.loop.parallel_accesses", !1}}
5666   !1 = distinct !{}
5667
5668It is also possible to have nested parallel loops:
5669
5670.. code-block:: llvm
5671
5672   outer.for.body:
5673     ...
5674     %val1 = load i32, i32* %arrayidx3, !llvm.access.group !4
5675     ...
5676     br label %inner.for.body
5677
5678   inner.for.body:
5679     ...
5680     %val0 = load i32, i32* %arrayidx1, !llvm.access.group !3
5681     ...
5682     store i32 %val0, i32* %arrayidx2, !llvm.access.group !3
5683     ...
5684     br i1 %exitcond, label %inner.for.end, label %inner.for.body, !llvm.loop !1
5685
5686   inner.for.end:
5687     ...
5688     store i32 %val1, i32* %arrayidx4, !llvm.access.group !4
5689     ...
5690     br i1 %exitcond, label %outer.for.end, label %outer.for.body, !llvm.loop !2
5691
5692   outer.for.end:                                          ; preds = %for.body
5693   ...
5694   !1 = distinct !{!1, !{!"llvm.loop.parallel_accesses", !3}}     ; metadata for the inner loop
5695   !2 = distinct !{!2, !{!"llvm.loop.parallel_accesses", !3, !4}} ; metadata for the outer loop
5696   !3 = distinct !{} ; access group for instructions in the inner loop (which are implicitly contained in outer loop as well)
5697   !4 = distinct !{} ; access group for instructions in the outer, but not the inner loop
5698
5699'``irr_loop``' Metadata
5700^^^^^^^^^^^^^^^^^^^^^^^
5701
5702``irr_loop`` metadata may be attached to the terminator instruction of a basic
5703block that's an irreducible loop header (note that an irreducible loop has more
5704than once header basic blocks.) If ``irr_loop`` metadata is attached to the
5705terminator instruction of a basic block that is not really an irreducible loop
5706header, the behavior is undefined. The intent of this metadata is to improve the
5707accuracy of the block frequency propagation. For example, in the code below, the
5708block ``header0`` may have a loop header weight (relative to the other headers of
5709the irreducible loop) of 100:
5710
5711.. code-block:: llvm
5712
5713    header0:
5714    ...
5715    br i1 %cmp, label %t1, label %t2, !irr_loop !0
5716
5717    ...
5718    !0 = !{"loop_header_weight", i64 100}
5719
5720Irreducible loop header weights are typically based on profile data.
5721
5722'``invariant.group``' Metadata
5723^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5724
5725The experimental ``invariant.group`` metadata may be attached to
5726``load``/``store`` instructions referencing a single metadata with no entries.
5727The existence of the ``invariant.group`` metadata on the instruction tells
5728the optimizer that every ``load`` and ``store`` to the same pointer operand
5729can be assumed to load or store the same
5730value (but see the ``llvm.launder.invariant.group`` intrinsic which affects
5731when two pointers are considered the same). Pointers returned by bitcast or
5732getelementptr with only zero indices are considered the same.
5733
5734Examples:
5735
5736.. code-block:: llvm
5737
5738   @unknownPtr = external global i8
5739   ...
5740   %ptr = alloca i8
5741   store i8 42, i8* %ptr, !invariant.group !0
5742   call void @foo(i8* %ptr)
5743
5744   %a = load i8, i8* %ptr, !invariant.group !0 ; Can assume that value under %ptr didn't change
5745   call void @foo(i8* %ptr)
5746
5747   %newPtr = call i8* @getPointer(i8* %ptr)
5748   %c = load i8, i8* %newPtr, !invariant.group !0 ; Can't assume anything, because we only have information about %ptr
5749
5750   %unknownValue = load i8, i8* @unknownPtr
5751   store i8 %unknownValue, i8* %ptr, !invariant.group !0 ; Can assume that %unknownValue == 42
5752
5753   call void @foo(i8* %ptr)
5754   %newPtr2 = call i8* @llvm.launder.invariant.group(i8* %ptr)
5755   %d = load i8, i8* %newPtr2, !invariant.group !0  ; Can't step through launder.invariant.group to get value of %ptr
5756
5757   ...
5758   declare void @foo(i8*)
5759   declare i8* @getPointer(i8*)
5760   declare i8* @llvm.launder.invariant.group(i8*)
5761
5762   !0 = !{}
5763
5764The invariant.group metadata must be dropped when replacing one pointer by
5765another based on aliasing information. This is because invariant.group is tied
5766to the SSA value of the pointer operand.
5767
5768.. code-block:: llvm
5769
5770  %v = load i8, i8* %x, !invariant.group !0
5771  ; if %x mustalias %y then we can replace the above instruction with
5772  %v = load i8, i8* %y
5773
5774Note that this is an experimental feature, which means that its semantics might
5775change in the future.
5776
5777'``type``' Metadata
5778^^^^^^^^^^^^^^^^^^^
5779
5780See :doc:`TypeMetadata`.
5781
5782'``associated``' Metadata
5783^^^^^^^^^^^^^^^^^^^^^^^^^
5784
5785The ``associated`` metadata may be attached to a global object
5786declaration with a single argument that references another global object.
5787
5788This metadata prevents discarding of the global object in linker GC
5789unless the referenced object is also discarded. The linker support for
5790this feature is spotty. For best compatibility, globals carrying this
5791metadata may also:
5792
5793- Be in a comdat with the referenced global.
5794- Be in @llvm.compiler.used.
5795- Have an explicit section with a name which is a valid C identifier.
5796
5797It does not have any effect on non-ELF targets.
5798
5799Example:
5800
5801.. code-block:: text
5802
5803    $a = comdat any
5804    @a = global i32 1, comdat $a
5805    @b = internal global i32 2, comdat $a, section "abc", !associated !0
5806    !0 = !{i32* @a}
5807
5808
5809'``prof``' Metadata
5810^^^^^^^^^^^^^^^^^^^
5811
5812The ``prof`` metadata is used to record profile data in the IR.
5813The first operand of the metadata node indicates the profile metadata
5814type. There are currently 3 types:
5815:ref:`branch_weights<prof_node_branch_weights>`,
5816:ref:`function_entry_count<prof_node_function_entry_count>`, and
5817:ref:`VP<prof_node_VP>`.
5818
5819.. _prof_node_branch_weights:
5820
5821branch_weights
5822""""""""""""""
5823
5824Branch weight metadata attached to a branch, select, switch or call instruction
5825represents the likeliness of the associated branch being taken.
5826For more information, see :doc:`BranchWeightMetadata`.
5827
5828.. _prof_node_function_entry_count:
5829
5830function_entry_count
5831""""""""""""""""""""
5832
5833Function entry count metadata can be attached to function definitions
5834to record the number of times the function is called. Used with BFI
5835information, it is also used to derive the basic block profile count.
5836For more information, see :doc:`BranchWeightMetadata`.
5837
5838.. _prof_node_VP:
5839
5840VP
5841""
5842
5843VP (value profile) metadata can be attached to instructions that have
5844value profile information. Currently this is indirect calls (where it
5845records the hottest callees) and calls to memory intrinsics such as memcpy,
5846memmove, and memset (where it records the hottest byte lengths).
5847
5848Each VP metadata node contains "VP" string, then a uint32_t value for the value
5849profiling kind, a uint64_t value for the total number of times the instruction
5850is executed, followed by uint64_t value and execution count pairs.
5851The value profiling kind is 0 for indirect call targets and 1 for memory
5852operations. For indirect call targets, each profile value is a hash
5853of the callee function name, and for memory operations each value is the
5854byte length.
5855
5856Note that the value counts do not need to add up to the total count
5857listed in the third operand (in practice only the top hottest values
5858are tracked and reported).
5859
5860Indirect call example:
5861
5862.. code-block:: llvm
5863
5864    call void %f(), !prof !1
5865    !1 = !{!"VP", i32 0, i64 1600, i64 7651369219802541373, i64 1030, i64 -4377547752858689819, i64 410}
5866
5867Note that the VP type is 0 (the second operand), which indicates this is
5868an indirect call value profile data. The third operand indicates that the
5869indirect call executed 1600 times. The 4th and 6th operands give the
5870hashes of the 2 hottest target functions' names (this is the same hash used
5871to represent function names in the profile database), and the 5th and 7th
5872operands give the execution count that each of the respective prior target
5873functions was called.
5874
5875Module Flags Metadata
5876=====================
5877
5878Information about the module as a whole is difficult to convey to LLVM's
5879subsystems. The LLVM IR isn't sufficient to transmit this information.
5880The ``llvm.module.flags`` named metadata exists in order to facilitate
5881this. These flags are in the form of key / value pairs --- much like a
5882dictionary --- making it easy for any subsystem who cares about a flag to
5883look it up.
5884
5885The ``llvm.module.flags`` metadata contains a list of metadata triplets.
5886Each triplet has the following form:
5887
5888-  The first element is a *behavior* flag, which specifies the behavior
5889   when two (or more) modules are merged together, and it encounters two
5890   (or more) metadata with the same ID. The supported behaviors are
5891   described below.
5892-  The second element is a metadata string that is a unique ID for the
5893   metadata. Each module may only have one flag entry for each unique ID (not
5894   including entries with the **Require** behavior).
5895-  The third element is the value of the flag.
5896
5897When two (or more) modules are merged together, the resulting
5898``llvm.module.flags`` metadata is the union of the modules' flags. That is, for
5899each unique metadata ID string, there will be exactly one entry in the merged
5900modules ``llvm.module.flags`` metadata table, and the value for that entry will
5901be determined by the merge behavior flag, as described below. The only exception
5902is that entries with the *Require* behavior are always preserved.
5903
5904The following behaviors are supported:
5905
5906.. list-table::
5907   :header-rows: 1
5908   :widths: 10 90
5909
5910   * - Value
5911     - Behavior
5912
5913   * - 1
5914     - **Error**
5915           Emits an error if two values disagree, otherwise the resulting value
5916           is that of the operands.
5917
5918   * - 2
5919     - **Warning**
5920           Emits a warning if two values disagree. The result value will be the
5921           operand for the flag from the first module being linked.
5922
5923   * - 3
5924     - **Require**
5925           Adds a requirement that another module flag be present and have a
5926           specified value after linking is performed. The value must be a
5927           metadata pair, where the first element of the pair is the ID of the
5928           module flag to be restricted, and the second element of the pair is
5929           the value the module flag should be restricted to. This behavior can
5930           be used to restrict the allowable results (via triggering of an
5931           error) of linking IDs with the **Override** behavior.
5932
5933   * - 4
5934     - **Override**
5935           Uses the specified value, regardless of the behavior or value of the
5936           other module. If both modules specify **Override**, but the values
5937           differ, an error will be emitted.
5938
5939   * - 5
5940     - **Append**
5941           Appends the two values, which are required to be metadata nodes.
5942
5943   * - 6
5944     - **AppendUnique**
5945           Appends the two values, which are required to be metadata
5946           nodes. However, duplicate entries in the second list are dropped
5947           during the append operation.
5948
5949   * - 7
5950     - **Max**
5951           Takes the max of the two values, which are required to be integers.
5952
5953It is an error for a particular unique flag ID to have multiple behaviors,
5954except in the case of **Require** (which adds restrictions on another metadata
5955value) or **Override**.
5956
5957An example of module flags:
5958
5959.. code-block:: llvm
5960
5961    !0 = !{ i32 1, !"foo", i32 1 }
5962    !1 = !{ i32 4, !"bar", i32 37 }
5963    !2 = !{ i32 2, !"qux", i32 42 }
5964    !3 = !{ i32 3, !"qux",
5965      !{
5966        !"foo", i32 1
5967      }
5968    }
5969    !llvm.module.flags = !{ !0, !1, !2, !3 }
5970
5971-  Metadata ``!0`` has the ID ``!"foo"`` and the value '1'. The behavior
5972   if two or more ``!"foo"`` flags are seen is to emit an error if their
5973   values are not equal.
5974
5975-  Metadata ``!1`` has the ID ``!"bar"`` and the value '37'. The
5976   behavior if two or more ``!"bar"`` flags are seen is to use the value
5977   '37'.
5978
5979-  Metadata ``!2`` has the ID ``!"qux"`` and the value '42'. The
5980   behavior if two or more ``!"qux"`` flags are seen is to emit a
5981   warning if their values are not equal.
5982
5983-  Metadata ``!3`` has the ID ``!"qux"`` and the value:
5984
5985   ::
5986
5987       !{ !"foo", i32 1 }
5988
5989   The behavior is to emit an error if the ``llvm.module.flags`` does not
5990   contain a flag with the ID ``!"foo"`` that has the value '1' after linking is
5991   performed.
5992
5993Objective-C Garbage Collection Module Flags Metadata
5994----------------------------------------------------
5995
5996On the Mach-O platform, Objective-C stores metadata about garbage
5997collection in a special section called "image info". The metadata
5998consists of a version number and a bitmask specifying what types of
5999garbage collection are supported (if any) by the file. If two or more
6000modules are linked together their garbage collection metadata needs to
6001be merged rather than appended together.
6002
6003The Objective-C garbage collection module flags metadata consists of the
6004following key-value pairs:
6005
6006.. list-table::
6007   :header-rows: 1
6008   :widths: 30 70
6009
6010   * - Key
6011     - Value
6012
6013   * - ``Objective-C Version``
6014     - **[Required]** --- The Objective-C ABI version. Valid values are 1 and 2.
6015
6016   * - ``Objective-C Image Info Version``
6017     - **[Required]** --- The version of the image info section. Currently
6018       always 0.
6019
6020   * - ``Objective-C Image Info Section``
6021     - **[Required]** --- The section to place the metadata. Valid values are
6022       ``"__OBJC, __image_info, regular"`` for Objective-C ABI version 1, and
6023       ``"__DATA,__objc_imageinfo, regular, no_dead_strip"`` for
6024       Objective-C ABI version 2.
6025
6026   * - ``Objective-C Garbage Collection``
6027     - **[Required]** --- Specifies whether garbage collection is supported or
6028       not. Valid values are 0, for no garbage collection, and 2, for garbage
6029       collection supported.
6030
6031   * - ``Objective-C GC Only``
6032     - **[Optional]** --- Specifies that only garbage collection is supported.
6033       If present, its value must be 6. This flag requires that the
6034       ``Objective-C Garbage Collection`` flag have the value 2.
6035
6036Some important flag interactions:
6037
6038-  If a module with ``Objective-C Garbage Collection`` set to 0 is
6039   merged with a module with ``Objective-C Garbage Collection`` set to
6040   2, then the resulting module has the
6041   ``Objective-C Garbage Collection`` flag set to 0.
6042-  A module with ``Objective-C Garbage Collection`` set to 0 cannot be
6043   merged with a module with ``Objective-C GC Only`` set to 6.
6044
6045C type width Module Flags Metadata
6046----------------------------------
6047
6048The ARM backend emits a section into each generated object file describing the
6049options that it was compiled with (in a compiler-independent way) to prevent
6050linking incompatible objects, and to allow automatic library selection. Some
6051of these options are not visible at the IR level, namely wchar_t width and enum
6052width.
6053
6054To pass this information to the backend, these options are encoded in module
6055flags metadata, using the following key-value pairs:
6056
6057.. list-table::
6058   :header-rows: 1
6059   :widths: 30 70
6060
6061   * - Key
6062     - Value
6063
6064   * - short_wchar
6065     - * 0 --- sizeof(wchar_t) == 4
6066       * 1 --- sizeof(wchar_t) == 2
6067
6068   * - short_enum
6069     - * 0 --- Enums are at least as large as an ``int``.
6070       * 1 --- Enums are stored in the smallest integer type which can
6071         represent all of its values.
6072
6073For example, the following metadata section specifies that the module was
6074compiled with a ``wchar_t`` width of 4 bytes, and the underlying type of an
6075enum is the smallest type which can represent all of its values::
6076
6077    !llvm.module.flags = !{!0, !1}
6078    !0 = !{i32 1, !"short_wchar", i32 1}
6079    !1 = !{i32 1, !"short_enum", i32 0}
6080
6081Automatic Linker Flags Named Metadata
6082=====================================
6083
6084Some targets support embedding flags to the linker inside individual object
6085files. Typically this is used in conjunction with language extensions which
6086allow source files to explicitly declare the libraries they depend on, and have
6087these automatically be transmitted to the linker via object files.
6088
6089These flags are encoded in the IR using named metadata with the name
6090``!llvm.linker.options``. Each operand is expected to be a metadata node
6091which should be a list of other metadata nodes, each of which should be a
6092list of metadata strings defining linker options.
6093
6094For example, the following metadata section specifies two separate sets of
6095linker options, presumably to link against ``libz`` and the ``Cocoa``
6096framework::
6097
6098    !0 = !{ !"-lz" },
6099    !1 = !{ !"-framework", !"Cocoa" } } }
6100    !llvm.linker.options = !{ !0, !1 }
6101
6102The metadata encoding as lists of lists of options, as opposed to a collapsed
6103list of options, is chosen so that the IR encoding can use multiple option
6104strings to specify e.g., a single library, while still having that specifier be
6105preserved as an atomic element that can be recognized by a target specific
6106assembly writer or object file emitter.
6107
6108Each individual option is required to be either a valid option for the target's
6109linker, or an option that is reserved by the target specific assembly writer or
6110object file emitter. No other aspect of these options is defined by the IR.
6111
6112.. _summary:
6113
6114ThinLTO Summary
6115===============
6116
6117Compiling with `ThinLTO <https://clang.llvm.org/docs/ThinLTO.html>`_
6118causes the building of a compact summary of the module that is emitted into
6119the bitcode. The summary is emitted into the LLVM assembly and identified
6120in syntax by a caret ('``^``').
6121
6122The summary is parsed into a bitcode output, along with the Module
6123IR, via the "``llvm-as``" tool. Tools that parse the Module IR for the purposes
6124of optimization (e.g. "``clang -x ir``" and "``opt``"), will ignore the
6125summary entries (just as they currently ignore summary entries in a bitcode
6126input file).
6127
6128Eventually, the summary will be parsed into a ModuleSummaryIndex object under
6129the same conditions where summary index is currently built from bitcode.
6130Specifically, tools that test the Thin Link portion of a ThinLTO compile
6131(i.e. llvm-lto and llvm-lto2), or when parsing a combined index
6132for a distributed ThinLTO backend via clang's "``-fthinlto-index=<>``" flag
6133(this part is not yet implemented, use llvm-as to create a bitcode object
6134before feeding into thin link tools for now).
6135
6136There are currently 3 types of summary entries in the LLVM assembly:
6137:ref:`module paths<module_path_summary>`,
6138:ref:`global values<gv_summary>`, and
6139:ref:`type identifiers<typeid_summary>`.
6140
6141.. _module_path_summary:
6142
6143Module Path Summary Entry
6144-------------------------
6145
6146Each module path summary entry lists a module containing global values included
6147in the summary. For a single IR module there will be one such entry, but
6148in a combined summary index produced during the thin link, there will be
6149one module path entry per linked module with summary.
6150
6151Example:
6152
6153.. code-block:: text
6154
6155    ^0 = module: (path: "/path/to/file.o", hash: (2468601609, 1329373163, 1565878005, 638838075, 3148790418))
6156
6157The ``path`` field is a string path to the bitcode file, and the ``hash``
6158field is the 160-bit SHA-1 hash of the IR bitcode contents, used for
6159incremental builds and caching.
6160
6161.. _gv_summary:
6162
6163Global Value Summary Entry
6164--------------------------
6165
6166Each global value summary entry corresponds to a global value defined or
6167referenced by a summarized module.
6168
6169Example:
6170
6171.. code-block:: text
6172
6173    ^4 = gv: (name: "f"[, summaries: (Summary)[, (Summary)]*]?) ; guid = 14740650423002898831
6174
6175For declarations, there will not be a summary list. For definitions, a
6176global value will contain a list of summaries, one per module containing
6177a definition. There can be multiple entries in a combined summary index
6178for symbols with weak linkage.
6179
6180Each ``Summary`` format will depend on whether the global value is a
6181:ref:`function<function_summary>`, :ref:`variable<variable_summary>`, or
6182:ref:`alias<alias_summary>`.
6183
6184.. _function_summary:
6185
6186Function Summary
6187^^^^^^^^^^^^^^^^
6188
6189If the global value is a function, the ``Summary`` entry will look like:
6190
6191.. code-block:: text
6192
6193    function: (module: ^0, flags: (linkage: external, notEligibleToImport: 0, live: 0, dsoLocal: 0), insts: 2[, FuncFlags]?[, Calls]?[, TypeIdInfo]?[, Refs]?
6194
6195The ``module`` field includes the summary entry id for the module containing
6196this definition, and the ``flags`` field contains information such as
6197the linkage type, a flag indicating whether it is legal to import the
6198definition, whether it is globally live and whether the linker resolved it
6199to a local definition (the latter two are populated during the thin link).
6200The ``insts`` field contains the number of IR instructions in the function.
6201Finally, there are several optional fields: :ref:`FuncFlags<funcflags_summary>`,
6202:ref:`Calls<calls_summary>`, :ref:`TypeIdInfo<typeidinfo_summary>`,
6203:ref:`Refs<refs_summary>`.
6204
6205.. _variable_summary:
6206
6207Global Variable Summary
6208^^^^^^^^^^^^^^^^^^^^^^^
6209
6210If the global value is a variable, the ``Summary`` entry will look like:
6211
6212.. code-block:: text
6213
6214    variable: (module: ^0, flags: (linkage: external, notEligibleToImport: 0, live: 0, dsoLocal: 0)[, Refs]?
6215
6216The variable entry contains a subset of the fields in a
6217:ref:`function summary <function_summary>`, see the descriptions there.
6218
6219.. _alias_summary:
6220
6221Alias Summary
6222^^^^^^^^^^^^^
6223
6224If the global value is an alias, the ``Summary`` entry will look like:
6225
6226.. code-block:: text
6227
6228    alias: (module: ^0, flags: (linkage: external, notEligibleToImport: 0, live: 0, dsoLocal: 0), aliasee: ^2)
6229
6230The ``module`` and ``flags`` fields are as described for a
6231:ref:`function summary <function_summary>`. The ``aliasee`` field
6232contains a reference to the global value summary entry of the aliasee.
6233
6234.. _funcflags_summary:
6235
6236Function Flags
6237^^^^^^^^^^^^^^
6238
6239The optional ``FuncFlags`` field looks like:
6240
6241.. code-block:: text
6242
6243    funcFlags: (readNone: 0, readOnly: 0, noRecurse: 0, returnDoesNotAlias: 0)
6244
6245If unspecified, flags are assumed to hold the conservative ``false`` value of
6246``0``.
6247
6248.. _calls_summary:
6249
6250Calls
6251^^^^^
6252
6253The optional ``Calls`` field looks like:
6254
6255.. code-block:: text
6256
6257    calls: ((Callee)[, (Callee)]*)
6258
6259where each ``Callee`` looks like:
6260
6261.. code-block:: text
6262
6263    callee: ^1[, hotness: None]?[, relbf: 0]?
6264
6265The ``callee`` refers to the summary entry id of the callee. At most one
6266of ``hotness`` (which can take the values ``Unknown``, ``Cold``, ``None``,
6267``Hot``, and ``Critical``), and ``relbf`` (which holds the integer
6268branch frequency relative to the entry frequency, scaled down by 2^8)
6269may be specified. The defaults are ``Unknown`` and ``0``, respectively.
6270
6271.. _refs_summary:
6272
6273Refs
6274^^^^
6275
6276The optional ``Refs`` field looks like:
6277
6278.. code-block:: text
6279
6280    refs: ((Ref)[, (Ref)]*)
6281
6282where each ``Ref`` contains a reference to the summary id of the referenced
6283value (e.g. ``^1``).
6284
6285.. _typeidinfo_summary:
6286
6287TypeIdInfo
6288^^^^^^^^^^
6289
6290The optional ``TypeIdInfo`` field, used for
6291`Control Flow Integrity <http://clang.llvm.org/docs/ControlFlowIntegrity.html>`_,
6292looks like:
6293
6294.. code-block:: text
6295
6296    typeIdInfo: [(TypeTests)]?[, (TypeTestAssumeVCalls)]?[, (TypeCheckedLoadVCalls)]?[, (TypeTestAssumeConstVCalls)]?[, (TypeCheckedLoadConstVCalls)]?
6297
6298These optional fields have the following forms:
6299
6300TypeTests
6301"""""""""
6302
6303.. code-block:: text
6304
6305    typeTests: (TypeIdRef[, TypeIdRef]*)
6306
6307Where each ``TypeIdRef`` refers to a :ref:`type id<typeid_summary>`
6308by summary id or ``GUID``.
6309
6310TypeTestAssumeVCalls
6311""""""""""""""""""""
6312
6313.. code-block:: text
6314
6315    typeTestAssumeVCalls: (VFuncId[, VFuncId]*)
6316
6317Where each VFuncId has the format:
6318
6319.. code-block:: text
6320
6321    vFuncId: (TypeIdRef, offset: 16)
6322
6323Where each ``TypeIdRef`` refers to a :ref:`type id<typeid_summary>`
6324by summary id or ``GUID`` preceeded by a ``guid:`` tag.
6325
6326TypeCheckedLoadVCalls
6327"""""""""""""""""""""
6328
6329.. code-block:: text
6330
6331    typeCheckedLoadVCalls: (VFuncId[, VFuncId]*)
6332
6333Where each VFuncId has the format described for ``TypeTestAssumeVCalls``.
6334
6335TypeTestAssumeConstVCalls
6336"""""""""""""""""""""""""
6337
6338.. code-block:: text
6339
6340    typeTestAssumeConstVCalls: (ConstVCall[, ConstVCall]*)
6341
6342Where each ConstVCall has the format:
6343
6344.. code-block:: text
6345
6346    (VFuncId, args: (Arg[, Arg]*))
6347
6348and where each VFuncId has the format described for ``TypeTestAssumeVCalls``,
6349and each Arg is an integer argument number.
6350
6351TypeCheckedLoadConstVCalls
6352""""""""""""""""""""""""""
6353
6354.. code-block:: text
6355
6356    typeCheckedLoadConstVCalls: (ConstVCall[, ConstVCall]*)
6357
6358Where each ConstVCall has the format described for
6359``TypeTestAssumeConstVCalls``.
6360
6361.. _typeid_summary:
6362
6363Type ID Summary Entry
6364---------------------
6365
6366Each type id summary entry corresponds to a type identifier resolution
6367which is generated during the LTO link portion of the compile when building
6368with `Control Flow Integrity <http://clang.llvm.org/docs/ControlFlowIntegrity.html>`_,
6369so these are only present in a combined summary index.
6370
6371Example:
6372
6373.. code-block:: text
6374
6375    ^4 = typeid: (name: "_ZTS1A", summary: (typeTestRes: (kind: allOnes, sizeM1BitWidth: 7[, alignLog2: 0]?[, sizeM1: 0]?[, bitMask: 0]?[, inlineBits: 0]?)[, WpdResolutions]?)) ; guid = 7004155349499253778
6376
6377The ``typeTestRes`` gives the type test resolution ``kind`` (which may
6378be ``unsat``, ``byteArray``, ``inline``, ``single``, or ``allOnes``), and
6379the ``size-1`` bit width. It is followed by optional flags, which default to 0,
6380and an optional WpdResolutions (whole program devirtualization resolution)
6381field that looks like:
6382
6383.. code-block:: text
6384
6385    wpdResolutions: ((offset: 0, WpdRes)[, (offset: 1, WpdRes)]*
6386
6387where each entry is a mapping from the given byte offset to the whole-program
6388devirtualization resolution WpdRes, that has one of the following formats:
6389
6390.. code-block:: text
6391
6392    wpdRes: (kind: branchFunnel)
6393    wpdRes: (kind: singleImpl, singleImplName: "_ZN1A1nEi")
6394    wpdRes: (kind: indir)
6395
6396Additionally, each wpdRes has an optional ``resByArg`` field, which
6397describes the resolutions for calls with all constant integer arguments:
6398
6399.. code-block:: text
6400
6401    resByArg: (ResByArg[, ResByArg]*)
6402
6403where ResByArg is:
6404
6405.. code-block:: text
6406
6407    args: (Arg[, Arg]*), byArg: (kind: UniformRetVal[, info: 0][, byte: 0][, bit: 0])
6408
6409Where the ``kind`` can be ``Indir``, ``UniformRetVal``, ``UniqueRetVal``
6410or ``VirtualConstProp``. The ``info`` field is only used if the kind
6411is ``UniformRetVal`` (indicates the uniform return value), or
6412``UniqueRetVal`` (holds the return value associated with the unique vtable
6413(0 or 1)). The ``byte`` and ``bit`` fields are only used if the target does
6414not support the use of absolute symbols to store constants.
6415
6416.. _intrinsicglobalvariables:
6417
6418Intrinsic Global Variables
6419==========================
6420
6421LLVM has a number of "magic" global variables that contain data that
6422affect code generation or other IR semantics. These are documented here.
6423All globals of this sort should have a section specified as
6424"``llvm.metadata``". This section and all globals that start with
6425"``llvm.``" are reserved for use by LLVM.
6426
6427.. _gv_llvmused:
6428
6429The '``llvm.used``' Global Variable
6430-----------------------------------
6431
6432The ``@llvm.used`` global is an array which has
6433:ref:`appending linkage <linkage_appending>`. This array contains a list of
6434pointers to named global variables, functions and aliases which may optionally
6435have a pointer cast formed of bitcast or getelementptr. For example, a legal
6436use of it is:
6437
6438.. code-block:: llvm
6439
6440    @X = global i8 4
6441    @Y = global i32 123
6442
6443    @llvm.used = appending global [2 x i8*] [
6444       i8* @X,
6445       i8* bitcast (i32* @Y to i8*)
6446    ], section "llvm.metadata"
6447
6448If a symbol appears in the ``@llvm.used`` list, then the compiler, assembler,
6449and linker are required to treat the symbol as if there is a reference to the
6450symbol that it cannot see (which is why they have to be named). For example, if
6451a variable has internal linkage and no references other than that from the
6452``@llvm.used`` list, it cannot be deleted. This is commonly used to represent
6453references from inline asms and other things the compiler cannot "see", and
6454corresponds to "``attribute((used))``" in GNU C.
6455
6456On some targets, the code generator must emit a directive to the
6457assembler or object file to prevent the assembler and linker from
6458molesting the symbol.
6459
6460.. _gv_llvmcompilerused:
6461
6462The '``llvm.compiler.used``' Global Variable
6463--------------------------------------------
6464
6465The ``@llvm.compiler.used`` directive is the same as the ``@llvm.used``
6466directive, except that it only prevents the compiler from touching the
6467symbol. On targets that support it, this allows an intelligent linker to
6468optimize references to the symbol without being impeded as it would be
6469by ``@llvm.used``.
6470
6471This is a rare construct that should only be used in rare circumstances,
6472and should not be exposed to source languages.
6473
6474.. _gv_llvmglobalctors:
6475
6476The '``llvm.global_ctors``' Global Variable
6477-------------------------------------------
6478
6479.. code-block:: llvm
6480
6481    %0 = type { i32, void ()*, i8* }
6482    @llvm.global_ctors = appending global [1 x %0] [%0 { i32 65535, void ()* @ctor, i8* @data }]
6483
6484The ``@llvm.global_ctors`` array contains a list of constructor
6485functions, priorities, and an optional associated global or function.
6486The functions referenced by this array will be called in ascending order
6487of priority (i.e. lowest first) when the module is loaded. The order of
6488functions with the same priority is not defined.
6489
6490If the third field is present, non-null, and points to a global variable
6491or function, the initializer function will only run if the associated
6492data from the current module is not discarded.
6493
6494.. _llvmglobaldtors:
6495
6496The '``llvm.global_dtors``' Global Variable
6497-------------------------------------------
6498
6499.. code-block:: llvm
6500
6501    %0 = type { i32, void ()*, i8* }
6502    @llvm.global_dtors = appending global [1 x %0] [%0 { i32 65535, void ()* @dtor, i8* @data }]
6503
6504The ``@llvm.global_dtors`` array contains a list of destructor
6505functions, priorities, and an optional associated global or function.
6506The functions referenced by this array will be called in descending
6507order of priority (i.e. highest first) when the module is unloaded. The
6508order of functions with the same priority is not defined.
6509
6510If the third field is present, non-null, and points to a global variable
6511or function, the destructor function will only run if the associated
6512data from the current module is not discarded.
6513
6514Instruction Reference
6515=====================
6516
6517The LLVM instruction set consists of several different classifications
6518of instructions: :ref:`terminator instructions <terminators>`, :ref:`binary
6519instructions <binaryops>`, :ref:`bitwise binary
6520instructions <bitwiseops>`, :ref:`memory instructions <memoryops>`, and
6521:ref:`other instructions <otherops>`.
6522
6523.. _terminators:
6524
6525Terminator Instructions
6526-----------------------
6527
6528As mentioned :ref:`previously <functionstructure>`, every basic block in a
6529program ends with a "Terminator" instruction, which indicates which
6530block should be executed after the current block is finished. These
6531terminator instructions typically yield a '``void``' value: they produce
6532control flow, not values (the one exception being the
6533':ref:`invoke <i_invoke>`' instruction).
6534
6535The terminator instructions are: ':ref:`ret <i_ret>`',
6536':ref:`br <i_br>`', ':ref:`switch <i_switch>`',
6537':ref:`indirectbr <i_indirectbr>`', ':ref:`invoke <i_invoke>`',
6538':ref:`callbr <i_callbr>`'
6539':ref:`resume <i_resume>`', ':ref:`catchswitch <i_catchswitch>`',
6540':ref:`catchret <i_catchret>`',
6541':ref:`cleanupret <i_cleanupret>`',
6542and ':ref:`unreachable <i_unreachable>`'.
6543
6544.. _i_ret:
6545
6546'``ret``' Instruction
6547^^^^^^^^^^^^^^^^^^^^^
6548
6549Syntax:
6550"""""""
6551
6552::
6553
6554      ret <type> <value>       ; Return a value from a non-void function
6555      ret void                 ; Return from void function
6556
6557Overview:
6558"""""""""
6559
6560The '``ret``' instruction is used to return control flow (and optionally
6561a value) from a function back to the caller.
6562
6563There are two forms of the '``ret``' instruction: one that returns a
6564value and then causes control flow, and one that just causes control
6565flow to occur.
6566
6567Arguments:
6568""""""""""
6569
6570The '``ret``' instruction optionally accepts a single argument, the
6571return value. The type of the return value must be a ':ref:`first
6572class <t_firstclass>`' type.
6573
6574A function is not :ref:`well formed <wellformed>` if it has a non-void
6575return type and contains a '``ret``' instruction with no return value or
6576a return value with a type that does not match its type, or if it has a
6577void return type and contains a '``ret``' instruction with a return
6578value.
6579
6580Semantics:
6581""""""""""
6582
6583When the '``ret``' instruction is executed, control flow returns back to
6584the calling function's context. If the caller is a
6585":ref:`call <i_call>`" instruction, execution continues at the
6586instruction after the call. If the caller was an
6587":ref:`invoke <i_invoke>`" instruction, execution continues at the
6588beginning of the "normal" destination block. If the instruction returns
6589a value, that value shall set the call or invoke instruction's return
6590value.
6591
6592Example:
6593""""""""
6594
6595.. code-block:: llvm
6596
6597      ret i32 5                       ; Return an integer value of 5
6598      ret void                        ; Return from a void function
6599      ret { i32, i8 } { i32 4, i8 2 } ; Return a struct of values 4 and 2
6600
6601.. _i_br:
6602
6603'``br``' Instruction
6604^^^^^^^^^^^^^^^^^^^^
6605
6606Syntax:
6607"""""""
6608
6609::
6610
6611      br i1 <cond>, label <iftrue>, label <iffalse>
6612      br label <dest>          ; Unconditional branch
6613
6614Overview:
6615"""""""""
6616
6617The '``br``' instruction is used to cause control flow to transfer to a
6618different basic block in the current function. There are two forms of
6619this instruction, corresponding to a conditional branch and an
6620unconditional branch.
6621
6622Arguments:
6623""""""""""
6624
6625The conditional branch form of the '``br``' instruction takes a single
6626'``i1``' value and two '``label``' values. The unconditional form of the
6627'``br``' instruction takes a single '``label``' value as a target.
6628
6629Semantics:
6630""""""""""
6631
6632Upon execution of a conditional '``br``' instruction, the '``i1``'
6633argument is evaluated. If the value is ``true``, control flows to the
6634'``iftrue``' ``label`` argument. If "cond" is ``false``, control flows
6635to the '``iffalse``' ``label`` argument.
6636
6637Example:
6638""""""""
6639
6640.. code-block:: llvm
6641
6642    Test:
6643      %cond = icmp eq i32 %a, %b
6644      br i1 %cond, label %IfEqual, label %IfUnequal
6645    IfEqual:
6646      ret i32 1
6647    IfUnequal:
6648      ret i32 0
6649
6650.. _i_switch:
6651
6652'``switch``' Instruction
6653^^^^^^^^^^^^^^^^^^^^^^^^
6654
6655Syntax:
6656"""""""
6657
6658::
6659
6660      switch <intty> <value>, label <defaultdest> [ <intty> <val>, label <dest> ... ]
6661
6662Overview:
6663"""""""""
6664
6665The '``switch``' instruction is used to transfer control flow to one of
6666several different places. It is a generalization of the '``br``'
6667instruction, allowing a branch to occur to one of many possible
6668destinations.
6669
6670Arguments:
6671""""""""""
6672
6673The '``switch``' instruction uses three parameters: an integer
6674comparison value '``value``', a default '``label``' destination, and an
6675array of pairs of comparison value constants and '``label``'s. The table
6676is not allowed to contain duplicate constant entries.
6677
6678Semantics:
6679""""""""""
6680
6681The ``switch`` instruction specifies a table of values and destinations.
6682When the '``switch``' instruction is executed, this table is searched
6683for the given value. If the value is found, control flow is transferred
6684to the corresponding destination; otherwise, control flow is transferred
6685to the default destination.
6686
6687Implementation:
6688"""""""""""""""
6689
6690Depending on properties of the target machine and the particular
6691``switch`` instruction, this instruction may be code generated in
6692different ways. For example, it could be generated as a series of
6693chained conditional branches or with a lookup table.
6694
6695Example:
6696""""""""
6697
6698.. code-block:: llvm
6699
6700     ; Emulate a conditional br instruction
6701     %Val = zext i1 %value to i32
6702     switch i32 %Val, label %truedest [ i32 0, label %falsedest ]
6703
6704     ; Emulate an unconditional br instruction
6705     switch i32 0, label %dest [ ]
6706
6707     ; Implement a jump table:
6708     switch i32 %val, label %otherwise [ i32 0, label %onzero
6709                                         i32 1, label %onone
6710                                         i32 2, label %ontwo ]
6711
6712.. _i_indirectbr:
6713
6714'``indirectbr``' Instruction
6715^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6716
6717Syntax:
6718"""""""
6719
6720::
6721
6722      indirectbr <somety>* <address>, [ label <dest1>, label <dest2>, ... ]
6723
6724Overview:
6725"""""""""
6726
6727The '``indirectbr``' instruction implements an indirect branch to a
6728label within the current function, whose address is specified by
6729"``address``". Address must be derived from a
6730:ref:`blockaddress <blockaddress>` constant.
6731
6732Arguments:
6733""""""""""
6734
6735The '``address``' argument is the address of the label to jump to. The
6736rest of the arguments indicate the full set of possible destinations
6737that the address may point to. Blocks are allowed to occur multiple
6738times in the destination list, though this isn't particularly useful.
6739
6740This destination list is required so that dataflow analysis has an
6741accurate understanding of the CFG.
6742
6743Semantics:
6744""""""""""
6745
6746Control transfers to the block specified in the address argument. All
6747possible destination blocks must be listed in the label list, otherwise
6748this instruction has undefined behavior. This implies that jumps to
6749labels defined in other functions have undefined behavior as well.
6750
6751Implementation:
6752"""""""""""""""
6753
6754This is typically implemented with a jump through a register.
6755
6756Example:
6757""""""""
6758
6759.. code-block:: llvm
6760
6761     indirectbr i8* %Addr, [ label %bb1, label %bb2, label %bb3 ]
6762
6763.. _i_invoke:
6764
6765'``invoke``' Instruction
6766^^^^^^^^^^^^^^^^^^^^^^^^
6767
6768Syntax:
6769"""""""
6770
6771::
6772
6773      <result> = invoke [cconv] [ret attrs] [addrspace(<num>)] [<ty>|<fnty> <fnptrval>(<function args>) [fn attrs]
6774                    [operand bundles] to label <normal label> unwind label <exception label>
6775
6776Overview:
6777"""""""""
6778
6779The '``invoke``' instruction causes control to transfer to a specified
6780function, with the possibility of control flow transfer to either the
6781'``normal``' label or the '``exception``' label. If the callee function
6782returns with the "``ret``" instruction, control flow will return to the
6783"normal" label. If the callee (or any indirect callees) returns via the
6784":ref:`resume <i_resume>`" instruction or other exception handling
6785mechanism, control is interrupted and continued at the dynamically
6786nearest "exception" label.
6787
6788The '``exception``' label is a `landing
6789pad <ExceptionHandling.html#overview>`_ for the exception. As such,
6790'``exception``' label is required to have the
6791":ref:`landingpad <i_landingpad>`" instruction, which contains the
6792information about the behavior of the program after unwinding happens,
6793as its first non-PHI instruction. The restrictions on the
6794"``landingpad``" instruction's tightly couples it to the "``invoke``"
6795instruction, so that the important information contained within the
6796"``landingpad``" instruction can't be lost through normal code motion.
6797
6798Arguments:
6799""""""""""
6800
6801This instruction requires several arguments:
6802
6803#. The optional "cconv" marker indicates which :ref:`calling
6804   convention <callingconv>` the call should use. If none is
6805   specified, the call defaults to using C calling conventions.
6806#. The optional :ref:`Parameter Attributes <paramattrs>` list for return
6807   values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
6808   are valid here.
6809#. The optional addrspace attribute can be used to indicate the address space
6810   of the called function. If it is not specified, the program address space
6811   from the :ref:`datalayout string<langref_datalayout>` will be used.
6812#. '``ty``': the type of the call instruction itself which is also the
6813   type of the return value. Functions that return no value are marked
6814   ``void``.
6815#. '``fnty``': shall be the signature of the function being invoked. The
6816   argument types must match the types implied by this signature. This
6817   type can be omitted if the function is not varargs.
6818#. '``fnptrval``': An LLVM value containing a pointer to a function to
6819   be invoked. In most cases, this is a direct function invocation, but
6820   indirect ``invoke``'s are just as possible, calling an arbitrary pointer
6821   to function value.
6822#. '``function args``': argument list whose types match the function
6823   signature argument types and parameter attributes. All arguments must
6824   be of :ref:`first class <t_firstclass>` type. If the function signature
6825   indicates the function accepts a variable number of arguments, the
6826   extra arguments can be specified.
6827#. '``normal label``': the label reached when the called function
6828   executes a '``ret``' instruction.
6829#. '``exception label``': the label reached when a callee returns via
6830   the :ref:`resume <i_resume>` instruction or other exception handling
6831   mechanism.
6832#. The optional :ref:`function attributes <fnattrs>` list.
6833#. The optional :ref:`operand bundles <opbundles>` list.
6834
6835Semantics:
6836""""""""""
6837
6838This instruction is designed to operate as a standard '``call``'
6839instruction in most regards. The primary difference is that it
6840establishes an association with a label, which is used by the runtime
6841library to unwind the stack.
6842
6843This instruction is used in languages with destructors to ensure that
6844proper cleanup is performed in the case of either a ``longjmp`` or a
6845thrown exception. Additionally, this is important for implementation of
6846'``catch``' clauses in high-level languages that support them.
6847
6848For the purposes of the SSA form, the definition of the value returned
6849by the '``invoke``' instruction is deemed to occur on the edge from the
6850current block to the "normal" label. If the callee unwinds then no
6851return value is available.
6852
6853Example:
6854""""""""
6855
6856.. code-block:: llvm
6857
6858      %retval = invoke i32 @Test(i32 15) to label %Continue
6859                  unwind label %TestCleanup              ; i32:retval set
6860      %retval = invoke coldcc i32 %Testfnptr(i32 15) to label %Continue
6861                  unwind label %TestCleanup              ; i32:retval set
6862
6863.. _i_callbr:
6864
6865'``callbr``' Instruction
6866^^^^^^^^^^^^^^^^^^^^^^^^
6867
6868Syntax:
6869"""""""
6870
6871::
6872
6873      <result> = callbr [cconv] [ret attrs] [addrspace(<num>)] [<ty>|<fnty> <fnptrval>(<function args>) [fn attrs]
6874                    [operand bundles] to label <normal label> or jump [other labels]
6875
6876Overview:
6877"""""""""
6878
6879The '``callbr``' instruction causes control to transfer to a specified
6880function, with the possibility of control flow transfer to either the
6881'``normal``' label or one of the '``other``' labels.
6882
6883This instruction should only be used to implement the "goto" feature of gcc
6884style inline assembly. Any other usage is an error in the IR verifier.
6885
6886Arguments:
6887""""""""""
6888
6889This instruction requires several arguments:
6890
6891#. The optional "cconv" marker indicates which :ref:`calling
6892   convention <callingconv>` the call should use. If none is
6893   specified, the call defaults to using C calling conventions.
6894#. The optional :ref:`Parameter Attributes <paramattrs>` list for return
6895   values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
6896   are valid here.
6897#. The optional addrspace attribute can be used to indicate the address space
6898   of the called function. If it is not specified, the program address space
6899   from the :ref:`datalayout string<langref_datalayout>` will be used.
6900#. '``ty``': the type of the call instruction itself which is also the
6901   type of the return value. Functions that return no value are marked
6902   ``void``.
6903#. '``fnty``': shall be the signature of the function being called. The
6904   argument types must match the types implied by this signature. This
6905   type can be omitted if the function is not varargs.
6906#. '``fnptrval``': An LLVM value containing a pointer to a function to
6907   be called. In most cases, this is a direct function call, but
6908   indirect ``callbr``'s are just as possible, calling an arbitrary pointer
6909   to function value.
6910#. '``function args``': argument list whose types match the function
6911   signature argument types and parameter attributes. All arguments must
6912   be of :ref:`first class <t_firstclass>` type. If the function signature
6913   indicates the function accepts a variable number of arguments, the
6914   extra arguments can be specified.
6915#. '``normal label``': the label reached when the called function
6916   executes a '``ret``' instruction.
6917#. '``other labels``': the labels reached when a callee transfers control
6918   to a location other than the normal '``normal label``'
6919#. The optional :ref:`function attributes <fnattrs>` list.
6920#. The optional :ref:`operand bundles <opbundles>` list.
6921
6922Semantics:
6923""""""""""
6924
6925This instruction is designed to operate as a standard '``call``'
6926instruction in most regards. The primary difference is that it
6927establishes an association with additional labels to define where control
6928flow goes after the call.
6929
6930The only use of this today is to implement the "goto" feature of gcc inline
6931assembly where additional labels can be provided as locations for the inline
6932assembly to jump to.
6933
6934Example:
6935""""""""
6936
6937.. code-block:: text
6938
6939      callbr void asm "", "r,x"(i32 %x, i8 *blockaddress(@foo, %fail))
6940                  to label %normal or jump [label %fail]
6941
6942.. _i_resume:
6943
6944'``resume``' Instruction
6945^^^^^^^^^^^^^^^^^^^^^^^^
6946
6947Syntax:
6948"""""""
6949
6950::
6951
6952      resume <type> <value>
6953
6954Overview:
6955"""""""""
6956
6957The '``resume``' instruction is a terminator instruction that has no
6958successors.
6959
6960Arguments:
6961""""""""""
6962
6963The '``resume``' instruction requires one argument, which must have the
6964same type as the result of any '``landingpad``' instruction in the same
6965function.
6966
6967Semantics:
6968""""""""""
6969
6970The '``resume``' instruction resumes propagation of an existing
6971(in-flight) exception whose unwinding was interrupted with a
6972:ref:`landingpad <i_landingpad>` instruction.
6973
6974Example:
6975""""""""
6976
6977.. code-block:: llvm
6978
6979      resume { i8*, i32 } %exn
6980
6981.. _i_catchswitch:
6982
6983'``catchswitch``' Instruction
6984^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6985
6986Syntax:
6987"""""""
6988
6989::
6990
6991      <resultval> = catchswitch within <parent> [ label <handler1>, label <handler2>, ... ] unwind to caller
6992      <resultval> = catchswitch within <parent> [ label <handler1>, label <handler2>, ... ] unwind label <default>
6993
6994Overview:
6995"""""""""
6996
6997The '``catchswitch``' instruction is used by `LLVM's exception handling system
6998<ExceptionHandling.html#overview>`_ to describe the set of possible catch handlers
6999that may be executed by the :ref:`EH personality routine <personalityfn>`.
7000
7001Arguments:
7002""""""""""
7003
7004The ``parent`` argument is the token of the funclet that contains the
7005``catchswitch`` instruction. If the ``catchswitch`` is not inside a funclet,
7006this operand may be the token ``none``.
7007
7008The ``default`` argument is the label of another basic block beginning with
7009either a ``cleanuppad`` or ``catchswitch`` instruction.  This unwind destination
7010must be a legal target with respect to the ``parent`` links, as described in
7011the `exception handling documentation\ <ExceptionHandling.html#wineh-constraints>`_.
7012
7013The ``handlers`` are a nonempty list of successor blocks that each begin with a
7014:ref:`catchpad <i_catchpad>` instruction.
7015
7016Semantics:
7017""""""""""
7018
7019Executing this instruction transfers control to one of the successors in
7020``handlers``, if appropriate, or continues to unwind via the unwind label if
7021present.
7022
7023The ``catchswitch`` is both a terminator and a "pad" instruction, meaning that
7024it must be both the first non-phi instruction and last instruction in the basic
7025block. Therefore, it must be the only non-phi instruction in the block.
7026
7027Example:
7028""""""""
7029
7030.. code-block:: text
7031
7032    dispatch1:
7033      %cs1 = catchswitch within none [label %handler0, label %handler1] unwind to caller
7034    dispatch2:
7035      %cs2 = catchswitch within %parenthandler [label %handler0] unwind label %cleanup
7036
7037.. _i_catchret:
7038
7039'``catchret``' Instruction
7040^^^^^^^^^^^^^^^^^^^^^^^^^^
7041
7042Syntax:
7043"""""""
7044
7045::
7046
7047      catchret from <token> to label <normal>
7048
7049Overview:
7050"""""""""
7051
7052The '``catchret``' instruction is a terminator instruction that has a
7053single successor.
7054
7055
7056Arguments:
7057""""""""""
7058
7059The first argument to a '``catchret``' indicates which ``catchpad`` it
7060exits.  It must be a :ref:`catchpad <i_catchpad>`.
7061The second argument to a '``catchret``' specifies where control will
7062transfer to next.
7063
7064Semantics:
7065""""""""""
7066
7067The '``catchret``' instruction ends an existing (in-flight) exception whose
7068unwinding was interrupted with a :ref:`catchpad <i_catchpad>` instruction.  The
7069:ref:`personality function <personalityfn>` gets a chance to execute arbitrary
7070code to, for example, destroy the active exception.  Control then transfers to
7071``normal``.
7072
7073The ``token`` argument must be a token produced by a ``catchpad`` instruction.
7074If the specified ``catchpad`` is not the most-recently-entered not-yet-exited
7075funclet pad (as described in the `EH documentation\ <ExceptionHandling.html#wineh-constraints>`_),
7076the ``catchret``'s behavior is undefined.
7077
7078Example:
7079""""""""
7080
7081.. code-block:: text
7082
7083      catchret from %catch label %continue
7084
7085.. _i_cleanupret:
7086
7087'``cleanupret``' Instruction
7088^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7089
7090Syntax:
7091"""""""
7092
7093::
7094
7095      cleanupret from <value> unwind label <continue>
7096      cleanupret from <value> unwind to caller
7097
7098Overview:
7099"""""""""
7100
7101The '``cleanupret``' instruction is a terminator instruction that has
7102an optional successor.
7103
7104
7105Arguments:
7106""""""""""
7107
7108The '``cleanupret``' instruction requires one argument, which indicates
7109which ``cleanuppad`` it exits, and must be a :ref:`cleanuppad <i_cleanuppad>`.
7110If the specified ``cleanuppad`` is not the most-recently-entered not-yet-exited
7111funclet pad (as described in the `EH documentation\ <ExceptionHandling.html#wineh-constraints>`_),
7112the ``cleanupret``'s behavior is undefined.
7113
7114The '``cleanupret``' instruction also has an optional successor, ``continue``,
7115which must be the label of another basic block beginning with either a
7116``cleanuppad`` or ``catchswitch`` instruction.  This unwind destination must
7117be a legal target with respect to the ``parent`` links, as described in the
7118`exception handling documentation\ <ExceptionHandling.html#wineh-constraints>`_.
7119
7120Semantics:
7121""""""""""
7122
7123The '``cleanupret``' instruction indicates to the
7124:ref:`personality function <personalityfn>` that one
7125:ref:`cleanuppad <i_cleanuppad>` it transferred control to has ended.
7126It transfers control to ``continue`` or unwinds out of the function.
7127
7128Example:
7129""""""""
7130
7131.. code-block:: text
7132
7133      cleanupret from %cleanup unwind to caller
7134      cleanupret from %cleanup unwind label %continue
7135
7136.. _i_unreachable:
7137
7138'``unreachable``' Instruction
7139^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7140
7141Syntax:
7142"""""""
7143
7144::
7145
7146      unreachable
7147
7148Overview:
7149"""""""""
7150
7151The '``unreachable``' instruction has no defined semantics. This
7152instruction is used to inform the optimizer that a particular portion of
7153the code is not reachable. This can be used to indicate that the code
7154after a no-return function cannot be reached, and other facts.
7155
7156Semantics:
7157""""""""""
7158
7159The '``unreachable``' instruction has no defined semantics.
7160
7161.. _unaryops:
7162
7163Unary Operations
7164-----------------
7165
7166Unary operators require a single operand, execute an operation on
7167it, and produce a single value. The operand might represent multiple
7168data, as is the case with the :ref:`vector <t_vector>` data type. The
7169result value has the same type as its operand.
7170
7171.. _i_fneg:
7172
7173'``fneg``' Instruction
7174^^^^^^^^^^^^^^^^^^^^^^
7175
7176Syntax:
7177"""""""
7178
7179::
7180
7181      <result> = fneg [fast-math flags]* <ty> <op1>   ; yields ty:result
7182
7183Overview:
7184"""""""""
7185
7186The '``fneg``' instruction returns the negation of its operand.
7187
7188Arguments:
7189""""""""""
7190
7191The argument to the '``fneg``' instruction must be a
7192:ref:`floating-point <t_floating>` or :ref:`vector <t_vector>` of
7193floating-point values.
7194
7195Semantics:
7196""""""""""
7197
7198The value produced is a copy of the operand with its sign bit flipped.
7199This instruction can also take any number of :ref:`fast-math
7200flags <fastmath>`, which are optimization hints to enable otherwise
7201unsafe floating-point optimizations:
7202
7203Example:
7204""""""""
7205
7206.. code-block:: text
7207
7208      <result> = fneg float %val          ; yields float:result = -%var
7209
7210.. _binaryops:
7211
7212Binary Operations
7213-----------------
7214
7215Binary operators are used to do most of the computation in a program.
7216They require two operands of the same type, execute an operation on
7217them, and produce a single value. The operands might represent multiple
7218data, as is the case with the :ref:`vector <t_vector>` data type. The
7219result value has the same type as its operands.
7220
7221There are several different binary operators:
7222
7223.. _i_add:
7224
7225'``add``' Instruction
7226^^^^^^^^^^^^^^^^^^^^^
7227
7228Syntax:
7229"""""""
7230
7231::
7232
7233      <result> = add <ty> <op1>, <op2>          ; yields ty:result
7234      <result> = add nuw <ty> <op1>, <op2>      ; yields ty:result
7235      <result> = add nsw <ty> <op1>, <op2>      ; yields ty:result
7236      <result> = add nuw nsw <ty> <op1>, <op2>  ; yields ty:result
7237
7238Overview:
7239"""""""""
7240
7241The '``add``' instruction returns the sum of its two operands.
7242
7243Arguments:
7244""""""""""
7245
7246The two arguments to the '``add``' instruction must be
7247:ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
7248arguments must have identical types.
7249
7250Semantics:
7251""""""""""
7252
7253The value produced is the integer sum of the two operands.
7254
7255If the sum has unsigned overflow, the result returned is the
7256mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
7257the result.
7258
7259Because LLVM integers use a two's complement representation, this
7260instruction is appropriate for both signed and unsigned integers.
7261
7262``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
7263respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
7264result value of the ``add`` is a :ref:`poison value <poisonvalues>` if
7265unsigned and/or signed overflow, respectively, occurs.
7266
7267Example:
7268""""""""
7269
7270.. code-block:: text
7271
7272      <result> = add i32 4, %var          ; yields i32:result = 4 + %var
7273
7274.. _i_fadd:
7275
7276'``fadd``' Instruction
7277^^^^^^^^^^^^^^^^^^^^^^
7278
7279Syntax:
7280"""""""
7281
7282::
7283
7284      <result> = fadd [fast-math flags]* <ty> <op1>, <op2>   ; yields ty:result
7285
7286Overview:
7287"""""""""
7288
7289The '``fadd``' instruction returns the sum of its two operands.
7290
7291Arguments:
7292""""""""""
7293
7294The two arguments to the '``fadd``' instruction must be
7295:ref:`floating-point <t_floating>` or :ref:`vector <t_vector>` of
7296floating-point values. Both arguments must have identical types.
7297
7298Semantics:
7299""""""""""
7300
7301The value produced is the floating-point sum of the two operands.
7302This instruction is assumed to execute in the default :ref:`floating-point
7303environment <floatenv>`.
7304This instruction can also take any number of :ref:`fast-math
7305flags <fastmath>`, which are optimization hints to enable otherwise
7306unsafe floating-point optimizations:
7307
7308Example:
7309""""""""
7310
7311.. code-block:: text
7312
7313      <result> = fadd float 4.0, %var          ; yields float:result = 4.0 + %var
7314
7315'``sub``' Instruction
7316^^^^^^^^^^^^^^^^^^^^^
7317
7318Syntax:
7319"""""""
7320
7321::
7322
7323      <result> = sub <ty> <op1>, <op2>          ; yields ty:result
7324      <result> = sub nuw <ty> <op1>, <op2>      ; yields ty:result
7325      <result> = sub nsw <ty> <op1>, <op2>      ; yields ty:result
7326      <result> = sub nuw nsw <ty> <op1>, <op2>  ; yields ty:result
7327
7328Overview:
7329"""""""""
7330
7331The '``sub``' instruction returns the difference of its two operands.
7332
7333Note that the '``sub``' instruction is used to represent the '``neg``'
7334instruction present in most other intermediate representations.
7335
7336Arguments:
7337""""""""""
7338
7339The two arguments to the '``sub``' instruction must be
7340:ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
7341arguments must have identical types.
7342
7343Semantics:
7344""""""""""
7345
7346The value produced is the integer difference of the two operands.
7347
7348If the difference has unsigned overflow, the result returned is the
7349mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
7350the result.
7351
7352Because LLVM integers use a two's complement representation, this
7353instruction is appropriate for both signed and unsigned integers.
7354
7355``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
7356respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
7357result value of the ``sub`` is a :ref:`poison value <poisonvalues>` if
7358unsigned and/or signed overflow, respectively, occurs.
7359
7360Example:
7361""""""""
7362
7363.. code-block:: text
7364
7365      <result> = sub i32 4, %var          ; yields i32:result = 4 - %var
7366      <result> = sub i32 0, %val          ; yields i32:result = -%var
7367
7368.. _i_fsub:
7369
7370'``fsub``' Instruction
7371^^^^^^^^^^^^^^^^^^^^^^
7372
7373Syntax:
7374"""""""
7375
7376::
7377
7378      <result> = fsub [fast-math flags]* <ty> <op1>, <op2>   ; yields ty:result
7379
7380Overview:
7381"""""""""
7382
7383The '``fsub``' instruction returns the difference of its two operands.
7384
7385Arguments:
7386""""""""""
7387
7388The two arguments to the '``fsub``' instruction must be
7389:ref:`floating-point <t_floating>` or :ref:`vector <t_vector>` of
7390floating-point values. Both arguments must have identical types.
7391
7392Semantics:
7393""""""""""
7394
7395The value produced is the floating-point difference of the two operands.
7396This instruction is assumed to execute in the default :ref:`floating-point
7397environment <floatenv>`.
7398This instruction can also take any number of :ref:`fast-math
7399flags <fastmath>`, which are optimization hints to enable otherwise
7400unsafe floating-point optimizations:
7401
7402Example:
7403""""""""
7404
7405.. code-block:: text
7406
7407      <result> = fsub float 4.0, %var           ; yields float:result = 4.0 - %var
7408      <result> = fsub float -0.0, %val          ; yields float:result = -%var
7409
7410'``mul``' Instruction
7411^^^^^^^^^^^^^^^^^^^^^
7412
7413Syntax:
7414"""""""
7415
7416::
7417
7418      <result> = mul <ty> <op1>, <op2>          ; yields ty:result
7419      <result> = mul nuw <ty> <op1>, <op2>      ; yields ty:result
7420      <result> = mul nsw <ty> <op1>, <op2>      ; yields ty:result
7421      <result> = mul nuw nsw <ty> <op1>, <op2>  ; yields ty:result
7422
7423Overview:
7424"""""""""
7425
7426The '``mul``' instruction returns the product of its two operands.
7427
7428Arguments:
7429""""""""""
7430
7431The two arguments to the '``mul``' instruction must be
7432:ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
7433arguments must have identical types.
7434
7435Semantics:
7436""""""""""
7437
7438The value produced is the integer product of the two operands.
7439
7440If the result of the multiplication has unsigned overflow, the result
7441returned is the mathematical result modulo 2\ :sup:`n`\ , where n is the
7442bit width of the result.
7443
7444Because LLVM integers use a two's complement representation, and the
7445result is the same width as the operands, this instruction returns the
7446correct result for both signed and unsigned integers. If a full product
7447(e.g. ``i32`` * ``i32`` -> ``i64``) is needed, the operands should be
7448sign-extended or zero-extended as appropriate to the width of the full
7449product.
7450
7451``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
7452respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
7453result value of the ``mul`` is a :ref:`poison value <poisonvalues>` if
7454unsigned and/or signed overflow, respectively, occurs.
7455
7456Example:
7457""""""""
7458
7459.. code-block:: text
7460
7461      <result> = mul i32 4, %var          ; yields i32:result = 4 * %var
7462
7463.. _i_fmul:
7464
7465'``fmul``' Instruction
7466^^^^^^^^^^^^^^^^^^^^^^
7467
7468Syntax:
7469"""""""
7470
7471::
7472
7473      <result> = fmul [fast-math flags]* <ty> <op1>, <op2>   ; yields ty:result
7474
7475Overview:
7476"""""""""
7477
7478The '``fmul``' instruction returns the product of its two operands.
7479
7480Arguments:
7481""""""""""
7482
7483The two arguments to the '``fmul``' instruction must be
7484:ref:`floating-point <t_floating>` or :ref:`vector <t_vector>` of
7485floating-point values. Both arguments must have identical types.
7486
7487Semantics:
7488""""""""""
7489
7490The value produced is the floating-point product of the two operands.
7491This instruction is assumed to execute in the default :ref:`floating-point
7492environment <floatenv>`.
7493This instruction can also take any number of :ref:`fast-math
7494flags <fastmath>`, which are optimization hints to enable otherwise
7495unsafe floating-point optimizations:
7496
7497Example:
7498""""""""
7499
7500.. code-block:: text
7501
7502      <result> = fmul float 4.0, %var          ; yields float:result = 4.0 * %var
7503
7504'``udiv``' Instruction
7505^^^^^^^^^^^^^^^^^^^^^^
7506
7507Syntax:
7508"""""""
7509
7510::
7511
7512      <result> = udiv <ty> <op1>, <op2>         ; yields ty:result
7513      <result> = udiv exact <ty> <op1>, <op2>   ; yields ty:result
7514
7515Overview:
7516"""""""""
7517
7518The '``udiv``' instruction returns the quotient of its two operands.
7519
7520Arguments:
7521""""""""""
7522
7523The two arguments to the '``udiv``' instruction must be
7524:ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
7525arguments must have identical types.
7526
7527Semantics:
7528""""""""""
7529
7530The value produced is the unsigned integer quotient of the two operands.
7531
7532Note that unsigned integer division and signed integer division are
7533distinct operations; for signed integer division, use '``sdiv``'.
7534
7535Division by zero is undefined behavior. For vectors, if any element
7536of the divisor is zero, the operation has undefined behavior.
7537
7538
7539If the ``exact`` keyword is present, the result value of the ``udiv`` is
7540a :ref:`poison value <poisonvalues>` if %op1 is not a multiple of %op2 (as
7541such, "((a udiv exact b) mul b) == a").
7542
7543Example:
7544""""""""
7545
7546.. code-block:: text
7547
7548      <result> = udiv i32 4, %var          ; yields i32:result = 4 / %var
7549
7550'``sdiv``' Instruction
7551^^^^^^^^^^^^^^^^^^^^^^
7552
7553Syntax:
7554"""""""
7555
7556::
7557
7558      <result> = sdiv <ty> <op1>, <op2>         ; yields ty:result
7559      <result> = sdiv exact <ty> <op1>, <op2>   ; yields ty:result
7560
7561Overview:
7562"""""""""
7563
7564The '``sdiv``' instruction returns the quotient of its two operands.
7565
7566Arguments:
7567""""""""""
7568
7569The two arguments to the '``sdiv``' instruction must be
7570:ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
7571arguments must have identical types.
7572
7573Semantics:
7574""""""""""
7575
7576The value produced is the signed integer quotient of the two operands
7577rounded towards zero.
7578
7579Note that signed integer division and unsigned integer division are
7580distinct operations; for unsigned integer division, use '``udiv``'.
7581
7582Division by zero is undefined behavior. For vectors, if any element
7583of the divisor is zero, the operation has undefined behavior.
7584Overflow also leads to undefined behavior; this is a rare case, but can
7585occur, for example, by doing a 32-bit division of -2147483648 by -1.
7586
7587If the ``exact`` keyword is present, the result value of the ``sdiv`` is
7588a :ref:`poison value <poisonvalues>` if the result would be rounded.
7589
7590Example:
7591""""""""
7592
7593.. code-block:: text
7594
7595      <result> = sdiv i32 4, %var          ; yields i32:result = 4 / %var
7596
7597.. _i_fdiv:
7598
7599'``fdiv``' Instruction
7600^^^^^^^^^^^^^^^^^^^^^^
7601
7602Syntax:
7603"""""""
7604
7605::
7606
7607      <result> = fdiv [fast-math flags]* <ty> <op1>, <op2>   ; yields ty:result
7608
7609Overview:
7610"""""""""
7611
7612The '``fdiv``' instruction returns the quotient of its two operands.
7613
7614Arguments:
7615""""""""""
7616
7617The two arguments to the '``fdiv``' instruction must be
7618:ref:`floating-point <t_floating>` or :ref:`vector <t_vector>` of
7619floating-point values. Both arguments must have identical types.
7620
7621Semantics:
7622""""""""""
7623
7624The value produced is the floating-point quotient of the two operands.
7625This instruction is assumed to execute in the default :ref:`floating-point
7626environment <floatenv>`.
7627This instruction can also take any number of :ref:`fast-math
7628flags <fastmath>`, which are optimization hints to enable otherwise
7629unsafe floating-point optimizations:
7630
7631Example:
7632""""""""
7633
7634.. code-block:: text
7635
7636      <result> = fdiv float 4.0, %var          ; yields float:result = 4.0 / %var
7637
7638'``urem``' Instruction
7639^^^^^^^^^^^^^^^^^^^^^^
7640
7641Syntax:
7642"""""""
7643
7644::
7645
7646      <result> = urem <ty> <op1>, <op2>   ; yields ty:result
7647
7648Overview:
7649"""""""""
7650
7651The '``urem``' instruction returns the remainder from the unsigned
7652division of its two arguments.
7653
7654Arguments:
7655""""""""""
7656
7657The two arguments to the '``urem``' instruction must be
7658:ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
7659arguments must have identical types.
7660
7661Semantics:
7662""""""""""
7663
7664This instruction returns the unsigned integer *remainder* of a division.
7665This instruction always performs an unsigned division to get the
7666remainder.
7667
7668Note that unsigned integer remainder and signed integer remainder are
7669distinct operations; for signed integer remainder, use '``srem``'.
7670
7671Taking the remainder of a division by zero is undefined behavior.
7672For vectors, if any element of the divisor is zero, the operation has
7673undefined behavior.
7674
7675Example:
7676""""""""
7677
7678.. code-block:: text
7679
7680      <result> = urem i32 4, %var          ; yields i32:result = 4 % %var
7681
7682'``srem``' Instruction
7683^^^^^^^^^^^^^^^^^^^^^^
7684
7685Syntax:
7686"""""""
7687
7688::
7689
7690      <result> = srem <ty> <op1>, <op2>   ; yields ty:result
7691
7692Overview:
7693"""""""""
7694
7695The '``srem``' instruction returns the remainder from the signed
7696division of its two operands. This instruction can also take
7697:ref:`vector <t_vector>` versions of the values in which case the elements
7698must be integers.
7699
7700Arguments:
7701""""""""""
7702
7703The two arguments to the '``srem``' instruction must be
7704:ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
7705arguments must have identical types.
7706
7707Semantics:
7708""""""""""
7709
7710This instruction returns the *remainder* of a division (where the result
7711is either zero or has the same sign as the dividend, ``op1``), not the
7712*modulo* operator (where the result is either zero or has the same sign
7713as the divisor, ``op2``) of a value. For more information about the
7714difference, see `The Math
7715Forum <http://mathforum.org/dr.math/problems/anne.4.28.99.html>`_. For a
7716table of how this is implemented in various languages, please see
7717`Wikipedia: modulo
7718operation <http://en.wikipedia.org/wiki/Modulo_operation>`_.
7719
7720Note that signed integer remainder and unsigned integer remainder are
7721distinct operations; for unsigned integer remainder, use '``urem``'.
7722
7723Taking the remainder of a division by zero is undefined behavior.
7724For vectors, if any element of the divisor is zero, the operation has
7725undefined behavior.
7726Overflow also leads to undefined behavior; this is a rare case, but can
7727occur, for example, by taking the remainder of a 32-bit division of
7728-2147483648 by -1. (The remainder doesn't actually overflow, but this
7729rule lets srem be implemented using instructions that return both the
7730result of the division and the remainder.)
7731
7732Example:
7733""""""""
7734
7735.. code-block:: text
7736
7737      <result> = srem i32 4, %var          ; yields i32:result = 4 % %var
7738
7739.. _i_frem:
7740
7741'``frem``' Instruction
7742^^^^^^^^^^^^^^^^^^^^^^
7743
7744Syntax:
7745"""""""
7746
7747::
7748
7749      <result> = frem [fast-math flags]* <ty> <op1>, <op2>   ; yields ty:result
7750
7751Overview:
7752"""""""""
7753
7754The '``frem``' instruction returns the remainder from the division of
7755its two operands.
7756
7757Arguments:
7758""""""""""
7759
7760The two arguments to the '``frem``' instruction must be
7761:ref:`floating-point <t_floating>` or :ref:`vector <t_vector>` of
7762floating-point values. Both arguments must have identical types.
7763
7764Semantics:
7765""""""""""
7766
7767The value produced is the floating-point remainder of the two operands.
7768This is the same output as a libm '``fmod``' function, but without any
7769possibility of setting ``errno``. The remainder has the same sign as the
7770dividend.
7771This instruction is assumed to execute in the default :ref:`floating-point
7772environment <floatenv>`.
7773This instruction can also take any number of :ref:`fast-math
7774flags <fastmath>`, which are optimization hints to enable otherwise
7775unsafe floating-point optimizations:
7776
7777Example:
7778""""""""
7779
7780.. code-block:: text
7781
7782      <result> = frem float 4.0, %var          ; yields float:result = 4.0 % %var
7783
7784.. _bitwiseops:
7785
7786Bitwise Binary Operations
7787-------------------------
7788
7789Bitwise binary operators are used to do various forms of bit-twiddling
7790in a program. They are generally very efficient instructions and can
7791commonly be strength reduced from other instructions. They require two
7792operands of the same type, execute an operation on them, and produce a
7793single value. The resulting value is the same type as its operands.
7794
7795'``shl``' Instruction
7796^^^^^^^^^^^^^^^^^^^^^
7797
7798Syntax:
7799"""""""
7800
7801::
7802
7803      <result> = shl <ty> <op1>, <op2>           ; yields ty:result
7804      <result> = shl nuw <ty> <op1>, <op2>       ; yields ty:result
7805      <result> = shl nsw <ty> <op1>, <op2>       ; yields ty:result
7806      <result> = shl nuw nsw <ty> <op1>, <op2>   ; yields ty:result
7807
7808Overview:
7809"""""""""
7810
7811The '``shl``' instruction returns the first operand shifted to the left
7812a specified number of bits.
7813
7814Arguments:
7815""""""""""
7816
7817Both arguments to the '``shl``' instruction must be the same
7818:ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
7819'``op2``' is treated as an unsigned value.
7820
7821Semantics:
7822""""""""""
7823
7824The value produced is ``op1`` \* 2\ :sup:`op2` mod 2\ :sup:`n`,
7825where ``n`` is the width of the result. If ``op2`` is (statically or
7826dynamically) equal to or larger than the number of bits in
7827``op1``, this instruction returns a :ref:`poison value <poisonvalues>`.
7828If the arguments are vectors, each vector element of ``op1`` is shifted
7829by the corresponding shift amount in ``op2``.
7830
7831If the ``nuw`` keyword is present, then the shift produces a poison
7832value if it shifts out any non-zero bits.
7833If the ``nsw`` keyword is present, then the shift produces a poison
7834value if it shifts out any bits that disagree with the resultant sign bit.
7835
7836Example:
7837""""""""
7838
7839.. code-block:: text
7840
7841      <result> = shl i32 4, %var   ; yields i32: 4 << %var
7842      <result> = shl i32 4, 2      ; yields i32: 16
7843      <result> = shl i32 1, 10     ; yields i32: 1024
7844      <result> = shl i32 1, 32     ; undefined
7845      <result> = shl <2 x i32> < i32 1, i32 1>, < i32 1, i32 2>   ; yields: result=<2 x i32> < i32 2, i32 4>
7846
7847'``lshr``' Instruction
7848^^^^^^^^^^^^^^^^^^^^^^
7849
7850Syntax:
7851"""""""
7852
7853::
7854
7855      <result> = lshr <ty> <op1>, <op2>         ; yields ty:result
7856      <result> = lshr exact <ty> <op1>, <op2>   ; yields ty:result
7857
7858Overview:
7859"""""""""
7860
7861The '``lshr``' instruction (logical shift right) returns the first
7862operand shifted to the right a specified number of bits with zero fill.
7863
7864Arguments:
7865""""""""""
7866
7867Both arguments to the '``lshr``' instruction must be the same
7868:ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
7869'``op2``' is treated as an unsigned value.
7870
7871Semantics:
7872""""""""""
7873
7874This instruction always performs a logical shift right operation. The
7875most significant bits of the result will be filled with zero bits after
7876the shift. If ``op2`` is (statically or dynamically) equal to or larger
7877than the number of bits in ``op1``, this instruction returns a :ref:`poison
7878value <poisonvalues>`. If the arguments are vectors, each vector element
7879of ``op1`` is shifted by the corresponding shift amount in ``op2``.
7880
7881If the ``exact`` keyword is present, the result value of the ``lshr`` is
7882a poison value if any of the bits shifted out are non-zero.
7883
7884Example:
7885""""""""
7886
7887.. code-block:: text
7888
7889      <result> = lshr i32 4, 1   ; yields i32:result = 2
7890      <result> = lshr i32 4, 2   ; yields i32:result = 1
7891      <result> = lshr i8  4, 3   ; yields i8:result = 0
7892      <result> = lshr i8 -2, 1   ; yields i8:result = 0x7F
7893      <result> = lshr i32 1, 32  ; undefined
7894      <result> = lshr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 2>   ; yields: result=<2 x i32> < i32 0x7FFFFFFF, i32 1>
7895
7896'``ashr``' Instruction
7897^^^^^^^^^^^^^^^^^^^^^^
7898
7899Syntax:
7900"""""""
7901
7902::
7903
7904      <result> = ashr <ty> <op1>, <op2>         ; yields ty:result
7905      <result> = ashr exact <ty> <op1>, <op2>   ; yields ty:result
7906
7907Overview:
7908"""""""""
7909
7910The '``ashr``' instruction (arithmetic shift right) returns the first
7911operand shifted to the right a specified number of bits with sign
7912extension.
7913
7914Arguments:
7915""""""""""
7916
7917Both arguments to the '``ashr``' instruction must be the same
7918:ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
7919'``op2``' is treated as an unsigned value.
7920
7921Semantics:
7922""""""""""
7923
7924This instruction always performs an arithmetic shift right operation,
7925The most significant bits of the result will be filled with the sign bit
7926of ``op1``. If ``op2`` is (statically or dynamically) equal to or larger
7927than the number of bits in ``op1``, this instruction returns a :ref:`poison
7928value <poisonvalues>`. If the arguments are vectors, each vector element
7929of ``op1`` is shifted by the corresponding shift amount in ``op2``.
7930
7931If the ``exact`` keyword is present, the result value of the ``ashr`` is
7932a poison value if any of the bits shifted out are non-zero.
7933
7934Example:
7935""""""""
7936
7937.. code-block:: text
7938
7939      <result> = ashr i32 4, 1   ; yields i32:result = 2
7940      <result> = ashr i32 4, 2   ; yields i32:result = 1
7941      <result> = ashr i8  4, 3   ; yields i8:result = 0
7942      <result> = ashr i8 -2, 1   ; yields i8:result = -1
7943      <result> = ashr i32 1, 32  ; undefined
7944      <result> = ashr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 3>   ; yields: result=<2 x i32> < i32 -1, i32 0>
7945
7946'``and``' Instruction
7947^^^^^^^^^^^^^^^^^^^^^
7948
7949Syntax:
7950"""""""
7951
7952::
7953
7954      <result> = and <ty> <op1>, <op2>   ; yields ty:result
7955
7956Overview:
7957"""""""""
7958
7959The '``and``' instruction returns the bitwise logical and of its two
7960operands.
7961
7962Arguments:
7963""""""""""
7964
7965The two arguments to the '``and``' instruction must be
7966:ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
7967arguments must have identical types.
7968
7969Semantics:
7970""""""""""
7971
7972The truth table used for the '``and``' instruction is:
7973
7974+-----+-----+-----+
7975| In0 | In1 | Out |
7976+-----+-----+-----+
7977|   0 |   0 |   0 |
7978+-----+-----+-----+
7979|   0 |   1 |   0 |
7980+-----+-----+-----+
7981|   1 |   0 |   0 |
7982+-----+-----+-----+
7983|   1 |   1 |   1 |
7984+-----+-----+-----+
7985
7986Example:
7987""""""""
7988
7989.. code-block:: text
7990
7991      <result> = and i32 4, %var         ; yields i32:result = 4 & %var
7992      <result> = and i32 15, 40          ; yields i32:result = 8
7993      <result> = and i32 4, 8            ; yields i32:result = 0
7994
7995'``or``' Instruction
7996^^^^^^^^^^^^^^^^^^^^
7997
7998Syntax:
7999"""""""
8000
8001::
8002
8003      <result> = or <ty> <op1>, <op2>   ; yields ty:result
8004
8005Overview:
8006"""""""""
8007
8008The '``or``' instruction returns the bitwise logical inclusive or of its
8009two operands.
8010
8011Arguments:
8012""""""""""
8013
8014The two arguments to the '``or``' instruction must be
8015:ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
8016arguments must have identical types.
8017
8018Semantics:
8019""""""""""
8020
8021The truth table used for the '``or``' instruction is:
8022
8023+-----+-----+-----+
8024| In0 | In1 | Out |
8025+-----+-----+-----+
8026|   0 |   0 |   0 |
8027+-----+-----+-----+
8028|   0 |   1 |   1 |
8029+-----+-----+-----+
8030|   1 |   0 |   1 |
8031+-----+-----+-----+
8032|   1 |   1 |   1 |
8033+-----+-----+-----+
8034
8035Example:
8036""""""""
8037
8038::
8039
8040      <result> = or i32 4, %var         ; yields i32:result = 4 | %var
8041      <result> = or i32 15, 40          ; yields i32:result = 47
8042      <result> = or i32 4, 8            ; yields i32:result = 12
8043
8044'``xor``' Instruction
8045^^^^^^^^^^^^^^^^^^^^^
8046
8047Syntax:
8048"""""""
8049
8050::
8051
8052      <result> = xor <ty> <op1>, <op2>   ; yields ty:result
8053
8054Overview:
8055"""""""""
8056
8057The '``xor``' instruction returns the bitwise logical exclusive or of
8058its two operands. The ``xor`` is used to implement the "one's
8059complement" operation, which is the "~" operator in C.
8060
8061Arguments:
8062""""""""""
8063
8064The two arguments to the '``xor``' instruction must be
8065:ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
8066arguments must have identical types.
8067
8068Semantics:
8069""""""""""
8070
8071The truth table used for the '``xor``' instruction is:
8072
8073+-----+-----+-----+
8074| In0 | In1 | Out |
8075+-----+-----+-----+
8076|   0 |   0 |   0 |
8077+-----+-----+-----+
8078|   0 |   1 |   1 |
8079+-----+-----+-----+
8080|   1 |   0 |   1 |
8081+-----+-----+-----+
8082|   1 |   1 |   0 |
8083+-----+-----+-----+
8084
8085Example:
8086""""""""
8087
8088.. code-block:: text
8089
8090      <result> = xor i32 4, %var         ; yields i32:result = 4 ^ %var
8091      <result> = xor i32 15, 40          ; yields i32:result = 39
8092      <result> = xor i32 4, 8            ; yields i32:result = 12
8093      <result> = xor i32 %V, -1          ; yields i32:result = ~%V
8094
8095Vector Operations
8096-----------------
8097
8098LLVM supports several instructions to represent vector operations in a
8099target-independent manner. These instructions cover the element-access
8100and vector-specific operations needed to process vectors effectively.
8101While LLVM does directly support these vector operations, many
8102sophisticated algorithms will want to use target-specific intrinsics to
8103take full advantage of a specific target.
8104
8105.. _i_extractelement:
8106
8107'``extractelement``' Instruction
8108^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8109
8110Syntax:
8111"""""""
8112
8113::
8114
8115      <result> = extractelement <n x <ty>> <val>, <ty2> <idx>  ; yields <ty>
8116
8117Overview:
8118"""""""""
8119
8120The '``extractelement``' instruction extracts a single scalar element
8121from a vector at a specified index.
8122
8123Arguments:
8124""""""""""
8125
8126The first operand of an '``extractelement``' instruction is a value of
8127:ref:`vector <t_vector>` type. The second operand is an index indicating
8128the position from which to extract the element. The index may be a
8129variable of any integer type.
8130
8131Semantics:
8132""""""""""
8133
8134The result is a scalar of the same type as the element type of ``val``.
8135Its value is the value at position ``idx`` of ``val``. If ``idx``
8136exceeds the length of ``val``, the result is a
8137:ref:`poison value <poisonvalues>`.
8138
8139Example:
8140""""""""
8141
8142.. code-block:: text
8143
8144      <result> = extractelement <4 x i32> %vec, i32 0    ; yields i32
8145
8146.. _i_insertelement:
8147
8148'``insertelement``' Instruction
8149^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8150
8151Syntax:
8152"""""""
8153
8154::
8155
8156      <result> = insertelement <n x <ty>> <val>, <ty> <elt>, <ty2> <idx>    ; yields <n x <ty>>
8157
8158Overview:
8159"""""""""
8160
8161The '``insertelement``' instruction inserts a scalar element into a
8162vector at a specified index.
8163
8164Arguments:
8165""""""""""
8166
8167The first operand of an '``insertelement``' instruction is a value of
8168:ref:`vector <t_vector>` type. The second operand is a scalar value whose
8169type must equal the element type of the first operand. The third operand
8170is an index indicating the position at which to insert the value. The
8171index may be a variable of any integer type.
8172
8173Semantics:
8174""""""""""
8175
8176The result is a vector of the same type as ``val``. Its element values
8177are those of ``val`` except at position ``idx``, where it gets the value
8178``elt``. If ``idx`` exceeds the length of ``val``, the result
8179is a :ref:`poison value <poisonvalues>`.
8180
8181Example:
8182""""""""
8183
8184.. code-block:: text
8185
8186      <result> = insertelement <4 x i32> %vec, i32 1, i32 0    ; yields <4 x i32>
8187
8188.. _i_shufflevector:
8189
8190'``shufflevector``' Instruction
8191^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8192
8193Syntax:
8194"""""""
8195
8196::
8197
8198      <result> = shufflevector <n x <ty>> <v1>, <n x <ty>> <v2>, <m x i32> <mask>    ; yields <m x <ty>>
8199
8200Overview:
8201"""""""""
8202
8203The '``shufflevector``' instruction constructs a permutation of elements
8204from two input vectors, returning a vector with the same element type as
8205the input and length that is the same as the shuffle mask.
8206
8207Arguments:
8208""""""""""
8209
8210The first two operands of a '``shufflevector``' instruction are vectors
8211with the same type. The third argument is a shuffle mask whose element
8212type is always 'i32'. The result of the instruction is a vector whose
8213length is the same as the shuffle mask and whose element type is the
8214same as the element type of the first two operands.
8215
8216The shuffle mask operand is required to be a constant vector with either
8217constant integer or undef values.
8218
8219Semantics:
8220""""""""""
8221
8222The elements of the two input vectors are numbered from left to right
8223across both of the vectors. The shuffle mask operand specifies, for each
8224element of the result vector, which element of the two input vectors the
8225result element gets. If the shuffle mask is undef, the result vector is
8226undef. If any element of the mask operand is undef, that element of the
8227result is undef. If the shuffle mask selects an undef element from one
8228of the input vectors, the resulting element is undef.
8229
8230Example:
8231""""""""
8232
8233.. code-block:: text
8234
8235      <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
8236                              <4 x i32> <i32 0, i32 4, i32 1, i32 5>  ; yields <4 x i32>
8237      <result> = shufflevector <4 x i32> %v1, <4 x i32> undef,
8238                              <4 x i32> <i32 0, i32 1, i32 2, i32 3>  ; yields <4 x i32> - Identity shuffle.
8239      <result> = shufflevector <8 x i32> %v1, <8 x i32> undef,
8240                              <4 x i32> <i32 0, i32 1, i32 2, i32 3>  ; yields <4 x i32>
8241      <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
8242                              <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7 >  ; yields <8 x i32>
8243
8244Aggregate Operations
8245--------------------
8246
8247LLVM supports several instructions for working with
8248:ref:`aggregate <t_aggregate>` values.
8249
8250.. _i_extractvalue:
8251
8252'``extractvalue``' Instruction
8253^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8254
8255Syntax:
8256"""""""
8257
8258::
8259
8260      <result> = extractvalue <aggregate type> <val>, <idx>{, <idx>}*
8261
8262Overview:
8263"""""""""
8264
8265The '``extractvalue``' instruction extracts the value of a member field
8266from an :ref:`aggregate <t_aggregate>` value.
8267
8268Arguments:
8269""""""""""
8270
8271The first operand of an '``extractvalue``' instruction is a value of
8272:ref:`struct <t_struct>` or :ref:`array <t_array>` type. The other operands are
8273constant indices to specify which value to extract in a similar manner
8274as indices in a '``getelementptr``' instruction.
8275
8276The major differences to ``getelementptr`` indexing are:
8277
8278-  Since the value being indexed is not a pointer, the first index is
8279   omitted and assumed to be zero.
8280-  At least one index must be specified.
8281-  Not only struct indices but also array indices must be in bounds.
8282
8283Semantics:
8284""""""""""
8285
8286The result is the value at the position in the aggregate specified by
8287the index operands.
8288
8289Example:
8290""""""""
8291
8292.. code-block:: text
8293
8294      <result> = extractvalue {i32, float} %agg, 0    ; yields i32
8295
8296.. _i_insertvalue:
8297
8298'``insertvalue``' Instruction
8299^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8300
8301Syntax:
8302"""""""
8303
8304::
8305
8306      <result> = insertvalue <aggregate type> <val>, <ty> <elt>, <idx>{, <idx>}*    ; yields <aggregate type>
8307
8308Overview:
8309"""""""""
8310
8311The '``insertvalue``' instruction inserts a value into a member field in
8312an :ref:`aggregate <t_aggregate>` value.
8313
8314Arguments:
8315""""""""""
8316
8317The first operand of an '``insertvalue``' instruction is a value of
8318:ref:`struct <t_struct>` or :ref:`array <t_array>` type. The second operand is
8319a first-class value to insert. The following operands are constant
8320indices indicating the position at which to insert the value in a
8321similar manner as indices in a '``extractvalue``' instruction. The value
8322to insert must have the same type as the value identified by the
8323indices.
8324
8325Semantics:
8326""""""""""
8327
8328The result is an aggregate of the same type as ``val``. Its value is
8329that of ``val`` except that the value at the position specified by the
8330indices is that of ``elt``.
8331
8332Example:
8333""""""""
8334
8335.. code-block:: llvm
8336
8337      %agg1 = insertvalue {i32, float} undef, i32 1, 0              ; yields {i32 1, float undef}
8338      %agg2 = insertvalue {i32, float} %agg1, float %val, 1         ; yields {i32 1, float %val}
8339      %agg3 = insertvalue {i32, {float}} undef, float %val, 1, 0    ; yields {i32 undef, {float %val}}
8340
8341.. _memoryops:
8342
8343Memory Access and Addressing Operations
8344---------------------------------------
8345
8346A key design point of an SSA-based representation is how it represents
8347memory. In LLVM, no memory locations are in SSA form, which makes things
8348very simple. This section describes how to read, write, and allocate
8349memory in LLVM.
8350
8351.. _i_alloca:
8352
8353'``alloca``' Instruction
8354^^^^^^^^^^^^^^^^^^^^^^^^
8355
8356Syntax:
8357"""""""
8358
8359::
8360
8361      <result> = alloca [inalloca] <type> [, <ty> <NumElements>] [, align <alignment>] [, addrspace(<num>)]     ; yields type addrspace(num)*:result
8362
8363Overview:
8364"""""""""
8365
8366The '``alloca``' instruction allocates memory on the stack frame of the
8367currently executing function, to be automatically released when this
8368function returns to its caller. The object is always allocated in the
8369address space for allocas indicated in the datalayout.
8370
8371Arguments:
8372""""""""""
8373
8374The '``alloca``' instruction allocates ``sizeof(<type>)*NumElements``
8375bytes of memory on the runtime stack, returning a pointer of the
8376appropriate type to the program. If "NumElements" is specified, it is
8377the number of elements allocated, otherwise "NumElements" is defaulted
8378to be one. If a constant alignment is specified, the value result of the
8379allocation is guaranteed to be aligned to at least that boundary. The
8380alignment may not be greater than ``1 << 29``. If not specified, or if
8381zero, the target can choose to align the allocation on any convenient
8382boundary compatible with the type.
8383
8384'``type``' may be any sized type.
8385
8386Semantics:
8387""""""""""
8388
8389Memory is allocated; a pointer is returned. The allocated memory is
8390uninitialized, and loading from uninitialized memory produces an undefined
8391value. The operation itself is undefined if there is insufficient stack
8392space for the allocation.'``alloca``'d memory is automatically released
8393when the function returns. The '``alloca``' instruction is commonly used
8394to represent automatic variables that must have an address available. When
8395the function returns (either with the ``ret`` or ``resume`` instructions),
8396the memory is reclaimed. Allocating zero bytes is legal, but the returned
8397pointer may not be unique. The order in which memory is allocated (ie.,
8398which way the stack grows) is not specified.
8399
8400Example:
8401""""""""
8402
8403.. code-block:: llvm
8404
8405      %ptr = alloca i32                             ; yields i32*:ptr
8406      %ptr = alloca i32, i32 4                      ; yields i32*:ptr
8407      %ptr = alloca i32, i32 4, align 1024          ; yields i32*:ptr
8408      %ptr = alloca i32, align 1024                 ; yields i32*:ptr
8409
8410.. _i_load:
8411
8412'``load``' Instruction
8413^^^^^^^^^^^^^^^^^^^^^^
8414
8415Syntax:
8416"""""""
8417
8418::
8419
8420      <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>]
8421      <result> = load atomic [volatile] <ty>, <ty>* <pointer> [syncscope("<target-scope>")] <ordering>, align <alignment> [, !invariant.group !<index>]
8422      !<index> = !{ i32 1 }
8423      !<deref_bytes_node> = !{i64 <dereferenceable_bytes>}
8424      !<align_node> = !{ i64 <value_alignment> }
8425
8426Overview:
8427"""""""""
8428
8429The '``load``' instruction is used to read from memory.
8430
8431Arguments:
8432""""""""""
8433
8434The argument to the ``load`` instruction specifies the memory address from which
8435to load. The type specified must be a :ref:`first class <t_firstclass>` type of
8436known size (i.e. not containing an :ref:`opaque structural type <t_opaque>`). If
8437the ``load`` is marked as ``volatile``, then the optimizer is not allowed to
8438modify the number or order of execution of this ``load`` with other
8439:ref:`volatile operations <volatile>`.
8440
8441If the ``load`` is marked as ``atomic``, it takes an extra :ref:`ordering
8442<ordering>` and optional ``syncscope("<target-scope>")`` argument. The
8443``release`` and ``acq_rel`` orderings are not valid on ``load`` instructions.
8444Atomic loads produce :ref:`defined <memmodel>` results when they may see
8445multiple atomic stores. The type of the pointee must be an integer, pointer, or
8446floating-point type whose bit width is a power of two greater than or equal to
8447eight and less than or equal to a target-specific size limit.  ``align`` must be
8448explicitly specified on atomic loads, and the load has undefined behavior if the
8449alignment is not set to a value which is at least the size in bytes of the
8450pointee. ``!nontemporal`` does not have any defined semantics for atomic loads.
8451
8452The optional constant ``align`` argument specifies the alignment of the
8453operation (that is, the alignment of the memory address). A value of 0
8454or an omitted ``align`` argument means that the operation has the ABI
8455alignment for the target. It is the responsibility of the code emitter
8456to ensure that the alignment information is correct. Overestimating the
8457alignment results in undefined behavior. Underestimating the alignment
8458may produce less efficient code. An alignment of 1 is always safe. The
8459maximum possible alignment is ``1 << 29``. An alignment value higher
8460than the size of the loaded type implies memory up to the alignment
8461value bytes can be safely loaded without trapping in the default
8462address space. Access of the high bytes can interfere with debugging
8463tools, so should not be accessed if the function has the
8464``sanitize_thread`` or ``sanitize_address`` attributes.
8465
8466The optional ``!nontemporal`` metadata must reference a single
8467metadata name ``<index>`` corresponding to a metadata node with one
8468``i32`` entry of value 1. The existence of the ``!nontemporal``
8469metadata on the instruction tells the optimizer and code generator
8470that this load is not expected to be reused in the cache. The code
8471generator may select special instructions to save cache bandwidth, such
8472as the ``MOVNT`` instruction on x86.
8473
8474The optional ``!invariant.load`` metadata must reference a single
8475metadata name ``<index>`` corresponding to a metadata node with no
8476entries. If a load instruction tagged with the ``!invariant.load``
8477metadata is executed, the optimizer may assume the memory location
8478referenced by the load contains the same value at all points in the
8479program where the memory location is known to be dereferenceable;
8480otherwise, the behavior is undefined.
8481
8482The optional ``!invariant.group`` metadata must reference a single metadata name
8483 ``<index>`` corresponding to a metadata node with no entries.
8484 See ``invariant.group`` metadata.
8485
8486The optional ``!nonnull`` metadata must reference a single
8487metadata name ``<index>`` corresponding to a metadata node with no
8488entries. The existence of the ``!nonnull`` metadata on the
8489instruction tells the optimizer that the value loaded is known to
8490never be null. If the value is null at runtime, the behavior is undefined.
8491This is analogous to the ``nonnull`` attribute on parameters and return
8492values. This metadata can only be applied to loads of a pointer type.
8493
8494The optional ``!dereferenceable`` metadata must reference a single metadata
8495name ``<deref_bytes_node>`` corresponding to a metadata node with one ``i64``
8496entry. The existence of the ``!dereferenceable`` metadata on the instruction
8497tells the optimizer that the value loaded is known to be dereferenceable.
8498The number of bytes known to be dereferenceable is specified by the integer
8499value in the metadata node. This is analogous to the ''dereferenceable''
8500attribute on parameters and return values. This metadata can only be applied
8501to loads of a pointer type.
8502
8503The optional ``!dereferenceable_or_null`` metadata must reference a single
8504metadata name ``<deref_bytes_node>`` corresponding to a metadata node with one
8505``i64`` entry. The existence of the ``!dereferenceable_or_null`` metadata on the
8506instruction tells the optimizer that the value loaded is known to be either
8507dereferenceable or null.
8508The number of bytes known to be dereferenceable is specified by the integer
8509value in the metadata node. This is analogous to the ''dereferenceable_or_null''
8510attribute on parameters and return values. This metadata can only be applied
8511to loads of a pointer type.
8512
8513The optional ``!align`` metadata must reference a single metadata name
8514``<align_node>`` corresponding to a metadata node with one ``i64`` entry.
8515The existence of the ``!align`` metadata on the instruction tells the
8516optimizer that the value loaded is known to be aligned to a boundary specified
8517by the integer value in the metadata node. The alignment must be a power of 2.
8518This is analogous to the ''align'' attribute on parameters and return values.
8519This metadata can only be applied to loads of a pointer type. If the returned
8520value is not appropriately aligned at runtime, the behavior is undefined.
8521
8522Semantics:
8523""""""""""
8524
8525The location of memory pointed to is loaded. If the value being loaded
8526is of scalar type then the number of bytes read does not exceed the
8527minimum number of bytes needed to hold all bits of the type. For
8528example, loading an ``i24`` reads at most three bytes. When loading a
8529value of a type like ``i20`` with a size that is not an integral number
8530of bytes, the result is undefined if the value was not originally
8531written using a store of the same type.
8532
8533Examples:
8534"""""""""
8535
8536.. code-block:: llvm
8537
8538      %ptr = alloca i32                               ; yields i32*:ptr
8539      store i32 3, i32* %ptr                          ; yields void
8540      %val = load i32, i32* %ptr                      ; yields i32:val = i32 3
8541
8542.. _i_store:
8543
8544'``store``' Instruction
8545^^^^^^^^^^^^^^^^^^^^^^^
8546
8547Syntax:
8548"""""""
8549
8550::
8551
8552      store [volatile] <ty> <value>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>][, !invariant.group !<index>]        ; yields void
8553      store atomic [volatile] <ty> <value>, <ty>* <pointer> [syncscope("<target-scope>")] <ordering>, align <alignment> [, !invariant.group !<index>] ; yields void
8554
8555Overview:
8556"""""""""
8557
8558The '``store``' instruction is used to write to memory.
8559
8560Arguments:
8561""""""""""
8562
8563There are two arguments to the ``store`` instruction: a value to store and an
8564address at which to store it. The type of the ``<pointer>`` operand must be a
8565pointer to the :ref:`first class <t_firstclass>` type of the ``<value>``
8566operand. If the ``store`` is marked as ``volatile``, then the optimizer is not
8567allowed to modify the number or order of execution of this ``store`` with other
8568:ref:`volatile operations <volatile>`.  Only values of :ref:`first class
8569<t_firstclass>` types of known size (i.e. not containing an :ref:`opaque
8570structural type <t_opaque>`) can be stored.
8571
8572If the ``store`` is marked as ``atomic``, it takes an extra :ref:`ordering
8573<ordering>` and optional ``syncscope("<target-scope>")`` argument. The
8574``acquire`` and ``acq_rel`` orderings aren't valid on ``store`` instructions.
8575Atomic loads produce :ref:`defined <memmodel>` results when they may see
8576multiple atomic stores. The type of the pointee must be an integer, pointer, or
8577floating-point type whose bit width is a power of two greater than or equal to
8578eight and less than or equal to a target-specific size limit.  ``align`` must be
8579explicitly specified on atomic stores, and the store has undefined behavior if
8580the alignment is not set to a value which is at least the size in bytes of the
8581pointee. ``!nontemporal`` does not have any defined semantics for atomic stores.
8582
8583The optional constant ``align`` argument specifies the alignment of the
8584operation (that is, the alignment of the memory address). A value of 0
8585or an omitted ``align`` argument means that the operation has the ABI
8586alignment for the target. It is the responsibility of the code emitter
8587to ensure that the alignment information is correct. Overestimating the
8588alignment results in undefined behavior. Underestimating the
8589alignment may produce less efficient code. An alignment of 1 is always
8590safe. The maximum possible alignment is ``1 << 29``. An alignment
8591value higher than the size of the stored type implies memory up to the
8592alignment value bytes can be stored to without trapping in the default
8593address space. Storing to the higher bytes however may result in data
8594races if another thread can access the same address. Introducing a
8595data race is not allowed. Storing to the extra bytes is not allowed
8596even in situations where a data race is known to not exist if the
8597function has the ``sanitize_address`` attribute.
8598
8599The optional ``!nontemporal`` metadata must reference a single metadata
8600name ``<index>`` corresponding to a metadata node with one ``i32`` entry of
8601value 1. The existence of the ``!nontemporal`` metadata on the instruction
8602tells the optimizer and code generator that this load is not expected to
8603be reused in the cache. The code generator may select special
8604instructions to save cache bandwidth, such as the ``MOVNT`` instruction on
8605x86.
8606
8607The optional ``!invariant.group`` metadata must reference a
8608single metadata name ``<index>``. See ``invariant.group`` metadata.
8609
8610Semantics:
8611""""""""""
8612
8613The contents of memory are updated to contain ``<value>`` at the
8614location specified by the ``<pointer>`` operand. If ``<value>`` is
8615of scalar type then the number of bytes written does not exceed the
8616minimum number of bytes needed to hold all bits of the type. For
8617example, storing an ``i24`` writes at most three bytes. When writing a
8618value of a type like ``i20`` with a size that is not an integral number
8619of bytes, it is unspecified what happens to the extra bits that do not
8620belong to the type, but they will typically be overwritten.
8621
8622Example:
8623""""""""
8624
8625.. code-block:: llvm
8626
8627      %ptr = alloca i32                               ; yields i32*:ptr
8628      store i32 3, i32* %ptr                          ; yields void
8629      %val = load i32, i32* %ptr                      ; yields i32:val = i32 3
8630
8631.. _i_fence:
8632
8633'``fence``' Instruction
8634^^^^^^^^^^^^^^^^^^^^^^^
8635
8636Syntax:
8637"""""""
8638
8639::
8640
8641      fence [syncscope("<target-scope>")] <ordering>  ; yields void
8642
8643Overview:
8644"""""""""
8645
8646The '``fence``' instruction is used to introduce happens-before edges
8647between operations.
8648
8649Arguments:
8650""""""""""
8651
8652'``fence``' instructions take an :ref:`ordering <ordering>` argument which
8653defines what *synchronizes-with* edges they add. They can only be given
8654``acquire``, ``release``, ``acq_rel``, and ``seq_cst`` orderings.
8655
8656Semantics:
8657""""""""""
8658
8659A fence A which has (at least) ``release`` ordering semantics
8660*synchronizes with* a fence B with (at least) ``acquire`` ordering
8661semantics if and only if there exist atomic operations X and Y, both
8662operating on some atomic object M, such that A is sequenced before X, X
8663modifies M (either directly or through some side effect of a sequence
8664headed by X), Y is sequenced before B, and Y observes M. This provides a
8665*happens-before* dependency between A and B. Rather than an explicit
8666``fence``, one (but not both) of the atomic operations X or Y might
8667provide a ``release`` or ``acquire`` (resp.) ordering constraint and
8668still *synchronize-with* the explicit ``fence`` and establish the
8669*happens-before* edge.
8670
8671A ``fence`` which has ``seq_cst`` ordering, in addition to having both
8672``acquire`` and ``release`` semantics specified above, participates in
8673the global program order of other ``seq_cst`` operations and/or fences.
8674
8675A ``fence`` instruction can also take an optional
8676":ref:`syncscope <syncscope>`" argument.
8677
8678Example:
8679""""""""
8680
8681.. code-block:: text
8682
8683      fence acquire                                        ; yields void
8684      fence syncscope("singlethread") seq_cst              ; yields void
8685      fence syncscope("agent") seq_cst                     ; yields void
8686
8687.. _i_cmpxchg:
8688
8689'``cmpxchg``' Instruction
8690^^^^^^^^^^^^^^^^^^^^^^^^^
8691
8692Syntax:
8693"""""""
8694
8695::
8696
8697      cmpxchg [weak] [volatile] <ty>* <pointer>, <ty> <cmp>, <ty> <new> [syncscope("<target-scope>")] <success ordering> <failure ordering> ; yields  { ty, i1 }
8698
8699Overview:
8700"""""""""
8701
8702The '``cmpxchg``' instruction is used to atomically modify memory. It
8703loads a value in memory and compares it to a given value. If they are
8704equal, it tries to store a new value into the memory.
8705
8706Arguments:
8707""""""""""
8708
8709There are three arguments to the '``cmpxchg``' instruction: an address
8710to operate on, a value to compare to the value currently be at that
8711address, and a new value to place at that address if the compared values
8712are equal. The type of '<cmp>' must be an integer or pointer type whose
8713bit width is a power of two greater than or equal to eight and less
8714than or equal to a target-specific size limit. '<cmp>' and '<new>' must
8715have the same type, and the type of '<pointer>' must be a pointer to
8716that type. If the ``cmpxchg`` is marked as ``volatile``, then the
8717optimizer is not allowed to modify the number or order of execution of
8718this ``cmpxchg`` with other :ref:`volatile operations <volatile>`.
8719
8720The success and failure :ref:`ordering <ordering>` arguments specify how this
8721``cmpxchg`` synchronizes with other atomic operations. Both ordering parameters
8722must be at least ``monotonic``, the ordering constraint on failure must be no
8723stronger than that on success, and the failure ordering cannot be either
8724``release`` or ``acq_rel``.
8725
8726A ``cmpxchg`` instruction can also take an optional
8727":ref:`syncscope <syncscope>`" argument.
8728
8729The pointer passed into cmpxchg must have alignment greater than or
8730equal to the size in memory of the operand.
8731
8732Semantics:
8733""""""""""
8734
8735The contents of memory at the location specified by the '``<pointer>``' operand
8736is read and compared to '``<cmp>``'; if the values are equal, '``<new>``' is
8737written to the location. The original value at the location is returned,
8738together with a flag indicating success (true) or failure (false).
8739
8740If the cmpxchg operation is marked as ``weak`` then a spurious failure is
8741permitted: the operation may not write ``<new>`` even if the comparison
8742matched.
8743
8744If the cmpxchg operation is strong (the default), the i1 value is 1 if and only
8745if the value loaded equals ``cmp``.
8746
8747A successful ``cmpxchg`` is a read-modify-write instruction for the purpose of
8748identifying release sequences. A failed ``cmpxchg`` is equivalent to an atomic
8749load with an ordering parameter determined the second ordering parameter.
8750
8751Example:
8752""""""""
8753
8754.. code-block:: llvm
8755
8756    entry:
8757      %orig = load atomic i32, i32* %ptr unordered, align 4                      ; yields i32
8758      br label %loop
8759
8760    loop:
8761      %cmp = phi i32 [ %orig, %entry ], [%value_loaded, %loop]
8762      %squared = mul i32 %cmp, %cmp
8763      %val_success = cmpxchg i32* %ptr, i32 %cmp, i32 %squared acq_rel monotonic ; yields  { i32, i1 }
8764      %value_loaded = extractvalue { i32, i1 } %val_success, 0
8765      %success = extractvalue { i32, i1 } %val_success, 1
8766      br i1 %success, label %done, label %loop
8767
8768    done:
8769      ...
8770
8771.. _i_atomicrmw:
8772
8773'``atomicrmw``' Instruction
8774^^^^^^^^^^^^^^^^^^^^^^^^^^^
8775
8776Syntax:
8777"""""""
8778
8779::
8780
8781      atomicrmw [volatile] <operation> <ty>* <pointer>, <ty> <value> [syncscope("<target-scope>")] <ordering>                   ; yields ty
8782
8783Overview:
8784"""""""""
8785
8786The '``atomicrmw``' instruction is used to atomically modify memory.
8787
8788Arguments:
8789""""""""""
8790
8791There are three arguments to the '``atomicrmw``' instruction: an
8792operation to apply, an address whose value to modify, an argument to the
8793operation. The operation must be one of the following keywords:
8794
8795-  xchg
8796-  add
8797-  sub
8798-  and
8799-  nand
8800-  or
8801-  xor
8802-  max
8803-  min
8804-  umax
8805-  umin
8806-  fadd
8807-  fsub
8808
8809For most of these operations, the type of '<value>' must be an integer
8810type whose bit width is a power of two greater than or equal to eight
8811and less than or equal to a target-specific size limit. For xchg, this
8812may also be a floating point type with the same size constraints as
8813integers.  For fadd/fsub, this must be a floating point type.  The
8814type of the '``<pointer>``' operand must be a pointer to that type. If
8815the ``atomicrmw`` is marked as ``volatile``, then the optimizer is not
8816allowed to modify the number or order of execution of this
8817``atomicrmw`` with other :ref:`volatile operations <volatile>`.
8818
8819A ``atomicrmw`` instruction can also take an optional
8820":ref:`syncscope <syncscope>`" argument.
8821
8822Semantics:
8823""""""""""
8824
8825The contents of memory at the location specified by the '``<pointer>``'
8826operand are atomically read, modified, and written back. The original
8827value at the location is returned. The modification is specified by the
8828operation argument:
8829
8830-  xchg: ``*ptr = val``
8831-  add: ``*ptr = *ptr + val``
8832-  sub: ``*ptr = *ptr - val``
8833-  and: ``*ptr = *ptr & val``
8834-  nand: ``*ptr = ~(*ptr & val)``
8835-  or: ``*ptr = *ptr | val``
8836-  xor: ``*ptr = *ptr ^ val``
8837-  max: ``*ptr = *ptr > val ? *ptr : val`` (using a signed comparison)
8838-  min: ``*ptr = *ptr < val ? *ptr : val`` (using a signed comparison)
8839-  umax: ``*ptr = *ptr > val ? *ptr : val`` (using an unsigned
8840   comparison)
8841-  umin: ``*ptr = *ptr < val ? *ptr : val`` (using an unsigned
8842   comparison)
8843- fadd: ``*ptr = *ptr + val`` (using floating point arithmetic)
8844- fsub: ``*ptr = *ptr - val`` (using floating point arithmetic)
8845
8846Example:
8847""""""""
8848
8849.. code-block:: llvm
8850
8851      %old = atomicrmw add i32* %ptr, i32 1 acquire                        ; yields i32
8852
8853.. _i_getelementptr:
8854
8855'``getelementptr``' Instruction
8856^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8857
8858Syntax:
8859"""""""
8860
8861::
8862
8863      <result> = getelementptr <ty>, <ty>* <ptrval>{, [inrange] <ty> <idx>}*
8864      <result> = getelementptr inbounds <ty>, <ty>* <ptrval>{, [inrange] <ty> <idx>}*
8865      <result> = getelementptr <ty>, <ptr vector> <ptrval>, [inrange] <vector index type> <idx>
8866
8867Overview:
8868"""""""""
8869
8870The '``getelementptr``' instruction is used to get the address of a
8871subelement of an :ref:`aggregate <t_aggregate>` data structure. It performs
8872address calculation only and does not access memory. The instruction can also
8873be used to calculate a vector of such addresses.
8874
8875Arguments:
8876""""""""""
8877
8878The first argument is always a type used as the basis for the calculations.
8879The second argument is always a pointer or a vector of pointers, and is the
8880base address to start from. The remaining arguments are indices
8881that indicate which of the elements of the aggregate object are indexed.
8882The interpretation of each index is dependent on the type being indexed
8883into. The first index always indexes the pointer value given as the
8884second argument, the second index indexes a value of the type pointed to
8885(not necessarily the value directly pointed to, since the first index
8886can be non-zero), etc. The first type indexed into must be a pointer
8887value, subsequent types can be arrays, vectors, and structs. Note that
8888subsequent types being indexed into can never be pointers, since that
8889would require loading the pointer before continuing calculation.
8890
8891The type of each index argument depends on the type it is indexing into.
8892When indexing into a (optionally packed) structure, only ``i32`` integer
8893**constants** are allowed (when using a vector of indices they must all
8894be the **same** ``i32`` integer constant). When indexing into an array,
8895pointer or vector, integers of any width are allowed, and they are not
8896required to be constant. These integers are treated as signed values
8897where relevant.
8898
8899For example, let's consider a C code fragment and how it gets compiled
8900to LLVM:
8901
8902.. code-block:: c
8903
8904    struct RT {
8905      char A;
8906      int B[10][20];
8907      char C;
8908    };
8909    struct ST {
8910      int X;
8911      double Y;
8912      struct RT Z;
8913    };
8914
8915    int *foo(struct ST *s) {
8916      return &s[1].Z.B[5][13];
8917    }
8918
8919The LLVM code generated by Clang is:
8920
8921.. code-block:: llvm
8922
8923    %struct.RT = type { i8, [10 x [20 x i32]], i8 }
8924    %struct.ST = type { i32, double, %struct.RT }
8925
8926    define i32* @foo(%struct.ST* %s) nounwind uwtable readnone optsize ssp {
8927    entry:
8928      %arrayidx = getelementptr inbounds %struct.ST, %struct.ST* %s, i64 1, i32 2, i32 1, i64 5, i64 13
8929      ret i32* %arrayidx
8930    }
8931
8932Semantics:
8933""""""""""
8934
8935In the example above, the first index is indexing into the
8936'``%struct.ST*``' type, which is a pointer, yielding a '``%struct.ST``'
8937= '``{ i32, double, %struct.RT }``' type, a structure. The second index
8938indexes into the third element of the structure, yielding a
8939'``%struct.RT``' = '``{ i8 , [10 x [20 x i32]], i8 }``' type, another
8940structure. The third index indexes into the second element of the
8941structure, yielding a '``[10 x [20 x i32]]``' type, an array. The two
8942dimensions of the array are subscripted into, yielding an '``i32``'
8943type. The '``getelementptr``' instruction returns a pointer to this
8944element, thus computing a value of '``i32*``' type.
8945
8946Note that it is perfectly legal to index partially through a structure,
8947returning a pointer to an inner element. Because of this, the LLVM code
8948for the given testcase is equivalent to:
8949
8950.. code-block:: llvm
8951
8952    define i32* @foo(%struct.ST* %s) {
8953      %t1 = getelementptr %struct.ST, %struct.ST* %s, i32 1                        ; yields %struct.ST*:%t1
8954      %t2 = getelementptr %struct.ST, %struct.ST* %t1, i32 0, i32 2                ; yields %struct.RT*:%t2
8955      %t3 = getelementptr %struct.RT, %struct.RT* %t2, i32 0, i32 1                ; yields [10 x [20 x i32]]*:%t3
8956      %t4 = getelementptr [10 x [20 x i32]], [10 x [20 x i32]]* %t3, i32 0, i32 5  ; yields [20 x i32]*:%t4
8957      %t5 = getelementptr [20 x i32], [20 x i32]* %t4, i32 0, i32 13               ; yields i32*:%t5
8958      ret i32* %t5
8959    }
8960
8961If the ``inbounds`` keyword is present, the result value of the
8962``getelementptr`` is a :ref:`poison value <poisonvalues>` if the base
8963pointer is not an *in bounds* address of an allocated object, or if any
8964of the addresses that would be formed by successive addition of the
8965offsets implied by the indices to the base address with infinitely
8966precise signed arithmetic are not an *in bounds* address of that
8967allocated object. The *in bounds* addresses for an allocated object are
8968all the addresses that point into the object, plus the address one byte
8969past the end. The only *in bounds* address for a null pointer in the
8970default address-space is the null pointer itself. In cases where the
8971base is a vector of pointers the ``inbounds`` keyword applies to each
8972of the computations element-wise.
8973
8974If the ``inbounds`` keyword is not present, the offsets are added to the
8975base address with silently-wrapping two's complement arithmetic. If the
8976offsets have a different width from the pointer, they are sign-extended
8977or truncated to the width of the pointer. The result value of the
8978``getelementptr`` may be outside the object pointed to by the base
8979pointer. The result value may not necessarily be used to access memory
8980though, even if it happens to point into allocated storage. See the
8981:ref:`Pointer Aliasing Rules <pointeraliasing>` section for more
8982information.
8983
8984If the ``inrange`` keyword is present before any index, loading from or
8985storing to any pointer derived from the ``getelementptr`` has undefined
8986behavior if the load or store would access memory outside of the bounds of
8987the element selected by the index marked as ``inrange``. The result of a
8988pointer comparison or ``ptrtoint`` (including ``ptrtoint``-like operations
8989involving memory) involving a pointer derived from a ``getelementptr`` with
8990the ``inrange`` keyword is undefined, with the exception of comparisons
8991in the case where both operands are in the range of the element selected
8992by the ``inrange`` keyword, inclusive of the address one past the end of
8993that element. Note that the ``inrange`` keyword is currently only allowed
8994in constant ``getelementptr`` expressions.
8995
8996The getelementptr instruction is often confusing. For some more insight
8997into how it works, see :doc:`the getelementptr FAQ <GetElementPtr>`.
8998
8999Example:
9000""""""""
9001
9002.. code-block:: llvm
9003
9004        ; yields [12 x i8]*:aptr
9005        %aptr = getelementptr {i32, [12 x i8]}, {i32, [12 x i8]}* %saptr, i64 0, i32 1
9006        ; yields i8*:vptr
9007        %vptr = getelementptr {i32, <2 x i8>}, {i32, <2 x i8>}* %svptr, i64 0, i32 1, i32 1
9008        ; yields i8*:eptr
9009        %eptr = getelementptr [12 x i8], [12 x i8]* %aptr, i64 0, i32 1
9010        ; yields i32*:iptr
9011        %iptr = getelementptr [10 x i32], [10 x i32]* @arr, i16 0, i16 0
9012
9013Vector of pointers:
9014"""""""""""""""""""
9015
9016The ``getelementptr`` returns a vector of pointers, instead of a single address,
9017when one or more of its arguments is a vector. In such cases, all vector
9018arguments should have the same number of elements, and every scalar argument
9019will be effectively broadcast into a vector during address calculation.
9020
9021.. code-block:: llvm
9022
9023     ; All arguments are vectors:
9024     ;   A[i] = ptrs[i] + offsets[i]*sizeof(i8)
9025     %A = getelementptr i8, <4 x i8*> %ptrs, <4 x i64> %offsets
9026
9027     ; Add the same scalar offset to each pointer of a vector:
9028     ;   A[i] = ptrs[i] + offset*sizeof(i8)
9029     %A = getelementptr i8, <4 x i8*> %ptrs, i64 %offset
9030
9031     ; Add distinct offsets to the same pointer:
9032     ;   A[i] = ptr + offsets[i]*sizeof(i8)
9033     %A = getelementptr i8, i8* %ptr, <4 x i64> %offsets
9034
9035     ; In all cases described above the type of the result is <4 x i8*>
9036
9037The two following instructions are equivalent:
9038
9039.. code-block:: llvm
9040
9041     getelementptr  %struct.ST, <4 x %struct.ST*> %s, <4 x i64> %ind1,
9042       <4 x i32> <i32 2, i32 2, i32 2, i32 2>,
9043       <4 x i32> <i32 1, i32 1, i32 1, i32 1>,
9044       <4 x i32> %ind4,
9045       <4 x i64> <i64 13, i64 13, i64 13, i64 13>
9046
9047     getelementptr  %struct.ST, <4 x %struct.ST*> %s, <4 x i64> %ind1,
9048       i32 2, i32 1, <4 x i32> %ind4, i64 13
9049
9050Let's look at the C code, where the vector version of ``getelementptr``
9051makes sense:
9052
9053.. code-block:: c
9054
9055    // Let's assume that we vectorize the following loop:
9056    double *A, *B; int *C;
9057    for (int i = 0; i < size; ++i) {
9058      A[i] = B[C[i]];
9059    }
9060
9061.. code-block:: llvm
9062
9063    ; get pointers for 8 elements from array B
9064    %ptrs = getelementptr double, double* %B, <8 x i32> %C
9065    ; load 8 elements from array B into A
9066    %A = call <8 x double> @llvm.masked.gather.v8f64.v8p0f64(<8 x double*> %ptrs,
9067         i32 8, <8 x i1> %mask, <8 x double> %passthru)
9068
9069Conversion Operations
9070---------------------
9071
9072The instructions in this category are the conversion instructions
9073(casting) which all take a single operand and a type. They perform
9074various bit conversions on the operand.
9075
9076.. _i_trunc:
9077
9078'``trunc .. to``' Instruction
9079^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9080
9081Syntax:
9082"""""""
9083
9084::
9085
9086      <result> = trunc <ty> <value> to <ty2>             ; yields ty2
9087
9088Overview:
9089"""""""""
9090
9091The '``trunc``' instruction truncates its operand to the type ``ty2``.
9092
9093Arguments:
9094""""""""""
9095
9096The '``trunc``' instruction takes a value to trunc, and a type to trunc
9097it to. Both types must be of :ref:`integer <t_integer>` types, or vectors
9098of the same number of integers. The bit size of the ``value`` must be
9099larger than the bit size of the destination type, ``ty2``. Equal sized
9100types are not allowed.
9101
9102Semantics:
9103""""""""""
9104
9105The '``trunc``' instruction truncates the high order bits in ``value``
9106and converts the remaining bits to ``ty2``. Since the source size must
9107be larger than the destination size, ``trunc`` cannot be a *no-op cast*.
9108It will always truncate bits.
9109
9110Example:
9111""""""""
9112
9113.. code-block:: llvm
9114
9115      %X = trunc i32 257 to i8                        ; yields i8:1
9116      %Y = trunc i32 123 to i1                        ; yields i1:true
9117      %Z = trunc i32 122 to i1                        ; yields i1:false
9118      %W = trunc <2 x i16> <i16 8, i16 7> to <2 x i8> ; yields <i8 8, i8 7>
9119
9120.. _i_zext:
9121
9122'``zext .. to``' Instruction
9123^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9124
9125Syntax:
9126"""""""
9127
9128::
9129
9130      <result> = zext <ty> <value> to <ty2>             ; yields ty2
9131
9132Overview:
9133"""""""""
9134
9135The '``zext``' instruction zero extends its operand to type ``ty2``.
9136
9137Arguments:
9138""""""""""
9139
9140The '``zext``' instruction takes a value to cast, and a type to cast it
9141to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
9142the same number of integers. The bit size of the ``value`` must be
9143smaller than the bit size of the destination type, ``ty2``.
9144
9145Semantics:
9146""""""""""
9147
9148The ``zext`` fills the high order bits of the ``value`` with zero bits
9149until it reaches the size of the destination type, ``ty2``.
9150
9151When zero extending from i1, the result will always be either 0 or 1.
9152
9153Example:
9154""""""""
9155
9156.. code-block:: llvm
9157
9158      %X = zext i32 257 to i64              ; yields i64:257
9159      %Y = zext i1 true to i32              ; yields i32:1
9160      %Z = zext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
9161
9162.. _i_sext:
9163
9164'``sext .. to``' Instruction
9165^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9166
9167Syntax:
9168"""""""
9169
9170::
9171
9172      <result> = sext <ty> <value> to <ty2>             ; yields ty2
9173
9174Overview:
9175"""""""""
9176
9177The '``sext``' sign extends ``value`` to the type ``ty2``.
9178
9179Arguments:
9180""""""""""
9181
9182The '``sext``' instruction takes a value to cast, and a type to cast it
9183to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
9184the same number of integers. The bit size of the ``value`` must be
9185smaller than the bit size of the destination type, ``ty2``.
9186
9187Semantics:
9188""""""""""
9189
9190The '``sext``' instruction performs a sign extension by copying the sign
9191bit (highest order bit) of the ``value`` until it reaches the bit size
9192of the type ``ty2``.
9193
9194When sign extending from i1, the extension always results in -1 or 0.
9195
9196Example:
9197""""""""
9198
9199.. code-block:: llvm
9200
9201      %X = sext i8  -1 to i16              ; yields i16   :65535
9202      %Y = sext i1 true to i32             ; yields i32:-1
9203      %Z = sext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
9204
9205'``fptrunc .. to``' Instruction
9206^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9207
9208Syntax:
9209"""""""
9210
9211::
9212
9213      <result> = fptrunc <ty> <value> to <ty2>             ; yields ty2
9214
9215Overview:
9216"""""""""
9217
9218The '``fptrunc``' instruction truncates ``value`` to type ``ty2``.
9219
9220Arguments:
9221""""""""""
9222
9223The '``fptrunc``' instruction takes a :ref:`floating-point <t_floating>`
9224value to cast and a :ref:`floating-point <t_floating>` type to cast it to.
9225The size of ``value`` must be larger than the size of ``ty2``. This
9226implies that ``fptrunc`` cannot be used to make a *no-op cast*.
9227
9228Semantics:
9229""""""""""
9230
9231The '``fptrunc``' instruction casts a ``value`` from a larger
9232:ref:`floating-point <t_floating>` type to a smaller :ref:`floating-point
9233<t_floating>` type.
9234This instruction is assumed to execute in the default :ref:`floating-point
9235environment <floatenv>`.
9236
9237Example:
9238""""""""
9239
9240.. code-block:: llvm
9241
9242      %X = fptrunc double 16777217.0 to float    ; yields float:16777216.0
9243      %Y = fptrunc double 1.0E+300 to half       ; yields half:+infinity
9244
9245'``fpext .. to``' Instruction
9246^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9247
9248Syntax:
9249"""""""
9250
9251::
9252
9253      <result> = fpext <ty> <value> to <ty2>             ; yields ty2
9254
9255Overview:
9256"""""""""
9257
9258The '``fpext``' extends a floating-point ``value`` to a larger floating-point
9259value.
9260
9261Arguments:
9262""""""""""
9263
9264The '``fpext``' instruction takes a :ref:`floating-point <t_floating>`
9265``value`` to cast, and a :ref:`floating-point <t_floating>` type to cast it
9266to. The source type must be smaller than the destination type.
9267
9268Semantics:
9269""""""""""
9270
9271The '``fpext``' instruction extends the ``value`` from a smaller
9272:ref:`floating-point <t_floating>` type to a larger :ref:`floating-point
9273<t_floating>` type. The ``fpext`` cannot be used to make a
9274*no-op cast* because it always changes bits. Use ``bitcast`` to make a
9275*no-op cast* for a floating-point cast.
9276
9277Example:
9278""""""""
9279
9280.. code-block:: llvm
9281
9282      %X = fpext float 3.125 to double         ; yields double:3.125000e+00
9283      %Y = fpext double %X to fp128            ; yields fp128:0xL00000000000000004000900000000000
9284
9285'``fptoui .. to``' Instruction
9286^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9287
9288Syntax:
9289"""""""
9290
9291::
9292
9293      <result> = fptoui <ty> <value> to <ty2>             ; yields ty2
9294
9295Overview:
9296"""""""""
9297
9298The '``fptoui``' converts a floating-point ``value`` to its unsigned
9299integer equivalent of type ``ty2``.
9300
9301Arguments:
9302""""""""""
9303
9304The '``fptoui``' instruction takes a value to cast, which must be a
9305scalar or vector :ref:`floating-point <t_floating>` value, and a type to
9306cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
9307``ty`` is a vector floating-point type, ``ty2`` must be a vector integer
9308type with the same number of elements as ``ty``
9309
9310Semantics:
9311""""""""""
9312
9313The '``fptoui``' instruction converts its :ref:`floating-point
9314<t_floating>` operand into the nearest (rounding towards zero)
9315unsigned integer value. If the value cannot fit in ``ty2``, the result
9316is a :ref:`poison value <poisonvalues>`.
9317
9318Example:
9319""""""""
9320
9321.. code-block:: llvm
9322
9323      %X = fptoui double 123.0 to i32      ; yields i32:123
9324      %Y = fptoui float 1.0E+300 to i1     ; yields undefined:1
9325      %Z = fptoui float 1.04E+17 to i8     ; yields undefined:1
9326
9327'``fptosi .. to``' Instruction
9328^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9329
9330Syntax:
9331"""""""
9332
9333::
9334
9335      <result> = fptosi <ty> <value> to <ty2>             ; yields ty2
9336
9337Overview:
9338"""""""""
9339
9340The '``fptosi``' instruction converts :ref:`floating-point <t_floating>`
9341``value`` to type ``ty2``.
9342
9343Arguments:
9344""""""""""
9345
9346The '``fptosi``' instruction takes a value to cast, which must be a
9347scalar or vector :ref:`floating-point <t_floating>` value, and a type to
9348cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
9349``ty`` is a vector floating-point type, ``ty2`` must be a vector integer
9350type with the same number of elements as ``ty``
9351
9352Semantics:
9353""""""""""
9354
9355The '``fptosi``' instruction converts its :ref:`floating-point
9356<t_floating>` operand into the nearest (rounding towards zero)
9357signed integer value. If the value cannot fit in ``ty2``, the result
9358is a :ref:`poison value <poisonvalues>`.
9359
9360Example:
9361""""""""
9362
9363.. code-block:: llvm
9364
9365      %X = fptosi double -123.0 to i32      ; yields i32:-123
9366      %Y = fptosi float 1.0E-247 to i1      ; yields undefined:1
9367      %Z = fptosi float 1.04E+17 to i8      ; yields undefined:1
9368
9369'``uitofp .. to``' Instruction
9370^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9371
9372Syntax:
9373"""""""
9374
9375::
9376
9377      <result> = uitofp <ty> <value> to <ty2>             ; yields ty2
9378
9379Overview:
9380"""""""""
9381
9382The '``uitofp``' instruction regards ``value`` as an unsigned integer
9383and converts that value to the ``ty2`` type.
9384
9385Arguments:
9386""""""""""
9387
9388The '``uitofp``' instruction takes a value to cast, which must be a
9389scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
9390``ty2``, which must be an :ref:`floating-point <t_floating>` type. If
9391``ty`` is a vector integer type, ``ty2`` must be a vector floating-point
9392type with the same number of elements as ``ty``
9393
9394Semantics:
9395""""""""""
9396
9397The '``uitofp``' instruction interprets its operand as an unsigned
9398integer quantity and converts it to the corresponding floating-point
9399value. If the value cannot be exactly represented, it is rounded using
9400the default rounding mode.
9401
9402
9403Example:
9404""""""""
9405
9406.. code-block:: llvm
9407
9408      %X = uitofp i32 257 to float         ; yields float:257.0
9409      %Y = uitofp i8 -1 to double          ; yields double:255.0
9410
9411'``sitofp .. to``' Instruction
9412^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9413
9414Syntax:
9415"""""""
9416
9417::
9418
9419      <result> = sitofp <ty> <value> to <ty2>             ; yields ty2
9420
9421Overview:
9422"""""""""
9423
9424The '``sitofp``' instruction regards ``value`` as a signed integer and
9425converts that value to the ``ty2`` type.
9426
9427Arguments:
9428""""""""""
9429
9430The '``sitofp``' instruction takes a value to cast, which must be a
9431scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
9432``ty2``, which must be an :ref:`floating-point <t_floating>` type. If
9433``ty`` is a vector integer type, ``ty2`` must be a vector floating-point
9434type with the same number of elements as ``ty``
9435
9436Semantics:
9437""""""""""
9438
9439The '``sitofp``' instruction interprets its operand as a signed integer
9440quantity and converts it to the corresponding floating-point value. If the
9441value cannot be exactly represented, it is rounded using the default rounding
9442mode.
9443
9444Example:
9445""""""""
9446
9447.. code-block:: llvm
9448
9449      %X = sitofp i32 257 to float         ; yields float:257.0
9450      %Y = sitofp i8 -1 to double          ; yields double:-1.0
9451
9452.. _i_ptrtoint:
9453
9454'``ptrtoint .. to``' Instruction
9455^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9456
9457Syntax:
9458"""""""
9459
9460::
9461
9462      <result> = ptrtoint <ty> <value> to <ty2>             ; yields ty2
9463
9464Overview:
9465"""""""""
9466
9467The '``ptrtoint``' instruction converts the pointer or a vector of
9468pointers ``value`` to the integer (or vector of integers) type ``ty2``.
9469
9470Arguments:
9471""""""""""
9472
9473The '``ptrtoint``' instruction takes a ``value`` to cast, which must be
9474a value of type :ref:`pointer <t_pointer>` or a vector of pointers, and a
9475type to cast it to ``ty2``, which must be an :ref:`integer <t_integer>` or
9476a vector of integers type.
9477
9478Semantics:
9479""""""""""
9480
9481The '``ptrtoint``' instruction converts ``value`` to integer type
9482``ty2`` by interpreting the pointer value as an integer and either
9483truncating or zero extending that value to the size of the integer type.
9484If ``value`` is smaller than ``ty2`` then a zero extension is done. If
9485``value`` is larger than ``ty2`` then a truncation is done. If they are
9486the same size, then nothing is done (*no-op cast*) other than a type
9487change.
9488
9489Example:
9490""""""""
9491
9492.. code-block:: llvm
9493
9494      %X = ptrtoint i32* %P to i8                         ; yields truncation on 32-bit architecture
9495      %Y = ptrtoint i32* %P to i64                        ; yields zero extension on 32-bit architecture
9496      %Z = ptrtoint <4 x i32*> %P to <4 x i64>; yields vector zero extension for a vector of addresses on 32-bit architecture
9497
9498.. _i_inttoptr:
9499
9500'``inttoptr .. to``' Instruction
9501^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9502
9503Syntax:
9504"""""""
9505
9506::
9507
9508      <result> = inttoptr <ty> <value> to <ty2>             ; yields ty2
9509
9510Overview:
9511"""""""""
9512
9513The '``inttoptr``' instruction converts an integer ``value`` to a
9514pointer type, ``ty2``.
9515
9516Arguments:
9517""""""""""
9518
9519The '``inttoptr``' instruction takes an :ref:`integer <t_integer>` value to
9520cast, and a type to cast it to, which must be a :ref:`pointer <t_pointer>`
9521type.
9522
9523Semantics:
9524""""""""""
9525
9526The '``inttoptr``' instruction converts ``value`` to type ``ty2`` by
9527applying either a zero extension or a truncation depending on the size
9528of the integer ``value``. If ``value`` is larger than the size of a
9529pointer then a truncation is done. If ``value`` is smaller than the size
9530of a pointer then a zero extension is done. If they are the same size,
9531nothing is done (*no-op cast*).
9532
9533Example:
9534""""""""
9535
9536.. code-block:: llvm
9537
9538      %X = inttoptr i32 255 to i32*          ; yields zero extension on 64-bit architecture
9539      %Y = inttoptr i32 255 to i32*          ; yields no-op on 32-bit architecture
9540      %Z = inttoptr i64 0 to i32*            ; yields truncation on 32-bit architecture
9541      %Z = inttoptr <4 x i32> %G to <4 x i8*>; yields truncation of vector G to four pointers
9542
9543.. _i_bitcast:
9544
9545'``bitcast .. to``' Instruction
9546^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9547
9548Syntax:
9549"""""""
9550
9551::
9552
9553      <result> = bitcast <ty> <value> to <ty2>             ; yields ty2
9554
9555Overview:
9556"""""""""
9557
9558The '``bitcast``' instruction converts ``value`` to type ``ty2`` without
9559changing any bits.
9560
9561Arguments:
9562""""""""""
9563
9564The '``bitcast``' instruction takes a value to cast, which must be a
9565non-aggregate first class value, and a type to cast it to, which must
9566also be a non-aggregate :ref:`first class <t_firstclass>` type. The
9567bit sizes of ``value`` and the destination type, ``ty2``, must be
9568identical. If the source type is a pointer, the destination type must
9569also be a pointer of the same size. This instruction supports bitwise
9570conversion of vectors to integers and to vectors of other types (as
9571long as they have the same size).
9572
9573Semantics:
9574""""""""""
9575
9576The '``bitcast``' instruction converts ``value`` to type ``ty2``. It
9577is always a *no-op cast* because no bits change with this
9578conversion. The conversion is done as if the ``value`` had been stored
9579to memory and read back as type ``ty2``. Pointer (or vector of
9580pointers) types may only be converted to other pointer (or vector of
9581pointers) types with the same address space through this instruction.
9582To convert pointers to other types, use the :ref:`inttoptr <i_inttoptr>`
9583or :ref:`ptrtoint <i_ptrtoint>` instructions first.
9584
9585Example:
9586""""""""
9587
9588.. code-block:: text
9589
9590      %X = bitcast i8 255 to i8              ; yields i8 :-1
9591      %Y = bitcast i32* %x to sint*          ; yields sint*:%x
9592      %Z = bitcast <2 x int> %V to i64;        ; yields i64: %V
9593      %Z = bitcast <2 x i32*> %V to <2 x i64*> ; yields <2 x i64*>
9594
9595.. _i_addrspacecast:
9596
9597'``addrspacecast .. to``' Instruction
9598^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9599
9600Syntax:
9601"""""""
9602
9603::
9604
9605      <result> = addrspacecast <pty> <ptrval> to <pty2>       ; yields pty2
9606
9607Overview:
9608"""""""""
9609
9610The '``addrspacecast``' instruction converts ``ptrval`` from ``pty`` in
9611address space ``n`` to type ``pty2`` in address space ``m``.
9612
9613Arguments:
9614""""""""""
9615
9616The '``addrspacecast``' instruction takes a pointer or vector of pointer value
9617to cast and a pointer type to cast it to, which must have a different
9618address space.
9619
9620Semantics:
9621""""""""""
9622
9623The '``addrspacecast``' instruction converts the pointer value
9624``ptrval`` to type ``pty2``. It can be a *no-op cast* or a complex
9625value modification, depending on the target and the address space
9626pair. Pointer conversions within the same address space must be
9627performed with the ``bitcast`` instruction. Note that if the address space
9628conversion is legal then both result and operand refer to the same memory
9629location.
9630
9631Example:
9632""""""""
9633
9634.. code-block:: llvm
9635
9636      %X = addrspacecast i32* %x to i32 addrspace(1)*    ; yields i32 addrspace(1)*:%x
9637      %Y = addrspacecast i32 addrspace(1)* %y to i64 addrspace(2)*    ; yields i64 addrspace(2)*:%y
9638      %Z = addrspacecast <4 x i32*> %z to <4 x float addrspace(3)*>   ; yields <4 x float addrspace(3)*>:%z
9639
9640.. _otherops:
9641
9642Other Operations
9643----------------
9644
9645The instructions in this category are the "miscellaneous" instructions,
9646which defy better classification.
9647
9648.. _i_icmp:
9649
9650'``icmp``' Instruction
9651^^^^^^^^^^^^^^^^^^^^^^
9652
9653Syntax:
9654"""""""
9655
9656::
9657
9658      <result> = icmp <cond> <ty> <op1>, <op2>   ; yields i1 or <N x i1>:result
9659
9660Overview:
9661"""""""""
9662
9663The '``icmp``' instruction returns a boolean value or a vector of
9664boolean values based on comparison of its two integer, integer vector,
9665pointer, or pointer vector operands.
9666
9667Arguments:
9668""""""""""
9669
9670The '``icmp``' instruction takes three operands. The first operand is
9671the condition code indicating the kind of comparison to perform. It is
9672not a value, just a keyword. The possible condition codes are:
9673
9674#. ``eq``: equal
9675#. ``ne``: not equal
9676#. ``ugt``: unsigned greater than
9677#. ``uge``: unsigned greater or equal
9678#. ``ult``: unsigned less than
9679#. ``ule``: unsigned less or equal
9680#. ``sgt``: signed greater than
9681#. ``sge``: signed greater or equal
9682#. ``slt``: signed less than
9683#. ``sle``: signed less or equal
9684
9685The remaining two arguments must be :ref:`integer <t_integer>` or
9686:ref:`pointer <t_pointer>` or integer :ref:`vector <t_vector>` typed. They
9687must also be identical types.
9688
9689Semantics:
9690""""""""""
9691
9692The '``icmp``' compares ``op1`` and ``op2`` according to the condition
9693code given as ``cond``. The comparison performed always yields either an
9694:ref:`i1 <t_integer>` or vector of ``i1`` result, as follows:
9695
9696#. ``eq``: yields ``true`` if the operands are equal, ``false``
9697   otherwise. No sign interpretation is necessary or performed.
9698#. ``ne``: yields ``true`` if the operands are unequal, ``false``
9699   otherwise. No sign interpretation is necessary or performed.
9700#. ``ugt``: interprets the operands as unsigned values and yields
9701   ``true`` if ``op1`` is greater than ``op2``.
9702#. ``uge``: interprets the operands as unsigned values and yields
9703   ``true`` if ``op1`` is greater than or equal to ``op2``.
9704#. ``ult``: interprets the operands as unsigned values and yields
9705   ``true`` if ``op1`` is less than ``op2``.
9706#. ``ule``: interprets the operands as unsigned values and yields
9707   ``true`` if ``op1`` is less than or equal to ``op2``.
9708#. ``sgt``: interprets the operands as signed values and yields ``true``
9709   if ``op1`` is greater than ``op2``.
9710#. ``sge``: interprets the operands as signed values and yields ``true``
9711   if ``op1`` is greater than or equal to ``op2``.
9712#. ``slt``: interprets the operands as signed values and yields ``true``
9713   if ``op1`` is less than ``op2``.
9714#. ``sle``: interprets the operands as signed values and yields ``true``
9715   if ``op1`` is less than or equal to ``op2``.
9716
9717If the operands are :ref:`pointer <t_pointer>` typed, the pointer values
9718are compared as if they were integers.
9719
9720If the operands are integer vectors, then they are compared element by
9721element. The result is an ``i1`` vector with the same number of elements
9722as the values being compared. Otherwise, the result is an ``i1``.
9723
9724Example:
9725""""""""
9726
9727.. code-block:: text
9728
9729      <result> = icmp eq i32 4, 5          ; yields: result=false
9730      <result> = icmp ne float* %X, %X     ; yields: result=false
9731      <result> = icmp ult i16  4, 5        ; yields: result=true
9732      <result> = icmp sgt i16  4, 5        ; yields: result=false
9733      <result> = icmp ule i16 -4, 5        ; yields: result=false
9734      <result> = icmp sge i16  4, 5        ; yields: result=false
9735
9736.. _i_fcmp:
9737
9738'``fcmp``' Instruction
9739^^^^^^^^^^^^^^^^^^^^^^
9740
9741Syntax:
9742"""""""
9743
9744::
9745
9746      <result> = fcmp [fast-math flags]* <cond> <ty> <op1>, <op2>     ; yields i1 or <N x i1>:result
9747
9748Overview:
9749"""""""""
9750
9751The '``fcmp``' instruction returns a boolean value or vector of boolean
9752values based on comparison of its operands.
9753
9754If the operands are floating-point scalars, then the result type is a
9755boolean (:ref:`i1 <t_integer>`).
9756
9757If the operands are floating-point vectors, then the result type is a
9758vector of boolean with the same number of elements as the operands being
9759compared.
9760
9761Arguments:
9762""""""""""
9763
9764The '``fcmp``' instruction takes three operands. The first operand is
9765the condition code indicating the kind of comparison to perform. It is
9766not a value, just a keyword. The possible condition codes are:
9767
9768#. ``false``: no comparison, always returns false
9769#. ``oeq``: ordered and equal
9770#. ``ogt``: ordered and greater than
9771#. ``oge``: ordered and greater than or equal
9772#. ``olt``: ordered and less than
9773#. ``ole``: ordered and less than or equal
9774#. ``one``: ordered and not equal
9775#. ``ord``: ordered (no nans)
9776#. ``ueq``: unordered or equal
9777#. ``ugt``: unordered or greater than
9778#. ``uge``: unordered or greater than or equal
9779#. ``ult``: unordered or less than
9780#. ``ule``: unordered or less than or equal
9781#. ``une``: unordered or not equal
9782#. ``uno``: unordered (either nans)
9783#. ``true``: no comparison, always returns true
9784
9785*Ordered* means that neither operand is a QNAN while *unordered* means
9786that either operand may be a QNAN.
9787
9788Each of ``val1`` and ``val2`` arguments must be either a :ref:`floating-point
9789<t_floating>` type or a :ref:`vector <t_vector>` of floating-point type.
9790They must have identical types.
9791
9792Semantics:
9793""""""""""
9794
9795The '``fcmp``' instruction compares ``op1`` and ``op2`` according to the
9796condition code given as ``cond``. If the operands are vectors, then the
9797vectors are compared element by element. Each comparison performed
9798always yields an :ref:`i1 <t_integer>` result, as follows:
9799
9800#. ``false``: always yields ``false``, regardless of operands.
9801#. ``oeq``: yields ``true`` if both operands are not a QNAN and ``op1``
9802   is equal to ``op2``.
9803#. ``ogt``: yields ``true`` if both operands are not a QNAN and ``op1``
9804   is greater than ``op2``.
9805#. ``oge``: yields ``true`` if both operands are not a QNAN and ``op1``
9806   is greater than or equal to ``op2``.
9807#. ``olt``: yields ``true`` if both operands are not a QNAN and ``op1``
9808   is less than ``op2``.
9809#. ``ole``: yields ``true`` if both operands are not a QNAN and ``op1``
9810   is less than or equal to ``op2``.
9811#. ``one``: yields ``true`` if both operands are not a QNAN and ``op1``
9812   is not equal to ``op2``.
9813#. ``ord``: yields ``true`` if both operands are not a QNAN.
9814#. ``ueq``: yields ``true`` if either operand is a QNAN or ``op1`` is
9815   equal to ``op2``.
9816#. ``ugt``: yields ``true`` if either operand is a QNAN or ``op1`` is
9817   greater than ``op2``.
9818#. ``uge``: yields ``true`` if either operand is a QNAN or ``op1`` is
9819   greater than or equal to ``op2``.
9820#. ``ult``: yields ``true`` if either operand is a QNAN or ``op1`` is
9821   less than ``op2``.
9822#. ``ule``: yields ``true`` if either operand is a QNAN or ``op1`` is
9823   less than or equal to ``op2``.
9824#. ``une``: yields ``true`` if either operand is a QNAN or ``op1`` is
9825   not equal to ``op2``.
9826#. ``uno``: yields ``true`` if either operand is a QNAN.
9827#. ``true``: always yields ``true``, regardless of operands.
9828
9829The ``fcmp`` instruction can also optionally take any number of
9830:ref:`fast-math flags <fastmath>`, which are optimization hints to enable
9831otherwise unsafe floating-point optimizations.
9832
9833Any set of fast-math flags are legal on an ``fcmp`` instruction, but the
9834only flags that have any effect on its semantics are those that allow
9835assumptions to be made about the values of input arguments; namely
9836``nnan``, ``ninf``, and ``reassoc``. See :ref:`fastmath` for more information.
9837
9838Example:
9839""""""""
9840
9841.. code-block:: text
9842
9843      <result> = fcmp oeq float 4.0, 5.0    ; yields: result=false
9844      <result> = fcmp one float 4.0, 5.0    ; yields: result=true
9845      <result> = fcmp olt float 4.0, 5.0    ; yields: result=true
9846      <result> = fcmp ueq double 1.0, 2.0   ; yields: result=false
9847
9848.. _i_phi:
9849
9850'``phi``' Instruction
9851^^^^^^^^^^^^^^^^^^^^^
9852
9853Syntax:
9854"""""""
9855
9856::
9857
9858      <result> = phi <ty> [ <val0>, <label0>], ...
9859
9860Overview:
9861"""""""""
9862
9863The '``phi``' instruction is used to implement the φ node in the SSA
9864graph representing the function.
9865
9866Arguments:
9867""""""""""
9868
9869The type of the incoming values is specified with the first type field.
9870After this, the '``phi``' instruction takes a list of pairs as
9871arguments, with one pair for each predecessor basic block of the current
9872block. Only values of :ref:`first class <t_firstclass>` type may be used as
9873the value arguments to the PHI node. Only labels may be used as the
9874label arguments.
9875
9876There must be no non-phi instructions between the start of a basic block
9877and the PHI instructions: i.e. PHI instructions must be first in a basic
9878block.
9879
9880For the purposes of the SSA form, the use of each incoming value is
9881deemed to occur on the edge from the corresponding predecessor block to
9882the current block (but after any definition of an '``invoke``'
9883instruction's return value on the same edge).
9884
9885Semantics:
9886""""""""""
9887
9888At runtime, the '``phi``' instruction logically takes on the value
9889specified by the pair corresponding to the predecessor basic block that
9890executed just prior to the current block.
9891
9892Example:
9893""""""""
9894
9895.. code-block:: llvm
9896
9897    Loop:       ; Infinite loop that counts from 0 on up...
9898      %indvar = phi i32 [ 0, %LoopHeader ], [ %nextindvar, %Loop ]
9899      %nextindvar = add i32 %indvar, 1
9900      br label %Loop
9901
9902.. _i_select:
9903
9904'``select``' Instruction
9905^^^^^^^^^^^^^^^^^^^^^^^^
9906
9907Syntax:
9908"""""""
9909
9910::
9911
9912      <result> = select selty <cond>, <ty> <val1>, <ty> <val2>             ; yields ty
9913
9914      selty is either i1 or {<N x i1>}
9915
9916Overview:
9917"""""""""
9918
9919The '``select``' instruction is used to choose one value based on a
9920condition, without IR-level branching.
9921
9922Arguments:
9923""""""""""
9924
9925The '``select``' instruction requires an 'i1' value or a vector of 'i1'
9926values indicating the condition, and two values of the same :ref:`first
9927class <t_firstclass>` type.
9928
9929Semantics:
9930""""""""""
9931
9932If the condition is an i1 and it evaluates to 1, the instruction returns
9933the first value argument; otherwise, it returns the second value
9934argument.
9935
9936If the condition is a vector of i1, then the value arguments must be
9937vectors of the same size, and the selection is done element by element.
9938
9939If the condition is an i1 and the value arguments are vectors of the
9940same size, then an entire vector is selected.
9941
9942Example:
9943""""""""
9944
9945.. code-block:: llvm
9946
9947      %X = select i1 true, i8 17, i8 42          ; yields i8:17
9948
9949.. _i_call:
9950
9951'``call``' Instruction
9952^^^^^^^^^^^^^^^^^^^^^^
9953
9954Syntax:
9955"""""""
9956
9957::
9958
9959      <result> = [tail | musttail | notail ] call [fast-math flags] [cconv] [ret attrs] [addrspace(<num>)]
9960                 [<ty>|<fnty> <fnptrval>(<function args>) [fn attrs] [ operand bundles ]
9961
9962Overview:
9963"""""""""
9964
9965The '``call``' instruction represents a simple function call.
9966
9967Arguments:
9968""""""""""
9969
9970This instruction requires several arguments:
9971
9972#. The optional ``tail`` and ``musttail`` markers indicate that the optimizers
9973   should perform tail call optimization. The ``tail`` marker is a hint that
9974   `can be ignored <CodeGenerator.html#sibcallopt>`_. The ``musttail`` marker
9975   means that the call must be tail call optimized in order for the program to
9976   be correct. The ``musttail`` marker provides these guarantees:
9977
9978   #. The call will not cause unbounded stack growth if it is part of a
9979      recursive cycle in the call graph.
9980   #. Arguments with the :ref:`inalloca <attr_inalloca>` attribute are
9981      forwarded in place.
9982
9983   Both markers imply that the callee does not access allocas from the caller.
9984   The ``tail`` marker additionally implies that the callee does not access
9985   varargs from the caller, while ``musttail`` implies that varargs from the
9986   caller are passed to the callee. Calls marked ``musttail`` must obey the
9987   following additional  rules:
9988
9989   - The call must immediately precede a :ref:`ret <i_ret>` instruction,
9990     or a pointer bitcast followed by a ret instruction.
9991   - The ret instruction must return the (possibly bitcasted) value
9992     produced by the call or void.
9993   - The caller and callee prototypes must match. Pointer types of
9994     parameters or return types may differ in pointee type, but not
9995     in address space.
9996   - The calling conventions of the caller and callee must match.
9997   - All ABI-impacting function attributes, such as sret, byval, inreg,
9998     returned, and inalloca, must match.
9999   - The callee must be varargs iff the caller is varargs. Bitcasting a
10000     non-varargs function to the appropriate varargs type is legal so
10001     long as the non-varargs prefixes obey the other rules.
10002
10003   Tail call optimization for calls marked ``tail`` is guaranteed to occur if
10004   the following conditions are met:
10005
10006   -  Caller and callee both have the calling convention ``fastcc``.
10007   -  The call is in tail position (ret immediately follows call and ret
10008      uses value of call or is void).
10009   -  Option ``-tailcallopt`` is enabled, or
10010      ``llvm::GuaranteedTailCallOpt`` is ``true``.
10011   -  `Platform-specific constraints are
10012      met. <CodeGenerator.html#tailcallopt>`_
10013
10014#. The optional ``notail`` marker indicates that the optimizers should not add
10015   ``tail`` or ``musttail`` markers to the call. It is used to prevent tail
10016   call optimization from being performed on the call.
10017
10018#. The optional ``fast-math flags`` marker indicates that the call has one or more
10019   :ref:`fast-math flags <fastmath>`, which are optimization hints to enable
10020   otherwise unsafe floating-point optimizations. Fast-math flags are only valid
10021   for calls that return a floating-point scalar or vector type.
10022
10023#. The optional "cconv" marker indicates which :ref:`calling
10024   convention <callingconv>` the call should use. If none is
10025   specified, the call defaults to using C calling conventions. The
10026   calling convention of the call must match the calling convention of
10027   the target function, or else the behavior is undefined.
10028#. The optional :ref:`Parameter Attributes <paramattrs>` list for return
10029   values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
10030   are valid here.
10031#. The optional addrspace attribute can be used to indicate the address space
10032   of the called function. If it is not specified, the program address space
10033   from the :ref:`datalayout string<langref_datalayout>` will be used.
10034#. '``ty``': the type of the call instruction itself which is also the
10035   type of the return value. Functions that return no value are marked
10036   ``void``.
10037#. '``fnty``': shall be the signature of the function being called. The
10038   argument types must match the types implied by this signature. This
10039   type can be omitted if the function is not varargs.
10040#. '``fnptrval``': An LLVM value containing a pointer to a function to
10041   be called. In most cases, this is a direct function call, but
10042   indirect ``call``'s are just as possible, calling an arbitrary pointer
10043   to function value.
10044#. '``function args``': argument list whose types match the function
10045   signature argument types and parameter attributes. All arguments must
10046   be of :ref:`first class <t_firstclass>` type. If the function signature
10047   indicates the function accepts a variable number of arguments, the
10048   extra arguments can be specified.
10049#. The optional :ref:`function attributes <fnattrs>` list.
10050#. The optional :ref:`operand bundles <opbundles>` list.
10051
10052Semantics:
10053""""""""""
10054
10055The '``call``' instruction is used to cause control flow to transfer to
10056a specified function, with its incoming arguments bound to the specified
10057values. Upon a '``ret``' instruction in the called function, control
10058flow continues with the instruction after the function call, and the
10059return value of the function is bound to the result argument.
10060
10061Example:
10062""""""""
10063
10064.. code-block:: llvm
10065
10066      %retval = call i32 @test(i32 %argc)
10067      call i32 (i8*, ...)* @printf(i8* %msg, i32 12, i8 42)        ; yields i32
10068      %X = tail call i32 @foo()                                    ; yields i32
10069      %Y = tail call fastcc i32 @foo()  ; yields i32
10070      call void %foo(i8 97 signext)
10071
10072      %struct.A = type { i32, i8 }
10073      %r = call %struct.A @foo()                        ; yields { i32, i8 }
10074      %gr = extractvalue %struct.A %r, 0                ; yields i32
10075      %gr1 = extractvalue %struct.A %r, 1               ; yields i8
10076      %Z = call void @foo() noreturn                    ; indicates that %foo never returns normally
10077      %ZZ = call zeroext i32 @bar()                     ; Return value is %zero extended
10078
10079llvm treats calls to some functions with names and arguments that match
10080the standard C99 library as being the C99 library functions, and may
10081perform optimizations or generate code for them under that assumption.
10082This is something we'd like to change in the future to provide better
10083support for freestanding environments and non-C-based languages.
10084
10085.. _i_va_arg:
10086
10087'``va_arg``' Instruction
10088^^^^^^^^^^^^^^^^^^^^^^^^
10089
10090Syntax:
10091"""""""
10092
10093::
10094
10095      <resultval> = va_arg <va_list*> <arglist>, <argty>
10096
10097Overview:
10098"""""""""
10099
10100The '``va_arg``' instruction is used to access arguments passed through
10101the "variable argument" area of a function call. It is used to implement
10102the ``va_arg`` macro in C.
10103
10104Arguments:
10105""""""""""
10106
10107This instruction takes a ``va_list*`` value and the type of the
10108argument. It returns a value of the specified argument type and
10109increments the ``va_list`` to point to the next argument. The actual
10110type of ``va_list`` is target specific.
10111
10112Semantics:
10113""""""""""
10114
10115The '``va_arg``' instruction loads an argument of the specified type
10116from the specified ``va_list`` and causes the ``va_list`` to point to
10117the next argument. For more information, see the variable argument
10118handling :ref:`Intrinsic Functions <int_varargs>`.
10119
10120It is legal for this instruction to be called in a function which does
10121not take a variable number of arguments, for example, the ``vfprintf``
10122function.
10123
10124``va_arg`` is an LLVM instruction instead of an :ref:`intrinsic
10125function <intrinsics>` because it takes a type as an argument.
10126
10127Example:
10128""""""""
10129
10130See the :ref:`variable argument processing <int_varargs>` section.
10131
10132Note that the code generator does not yet fully support va\_arg on many
10133targets. Also, it does not currently support va\_arg with aggregate
10134types on any target.
10135
10136.. _i_landingpad:
10137
10138'``landingpad``' Instruction
10139^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10140
10141Syntax:
10142"""""""
10143
10144::
10145
10146      <resultval> = landingpad <resultty> <clause>+
10147      <resultval> = landingpad <resultty> cleanup <clause>*
10148
10149      <clause> := catch <type> <value>
10150      <clause> := filter <array constant type> <array constant>
10151
10152Overview:
10153"""""""""
10154
10155The '``landingpad``' instruction is used by `LLVM's exception handling
10156system <ExceptionHandling.html#overview>`_ to specify that a basic block
10157is a landing pad --- one where the exception lands, and corresponds to the
10158code found in the ``catch`` portion of a ``try``/``catch`` sequence. It
10159defines values supplied by the :ref:`personality function <personalityfn>` upon
10160re-entry to the function. The ``resultval`` has the type ``resultty``.
10161
10162Arguments:
10163""""""""""
10164
10165The optional
10166``cleanup`` flag indicates that the landing pad block is a cleanup.
10167
10168A ``clause`` begins with the clause type --- ``catch`` or ``filter`` --- and
10169contains the global variable representing the "type" that may be caught
10170or filtered respectively. Unlike the ``catch`` clause, the ``filter``
10171clause takes an array constant as its argument. Use
10172"``[0 x i8**] undef``" for a filter which cannot throw. The
10173'``landingpad``' instruction must contain *at least* one ``clause`` or
10174the ``cleanup`` flag.
10175
10176Semantics:
10177""""""""""
10178
10179The '``landingpad``' instruction defines the values which are set by the
10180:ref:`personality function <personalityfn>` upon re-entry to the function, and
10181therefore the "result type" of the ``landingpad`` instruction. As with
10182calling conventions, how the personality function results are
10183represented in LLVM IR is target specific.
10184
10185The clauses are applied in order from top to bottom. If two
10186``landingpad`` instructions are merged together through inlining, the
10187clauses from the calling function are appended to the list of clauses.
10188When the call stack is being unwound due to an exception being thrown,
10189the exception is compared against each ``clause`` in turn. If it doesn't
10190match any of the clauses, and the ``cleanup`` flag is not set, then
10191unwinding continues further up the call stack.
10192
10193The ``landingpad`` instruction has several restrictions:
10194
10195-  A landing pad block is a basic block which is the unwind destination
10196   of an '``invoke``' instruction.
10197-  A landing pad block must have a '``landingpad``' instruction as its
10198   first non-PHI instruction.
10199-  There can be only one '``landingpad``' instruction within the landing
10200   pad block.
10201-  A basic block that is not a landing pad block may not include a
10202   '``landingpad``' instruction.
10203
10204Example:
10205""""""""
10206
10207.. code-block:: llvm
10208
10209      ;; A landing pad which can catch an integer.
10210      %res = landingpad { i8*, i32 }
10211               catch i8** @_ZTIi
10212      ;; A landing pad that is a cleanup.
10213      %res = landingpad { i8*, i32 }
10214               cleanup
10215      ;; A landing pad which can catch an integer and can only throw a double.
10216      %res = landingpad { i8*, i32 }
10217               catch i8** @_ZTIi
10218               filter [1 x i8**] [@_ZTId]
10219
10220.. _i_catchpad:
10221
10222'``catchpad``' Instruction
10223^^^^^^^^^^^^^^^^^^^^^^^^^^
10224
10225Syntax:
10226"""""""
10227
10228::
10229
10230      <resultval> = catchpad within <catchswitch> [<args>*]
10231
10232Overview:
10233"""""""""
10234
10235The '``catchpad``' instruction is used by `LLVM's exception handling
10236system <ExceptionHandling.html#overview>`_ to specify that a basic block
10237begins a catch handler --- one where a personality routine attempts to transfer
10238control to catch an exception.
10239
10240Arguments:
10241""""""""""
10242
10243The ``catchswitch`` operand must always be a token produced by a
10244:ref:`catchswitch <i_catchswitch>` instruction in a predecessor block. This
10245ensures that each ``catchpad`` has exactly one predecessor block, and it always
10246terminates in a ``catchswitch``.
10247
10248The ``args`` correspond to whatever information the personality routine
10249requires to know if this is an appropriate handler for the exception. Control
10250will transfer to the ``catchpad`` if this is the first appropriate handler for
10251the exception.
10252
10253The ``resultval`` has the type :ref:`token <t_token>` and is used to match the
10254``catchpad`` to corresponding :ref:`catchrets <i_catchret>` and other nested EH
10255pads.
10256
10257Semantics:
10258""""""""""
10259
10260When the call stack is being unwound due to an exception being thrown, the
10261exception is compared against the ``args``. If it doesn't match, control will
10262not reach the ``catchpad`` instruction.  The representation of ``args`` is
10263entirely target and personality function-specific.
10264
10265Like the :ref:`landingpad <i_landingpad>` instruction, the ``catchpad``
10266instruction must be the first non-phi of its parent basic block.
10267
10268The meaning of the tokens produced and consumed by ``catchpad`` and other "pad"
10269instructions is described in the
10270`Windows exception handling documentation\ <ExceptionHandling.html#wineh>`_.
10271
10272When a ``catchpad`` has been "entered" but not yet "exited" (as
10273described in the `EH documentation\ <ExceptionHandling.html#wineh-constraints>`_),
10274it is undefined behavior to execute a :ref:`call <i_call>` or :ref:`invoke <i_invoke>`
10275that does not carry an appropriate :ref:`"funclet" bundle <ob_funclet>`.
10276
10277Example:
10278""""""""
10279
10280.. code-block:: text
10281
10282    dispatch:
10283      %cs = catchswitch within none [label %handler0] unwind to caller
10284      ;; A catch block which can catch an integer.
10285    handler0:
10286      %tok = catchpad within %cs [i8** @_ZTIi]
10287
10288.. _i_cleanuppad:
10289
10290'``cleanuppad``' Instruction
10291^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10292
10293Syntax:
10294"""""""
10295
10296::
10297
10298      <resultval> = cleanuppad within <parent> [<args>*]
10299
10300Overview:
10301"""""""""
10302
10303The '``cleanuppad``' instruction is used by `LLVM's exception handling
10304system <ExceptionHandling.html#overview>`_ to specify that a basic block
10305is a cleanup block --- one where a personality routine attempts to
10306transfer control to run cleanup actions.
10307The ``args`` correspond to whatever additional
10308information the :ref:`personality function <personalityfn>` requires to
10309execute the cleanup.
10310The ``resultval`` has the type :ref:`token <t_token>` and is used to
10311match the ``cleanuppad`` to corresponding :ref:`cleanuprets <i_cleanupret>`.
10312The ``parent`` argument is the token of the funclet that contains the
10313``cleanuppad`` instruction. If the ``cleanuppad`` is not inside a funclet,
10314this operand may be the token ``none``.
10315
10316Arguments:
10317""""""""""
10318
10319The instruction takes a list of arbitrary values which are interpreted
10320by the :ref:`personality function <personalityfn>`.
10321
10322Semantics:
10323""""""""""
10324
10325When the call stack is being unwound due to an exception being thrown,
10326the :ref:`personality function <personalityfn>` transfers control to the
10327``cleanuppad`` with the aid of the personality-specific arguments.
10328As with calling conventions, how the personality function results are
10329represented in LLVM IR is target specific.
10330
10331The ``cleanuppad`` instruction has several restrictions:
10332
10333-  A cleanup block is a basic block which is the unwind destination of
10334   an exceptional instruction.
10335-  A cleanup block must have a '``cleanuppad``' instruction as its
10336   first non-PHI instruction.
10337-  There can be only one '``cleanuppad``' instruction within the
10338   cleanup block.
10339-  A basic block that is not a cleanup block may not include a
10340   '``cleanuppad``' instruction.
10341
10342When a ``cleanuppad`` has been "entered" but not yet "exited" (as
10343described in the `EH documentation\ <ExceptionHandling.html#wineh-constraints>`_),
10344it is undefined behavior to execute a :ref:`call <i_call>` or :ref:`invoke <i_invoke>`
10345that does not carry an appropriate :ref:`"funclet" bundle <ob_funclet>`.
10346
10347Example:
10348""""""""
10349
10350.. code-block:: text
10351
10352      %tok = cleanuppad within %cs []
10353
10354.. _intrinsics:
10355
10356Intrinsic Functions
10357===================
10358
10359LLVM supports the notion of an "intrinsic function". These functions
10360have well known names and semantics and are required to follow certain
10361restrictions. Overall, these intrinsics represent an extension mechanism
10362for the LLVM language that does not require changing all of the
10363transformations in LLVM when adding to the language (or the bitcode
10364reader/writer, the parser, etc...).
10365
10366Intrinsic function names must all start with an "``llvm.``" prefix. This
10367prefix is reserved in LLVM for intrinsic names; thus, function names may
10368not begin with this prefix. Intrinsic functions must always be external
10369functions: you cannot define the body of intrinsic functions. Intrinsic
10370functions may only be used in call or invoke instructions: it is illegal
10371to take the address of an intrinsic function. Additionally, because
10372intrinsic functions are part of the LLVM language, it is required if any
10373are added that they be documented here.
10374
10375Some intrinsic functions can be overloaded, i.e., the intrinsic
10376represents a family of functions that perform the same operation but on
10377different data types. Because LLVM can represent over 8 million
10378different integer types, overloading is used commonly to allow an
10379intrinsic function to operate on any integer type. One or more of the
10380argument types or the result type can be overloaded to accept any
10381integer type. Argument types may also be defined as exactly matching a
10382previous argument's type or the result type. This allows an intrinsic
10383function which accepts multiple arguments, but needs all of them to be
10384of the same type, to only be overloaded with respect to a single
10385argument or the result.
10386
10387Overloaded intrinsics will have the names of its overloaded argument
10388types encoded into its function name, each preceded by a period. Only
10389those types which are overloaded result in a name suffix. Arguments
10390whose type is matched against another type do not. For example, the
10391``llvm.ctpop`` function can take an integer of any width and returns an
10392integer of exactly the same integer width. This leads to a family of
10393functions such as ``i8 @llvm.ctpop.i8(i8 %val)`` and
10394``i29 @llvm.ctpop.i29(i29 %val)``. Only one type, the return type, is
10395overloaded, and only one type suffix is required. Because the argument's
10396type is matched against the return type, it does not require its own
10397name suffix.
10398
10399To learn how to add an intrinsic function, please see the `Extending
10400LLVM Guide <ExtendingLLVM.html>`_.
10401
10402.. _int_varargs:
10403
10404Variable Argument Handling Intrinsics
10405-------------------------------------
10406
10407Variable argument support is defined in LLVM with the
10408:ref:`va_arg <i_va_arg>` instruction and these three intrinsic
10409functions. These functions are related to the similarly named macros
10410defined in the ``<stdarg.h>`` header file.
10411
10412All of these functions operate on arguments that use a target-specific
10413value type "``va_list``". The LLVM assembly language reference manual
10414does not define what this type is, so all transformations should be
10415prepared to handle these functions regardless of the type used.
10416
10417This example shows how the :ref:`va_arg <i_va_arg>` instruction and the
10418variable argument handling intrinsic functions are used.
10419
10420.. code-block:: llvm
10421
10422    ; This struct is different for every platform. For most platforms,
10423    ; it is merely an i8*.
10424    %struct.va_list = type { i8* }
10425
10426    ; For Unix x86_64 platforms, va_list is the following struct:
10427    ; %struct.va_list = type { i32, i32, i8*, i8* }
10428
10429    define i32 @test(i32 %X, ...) {
10430      ; Initialize variable argument processing
10431      %ap = alloca %struct.va_list
10432      %ap2 = bitcast %struct.va_list* %ap to i8*
10433      call void @llvm.va_start(i8* %ap2)
10434
10435      ; Read a single integer argument
10436      %tmp = va_arg i8* %ap2, i32
10437
10438      ; Demonstrate usage of llvm.va_copy and llvm.va_end
10439      %aq = alloca i8*
10440      %aq2 = bitcast i8** %aq to i8*
10441      call void @llvm.va_copy(i8* %aq2, i8* %ap2)
10442      call void @llvm.va_end(i8* %aq2)
10443
10444      ; Stop processing of arguments.
10445      call void @llvm.va_end(i8* %ap2)
10446      ret i32 %tmp
10447    }
10448
10449    declare void @llvm.va_start(i8*)
10450    declare void @llvm.va_copy(i8*, i8*)
10451    declare void @llvm.va_end(i8*)
10452
10453.. _int_va_start:
10454
10455'``llvm.va_start``' Intrinsic
10456^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10457
10458Syntax:
10459"""""""
10460
10461::
10462
10463      declare void @llvm.va_start(i8* <arglist>)
10464
10465Overview:
10466"""""""""
10467
10468The '``llvm.va_start``' intrinsic initializes ``*<arglist>`` for
10469subsequent use by ``va_arg``.
10470
10471Arguments:
10472""""""""""
10473
10474The argument is a pointer to a ``va_list`` element to initialize.
10475
10476Semantics:
10477""""""""""
10478
10479The '``llvm.va_start``' intrinsic works just like the ``va_start`` macro
10480available in C. In a target-dependent way, it initializes the
10481``va_list`` element to which the argument points, so that the next call
10482to ``va_arg`` will produce the first variable argument passed to the
10483function. Unlike the C ``va_start`` macro, this intrinsic does not need
10484to know the last argument of the function as the compiler can figure
10485that out.
10486
10487'``llvm.va_end``' Intrinsic
10488^^^^^^^^^^^^^^^^^^^^^^^^^^^
10489
10490Syntax:
10491"""""""
10492
10493::
10494
10495      declare void @llvm.va_end(i8* <arglist>)
10496
10497Overview:
10498"""""""""
10499
10500The '``llvm.va_end``' intrinsic destroys ``*<arglist>``, which has been
10501initialized previously with ``llvm.va_start`` or ``llvm.va_copy``.
10502
10503Arguments:
10504""""""""""
10505
10506The argument is a pointer to a ``va_list`` to destroy.
10507
10508Semantics:
10509""""""""""
10510
10511The '``llvm.va_end``' intrinsic works just like the ``va_end`` macro
10512available in C. In a target-dependent way, it destroys the ``va_list``
10513element to which the argument points. Calls to
10514:ref:`llvm.va_start <int_va_start>` and
10515:ref:`llvm.va_copy <int_va_copy>` must be matched exactly with calls to
10516``llvm.va_end``.
10517
10518.. _int_va_copy:
10519
10520'``llvm.va_copy``' Intrinsic
10521^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10522
10523Syntax:
10524"""""""
10525
10526::
10527
10528      declare void @llvm.va_copy(i8* <destarglist>, i8* <srcarglist>)
10529
10530Overview:
10531"""""""""
10532
10533The '``llvm.va_copy``' intrinsic copies the current argument position
10534from the source argument list to the destination argument list.
10535
10536Arguments:
10537""""""""""
10538
10539The first argument is a pointer to a ``va_list`` element to initialize.
10540The second argument is a pointer to a ``va_list`` element to copy from.
10541
10542Semantics:
10543""""""""""
10544
10545The '``llvm.va_copy``' intrinsic works just like the ``va_copy`` macro
10546available in C. In a target-dependent way, it copies the source
10547``va_list`` element into the destination ``va_list`` element. This
10548intrinsic is necessary because the `` llvm.va_start`` intrinsic may be
10549arbitrarily complex and require, for example, memory allocation.
10550
10551Accurate Garbage Collection Intrinsics
10552--------------------------------------
10553
10554LLVM's support for `Accurate Garbage Collection <GarbageCollection.html>`_
10555(GC) requires the frontend to generate code containing appropriate intrinsic
10556calls and select an appropriate GC strategy which knows how to lower these
10557intrinsics in a manner which is appropriate for the target collector.
10558
10559These intrinsics allow identification of :ref:`GC roots on the
10560stack <int_gcroot>`, as well as garbage collector implementations that
10561require :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers.
10562Frontends for type-safe garbage collected languages should generate
10563these intrinsics to make use of the LLVM garbage collectors. For more
10564details, see `Garbage Collection with LLVM <GarbageCollection.html>`_.
10565
10566Experimental Statepoint Intrinsics
10567^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10568
10569LLVM provides an second experimental set of intrinsics for describing garbage
10570collection safepoints in compiled code. These intrinsics are an alternative
10571to the ``llvm.gcroot`` intrinsics, but are compatible with the ones for
10572:ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers. The
10573differences in approach are covered in the `Garbage Collection with LLVM
10574<GarbageCollection.html>`_ documentation. The intrinsics themselves are
10575described in :doc:`Statepoints`.
10576
10577.. _int_gcroot:
10578
10579'``llvm.gcroot``' Intrinsic
10580^^^^^^^^^^^^^^^^^^^^^^^^^^^
10581
10582Syntax:
10583"""""""
10584
10585::
10586
10587      declare void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
10588
10589Overview:
10590"""""""""
10591
10592The '``llvm.gcroot``' intrinsic declares the existence of a GC root to
10593the code generator, and allows some metadata to be associated with it.
10594
10595Arguments:
10596""""""""""
10597
10598The first argument specifies the address of a stack object that contains
10599the root pointer. The second pointer (which must be either a constant or
10600a global value address) contains the meta-data to be associated with the
10601root.
10602
10603Semantics:
10604""""""""""
10605
10606At runtime, a call to this intrinsic stores a null pointer into the
10607"ptrloc" location. At compile-time, the code generator generates
10608information to allow the runtime to find the pointer at GC safe points.
10609The '``llvm.gcroot``' intrinsic may only be used in a function which
10610:ref:`specifies a GC algorithm <gc>`.
10611
10612.. _int_gcread:
10613
10614'``llvm.gcread``' Intrinsic
10615^^^^^^^^^^^^^^^^^^^^^^^^^^^
10616
10617Syntax:
10618"""""""
10619
10620::
10621
10622      declare i8* @llvm.gcread(i8* %ObjPtr, i8** %Ptr)
10623
10624Overview:
10625"""""""""
10626
10627The '``llvm.gcread``' intrinsic identifies reads of references from heap
10628locations, allowing garbage collector implementations that require read
10629barriers.
10630
10631Arguments:
10632""""""""""
10633
10634The second argument is the address to read from, which should be an
10635address allocated from the garbage collector. The first object is a
10636pointer to the start of the referenced object, if needed by the language
10637runtime (otherwise null).
10638
10639Semantics:
10640""""""""""
10641
10642The '``llvm.gcread``' intrinsic has the same semantics as a load
10643instruction, but may be replaced with substantially more complex code by
10644the garbage collector runtime, as needed. The '``llvm.gcread``'
10645intrinsic may only be used in a function which :ref:`specifies a GC
10646algorithm <gc>`.
10647
10648.. _int_gcwrite:
10649
10650'``llvm.gcwrite``' Intrinsic
10651^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10652
10653Syntax:
10654"""""""
10655
10656::
10657
10658      declare void @llvm.gcwrite(i8* %P1, i8* %Obj, i8** %P2)
10659
10660Overview:
10661"""""""""
10662
10663The '``llvm.gcwrite``' intrinsic identifies writes of references to heap
10664locations, allowing garbage collector implementations that require write
10665barriers (such as generational or reference counting collectors).
10666
10667Arguments:
10668""""""""""
10669
10670The first argument is the reference to store, the second is the start of
10671the object to store it to, and the third is the address of the field of
10672Obj to store to. If the runtime does not require a pointer to the
10673object, Obj may be null.
10674
10675Semantics:
10676""""""""""
10677
10678The '``llvm.gcwrite``' intrinsic has the same semantics as a store
10679instruction, but may be replaced with substantially more complex code by
10680the garbage collector runtime, as needed. The '``llvm.gcwrite``'
10681intrinsic may only be used in a function which :ref:`specifies a GC
10682algorithm <gc>`.
10683
10684Code Generator Intrinsics
10685-------------------------
10686
10687These intrinsics are provided by LLVM to expose special features that
10688may only be implemented with code generator support.
10689
10690'``llvm.returnaddress``' Intrinsic
10691^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10692
10693Syntax:
10694"""""""
10695
10696::
10697
10698      declare i8* @llvm.returnaddress(i32 <level>)
10699
10700Overview:
10701"""""""""
10702
10703The '``llvm.returnaddress``' intrinsic attempts to compute a
10704target-specific value indicating the return address of the current
10705function or one of its callers.
10706
10707Arguments:
10708""""""""""
10709
10710The argument to this intrinsic indicates which function to return the
10711address for. Zero indicates the calling function, one indicates its
10712caller, etc. The argument is **required** to be a constant integer
10713value.
10714
10715Semantics:
10716""""""""""
10717
10718The '``llvm.returnaddress``' intrinsic either returns a pointer
10719indicating the return address of the specified call frame, or zero if it
10720cannot be identified. The value returned by this intrinsic is likely to
10721be incorrect or 0 for arguments other than zero, so it should only be
10722used for debugging purposes.
10723
10724Note that calling this intrinsic does not prevent function inlining or
10725other aggressive transformations, so the value returned may not be that
10726of the obvious source-language caller.
10727
10728'``llvm.addressofreturnaddress``' Intrinsic
10729^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10730
10731Syntax:
10732"""""""
10733
10734::
10735
10736      declare i8* @llvm.addressofreturnaddress()
10737
10738Overview:
10739"""""""""
10740
10741The '``llvm.addressofreturnaddress``' intrinsic returns a target-specific
10742pointer to the place in the stack frame where the return address of the
10743current function is stored.
10744
10745Semantics:
10746""""""""""
10747
10748Note that calling this intrinsic does not prevent function inlining or
10749other aggressive transformations, so the value returned may not be that
10750of the obvious source-language caller.
10751
10752This intrinsic is only implemented for x86 and aarch64.
10753
10754'``llvm.sponentry``' Intrinsic
10755^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10756
10757Syntax:
10758"""""""
10759
10760::
10761
10762      declare i8* @llvm.sponentry()
10763
10764Overview:
10765"""""""""
10766
10767The '``llvm.sponentry``' intrinsic returns the stack pointer value at
10768the entry of the current function calling this intrinsic.
10769
10770Semantics:
10771""""""""""
10772
10773Note this intrinsic is only verified on AArch64.
10774
10775'``llvm.frameaddress``' Intrinsic
10776^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10777
10778Syntax:
10779"""""""
10780
10781::
10782
10783      declare i8* @llvm.frameaddress(i32 <level>)
10784
10785Overview:
10786"""""""""
10787
10788The '``llvm.frameaddress``' intrinsic attempts to return the
10789target-specific frame pointer value for the specified stack frame.
10790
10791Arguments:
10792""""""""""
10793
10794The argument to this intrinsic indicates which function to return the
10795frame pointer for. Zero indicates the calling function, one indicates
10796its caller, etc. The argument is **required** to be a constant integer
10797value.
10798
10799Semantics:
10800""""""""""
10801
10802The '``llvm.frameaddress``' intrinsic either returns a pointer
10803indicating the frame address of the specified call frame, or zero if it
10804cannot be identified. The value returned by this intrinsic is likely to
10805be incorrect or 0 for arguments other than zero, so it should only be
10806used for debugging purposes.
10807
10808Note that calling this intrinsic does not prevent function inlining or
10809other aggressive transformations, so the value returned may not be that
10810of the obvious source-language caller.
10811
10812'``llvm.localescape``' and '``llvm.localrecover``' Intrinsics
10813^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10814
10815Syntax:
10816"""""""
10817
10818::
10819
10820      declare void @llvm.localescape(...)
10821      declare i8* @llvm.localrecover(i8* %func, i8* %fp, i32 %idx)
10822
10823Overview:
10824"""""""""
10825
10826The '``llvm.localescape``' intrinsic escapes offsets of a collection of static
10827allocas, and the '``llvm.localrecover``' intrinsic applies those offsets to a
10828live frame pointer to recover the address of the allocation. The offset is
10829computed during frame layout of the caller of ``llvm.localescape``.
10830
10831Arguments:
10832""""""""""
10833
10834All arguments to '``llvm.localescape``' must be pointers to static allocas or
10835casts of static allocas. Each function can only call '``llvm.localescape``'
10836once, and it can only do so from the entry block.
10837
10838The ``func`` argument to '``llvm.localrecover``' must be a constant
10839bitcasted pointer to a function defined in the current module. The code
10840generator cannot determine the frame allocation offset of functions defined in
10841other modules.
10842
10843The ``fp`` argument to '``llvm.localrecover``' must be a frame pointer of a
10844call frame that is currently live. The return value of '``llvm.localaddress``'
10845is one way to produce such a value, but various runtimes also expose a suitable
10846pointer in platform-specific ways.
10847
10848The ``idx`` argument to '``llvm.localrecover``' indicates which alloca passed to
10849'``llvm.localescape``' to recover. It is zero-indexed.
10850
10851Semantics:
10852""""""""""
10853
10854These intrinsics allow a group of functions to share access to a set of local
10855stack allocations of a one parent function. The parent function may call the
10856'``llvm.localescape``' intrinsic once from the function entry block, and the
10857child functions can use '``llvm.localrecover``' to access the escaped allocas.
10858The '``llvm.localescape``' intrinsic blocks inlining, as inlining changes where
10859the escaped allocas are allocated, which would break attempts to use
10860'``llvm.localrecover``'.
10861
10862.. _int_read_register:
10863.. _int_write_register:
10864
10865'``llvm.read_register``' and '``llvm.write_register``' Intrinsics
10866^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10867
10868Syntax:
10869"""""""
10870
10871::
10872
10873      declare i32 @llvm.read_register.i32(metadata)
10874      declare i64 @llvm.read_register.i64(metadata)
10875      declare void @llvm.write_register.i32(metadata, i32 @value)
10876      declare void @llvm.write_register.i64(metadata, i64 @value)
10877      !0 = !{!"sp\00"}
10878
10879Overview:
10880"""""""""
10881
10882The '``llvm.read_register``' and '``llvm.write_register``' intrinsics
10883provides access to the named register. The register must be valid on
10884the architecture being compiled to. The type needs to be compatible
10885with the register being read.
10886
10887Semantics:
10888""""""""""
10889
10890The '``llvm.read_register``' intrinsic returns the current value of the
10891register, where possible. The '``llvm.write_register``' intrinsic sets
10892the current value of the register, where possible.
10893
10894This is useful to implement named register global variables that need
10895to always be mapped to a specific register, as is common practice on
10896bare-metal programs including OS kernels.
10897
10898The compiler doesn't check for register availability or use of the used
10899register in surrounding code, including inline assembly. Because of that,
10900allocatable registers are not supported.
10901
10902Warning: So far it only works with the stack pointer on selected
10903architectures (ARM, AArch64, PowerPC and x86_64). Significant amount of
10904work is needed to support other registers and even more so, allocatable
10905registers.
10906
10907.. _int_stacksave:
10908
10909'``llvm.stacksave``' Intrinsic
10910^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10911
10912Syntax:
10913"""""""
10914
10915::
10916
10917      declare i8* @llvm.stacksave()
10918
10919Overview:
10920"""""""""
10921
10922The '``llvm.stacksave``' intrinsic is used to remember the current state
10923of the function stack, for use with
10924:ref:`llvm.stackrestore <int_stackrestore>`. This is useful for
10925implementing language features like scoped automatic variable sized
10926arrays in C99.
10927
10928Semantics:
10929""""""""""
10930
10931This intrinsic returns a opaque pointer value that can be passed to
10932:ref:`llvm.stackrestore <int_stackrestore>`. When an
10933``llvm.stackrestore`` intrinsic is executed with a value saved from
10934``llvm.stacksave``, it effectively restores the state of the stack to
10935the state it was in when the ``llvm.stacksave`` intrinsic executed. In
10936practice, this pops any :ref:`alloca <i_alloca>` blocks from the stack that
10937were allocated after the ``llvm.stacksave`` was executed.
10938
10939.. _int_stackrestore:
10940
10941'``llvm.stackrestore``' Intrinsic
10942^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10943
10944Syntax:
10945"""""""
10946
10947::
10948
10949      declare void @llvm.stackrestore(i8* %ptr)
10950
10951Overview:
10952"""""""""
10953
10954The '``llvm.stackrestore``' intrinsic is used to restore the state of
10955the function stack to the state it was in when the corresponding
10956:ref:`llvm.stacksave <int_stacksave>` intrinsic executed. This is
10957useful for implementing language features like scoped automatic variable
10958sized arrays in C99.
10959
10960Semantics:
10961""""""""""
10962
10963See the description for :ref:`llvm.stacksave <int_stacksave>`.
10964
10965.. _int_get_dynamic_area_offset:
10966
10967'``llvm.get.dynamic.area.offset``' Intrinsic
10968^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10969
10970Syntax:
10971"""""""
10972
10973::
10974
10975      declare i32 @llvm.get.dynamic.area.offset.i32()
10976      declare i64 @llvm.get.dynamic.area.offset.i64()
10977
10978Overview:
10979"""""""""
10980
10981      The '``llvm.get.dynamic.area.offset.*``' intrinsic family is used to
10982      get the offset from native stack pointer to the address of the most
10983      recent dynamic alloca on the caller's stack. These intrinsics are
10984      intendend for use in combination with
10985      :ref:`llvm.stacksave <int_stacksave>` to get a
10986      pointer to the most recent dynamic alloca. This is useful, for example,
10987      for AddressSanitizer's stack unpoisoning routines.
10988
10989Semantics:
10990""""""""""
10991
10992      These intrinsics return a non-negative integer value that can be used to
10993      get the address of the most recent dynamic alloca, allocated by :ref:`alloca <i_alloca>`
10994      on the caller's stack. In particular, for targets where stack grows downwards,
10995      adding this offset to the native stack pointer would get the address of the most
10996      recent dynamic alloca. For targets where stack grows upwards, the situation is a bit more
10997      complicated, because subtracting this value from stack pointer would get the address
10998      one past the end of the most recent dynamic alloca.
10999
11000      Although for most targets `llvm.get.dynamic.area.offset <int_get_dynamic_area_offset>`
11001      returns just a zero, for others, such as PowerPC and PowerPC64, it returns a
11002      compile-time-known constant value.
11003
11004      The return value type of :ref:`llvm.get.dynamic.area.offset <int_get_dynamic_area_offset>`
11005      must match the target's default address space's (address space 0) pointer type.
11006
11007'``llvm.prefetch``' Intrinsic
11008^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11009
11010Syntax:
11011"""""""
11012
11013::
11014
11015      declare void @llvm.prefetch(i8* <address>, i32 <rw>, i32 <locality>, i32 <cache type>)
11016
11017Overview:
11018"""""""""
11019
11020The '``llvm.prefetch``' intrinsic is a hint to the code generator to
11021insert a prefetch instruction if supported; otherwise, it is a noop.
11022Prefetches have no effect on the behavior of the program but can change
11023its performance characteristics.
11024
11025Arguments:
11026""""""""""
11027
11028``address`` is the address to be prefetched, ``rw`` is the specifier
11029determining if the fetch should be for a read (0) or write (1), and
11030``locality`` is a temporal locality specifier ranging from (0) - no
11031locality, to (3) - extremely local keep in cache. The ``cache type``
11032specifies whether the prefetch is performed on the data (1) or
11033instruction (0) cache. The ``rw``, ``locality`` and ``cache type``
11034arguments must be constant integers.
11035
11036Semantics:
11037""""""""""
11038
11039This intrinsic does not modify the behavior of the program. In
11040particular, prefetches cannot trap and do not produce a value. On
11041targets that support this intrinsic, the prefetch can provide hints to
11042the processor cache for better performance.
11043
11044'``llvm.pcmarker``' Intrinsic
11045^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11046
11047Syntax:
11048"""""""
11049
11050::
11051
11052      declare void @llvm.pcmarker(i32 <id>)
11053
11054Overview:
11055"""""""""
11056
11057The '``llvm.pcmarker``' intrinsic is a method to export a Program
11058Counter (PC) in a region of code to simulators and other tools. The
11059method is target specific, but it is expected that the marker will use
11060exported symbols to transmit the PC of the marker. The marker makes no
11061guarantees that it will remain with any specific instruction after
11062optimizations. It is possible that the presence of a marker will inhibit
11063optimizations. The intended use is to be inserted after optimizations to
11064allow correlations of simulation runs.
11065
11066Arguments:
11067""""""""""
11068
11069``id`` is a numerical id identifying the marker.
11070
11071Semantics:
11072""""""""""
11073
11074This intrinsic does not modify the behavior of the program. Backends
11075that do not support this intrinsic may ignore it.
11076
11077'``llvm.readcyclecounter``' Intrinsic
11078^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11079
11080Syntax:
11081"""""""
11082
11083::
11084
11085      declare i64 @llvm.readcyclecounter()
11086
11087Overview:
11088"""""""""
11089
11090The '``llvm.readcyclecounter``' intrinsic provides access to the cycle
11091counter register (or similar low latency, high accuracy clocks) on those
11092targets that support it. On X86, it should map to RDTSC. On Alpha, it
11093should map to RPCC. As the backing counters overflow quickly (on the
11094order of 9 seconds on alpha), this should only be used for small
11095timings.
11096
11097Semantics:
11098""""""""""
11099
11100When directly supported, reading the cycle counter should not modify any
11101memory. Implementations are allowed to either return a application
11102specific value or a system wide value. On backends without support, this
11103is lowered to a constant 0.
11104
11105Note that runtime support may be conditional on the privilege-level code is
11106running at and the host platform.
11107
11108'``llvm.clear_cache``' Intrinsic
11109^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11110
11111Syntax:
11112"""""""
11113
11114::
11115
11116      declare void @llvm.clear_cache(i8*, i8*)
11117
11118Overview:
11119"""""""""
11120
11121The '``llvm.clear_cache``' intrinsic ensures visibility of modifications
11122in the specified range to the execution unit of the processor. On
11123targets with non-unified instruction and data cache, the implementation
11124flushes the instruction cache.
11125
11126Semantics:
11127""""""""""
11128
11129On platforms with coherent instruction and data caches (e.g. x86), this
11130intrinsic is a nop. On platforms with non-coherent instruction and data
11131cache (e.g. ARM, MIPS), the intrinsic is lowered either to appropriate
11132instructions or a system call, if cache flushing requires special
11133privileges.
11134
11135The default behavior is to emit a call to ``__clear_cache`` from the run
11136time library.
11137
11138This instrinsic does *not* empty the instruction pipeline. Modifications
11139of the current function are outside the scope of the intrinsic.
11140
11141'``llvm.instrprof.increment``' Intrinsic
11142^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11143
11144Syntax:
11145"""""""
11146
11147::
11148
11149      declare void @llvm.instrprof.increment(i8* <name>, i64 <hash>,
11150                                             i32 <num-counters>, i32 <index>)
11151
11152Overview:
11153"""""""""
11154
11155The '``llvm.instrprof.increment``' intrinsic can be emitted by a
11156frontend for use with instrumentation based profiling. These will be
11157lowered by the ``-instrprof`` pass to generate execution counts of a
11158program at runtime.
11159
11160Arguments:
11161""""""""""
11162
11163The first argument is a pointer to a global variable containing the
11164name of the entity being instrumented. This should generally be the
11165(mangled) function name for a set of counters.
11166
11167The second argument is a hash value that can be used by the consumer
11168of the profile data to detect changes to the instrumented source, and
11169the third is the number of counters associated with ``name``. It is an
11170error if ``hash`` or ``num-counters`` differ between two instances of
11171``instrprof.increment`` that refer to the same name.
11172
11173The last argument refers to which of the counters for ``name`` should
11174be incremented. It should be a value between 0 and ``num-counters``.
11175
11176Semantics:
11177""""""""""
11178
11179This intrinsic represents an increment of a profiling counter. It will
11180cause the ``-instrprof`` pass to generate the appropriate data
11181structures and the code to increment the appropriate value, in a
11182format that can be written out by a compiler runtime and consumed via
11183the ``llvm-profdata`` tool.
11184
11185'``llvm.instrprof.increment.step``' Intrinsic
11186^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11187
11188Syntax:
11189"""""""
11190
11191::
11192
11193      declare void @llvm.instrprof.increment.step(i8* <name>, i64 <hash>,
11194                                                  i32 <num-counters>,
11195                                                  i32 <index>, i64 <step>)
11196
11197Overview:
11198"""""""""
11199
11200The '``llvm.instrprof.increment.step``' intrinsic is an extension to
11201the '``llvm.instrprof.increment``' intrinsic with an additional fifth
11202argument to specify the step of the increment.
11203
11204Arguments:
11205""""""""""
11206The first four arguments are the same as '``llvm.instrprof.increment``'
11207intrinsic.
11208
11209The last argument specifies the value of the increment of the counter variable.
11210
11211Semantics:
11212""""""""""
11213See description of '``llvm.instrprof.increment``' instrinsic.
11214
11215
11216'``llvm.instrprof.value.profile``' Intrinsic
11217^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11218
11219Syntax:
11220"""""""
11221
11222::
11223
11224      declare void @llvm.instrprof.value.profile(i8* <name>, i64 <hash>,
11225                                                 i64 <value>, i32 <value_kind>,
11226                                                 i32 <index>)
11227
11228Overview:
11229"""""""""
11230
11231The '``llvm.instrprof.value.profile``' intrinsic can be emitted by a
11232frontend for use with instrumentation based profiling. This will be
11233lowered by the ``-instrprof`` pass to find out the target values,
11234instrumented expressions take in a program at runtime.
11235
11236Arguments:
11237""""""""""
11238
11239The first argument is a pointer to a global variable containing the
11240name of the entity being instrumented. ``name`` should generally be the
11241(mangled) function name for a set of counters.
11242
11243The second argument is a hash value that can be used by the consumer
11244of the profile data to detect changes to the instrumented source. It
11245is an error if ``hash`` differs between two instances of
11246``llvm.instrprof.*`` that refer to the same name.
11247
11248The third argument is the value of the expression being profiled. The profiled
11249expression's value should be representable as an unsigned 64-bit value. The
11250fourth argument represents the kind of value profiling that is being done. The
11251supported value profiling kinds are enumerated through the
11252``InstrProfValueKind`` type declared in the
11253``<include/llvm/ProfileData/InstrProf.h>`` header file. The last argument is the
11254index of the instrumented expression within ``name``. It should be >= 0.
11255
11256Semantics:
11257""""""""""
11258
11259This intrinsic represents the point where a call to a runtime routine
11260should be inserted for value profiling of target expressions. ``-instrprof``
11261pass will generate the appropriate data structures and replace the
11262``llvm.instrprof.value.profile`` intrinsic with the call to the profile
11263runtime library with proper arguments.
11264
11265'``llvm.thread.pointer``' Intrinsic
11266^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11267
11268Syntax:
11269"""""""
11270
11271::
11272
11273      declare i8* @llvm.thread.pointer()
11274
11275Overview:
11276"""""""""
11277
11278The '``llvm.thread.pointer``' intrinsic returns the value of the thread
11279pointer.
11280
11281Semantics:
11282""""""""""
11283
11284The '``llvm.thread.pointer``' intrinsic returns a pointer to the TLS area
11285for the current thread.  The exact semantics of this value are target
11286specific: it may point to the start of TLS area, to the end, or somewhere
11287in the middle.  Depending on the target, this intrinsic may read a register,
11288call a helper function, read from an alternate memory space, or perform
11289other operations necessary to locate the TLS area.  Not all targets support
11290this intrinsic.
11291
11292Standard C Library Intrinsics
11293-----------------------------
11294
11295LLVM provides intrinsics for a few important standard C library
11296functions. These intrinsics allow source-language front-ends to pass
11297information about the alignment of the pointer arguments to the code
11298generator, providing opportunity for more efficient code generation.
11299
11300.. _int_memcpy:
11301
11302'``llvm.memcpy``' Intrinsic
11303^^^^^^^^^^^^^^^^^^^^^^^^^^^
11304
11305Syntax:
11306"""""""
11307
11308This is an overloaded intrinsic. You can use ``llvm.memcpy`` on any
11309integer bit width and for different address spaces. Not all targets
11310support all bit widths however.
11311
11312::
11313
11314      declare void @llvm.memcpy.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
11315                                              i32 <len>, i1 <isvolatile>)
11316      declare void @llvm.memcpy.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
11317                                              i64 <len>, i1 <isvolatile>)
11318
11319Overview:
11320"""""""""
11321
11322The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
11323source location to the destination location.
11324
11325Note that, unlike the standard libc function, the ``llvm.memcpy.*``
11326intrinsics do not return a value, takes extra isvolatile
11327arguments and the pointers can be in specified address spaces.
11328
11329Arguments:
11330""""""""""
11331
11332The first argument is a pointer to the destination, the second is a
11333pointer to the source. The third argument is an integer argument
11334specifying the number of bytes to copy, and the fourth is a
11335boolean indicating a volatile access.
11336
11337The :ref:`align <attr_align>` parameter attribute can be provided
11338for the first and second arguments.
11339
11340If the ``isvolatile`` parameter is ``true``, the ``llvm.memcpy`` call is
11341a :ref:`volatile operation <volatile>`. The detailed access behavior is not
11342very cleanly specified and it is unwise to depend on it.
11343
11344Semantics:
11345""""""""""
11346
11347The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
11348source location to the destination location, which are not allowed to
11349overlap. It copies "len" bytes of memory over. If the argument is known
11350to be aligned to some boundary, this can be specified as an attribute on
11351the argument.
11352
11353If "len" is 0, the pointers may be NULL or dangling. However, they must still
11354be appropriately aligned.
11355
11356.. _int_memmove:
11357
11358'``llvm.memmove``' Intrinsic
11359^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11360
11361Syntax:
11362"""""""
11363
11364This is an overloaded intrinsic. You can use llvm.memmove on any integer
11365bit width and for different address space. Not all targets support all
11366bit widths however.
11367
11368::
11369
11370      declare void @llvm.memmove.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
11371                                               i32 <len>, i1 <isvolatile>)
11372      declare void @llvm.memmove.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
11373                                               i64 <len>, i1 <isvolatile>)
11374
11375Overview:
11376"""""""""
11377
11378The '``llvm.memmove.*``' intrinsics move a block of memory from the
11379source location to the destination location. It is similar to the
11380'``llvm.memcpy``' intrinsic but allows the two memory locations to
11381overlap.
11382
11383Note that, unlike the standard libc function, the ``llvm.memmove.*``
11384intrinsics do not return a value, takes an extra isvolatile
11385argument and the pointers can be in specified address spaces.
11386
11387Arguments:
11388""""""""""
11389
11390The first argument is a pointer to the destination, the second is a
11391pointer to the source. The third argument is an integer argument
11392specifying the number of bytes to copy, and the fourth is a
11393boolean indicating a volatile access.
11394
11395The :ref:`align <attr_align>` parameter attribute can be provided
11396for the first and second arguments.
11397
11398If the ``isvolatile`` parameter is ``true``, the ``llvm.memmove`` call
11399is a :ref:`volatile operation <volatile>`. The detailed access behavior is
11400not very cleanly specified and it is unwise to depend on it.
11401
11402Semantics:
11403""""""""""
11404
11405The '``llvm.memmove.*``' intrinsics copy a block of memory from the
11406source location to the destination location, which may overlap. It
11407copies "len" bytes of memory over. If the argument is known to be
11408aligned to some boundary, this can be specified as an attribute on
11409the argument.
11410
11411If "len" is 0, the pointers may be NULL or dangling. However, they must still
11412be appropriately aligned.
11413
11414.. _int_memset:
11415
11416'``llvm.memset.*``' Intrinsics
11417^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11418
11419Syntax:
11420"""""""
11421
11422This is an overloaded intrinsic. You can use llvm.memset on any integer
11423bit width and for different address spaces. However, not all targets
11424support all bit widths.
11425
11426::
11427
11428      declare void @llvm.memset.p0i8.i32(i8* <dest>, i8 <val>,
11429                                         i32 <len>, i1 <isvolatile>)
11430      declare void @llvm.memset.p0i8.i64(i8* <dest>, i8 <val>,
11431                                         i64 <len>, i1 <isvolatile>)
11432
11433Overview:
11434"""""""""
11435
11436The '``llvm.memset.*``' intrinsics fill a block of memory with a
11437particular byte value.
11438
11439Note that, unlike the standard libc function, the ``llvm.memset``
11440intrinsic does not return a value and takes an extra volatile
11441argument. Also, the destination can be in an arbitrary address space.
11442
11443Arguments:
11444""""""""""
11445
11446The first argument is a pointer to the destination to fill, the second
11447is the byte value with which to fill it, the third argument is an
11448integer argument specifying the number of bytes to fill, and the fourth
11449is a boolean indicating a volatile access.
11450
11451The :ref:`align <attr_align>` parameter attribute can be provided
11452for the first arguments.
11453
11454If the ``isvolatile`` parameter is ``true``, the ``llvm.memset`` call is
11455a :ref:`volatile operation <volatile>`. The detailed access behavior is not
11456very cleanly specified and it is unwise to depend on it.
11457
11458Semantics:
11459""""""""""
11460
11461The '``llvm.memset.*``' intrinsics fill "len" bytes of memory starting
11462at the destination location. If the argument is known to be
11463aligned to some boundary, this can be specified as an attribute on
11464the argument.
11465
11466If "len" is 0, the pointers may be NULL or dangling. However, they must still
11467be appropriately aligned.
11468
11469'``llvm.sqrt.*``' Intrinsic
11470^^^^^^^^^^^^^^^^^^^^^^^^^^^
11471
11472Syntax:
11473"""""""
11474
11475This is an overloaded intrinsic. You can use ``llvm.sqrt`` on any
11476floating-point or vector of floating-point type. Not all targets support
11477all types however.
11478
11479::
11480
11481      declare float     @llvm.sqrt.f32(float %Val)
11482      declare double    @llvm.sqrt.f64(double %Val)
11483      declare x86_fp80  @llvm.sqrt.f80(x86_fp80 %Val)
11484      declare fp128     @llvm.sqrt.f128(fp128 %Val)
11485      declare ppc_fp128 @llvm.sqrt.ppcf128(ppc_fp128 %Val)
11486
11487Overview:
11488"""""""""
11489
11490The '``llvm.sqrt``' intrinsics return the square root of the specified value.
11491
11492Arguments:
11493""""""""""
11494
11495The argument and return value are floating-point numbers of the same type.
11496
11497Semantics:
11498""""""""""
11499
11500Return the same value as a corresponding libm '``sqrt``' function but without
11501trapping or setting ``errno``. For types specified by IEEE-754, the result
11502matches a conforming libm implementation.
11503
11504When specified with the fast-math-flag 'afn', the result may be approximated
11505using a less accurate calculation.
11506
11507'``llvm.powi.*``' Intrinsic
11508^^^^^^^^^^^^^^^^^^^^^^^^^^^
11509
11510Syntax:
11511"""""""
11512
11513This is an overloaded intrinsic. You can use ``llvm.powi`` on any
11514floating-point or vector of floating-point type. Not all targets support
11515all types however.
11516
11517::
11518
11519      declare float     @llvm.powi.f32(float  %Val, i32 %power)
11520      declare double    @llvm.powi.f64(double %Val, i32 %power)
11521      declare x86_fp80  @llvm.powi.f80(x86_fp80  %Val, i32 %power)
11522      declare fp128     @llvm.powi.f128(fp128 %Val, i32 %power)
11523      declare ppc_fp128 @llvm.powi.ppcf128(ppc_fp128  %Val, i32 %power)
11524
11525Overview:
11526"""""""""
11527
11528The '``llvm.powi.*``' intrinsics return the first operand raised to the
11529specified (positive or negative) power. The order of evaluation of
11530multiplications is not defined. When a vector of floating-point type is
11531used, the second argument remains a scalar integer value.
11532
11533Arguments:
11534""""""""""
11535
11536The second argument is an integer power, and the first is a value to
11537raise to that power.
11538
11539Semantics:
11540""""""""""
11541
11542This function returns the first value raised to the second power with an
11543unspecified sequence of rounding operations.
11544
11545'``llvm.sin.*``' Intrinsic
11546^^^^^^^^^^^^^^^^^^^^^^^^^^
11547
11548Syntax:
11549"""""""
11550
11551This is an overloaded intrinsic. You can use ``llvm.sin`` on any
11552floating-point or vector of floating-point type. Not all targets support
11553all types however.
11554
11555::
11556
11557      declare float     @llvm.sin.f32(float  %Val)
11558      declare double    @llvm.sin.f64(double %Val)
11559      declare x86_fp80  @llvm.sin.f80(x86_fp80  %Val)
11560      declare fp128     @llvm.sin.f128(fp128 %Val)
11561      declare ppc_fp128 @llvm.sin.ppcf128(ppc_fp128  %Val)
11562
11563Overview:
11564"""""""""
11565
11566The '``llvm.sin.*``' intrinsics return the sine of the operand.
11567
11568Arguments:
11569""""""""""
11570
11571The argument and return value are floating-point numbers of the same type.
11572
11573Semantics:
11574""""""""""
11575
11576Return the same value as a corresponding libm '``sin``' function but without
11577trapping or setting ``errno``.
11578
11579When specified with the fast-math-flag 'afn', the result may be approximated
11580using a less accurate calculation.
11581
11582'``llvm.cos.*``' Intrinsic
11583^^^^^^^^^^^^^^^^^^^^^^^^^^
11584
11585Syntax:
11586"""""""
11587
11588This is an overloaded intrinsic. You can use ``llvm.cos`` on any
11589floating-point or vector of floating-point type. Not all targets support
11590all types however.
11591
11592::
11593
11594      declare float     @llvm.cos.f32(float  %Val)
11595      declare double    @llvm.cos.f64(double %Val)
11596      declare x86_fp80  @llvm.cos.f80(x86_fp80  %Val)
11597      declare fp128     @llvm.cos.f128(fp128 %Val)
11598      declare ppc_fp128 @llvm.cos.ppcf128(ppc_fp128  %Val)
11599
11600Overview:
11601"""""""""
11602
11603The '``llvm.cos.*``' intrinsics return the cosine of the operand.
11604
11605Arguments:
11606""""""""""
11607
11608The argument and return value are floating-point numbers of the same type.
11609
11610Semantics:
11611""""""""""
11612
11613Return the same value as a corresponding libm '``cos``' function but without
11614trapping or setting ``errno``.
11615
11616When specified with the fast-math-flag 'afn', the result may be approximated
11617using a less accurate calculation.
11618
11619'``llvm.pow.*``' Intrinsic
11620^^^^^^^^^^^^^^^^^^^^^^^^^^
11621
11622Syntax:
11623"""""""
11624
11625This is an overloaded intrinsic. You can use ``llvm.pow`` on any
11626floating-point or vector of floating-point type. Not all targets support
11627all types however.
11628
11629::
11630
11631      declare float     @llvm.pow.f32(float  %Val, float %Power)
11632      declare double    @llvm.pow.f64(double %Val, double %Power)
11633      declare x86_fp80  @llvm.pow.f80(x86_fp80  %Val, x86_fp80 %Power)
11634      declare fp128     @llvm.pow.f128(fp128 %Val, fp128 %Power)
11635      declare ppc_fp128 @llvm.pow.ppcf128(ppc_fp128  %Val, ppc_fp128 Power)
11636
11637Overview:
11638"""""""""
11639
11640The '``llvm.pow.*``' intrinsics return the first operand raised to the
11641specified (positive or negative) power.
11642
11643Arguments:
11644""""""""""
11645
11646The arguments and return value are floating-point numbers of the same type.
11647
11648Semantics:
11649""""""""""
11650
11651Return the same value as a corresponding libm '``pow``' function but without
11652trapping or setting ``errno``.
11653
11654When specified with the fast-math-flag 'afn', the result may be approximated
11655using a less accurate calculation.
11656
11657'``llvm.exp.*``' Intrinsic
11658^^^^^^^^^^^^^^^^^^^^^^^^^^
11659
11660Syntax:
11661"""""""
11662
11663This is an overloaded intrinsic. You can use ``llvm.exp`` on any
11664floating-point or vector of floating-point type. Not all targets support
11665all types however.
11666
11667::
11668
11669      declare float     @llvm.exp.f32(float  %Val)
11670      declare double    @llvm.exp.f64(double %Val)
11671      declare x86_fp80  @llvm.exp.f80(x86_fp80  %Val)
11672      declare fp128     @llvm.exp.f128(fp128 %Val)
11673      declare ppc_fp128 @llvm.exp.ppcf128(ppc_fp128  %Val)
11674
11675Overview:
11676"""""""""
11677
11678The '``llvm.exp.*``' intrinsics compute the base-e exponential of the specified
11679value.
11680
11681Arguments:
11682""""""""""
11683
11684The argument and return value are floating-point numbers of the same type.
11685
11686Semantics:
11687""""""""""
11688
11689Return the same value as a corresponding libm '``exp``' function but without
11690trapping or setting ``errno``.
11691
11692When specified with the fast-math-flag 'afn', the result may be approximated
11693using a less accurate calculation.
11694
11695'``llvm.exp2.*``' Intrinsic
11696^^^^^^^^^^^^^^^^^^^^^^^^^^^
11697
11698Syntax:
11699"""""""
11700
11701This is an overloaded intrinsic. You can use ``llvm.exp2`` on any
11702floating-point or vector of floating-point type. Not all targets support
11703all types however.
11704
11705::
11706
11707      declare float     @llvm.exp2.f32(float  %Val)
11708      declare double    @llvm.exp2.f64(double %Val)
11709      declare x86_fp80  @llvm.exp2.f80(x86_fp80  %Val)
11710      declare fp128     @llvm.exp2.f128(fp128 %Val)
11711      declare ppc_fp128 @llvm.exp2.ppcf128(ppc_fp128  %Val)
11712
11713Overview:
11714"""""""""
11715
11716The '``llvm.exp2.*``' intrinsics compute the base-2 exponential of the
11717specified value.
11718
11719Arguments:
11720""""""""""
11721
11722The argument and return value are floating-point numbers of the same type.
11723
11724Semantics:
11725""""""""""
11726
11727Return the same value as a corresponding libm '``exp2``' function but without
11728trapping or setting ``errno``.
11729
11730When specified with the fast-math-flag 'afn', the result may be approximated
11731using a less accurate calculation.
11732
11733'``llvm.log.*``' Intrinsic
11734^^^^^^^^^^^^^^^^^^^^^^^^^^
11735
11736Syntax:
11737"""""""
11738
11739This is an overloaded intrinsic. You can use ``llvm.log`` on any
11740floating-point or vector of floating-point type. Not all targets support
11741all types however.
11742
11743::
11744
11745      declare float     @llvm.log.f32(float  %Val)
11746      declare double    @llvm.log.f64(double %Val)
11747      declare x86_fp80  @llvm.log.f80(x86_fp80  %Val)
11748      declare fp128     @llvm.log.f128(fp128 %Val)
11749      declare ppc_fp128 @llvm.log.ppcf128(ppc_fp128  %Val)
11750
11751Overview:
11752"""""""""
11753
11754The '``llvm.log.*``' intrinsics compute the base-e logarithm of the specified
11755value.
11756
11757Arguments:
11758""""""""""
11759
11760The argument and return value are floating-point numbers of the same type.
11761
11762Semantics:
11763""""""""""
11764
11765Return the same value as a corresponding libm '``log``' function but without
11766trapping or setting ``errno``.
11767
11768When specified with the fast-math-flag 'afn', the result may be approximated
11769using a less accurate calculation.
11770
11771'``llvm.log10.*``' Intrinsic
11772^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11773
11774Syntax:
11775"""""""
11776
11777This is an overloaded intrinsic. You can use ``llvm.log10`` on any
11778floating-point or vector of floating-point type. Not all targets support
11779all types however.
11780
11781::
11782
11783      declare float     @llvm.log10.f32(float  %Val)
11784      declare double    @llvm.log10.f64(double %Val)
11785      declare x86_fp80  @llvm.log10.f80(x86_fp80  %Val)
11786      declare fp128     @llvm.log10.f128(fp128 %Val)
11787      declare ppc_fp128 @llvm.log10.ppcf128(ppc_fp128  %Val)
11788
11789Overview:
11790"""""""""
11791
11792The '``llvm.log10.*``' intrinsics compute the base-10 logarithm of the
11793specified value.
11794
11795Arguments:
11796""""""""""
11797
11798The argument and return value are floating-point numbers of the same type.
11799
11800Semantics:
11801""""""""""
11802
11803Return the same value as a corresponding libm '``log10``' function but without
11804trapping or setting ``errno``.
11805
11806When specified with the fast-math-flag 'afn', the result may be approximated
11807using a less accurate calculation.
11808
11809'``llvm.log2.*``' Intrinsic
11810^^^^^^^^^^^^^^^^^^^^^^^^^^^
11811
11812Syntax:
11813"""""""
11814
11815This is an overloaded intrinsic. You can use ``llvm.log2`` on any
11816floating-point or vector of floating-point type. Not all targets support
11817all types however.
11818
11819::
11820
11821      declare float     @llvm.log2.f32(float  %Val)
11822      declare double    @llvm.log2.f64(double %Val)
11823      declare x86_fp80  @llvm.log2.f80(x86_fp80  %Val)
11824      declare fp128     @llvm.log2.f128(fp128 %Val)
11825      declare ppc_fp128 @llvm.log2.ppcf128(ppc_fp128  %Val)
11826
11827Overview:
11828"""""""""
11829
11830The '``llvm.log2.*``' intrinsics compute the base-2 logarithm of the specified
11831value.
11832
11833Arguments:
11834""""""""""
11835
11836The argument and return value are floating-point numbers of the same type.
11837
11838Semantics:
11839""""""""""
11840
11841Return the same value as a corresponding libm '``log2``' function but without
11842trapping or setting ``errno``.
11843
11844When specified with the fast-math-flag 'afn', the result may be approximated
11845using a less accurate calculation.
11846
11847'``llvm.fma.*``' Intrinsic
11848^^^^^^^^^^^^^^^^^^^^^^^^^^
11849
11850Syntax:
11851"""""""
11852
11853This is an overloaded intrinsic. You can use ``llvm.fma`` on any
11854floating-point or vector of floating-point type. Not all targets support
11855all types however.
11856
11857::
11858
11859      declare float     @llvm.fma.f32(float  %a, float  %b, float  %c)
11860      declare double    @llvm.fma.f64(double %a, double %b, double %c)
11861      declare x86_fp80  @llvm.fma.f80(x86_fp80 %a, x86_fp80 %b, x86_fp80 %c)
11862      declare fp128     @llvm.fma.f128(fp128 %a, fp128 %b, fp128 %c)
11863      declare ppc_fp128 @llvm.fma.ppcf128(ppc_fp128 %a, ppc_fp128 %b, ppc_fp128 %c)
11864
11865Overview:
11866"""""""""
11867
11868The '``llvm.fma.*``' intrinsics perform the fused multiply-add operation.
11869
11870Arguments:
11871""""""""""
11872
11873The arguments and return value are floating-point numbers of the same type.
11874
11875Semantics:
11876""""""""""
11877
11878Return the same value as a corresponding libm '``fma``' function but without
11879trapping or setting ``errno``.
11880
11881When specified with the fast-math-flag 'afn', the result may be approximated
11882using a less accurate calculation.
11883
11884'``llvm.fabs.*``' Intrinsic
11885^^^^^^^^^^^^^^^^^^^^^^^^^^^
11886
11887Syntax:
11888"""""""
11889
11890This is an overloaded intrinsic. You can use ``llvm.fabs`` on any
11891floating-point or vector of floating-point type. Not all targets support
11892all types however.
11893
11894::
11895
11896      declare float     @llvm.fabs.f32(float  %Val)
11897      declare double    @llvm.fabs.f64(double %Val)
11898      declare x86_fp80  @llvm.fabs.f80(x86_fp80 %Val)
11899      declare fp128     @llvm.fabs.f128(fp128 %Val)
11900      declare ppc_fp128 @llvm.fabs.ppcf128(ppc_fp128 %Val)
11901
11902Overview:
11903"""""""""
11904
11905The '``llvm.fabs.*``' intrinsics return the absolute value of the
11906operand.
11907
11908Arguments:
11909""""""""""
11910
11911The argument and return value are floating-point numbers of the same
11912type.
11913
11914Semantics:
11915""""""""""
11916
11917This function returns the same values as the libm ``fabs`` functions
11918would, and handles error conditions in the same way.
11919
11920'``llvm.minnum.*``' Intrinsic
11921^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11922
11923Syntax:
11924"""""""
11925
11926This is an overloaded intrinsic. You can use ``llvm.minnum`` on any
11927floating-point or vector of floating-point type. Not all targets support
11928all types however.
11929
11930::
11931
11932      declare float     @llvm.minnum.f32(float %Val0, float %Val1)
11933      declare double    @llvm.minnum.f64(double %Val0, double %Val1)
11934      declare x86_fp80  @llvm.minnum.f80(x86_fp80 %Val0, x86_fp80 %Val1)
11935      declare fp128     @llvm.minnum.f128(fp128 %Val0, fp128 %Val1)
11936      declare ppc_fp128 @llvm.minnum.ppcf128(ppc_fp128 %Val0, ppc_fp128 %Val1)
11937
11938Overview:
11939"""""""""
11940
11941The '``llvm.minnum.*``' intrinsics return the minimum of the two
11942arguments.
11943
11944
11945Arguments:
11946""""""""""
11947
11948The arguments and return value are floating-point numbers of the same
11949type.
11950
11951Semantics:
11952""""""""""
11953
11954Follows the IEEE-754 semantics for minNum, except for handling of
11955signaling NaNs. This match's the behavior of libm's fmin.
11956
11957If either operand is a NaN, returns the other non-NaN operand. Returns
11958NaN only if both operands are NaN. The returned NaN is always
11959quiet. If the operands compare equal, returns a value that compares
11960equal to both operands. This means that fmin(+/-0.0, +/-0.0) could
11961return either -0.0 or 0.0.
11962
11963Unlike the IEEE-754 2008 behavior, this does not distinguish between
11964signaling and quiet NaN inputs. If a target's implementation follows
11965the standard and returns a quiet NaN if either input is a signaling
11966NaN, the intrinsic lowering is responsible for quieting the inputs to
11967correctly return the non-NaN input (e.g. by using the equivalent of
11968``llvm.canonicalize``).
11969
11970
11971'``llvm.maxnum.*``' Intrinsic
11972^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11973
11974Syntax:
11975"""""""
11976
11977This is an overloaded intrinsic. You can use ``llvm.maxnum`` on any
11978floating-point or vector of floating-point type. Not all targets support
11979all types however.
11980
11981::
11982
11983      declare float     @llvm.maxnum.f32(float  %Val0, float  %Val1l)
11984      declare double    @llvm.maxnum.f64(double %Val0, double %Val1)
11985      declare x86_fp80  @llvm.maxnum.f80(x86_fp80  %Val0, x86_fp80  %Val1)
11986      declare fp128     @llvm.maxnum.f128(fp128 %Val0, fp128 %Val1)
11987      declare ppc_fp128 @llvm.maxnum.ppcf128(ppc_fp128  %Val0, ppc_fp128  %Val1)
11988
11989Overview:
11990"""""""""
11991
11992The '``llvm.maxnum.*``' intrinsics return the maximum of the two
11993arguments.
11994
11995
11996Arguments:
11997""""""""""
11998
11999The arguments and return value are floating-point numbers of the same
12000type.
12001
12002Semantics:
12003""""""""""
12004Follows the IEEE-754 semantics for maxNum except for the handling of
12005signaling NaNs. This matches the behavior of libm's fmax.
12006
12007If either operand is a NaN, returns the other non-NaN operand. Returns
12008NaN only if both operands are NaN. The returned NaN is always
12009quiet. If the operands compare equal, returns a value that compares
12010equal to both operands. This means that fmax(+/-0.0, +/-0.0) could
12011return either -0.0 or 0.0.
12012
12013Unlike the IEEE-754 2008 behavior, this does not distinguish between
12014signaling and quiet NaN inputs. If a target's implementation follows
12015the standard and returns a quiet NaN if either input is a signaling
12016NaN, the intrinsic lowering is responsible for quieting the inputs to
12017correctly return the non-NaN input (e.g. by using the equivalent of
12018``llvm.canonicalize``).
12019
12020'``llvm.minimum.*``' Intrinsic
12021^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12022
12023Syntax:
12024"""""""
12025
12026This is an overloaded intrinsic. You can use ``llvm.minimum`` on any
12027floating-point or vector of floating-point type. Not all targets support
12028all types however.
12029
12030::
12031
12032      declare float     @llvm.minimum.f32(float %Val0, float %Val1)
12033      declare double    @llvm.minimum.f64(double %Val0, double %Val1)
12034      declare x86_fp80  @llvm.minimum.f80(x86_fp80 %Val0, x86_fp80 %Val1)
12035      declare fp128     @llvm.minimum.f128(fp128 %Val0, fp128 %Val1)
12036      declare ppc_fp128 @llvm.minimum.ppcf128(ppc_fp128 %Val0, ppc_fp128 %Val1)
12037
12038Overview:
12039"""""""""
12040
12041The '``llvm.minimum.*``' intrinsics return the minimum of the two
12042arguments, propagating NaNs and treating -0.0 as less than +0.0.
12043
12044
12045Arguments:
12046""""""""""
12047
12048The arguments and return value are floating-point numbers of the same
12049type.
12050
12051Semantics:
12052""""""""""
12053If either operand is a NaN, returns NaN. Otherwise returns the lesser
12054of the two arguments. -0.0 is considered to be less than +0.0 for this
12055intrinsic. Note that these are the semantics specified in the draft of
12056IEEE 754-2018.
12057
12058'``llvm.maximum.*``' Intrinsic
12059^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12060
12061Syntax:
12062"""""""
12063
12064This is an overloaded intrinsic. You can use ``llvm.maximum`` on any
12065floating-point or vector of floating-point type. Not all targets support
12066all types however.
12067
12068::
12069
12070      declare float     @llvm.maximum.f32(float %Val0, float %Val1)
12071      declare double    @llvm.maximum.f64(double %Val0, double %Val1)
12072      declare x86_fp80  @llvm.maximum.f80(x86_fp80 %Val0, x86_fp80 %Val1)
12073      declare fp128     @llvm.maximum.f128(fp128 %Val0, fp128 %Val1)
12074      declare ppc_fp128 @llvm.maximum.ppcf128(ppc_fp128 %Val0, ppc_fp128 %Val1)
12075
12076Overview:
12077"""""""""
12078
12079The '``llvm.maximum.*``' intrinsics return the maximum of the two
12080arguments, propagating NaNs and treating -0.0 as less than +0.0.
12081
12082
12083Arguments:
12084""""""""""
12085
12086The arguments and return value are floating-point numbers of the same
12087type.
12088
12089Semantics:
12090""""""""""
12091If either operand is a NaN, returns NaN. Otherwise returns the greater
12092of the two arguments. -0.0 is considered to be less than +0.0 for this
12093intrinsic. Note that these are the semantics specified in the draft of
12094IEEE 754-2018.
12095
12096'``llvm.copysign.*``' Intrinsic
12097^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12098
12099Syntax:
12100"""""""
12101
12102This is an overloaded intrinsic. You can use ``llvm.copysign`` on any
12103floating-point or vector of floating-point type. Not all targets support
12104all types however.
12105
12106::
12107
12108      declare float     @llvm.copysign.f32(float  %Mag, float  %Sgn)
12109      declare double    @llvm.copysign.f64(double %Mag, double %Sgn)
12110      declare x86_fp80  @llvm.copysign.f80(x86_fp80  %Mag, x86_fp80  %Sgn)
12111      declare fp128     @llvm.copysign.f128(fp128 %Mag, fp128 %Sgn)
12112      declare ppc_fp128 @llvm.copysign.ppcf128(ppc_fp128  %Mag, ppc_fp128  %Sgn)
12113
12114Overview:
12115"""""""""
12116
12117The '``llvm.copysign.*``' intrinsics return a value with the magnitude of the
12118first operand and the sign of the second operand.
12119
12120Arguments:
12121""""""""""
12122
12123The arguments and return value are floating-point numbers of the same
12124type.
12125
12126Semantics:
12127""""""""""
12128
12129This function returns the same values as the libm ``copysign``
12130functions would, and handles error conditions in the same way.
12131
12132'``llvm.floor.*``' Intrinsic
12133^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12134
12135Syntax:
12136"""""""
12137
12138This is an overloaded intrinsic. You can use ``llvm.floor`` on any
12139floating-point or vector of floating-point type. Not all targets support
12140all types however.
12141
12142::
12143
12144      declare float     @llvm.floor.f32(float  %Val)
12145      declare double    @llvm.floor.f64(double %Val)
12146      declare x86_fp80  @llvm.floor.f80(x86_fp80  %Val)
12147      declare fp128     @llvm.floor.f128(fp128 %Val)
12148      declare ppc_fp128 @llvm.floor.ppcf128(ppc_fp128  %Val)
12149
12150Overview:
12151"""""""""
12152
12153The '``llvm.floor.*``' intrinsics return the floor of the operand.
12154
12155Arguments:
12156""""""""""
12157
12158The argument and return value are floating-point numbers of the same
12159type.
12160
12161Semantics:
12162""""""""""
12163
12164This function returns the same values as the libm ``floor`` functions
12165would, and handles error conditions in the same way.
12166
12167'``llvm.ceil.*``' Intrinsic
12168^^^^^^^^^^^^^^^^^^^^^^^^^^^
12169
12170Syntax:
12171"""""""
12172
12173This is an overloaded intrinsic. You can use ``llvm.ceil`` on any
12174floating-point or vector of floating-point type. Not all targets support
12175all types however.
12176
12177::
12178
12179      declare float     @llvm.ceil.f32(float  %Val)
12180      declare double    @llvm.ceil.f64(double %Val)
12181      declare x86_fp80  @llvm.ceil.f80(x86_fp80  %Val)
12182      declare fp128     @llvm.ceil.f128(fp128 %Val)
12183      declare ppc_fp128 @llvm.ceil.ppcf128(ppc_fp128  %Val)
12184
12185Overview:
12186"""""""""
12187
12188The '``llvm.ceil.*``' intrinsics return the ceiling of the operand.
12189
12190Arguments:
12191""""""""""
12192
12193The argument and return value are floating-point numbers of the same
12194type.
12195
12196Semantics:
12197""""""""""
12198
12199This function returns the same values as the libm ``ceil`` functions
12200would, and handles error conditions in the same way.
12201
12202'``llvm.trunc.*``' Intrinsic
12203^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12204
12205Syntax:
12206"""""""
12207
12208This is an overloaded intrinsic. You can use ``llvm.trunc`` on any
12209floating-point or vector of floating-point type. Not all targets support
12210all types however.
12211
12212::
12213
12214      declare float     @llvm.trunc.f32(float  %Val)
12215      declare double    @llvm.trunc.f64(double %Val)
12216      declare x86_fp80  @llvm.trunc.f80(x86_fp80  %Val)
12217      declare fp128     @llvm.trunc.f128(fp128 %Val)
12218      declare ppc_fp128 @llvm.trunc.ppcf128(ppc_fp128  %Val)
12219
12220Overview:
12221"""""""""
12222
12223The '``llvm.trunc.*``' intrinsics returns the operand rounded to the
12224nearest integer not larger in magnitude than the operand.
12225
12226Arguments:
12227""""""""""
12228
12229The argument and return value are floating-point numbers of the same
12230type.
12231
12232Semantics:
12233""""""""""
12234
12235This function returns the same values as the libm ``trunc`` functions
12236would, and handles error conditions in the same way.
12237
12238'``llvm.rint.*``' Intrinsic
12239^^^^^^^^^^^^^^^^^^^^^^^^^^^
12240
12241Syntax:
12242"""""""
12243
12244This is an overloaded intrinsic. You can use ``llvm.rint`` on any
12245floating-point or vector of floating-point type. Not all targets support
12246all types however.
12247
12248::
12249
12250      declare float     @llvm.rint.f32(float  %Val)
12251      declare double    @llvm.rint.f64(double %Val)
12252      declare x86_fp80  @llvm.rint.f80(x86_fp80  %Val)
12253      declare fp128     @llvm.rint.f128(fp128 %Val)
12254      declare ppc_fp128 @llvm.rint.ppcf128(ppc_fp128  %Val)
12255
12256Overview:
12257"""""""""
12258
12259The '``llvm.rint.*``' intrinsics returns the operand rounded to the
12260nearest integer. It may raise an inexact floating-point exception if the
12261operand isn't an integer.
12262
12263Arguments:
12264""""""""""
12265
12266The argument and return value are floating-point numbers of the same
12267type.
12268
12269Semantics:
12270""""""""""
12271
12272This function returns the same values as the libm ``rint`` functions
12273would, and handles error conditions in the same way.
12274
12275'``llvm.nearbyint.*``' Intrinsic
12276^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12277
12278Syntax:
12279"""""""
12280
12281This is an overloaded intrinsic. You can use ``llvm.nearbyint`` on any
12282floating-point or vector of floating-point type. Not all targets support
12283all types however.
12284
12285::
12286
12287      declare float     @llvm.nearbyint.f32(float  %Val)
12288      declare double    @llvm.nearbyint.f64(double %Val)
12289      declare x86_fp80  @llvm.nearbyint.f80(x86_fp80  %Val)
12290      declare fp128     @llvm.nearbyint.f128(fp128 %Val)
12291      declare ppc_fp128 @llvm.nearbyint.ppcf128(ppc_fp128  %Val)
12292
12293Overview:
12294"""""""""
12295
12296The '``llvm.nearbyint.*``' intrinsics returns the operand rounded to the
12297nearest integer.
12298
12299Arguments:
12300""""""""""
12301
12302The argument and return value are floating-point numbers of the same
12303type.
12304
12305Semantics:
12306""""""""""
12307
12308This function returns the same values as the libm ``nearbyint``
12309functions would, and handles error conditions in the same way.
12310
12311'``llvm.round.*``' Intrinsic
12312^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12313
12314Syntax:
12315"""""""
12316
12317This is an overloaded intrinsic. You can use ``llvm.round`` on any
12318floating-point or vector of floating-point type. Not all targets support
12319all types however.
12320
12321::
12322
12323      declare float     @llvm.round.f32(float  %Val)
12324      declare double    @llvm.round.f64(double %Val)
12325      declare x86_fp80  @llvm.round.f80(x86_fp80  %Val)
12326      declare fp128     @llvm.round.f128(fp128 %Val)
12327      declare ppc_fp128 @llvm.round.ppcf128(ppc_fp128  %Val)
12328
12329Overview:
12330"""""""""
12331
12332The '``llvm.round.*``' intrinsics returns the operand rounded to the
12333nearest integer.
12334
12335Arguments:
12336""""""""""
12337
12338The argument and return value are floating-point numbers of the same
12339type.
12340
12341Semantics:
12342""""""""""
12343
12344This function returns the same values as the libm ``round``
12345functions would, and handles error conditions in the same way.
12346
12347Bit Manipulation Intrinsics
12348---------------------------
12349
12350LLVM provides intrinsics for a few important bit manipulation
12351operations. These allow efficient code generation for some algorithms.
12352
12353'``llvm.bitreverse.*``' Intrinsics
12354^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12355
12356Syntax:
12357"""""""
12358
12359This is an overloaded intrinsic function. You can use bitreverse on any
12360integer type.
12361
12362::
12363
12364      declare i16 @llvm.bitreverse.i16(i16 <id>)
12365      declare i32 @llvm.bitreverse.i32(i32 <id>)
12366      declare i64 @llvm.bitreverse.i64(i64 <id>)
12367      declare <4 x i32> @llvm.bitreverse.v4i32(<4 x i32> <id>)
12368
12369Overview:
12370"""""""""
12371
12372The '``llvm.bitreverse``' family of intrinsics is used to reverse the
12373bitpattern of an integer value or vector of integer values; for example
12374``0b10110110`` becomes ``0b01101101``.
12375
12376Semantics:
12377""""""""""
12378
12379The ``llvm.bitreverse.iN`` intrinsic returns an iN value that has bit
12380``M`` in the input moved to bit ``N-M`` in the output. The vector
12381intrinsics, such as ``llvm.bitreverse.v4i32``, operate on a per-element
12382basis and the element order is not affected.
12383
12384'``llvm.bswap.*``' Intrinsics
12385^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12386
12387Syntax:
12388"""""""
12389
12390This is an overloaded intrinsic function. You can use bswap on any
12391integer type that is an even number of bytes (i.e. BitWidth % 16 == 0).
12392
12393::
12394
12395      declare i16 @llvm.bswap.i16(i16 <id>)
12396      declare i32 @llvm.bswap.i32(i32 <id>)
12397      declare i64 @llvm.bswap.i64(i64 <id>)
12398      declare <4 x i32> @llvm.bswap.v4i32(<4 x i32> <id>)
12399
12400Overview:
12401"""""""""
12402
12403The '``llvm.bswap``' family of intrinsics is used to byte swap an integer
12404value or vector of integer values with an even number of bytes (positive
12405multiple of 16 bits).
12406
12407Semantics:
12408""""""""""
12409
12410The ``llvm.bswap.i16`` intrinsic returns an i16 value that has the high
12411and low byte of the input i16 swapped. Similarly, the ``llvm.bswap.i32``
12412intrinsic returns an i32 value that has the four bytes of the input i32
12413swapped, so that if the input bytes are numbered 0, 1, 2, 3 then the
12414returned i32 will have its bytes in 3, 2, 1, 0 order. The
12415``llvm.bswap.i48``, ``llvm.bswap.i64`` and other intrinsics extend this
12416concept to additional even-byte lengths (6 bytes, 8 bytes and more,
12417respectively). The vector intrinsics, such as ``llvm.bswap.v4i32``,
12418operate on a per-element basis and the element order is not affected.
12419
12420'``llvm.ctpop.*``' Intrinsic
12421^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12422
12423Syntax:
12424"""""""
12425
12426This is an overloaded intrinsic. You can use llvm.ctpop on any integer
12427bit width, or on any vector with integer elements. Not all targets
12428support all bit widths or vector types, however.
12429
12430::
12431
12432      declare i8 @llvm.ctpop.i8(i8  <src>)
12433      declare i16 @llvm.ctpop.i16(i16 <src>)
12434      declare i32 @llvm.ctpop.i32(i32 <src>)
12435      declare i64 @llvm.ctpop.i64(i64 <src>)
12436      declare i256 @llvm.ctpop.i256(i256 <src>)
12437      declare <2 x i32> @llvm.ctpop.v2i32(<2 x i32> <src>)
12438
12439Overview:
12440"""""""""
12441
12442The '``llvm.ctpop``' family of intrinsics counts the number of bits set
12443in a value.
12444
12445Arguments:
12446""""""""""
12447
12448The only argument is the value to be counted. The argument may be of any
12449integer type, or a vector with integer elements. The return type must
12450match the argument type.
12451
12452Semantics:
12453""""""""""
12454
12455The '``llvm.ctpop``' intrinsic counts the 1's in a variable, or within
12456each element of a vector.
12457
12458'``llvm.ctlz.*``' Intrinsic
12459^^^^^^^^^^^^^^^^^^^^^^^^^^^
12460
12461Syntax:
12462"""""""
12463
12464This is an overloaded intrinsic. You can use ``llvm.ctlz`` on any
12465integer bit width, or any vector whose elements are integers. Not all
12466targets support all bit widths or vector types, however.
12467
12468::
12469
12470      declare i8   @llvm.ctlz.i8  (i8   <src>, i1 <is_zero_undef>)
12471      declare i16  @llvm.ctlz.i16 (i16  <src>, i1 <is_zero_undef>)
12472      declare i32  @llvm.ctlz.i32 (i32  <src>, i1 <is_zero_undef>)
12473      declare i64  @llvm.ctlz.i64 (i64  <src>, i1 <is_zero_undef>)
12474      declare i256 @llvm.ctlz.i256(i256 <src>, i1 <is_zero_undef>)
12475      declare <2 x i32> @llvm.ctlz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
12476
12477Overview:
12478"""""""""
12479
12480The '``llvm.ctlz``' family of intrinsic functions counts the number of
12481leading zeros in a variable.
12482
12483Arguments:
12484""""""""""
12485
12486The first argument is the value to be counted. This argument may be of
12487any integer type, or a vector with integer element type. The return
12488type must match the first argument type.
12489
12490The second argument must be a constant and is a flag to indicate whether
12491the intrinsic should ensure that a zero as the first argument produces a
12492defined result. Historically some architectures did not provide a
12493defined result for zero values as efficiently, and many algorithms are
12494now predicated on avoiding zero-value inputs.
12495
12496Semantics:
12497""""""""""
12498
12499The '``llvm.ctlz``' intrinsic counts the leading (most significant)
12500zeros in a variable, or within each element of the vector. If
12501``src == 0`` then the result is the size in bits of the type of ``src``
12502if ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
12503``llvm.ctlz(i32 2) = 30``.
12504
12505'``llvm.cttz.*``' Intrinsic
12506^^^^^^^^^^^^^^^^^^^^^^^^^^^
12507
12508Syntax:
12509"""""""
12510
12511This is an overloaded intrinsic. You can use ``llvm.cttz`` on any
12512integer bit width, or any vector of integer elements. Not all targets
12513support all bit widths or vector types, however.
12514
12515::
12516
12517      declare i8   @llvm.cttz.i8  (i8   <src>, i1 <is_zero_undef>)
12518      declare i16  @llvm.cttz.i16 (i16  <src>, i1 <is_zero_undef>)
12519      declare i32  @llvm.cttz.i32 (i32  <src>, i1 <is_zero_undef>)
12520      declare i64  @llvm.cttz.i64 (i64  <src>, i1 <is_zero_undef>)
12521      declare i256 @llvm.cttz.i256(i256 <src>, i1 <is_zero_undef>)
12522      declare <2 x i32> @llvm.cttz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
12523
12524Overview:
12525"""""""""
12526
12527The '``llvm.cttz``' family of intrinsic functions counts the number of
12528trailing zeros.
12529
12530Arguments:
12531""""""""""
12532
12533The first argument is the value to be counted. This argument may be of
12534any integer type, or a vector with integer element type. The return
12535type must match the first argument type.
12536
12537The second argument must be a constant and is a flag to indicate whether
12538the intrinsic should ensure that a zero as the first argument produces a
12539defined result. Historically some architectures did not provide a
12540defined result for zero values as efficiently, and many algorithms are
12541now predicated on avoiding zero-value inputs.
12542
12543Semantics:
12544""""""""""
12545
12546The '``llvm.cttz``' intrinsic counts the trailing (least significant)
12547zeros in a variable, or within each element of a vector. If ``src == 0``
12548then the result is the size in bits of the type of ``src`` if
12549``is_zero_undef == 0`` and ``undef`` otherwise. For example,
12550``llvm.cttz(2) = 1``.
12551
12552.. _int_overflow:
12553
12554'``llvm.fshl.*``' Intrinsic
12555^^^^^^^^^^^^^^^^^^^^^^^^^^^
12556
12557Syntax:
12558"""""""
12559
12560This is an overloaded intrinsic. You can use ``llvm.fshl`` on any
12561integer bit width or any vector of integer elements. Not all targets
12562support all bit widths or vector types, however.
12563
12564::
12565
12566      declare i8  @llvm.fshl.i8 (i8 %a, i8 %b, i8 %c)
12567      declare i67 @llvm.fshl.i67(i67 %a, i67 %b, i67 %c)
12568      declare <2 x i32> @llvm.fshl.v2i32(<2 x i32> %a, <2 x i32> %b, <2 x i32> %c)
12569
12570Overview:
12571"""""""""
12572
12573The '``llvm.fshl``' family of intrinsic functions performs a funnel shift left:
12574the first two values are concatenated as { %a : %b } (%a is the most significant
12575bits of the wide value), the combined value is shifted left, and the most
12576significant bits are extracted to produce a result that is the same size as the
12577original arguments. If the first 2 arguments are identical, this is equivalent
12578to a rotate left operation. For vector types, the operation occurs for each
12579element of the vector. The shift argument is treated as an unsigned amount
12580modulo the element size of the arguments.
12581
12582Arguments:
12583""""""""""
12584
12585The first two arguments are the values to be concatenated. The third
12586argument is the shift amount. The arguments may be any integer type or a
12587vector with integer element type. All arguments and the return value must
12588have the same type.
12589
12590Example:
12591""""""""
12592
12593.. code-block:: text
12594
12595      %r = call i8 @llvm.fshl.i8(i8 %x, i8 %y, i8 %z)  ; %r = i8: msb_extract((concat(x, y) << (z % 8)), 8)
12596      %r = call i8 @llvm.fshl.i8(i8 255, i8 0, i8 15)  ; %r = i8: 128 (0b10000000)
12597      %r = call i8 @llvm.fshl.i8(i8 15, i8 15, i8 11)  ; %r = i8: 120 (0b01111000)
12598      %r = call i8 @llvm.fshl.i8(i8 0, i8 255, i8 8)   ; %r = i8: 0   (0b00000000)
12599
12600'``llvm.fshr.*``' Intrinsic
12601^^^^^^^^^^^^^^^^^^^^^^^^^^^
12602
12603Syntax:
12604"""""""
12605
12606This is an overloaded intrinsic. You can use ``llvm.fshr`` on any
12607integer bit width or any vector of integer elements. Not all targets
12608support all bit widths or vector types, however.
12609
12610::
12611
12612      declare i8  @llvm.fshr.i8 (i8 %a, i8 %b, i8 %c)
12613      declare i67 @llvm.fshr.i67(i67 %a, i67 %b, i67 %c)
12614      declare <2 x i32> @llvm.fshr.v2i32(<2 x i32> %a, <2 x i32> %b, <2 x i32> %c)
12615
12616Overview:
12617"""""""""
12618
12619The '``llvm.fshr``' family of intrinsic functions performs a funnel shift right:
12620the first two values are concatenated as { %a : %b } (%a is the most significant
12621bits of the wide value), the combined value is shifted right, and the least
12622significant bits are extracted to produce a result that is the same size as the
12623original arguments. If the first 2 arguments are identical, this is equivalent
12624to a rotate right operation. For vector types, the operation occurs for each
12625element of the vector. The shift argument is treated as an unsigned amount
12626modulo the element size of the arguments.
12627
12628Arguments:
12629""""""""""
12630
12631The first two arguments are the values to be concatenated. The third
12632argument is the shift amount. The arguments may be any integer type or a
12633vector with integer element type. All arguments and the return value must
12634have the same type.
12635
12636Example:
12637""""""""
12638
12639.. code-block:: text
12640
12641      %r = call i8 @llvm.fshr.i8(i8 %x, i8 %y, i8 %z)  ; %r = i8: lsb_extract((concat(x, y) >> (z % 8)), 8)
12642      %r = call i8 @llvm.fshr.i8(i8 255, i8 0, i8 15)  ; %r = i8: 254 (0b11111110)
12643      %r = call i8 @llvm.fshr.i8(i8 15, i8 15, i8 11)  ; %r = i8: 225 (0b11100001)
12644      %r = call i8 @llvm.fshr.i8(i8 0, i8 255, i8 8)   ; %r = i8: 255 (0b11111111)
12645
12646Arithmetic with Overflow Intrinsics
12647-----------------------------------
12648
12649LLVM provides intrinsics for fast arithmetic overflow checking.
12650
12651Each of these intrinsics returns a two-element struct. The first
12652element of this struct contains the result of the corresponding
12653arithmetic operation modulo 2\ :sup:`n`\ , where n is the bit width of
12654the result. Therefore, for example, the first element of the struct
12655returned by ``llvm.sadd.with.overflow.i32`` is always the same as the
12656result of a 32-bit ``add`` instruction with the same operands, where
12657the ``add`` is *not* modified by an ``nsw`` or ``nuw`` flag.
12658
12659The second element of the result is an ``i1`` that is 1 if the
12660arithmetic operation overflowed and 0 otherwise. An operation
12661overflows if, for any values of its operands ``A`` and ``B`` and for
12662any ``N`` larger than the operands' width, ``ext(A op B) to iN`` is
12663not equal to ``(ext(A) to iN) op (ext(B) to iN)`` where ``ext`` is
12664``sext`` for signed overflow and ``zext`` for unsigned overflow, and
12665``op`` is the underlying arithmetic operation.
12666
12667The behavior of these intrinsics is well-defined for all argument
12668values.
12669
12670'``llvm.sadd.with.overflow.*``' Intrinsics
12671^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12672
12673Syntax:
12674"""""""
12675
12676This is an overloaded intrinsic. You can use ``llvm.sadd.with.overflow``
12677on any integer bit width or vectors of integers.
12678
12679::
12680
12681      declare {i16, i1} @llvm.sadd.with.overflow.i16(i16 %a, i16 %b)
12682      declare {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
12683      declare {i64, i1} @llvm.sadd.with.overflow.i64(i64 %a, i64 %b)
12684      declare {<4 x i32>, <4 x i1>} @llvm.sadd.with.overflow.v4i32(<4 x i32> %a, <4 x i32> %b)
12685
12686Overview:
12687"""""""""
12688
12689The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
12690a signed addition of the two arguments, and indicate whether an overflow
12691occurred during the signed summation.
12692
12693Arguments:
12694""""""""""
12695
12696The arguments (%a and %b) and the first element of the result structure
12697may be of integer types of any bit width, but they must have the same
12698bit width. The second element of the result structure must be of type
12699``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
12700addition.
12701
12702Semantics:
12703""""""""""
12704
12705The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
12706a signed addition of the two variables. They return a structure --- the
12707first element of which is the signed summation, and the second element
12708of which is a bit specifying if the signed summation resulted in an
12709overflow.
12710
12711Examples:
12712"""""""""
12713
12714.. code-block:: llvm
12715
12716      %res = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
12717      %sum = extractvalue {i32, i1} %res, 0
12718      %obit = extractvalue {i32, i1} %res, 1
12719      br i1 %obit, label %overflow, label %normal
12720
12721'``llvm.uadd.with.overflow.*``' Intrinsics
12722^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12723
12724Syntax:
12725"""""""
12726
12727This is an overloaded intrinsic. You can use ``llvm.uadd.with.overflow``
12728on any integer bit width or vectors of integers.
12729
12730::
12731
12732      declare {i16, i1} @llvm.uadd.with.overflow.i16(i16 %a, i16 %b)
12733      declare {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
12734      declare {i64, i1} @llvm.uadd.with.overflow.i64(i64 %a, i64 %b)
12735      declare {<4 x i32>, <4 x i1>} @llvm.uadd.with.overflow.v4i32(<4 x i32> %a, <4 x i32> %b)
12736
12737Overview:
12738"""""""""
12739
12740The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
12741an unsigned addition of the two arguments, and indicate whether a carry
12742occurred during the unsigned summation.
12743
12744Arguments:
12745""""""""""
12746
12747The arguments (%a and %b) and the first element of the result structure
12748may be of integer types of any bit width, but they must have the same
12749bit width. The second element of the result structure must be of type
12750``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
12751addition.
12752
12753Semantics:
12754""""""""""
12755
12756The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
12757an unsigned addition of the two arguments. They return a structure --- the
12758first element of which is the sum, and the second element of which is a
12759bit specifying if the unsigned summation resulted in a carry.
12760
12761Examples:
12762"""""""""
12763
12764.. code-block:: llvm
12765
12766      %res = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
12767      %sum = extractvalue {i32, i1} %res, 0
12768      %obit = extractvalue {i32, i1} %res, 1
12769      br i1 %obit, label %carry, label %normal
12770
12771'``llvm.ssub.with.overflow.*``' Intrinsics
12772^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12773
12774Syntax:
12775"""""""
12776
12777This is an overloaded intrinsic. You can use ``llvm.ssub.with.overflow``
12778on any integer bit width or vectors of integers.
12779
12780::
12781
12782      declare {i16, i1} @llvm.ssub.with.overflow.i16(i16 %a, i16 %b)
12783      declare {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
12784      declare {i64, i1} @llvm.ssub.with.overflow.i64(i64 %a, i64 %b)
12785      declare {<4 x i32>, <4 x i1>} @llvm.ssub.with.overflow.v4i32(<4 x i32> %a, <4 x i32> %b)
12786
12787Overview:
12788"""""""""
12789
12790The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
12791a signed subtraction of the two arguments, and indicate whether an
12792overflow occurred during the signed subtraction.
12793
12794Arguments:
12795""""""""""
12796
12797The arguments (%a and %b) and the first element of the result structure
12798may be of integer types of any bit width, but they must have the same
12799bit width. The second element of the result structure must be of type
12800``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
12801subtraction.
12802
12803Semantics:
12804""""""""""
12805
12806The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
12807a signed subtraction of the two arguments. They return a structure --- the
12808first element of which is the subtraction, and the second element of
12809which is a bit specifying if the signed subtraction resulted in an
12810overflow.
12811
12812Examples:
12813"""""""""
12814
12815.. code-block:: llvm
12816
12817      %res = call {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
12818      %sum = extractvalue {i32, i1} %res, 0
12819      %obit = extractvalue {i32, i1} %res, 1
12820      br i1 %obit, label %overflow, label %normal
12821
12822'``llvm.usub.with.overflow.*``' Intrinsics
12823^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12824
12825Syntax:
12826"""""""
12827
12828This is an overloaded intrinsic. You can use ``llvm.usub.with.overflow``
12829on any integer bit width or vectors of integers.
12830
12831::
12832
12833      declare {i16, i1} @llvm.usub.with.overflow.i16(i16 %a, i16 %b)
12834      declare {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
12835      declare {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
12836      declare {<4 x i32>, <4 x i1>} @llvm.usub.with.overflow.v4i32(<4 x i32> %a, <4 x i32> %b)
12837
12838Overview:
12839"""""""""
12840
12841The '``llvm.usub.with.overflow``' family of intrinsic functions perform
12842an unsigned subtraction of the two arguments, and indicate whether an
12843overflow occurred during the unsigned subtraction.
12844
12845Arguments:
12846""""""""""
12847
12848The arguments (%a and %b) and the first element of the result structure
12849may be of integer types of any bit width, but they must have the same
12850bit width. The second element of the result structure must be of type
12851``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
12852subtraction.
12853
12854Semantics:
12855""""""""""
12856
12857The '``llvm.usub.with.overflow``' family of intrinsic functions perform
12858an unsigned subtraction of the two arguments. They return a structure ---
12859the first element of which is the subtraction, and the second element of
12860which is a bit specifying if the unsigned subtraction resulted in an
12861overflow.
12862
12863Examples:
12864"""""""""
12865
12866.. code-block:: llvm
12867
12868      %res = call {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
12869      %sum = extractvalue {i32, i1} %res, 0
12870      %obit = extractvalue {i32, i1} %res, 1
12871      br i1 %obit, label %overflow, label %normal
12872
12873'``llvm.smul.with.overflow.*``' Intrinsics
12874^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12875
12876Syntax:
12877"""""""
12878
12879This is an overloaded intrinsic. You can use ``llvm.smul.with.overflow``
12880on any integer bit width or vectors of integers.
12881
12882::
12883
12884      declare {i16, i1} @llvm.smul.with.overflow.i16(i16 %a, i16 %b)
12885      declare {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
12886      declare {i64, i1} @llvm.smul.with.overflow.i64(i64 %a, i64 %b)
12887      declare {<4 x i32>, <4 x i1>} @llvm.smul.with.overflow.v4i32(<4 x i32> %a, <4 x i32> %b)
12888
12889Overview:
12890"""""""""
12891
12892The '``llvm.smul.with.overflow``' family of intrinsic functions perform
12893a signed multiplication of the two arguments, and indicate whether an
12894overflow occurred during the signed multiplication.
12895
12896Arguments:
12897""""""""""
12898
12899The arguments (%a and %b) and the first element of the result structure
12900may be of integer types of any bit width, but they must have the same
12901bit width. The second element of the result structure must be of type
12902``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
12903multiplication.
12904
12905Semantics:
12906""""""""""
12907
12908The '``llvm.smul.with.overflow``' family of intrinsic functions perform
12909a signed multiplication of the two arguments. They return a structure ---
12910the first element of which is the multiplication, and the second element
12911of which is a bit specifying if the signed multiplication resulted in an
12912overflow.
12913
12914Examples:
12915"""""""""
12916
12917.. code-block:: llvm
12918
12919      %res = call {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
12920      %sum = extractvalue {i32, i1} %res, 0
12921      %obit = extractvalue {i32, i1} %res, 1
12922      br i1 %obit, label %overflow, label %normal
12923
12924'``llvm.umul.with.overflow.*``' Intrinsics
12925^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12926
12927Syntax:
12928"""""""
12929
12930This is an overloaded intrinsic. You can use ``llvm.umul.with.overflow``
12931on any integer bit width or vectors of integers.
12932
12933::
12934
12935      declare {i16, i1} @llvm.umul.with.overflow.i16(i16 %a, i16 %b)
12936      declare {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
12937      declare {i64, i1} @llvm.umul.with.overflow.i64(i64 %a, i64 %b)
12938      declare {<4 x i32>, <4 x i1>} @llvm.umul.with.overflow.v4i32(<4 x i32> %a, <4 x i32> %b)
12939
12940Overview:
12941"""""""""
12942
12943The '``llvm.umul.with.overflow``' family of intrinsic functions perform
12944a unsigned multiplication of the two arguments, and indicate whether an
12945overflow occurred during the unsigned multiplication.
12946
12947Arguments:
12948""""""""""
12949
12950The arguments (%a and %b) and the first element of the result structure
12951may be of integer types of any bit width, but they must have the same
12952bit width. The second element of the result structure must be of type
12953``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
12954multiplication.
12955
12956Semantics:
12957""""""""""
12958
12959The '``llvm.umul.with.overflow``' family of intrinsic functions perform
12960an unsigned multiplication of the two arguments. They return a structure ---
12961the first element of which is the multiplication, and the second
12962element of which is a bit specifying if the unsigned multiplication
12963resulted in an overflow.
12964
12965Examples:
12966"""""""""
12967
12968.. code-block:: llvm
12969
12970      %res = call {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
12971      %sum = extractvalue {i32, i1} %res, 0
12972      %obit = extractvalue {i32, i1} %res, 1
12973      br i1 %obit, label %overflow, label %normal
12974
12975Saturation Arithmetic Intrinsics
12976---------------------------------
12977
12978Saturation arithmetic is a version of arithmetic in which operations are
12979limited to a fixed range between a minimum and maximum value. If the result of
12980an operation is greater than the maximum value, the result is set (or
12981"clamped") to this maximum. If it is below the minimum, it is clamped to this
12982minimum.
12983
12984
12985'``llvm.sadd.sat.*``' Intrinsics
12986^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12987
12988Syntax
12989"""""""
12990
12991This is an overloaded intrinsic. You can use ``llvm.sadd.sat``
12992on any integer bit width or vectors of integers.
12993
12994::
12995
12996      declare i16 @llvm.sadd.sat.i16(i16 %a, i16 %b)
12997      declare i32 @llvm.sadd.sat.i32(i32 %a, i32 %b)
12998      declare i64 @llvm.sadd.sat.i64(i64 %a, i64 %b)
12999      declare <4 x i32> @llvm.sadd.sat.v4i32(<4 x i32> %a, <4 x i32> %b)
13000
13001Overview
13002"""""""""
13003
13004The '``llvm.sadd.sat``' family of intrinsic functions perform signed
13005saturation addition on the 2 arguments.
13006
13007Arguments
13008""""""""""
13009
13010The arguments (%a and %b) and the result may be of integer types of any bit
13011width, but they must have the same bit width. ``%a`` and ``%b`` are the two
13012values that will undergo signed addition.
13013
13014Semantics:
13015""""""""""
13016
13017The maximum value this operation can clamp to is the largest signed value
13018representable by the bit width of the arguments. The minimum value is the
13019smallest signed value representable by this bit width.
13020
13021
13022Examples
13023"""""""""
13024
13025.. code-block:: llvm
13026
13027      %res = call i4 @llvm.sadd.sat.i4(i4 1, i4 2)  ; %res = 3
13028      %res = call i4 @llvm.sadd.sat.i4(i4 5, i4 6)  ; %res = 7
13029      %res = call i4 @llvm.sadd.sat.i4(i4 -4, i4 2)  ; %res = -2
13030      %res = call i4 @llvm.sadd.sat.i4(i4 -4, i4 -5)  ; %res = -8
13031
13032
13033'``llvm.uadd.sat.*``' Intrinsics
13034^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13035
13036Syntax
13037"""""""
13038
13039This is an overloaded intrinsic. You can use ``llvm.uadd.sat``
13040on any integer bit width or vectors of integers.
13041
13042::
13043
13044      declare i16 @llvm.uadd.sat.i16(i16 %a, i16 %b)
13045      declare i32 @llvm.uadd.sat.i32(i32 %a, i32 %b)
13046      declare i64 @llvm.uadd.sat.i64(i64 %a, i64 %b)
13047      declare <4 x i32> @llvm.uadd.sat.v4i32(<4 x i32> %a, <4 x i32> %b)
13048
13049Overview
13050"""""""""
13051
13052The '``llvm.uadd.sat``' family of intrinsic functions perform unsigned
13053saturation addition on the 2 arguments.
13054
13055Arguments
13056""""""""""
13057
13058The arguments (%a and %b) and the result may be of integer types of any bit
13059width, but they must have the same bit width. ``%a`` and ``%b`` are the two
13060values that will undergo unsigned addition.
13061
13062Semantics:
13063""""""""""
13064
13065The maximum value this operation can clamp to is the largest unsigned value
13066representable by the bit width of the arguments. Because this is an unsigned
13067operation, the result will never saturate towards zero.
13068
13069
13070Examples
13071"""""""""
13072
13073.. code-block:: llvm
13074
13075      %res = call i4 @llvm.uadd.sat.i4(i4 1, i4 2)  ; %res = 3
13076      %res = call i4 @llvm.uadd.sat.i4(i4 5, i4 6)  ; %res = 11
13077      %res = call i4 @llvm.uadd.sat.i4(i4 8, i4 8)  ; %res = 15
13078
13079
13080'``llvm.ssub.sat.*``' Intrinsics
13081^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13082
13083Syntax
13084"""""""
13085
13086This is an overloaded intrinsic. You can use ``llvm.ssub.sat``
13087on any integer bit width or vectors of integers.
13088
13089::
13090
13091      declare i16 @llvm.ssub.sat.i16(i16 %a, i16 %b)
13092      declare i32 @llvm.ssub.sat.i32(i32 %a, i32 %b)
13093      declare i64 @llvm.ssub.sat.i64(i64 %a, i64 %b)
13094      declare <4 x i32> @llvm.ssub.sat.v4i32(<4 x i32> %a, <4 x i32> %b)
13095
13096Overview
13097"""""""""
13098
13099The '``llvm.ssub.sat``' family of intrinsic functions perform signed
13100saturation subtraction on the 2 arguments.
13101
13102Arguments
13103""""""""""
13104
13105The arguments (%a and %b) and the result may be of integer types of any bit
13106width, but they must have the same bit width. ``%a`` and ``%b`` are the two
13107values that will undergo signed subtraction.
13108
13109Semantics:
13110""""""""""
13111
13112The maximum value this operation can clamp to is the largest signed value
13113representable by the bit width of the arguments. The minimum value is the
13114smallest signed value representable by this bit width.
13115
13116
13117Examples
13118"""""""""
13119
13120.. code-block:: llvm
13121
13122      %res = call i4 @llvm.ssub.sat.i4(i4 2, i4 1)  ; %res = 1
13123      %res = call i4 @llvm.ssub.sat.i4(i4 2, i4 6)  ; %res = -4
13124      %res = call i4 @llvm.ssub.sat.i4(i4 -4, i4 5)  ; %res = -8
13125      %res = call i4 @llvm.ssub.sat.i4(i4 4, i4 -5)  ; %res = 7
13126
13127
13128'``llvm.usub.sat.*``' Intrinsics
13129^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13130
13131Syntax
13132"""""""
13133
13134This is an overloaded intrinsic. You can use ``llvm.usub.sat``
13135on any integer bit width or vectors of integers.
13136
13137::
13138
13139      declare i16 @llvm.usub.sat.i16(i16 %a, i16 %b)
13140      declare i32 @llvm.usub.sat.i32(i32 %a, i32 %b)
13141      declare i64 @llvm.usub.sat.i64(i64 %a, i64 %b)
13142      declare <4 x i32> @llvm.usub.sat.v4i32(<4 x i32> %a, <4 x i32> %b)
13143
13144Overview
13145"""""""""
13146
13147The '``llvm.usub.sat``' family of intrinsic functions perform unsigned
13148saturation subtraction on the 2 arguments.
13149
13150Arguments
13151""""""""""
13152
13153The arguments (%a and %b) and the result may be of integer types of any bit
13154width, but they must have the same bit width. ``%a`` and ``%b`` are the two
13155values that will undergo unsigned subtraction.
13156
13157Semantics:
13158""""""""""
13159
13160The minimum value this operation can clamp to is 0, which is the smallest
13161unsigned value representable by the bit width of the unsigned arguments.
13162Because this is an unsigned operation, the result will never saturate towards
13163the largest possible value representable by this bit width.
13164
13165
13166Examples
13167"""""""""
13168
13169.. code-block:: llvm
13170
13171      %res = call i4 @llvm.usub.sat.i4(i4 2, i4 1)  ; %res = 1
13172      %res = call i4 @llvm.usub.sat.i4(i4 2, i4 6)  ; %res = 0
13173
13174
13175Fixed Point Arithmetic Intrinsics
13176---------------------------------
13177
13178A fixed point number represents a real data type for a number that has a fixed
13179number of digits after a radix point (equivalent to the decimal point '.').
13180The number of digits after the radix point is referred as the ``scale``. These
13181are useful for representing fractional values to a specific precision. The
13182following intrinsics perform fixed point arithmetic operations on 2 operands
13183of the same scale, specified as the third argument.
13184
13185
13186'``llvm.smul.fix.*``' Intrinsics
13187^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13188
13189Syntax
13190"""""""
13191
13192This is an overloaded intrinsic. You can use ``llvm.smul.fix``
13193on any integer bit width or vectors of integers.
13194
13195::
13196
13197      declare i16 @llvm.smul.fix.i16(i16 %a, i16 %b, i32 %scale)
13198      declare i32 @llvm.smul.fix.i32(i32 %a, i32 %b, i32 %scale)
13199      declare i64 @llvm.smul.fix.i64(i64 %a, i64 %b, i32 %scale)
13200      declare <4 x i32> @llvm.smul.fix.v4i32(<4 x i32> %a, <4 x i32> %b, i32 %scale)
13201
13202Overview
13203"""""""""
13204
13205The '``llvm.smul.fix``' family of intrinsic functions perform signed
13206fixed point multiplication on 2 arguments of the same scale.
13207
13208Arguments
13209""""""""""
13210
13211The arguments (%a and %b) and the result may be of integer types of any bit
13212width, but they must have the same bit width. The arguments may also work with
13213int vectors of the same length and int size. ``%a`` and ``%b`` are the two
13214values that will undergo signed fixed point multiplication. The argument
13215``%scale`` represents the scale of both operands, and must be a constant
13216integer.
13217
13218Semantics:
13219""""""""""
13220
13221This operation performs fixed point multiplication on the 2 arguments of a
13222specified scale. The result will also be returned in the same scale specified
13223in the third argument.
13224
13225If the result value cannot be precisely represented in the given scale, the
13226value is rounded up or down to the closest representable value. The rounding
13227direction is unspecified.
13228
13229It is undefined behavior if the result value does not fit within the range of
13230the fixed point type.
13231
13232
13233Examples
13234"""""""""
13235
13236.. code-block:: llvm
13237
13238      %res = call i4 @llvm.smul.fix.i4(i4 3, i4 2, i32 0)  ; %res = 6 (2 x 3 = 6)
13239      %res = call i4 @llvm.smul.fix.i4(i4 3, i4 2, i32 1)  ; %res = 3 (1.5 x 1 = 1.5)
13240      %res = call i4 @llvm.smul.fix.i4(i4 3, i4 -2, i32 1)  ; %res = -3 (1.5 x -1 = -1.5)
13241
13242      ; The result in the following could be rounded up to -2 or down to -2.5
13243      %res = call i4 @llvm.smul.fix.i4(i4 3, i4 -3, i32 1)  ; %res = -5 (or -4) (1.5 x -1.5 = -2.25)
13244
13245
13246'``llvm.umul.fix.*``' Intrinsics
13247^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13248
13249Syntax
13250"""""""
13251
13252This is an overloaded intrinsic. You can use ``llvm.umul.fix``
13253on any integer bit width or vectors of integers.
13254
13255::
13256
13257      declare i16 @llvm.umul.fix.i16(i16 %a, i16 %b, i32 %scale)
13258      declare i32 @llvm.umul.fix.i32(i32 %a, i32 %b, i32 %scale)
13259      declare i64 @llvm.umul.fix.i64(i64 %a, i64 %b, i32 %scale)
13260      declare <4 x i32> @llvm.umul.fix.v4i32(<4 x i32> %a, <4 x i32> %b, i32 %scale)
13261
13262Overview
13263"""""""""
13264
13265The '``llvm.umul.fix``' family of intrinsic functions perform unsigned
13266fixed point multiplication on 2 arguments of the same scale.
13267
13268Arguments
13269""""""""""
13270
13271The arguments (%a and %b) and the result may be of integer types of any bit
13272width, but they must have the same bit width. The arguments may also work with
13273int vectors of the same length and int size. ``%a`` and ``%b`` are the two
13274values that will undergo unsigned fixed point multiplication. The argument
13275``%scale`` represents the scale of both operands, and must be a constant
13276integer.
13277
13278Semantics:
13279""""""""""
13280
13281This operation performs unsigned fixed point multiplication on the 2 arguments of a
13282specified scale. The result will also be returned in the same scale specified
13283in the third argument.
13284
13285If the result value cannot be precisely represented in the given scale, the
13286value is rounded up or down to the closest representable value. The rounding
13287direction is unspecified.
13288
13289It is undefined behavior if the result value does not fit within the range of
13290the fixed point type.
13291
13292
13293Examples
13294"""""""""
13295
13296.. code-block:: llvm
13297
13298      %res = call i4 @llvm.umul.fix.i4(i4 3, i4 2, i32 0)  ; %res = 6 (2 x 3 = 6)
13299      %res = call i4 @llvm.umul.fix.i4(i4 3, i4 2, i32 1)  ; %res = 3 (1.5 x 1 = 1.5)
13300
13301      ; The result in the following could be rounded down to 3.5 or up to 4
13302      %res = call i4 @llvm.umul.fix.i4(i4 15, i4 1, i32 1)  ; %res = 7 (or 8) (7.5 x 0.5 = 3.75)
13303
13304
13305Specialised Arithmetic Intrinsics
13306---------------------------------
13307
13308'``llvm.canonicalize.*``' Intrinsic
13309^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13310
13311Syntax:
13312"""""""
13313
13314::
13315
13316      declare float @llvm.canonicalize.f32(float %a)
13317      declare double @llvm.canonicalize.f64(double %b)
13318
13319Overview:
13320"""""""""
13321
13322The '``llvm.canonicalize.*``' intrinsic returns the platform specific canonical
13323encoding of a floating-point number. This canonicalization is useful for
13324implementing certain numeric primitives such as frexp. The canonical encoding is
13325defined by IEEE-754-2008 to be:
13326
13327::
13328
13329      2.1.8 canonical encoding: The preferred encoding of a floating-point
13330      representation in a format. Applied to declets, significands of finite
13331      numbers, infinities, and NaNs, especially in decimal formats.
13332
13333This operation can also be considered equivalent to the IEEE-754-2008
13334conversion of a floating-point value to the same format. NaNs are handled
13335according to section 6.2.
13336
13337Examples of non-canonical encodings:
13338
13339- x87 pseudo denormals, pseudo NaNs, pseudo Infinity, Unnormals. These are
13340  converted to a canonical representation per hardware-specific protocol.
13341- Many normal decimal floating-point numbers have non-canonical alternative
13342  encodings.
13343- Some machines, like GPUs or ARMv7 NEON, do not support subnormal values.
13344  These are treated as non-canonical encodings of zero and will be flushed to
13345  a zero of the same sign by this operation.
13346
13347Note that per IEEE-754-2008 6.2, systems that support signaling NaNs with
13348default exception handling must signal an invalid exception, and produce a
13349quiet NaN result.
13350
13351This function should always be implementable as multiplication by 1.0, provided
13352that the compiler does not constant fold the operation. Likewise, division by
133531.0 and ``llvm.minnum(x, x)`` are possible implementations. Addition with
13354-0.0 is also sufficient provided that the rounding mode is not -Infinity.
13355
13356``@llvm.canonicalize`` must preserve the equality relation. That is:
13357
13358- ``(@llvm.canonicalize(x) == x)`` is equivalent to ``(x == x)``
13359- ``(@llvm.canonicalize(x) == @llvm.canonicalize(y))`` is equivalent to
13360  to ``(x == y)``
13361
13362Additionally, the sign of zero must be conserved:
13363``@llvm.canonicalize(-0.0) = -0.0`` and ``@llvm.canonicalize(+0.0) = +0.0``
13364
13365The payload bits of a NaN must be conserved, with two exceptions.
13366First, environments which use only a single canonical representation of NaN
13367must perform said canonicalization. Second, SNaNs must be quieted per the
13368usual methods.
13369
13370The canonicalization operation may be optimized away if:
13371
13372- The input is known to be canonical. For example, it was produced by a
13373  floating-point operation that is required by the standard to be canonical.
13374- The result is consumed only by (or fused with) other floating-point
13375  operations. That is, the bits of the floating-point value are not examined.
13376
13377'``llvm.fmuladd.*``' Intrinsic
13378^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13379
13380Syntax:
13381"""""""
13382
13383::
13384
13385      declare float @llvm.fmuladd.f32(float %a, float %b, float %c)
13386      declare double @llvm.fmuladd.f64(double %a, double %b, double %c)
13387
13388Overview:
13389"""""""""
13390
13391The '``llvm.fmuladd.*``' intrinsic functions represent multiply-add
13392expressions that can be fused if the code generator determines that (a) the
13393target instruction set has support for a fused operation, and (b) that the
13394fused operation is more efficient than the equivalent, separate pair of mul
13395and add instructions.
13396
13397Arguments:
13398""""""""""
13399
13400The '``llvm.fmuladd.*``' intrinsics each take three arguments: two
13401multiplicands, a and b, and an addend c.
13402
13403Semantics:
13404""""""""""
13405
13406The expression:
13407
13408::
13409
13410      %0 = call float @llvm.fmuladd.f32(%a, %b, %c)
13411
13412is equivalent to the expression a \* b + c, except that rounding will
13413not be performed between the multiplication and addition steps if the
13414code generator fuses the operations. Fusion is not guaranteed, even if
13415the target platform supports it. If a fused multiply-add is required the
13416corresponding llvm.fma.\* intrinsic function should be used
13417instead. This never sets errno, just as '``llvm.fma.*``'.
13418
13419Examples:
13420"""""""""
13421
13422.. code-block:: llvm
13423
13424      %r2 = call float @llvm.fmuladd.f32(float %a, float %b, float %c) ; yields float:r2 = (a * b) + c
13425
13426
13427Experimental Vector Reduction Intrinsics
13428----------------------------------------
13429
13430Horizontal reductions of vectors can be expressed using the following
13431intrinsics. Each one takes a vector operand as an input and applies its
13432respective operation across all elements of the vector, returning a single
13433scalar result of the same element type.
13434
13435
13436'``llvm.experimental.vector.reduce.add.*``' Intrinsic
13437^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13438
13439Syntax:
13440"""""""
13441
13442::
13443
13444      declare i32 @llvm.experimental.vector.reduce.add.i32.v4i32(<4 x i32> %a)
13445      declare i64 @llvm.experimental.vector.reduce.add.i64.v2i64(<2 x i64> %a)
13446
13447Overview:
13448"""""""""
13449
13450The '``llvm.experimental.vector.reduce.add.*``' intrinsics do an integer ``ADD``
13451reduction of a vector, returning the result as a scalar. The return type matches
13452the element-type of the vector input.
13453
13454Arguments:
13455""""""""""
13456The argument to this intrinsic must be a vector of integer values.
13457
13458'``llvm.experimental.vector.reduce.fadd.*``' Intrinsic
13459^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13460
13461Syntax:
13462"""""""
13463
13464::
13465
13466      declare float @llvm.experimental.vector.reduce.fadd.f32.v4f32(float %acc, <4 x float> %a)
13467      declare double @llvm.experimental.vector.reduce.fadd.f64.v2f64(double %acc, <2 x double> %a)
13468
13469Overview:
13470"""""""""
13471
13472The '``llvm.experimental.vector.reduce.fadd.*``' intrinsics do a floating-point
13473``ADD`` reduction of a vector, returning the result as a scalar. The return type
13474matches the element-type of the vector input.
13475
13476If the intrinsic call has fast-math flags, then the reduction will not preserve
13477the associativity of an equivalent scalarized counterpart. If it does not have
13478fast-math flags, then the reduction will be *ordered*, implying that the
13479operation respects the associativity of a scalarized reduction.
13480
13481
13482Arguments:
13483""""""""""
13484The first argument to this intrinsic is a scalar accumulator value, which is
13485only used when there are no fast-math flags attached. This argument may be undef
13486when fast-math flags are used.
13487
13488The second argument must be a vector of floating-point values.
13489
13490Examples:
13491"""""""""
13492
13493.. code-block:: llvm
13494
13495      %fast = call fast float @llvm.experimental.vector.reduce.fadd.f32.v4f32(float undef, <4 x float> %input) ; fast reduction
13496      %ord = call float @llvm.experimental.vector.reduce.fadd.f32.v4f32(float %acc, <4 x float> %input) ; ordered reduction
13497
13498
13499'``llvm.experimental.vector.reduce.mul.*``' Intrinsic
13500^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13501
13502Syntax:
13503"""""""
13504
13505::
13506
13507      declare i32 @llvm.experimental.vector.reduce.mul.i32.v4i32(<4 x i32> %a)
13508      declare i64 @llvm.experimental.vector.reduce.mul.i64.v2i64(<2 x i64> %a)
13509
13510Overview:
13511"""""""""
13512
13513The '``llvm.experimental.vector.reduce.mul.*``' intrinsics do an integer ``MUL``
13514reduction of a vector, returning the result as a scalar. The return type matches
13515the element-type of the vector input.
13516
13517Arguments:
13518""""""""""
13519The argument to this intrinsic must be a vector of integer values.
13520
13521'``llvm.experimental.vector.reduce.fmul.*``' Intrinsic
13522^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13523
13524Syntax:
13525"""""""
13526
13527::
13528
13529      declare float @llvm.experimental.vector.reduce.fmul.f32.v4f32(float %acc, <4 x float> %a)
13530      declare double @llvm.experimental.vector.reduce.fmul.f64.v2f64(double %acc, <2 x double> %a)
13531
13532Overview:
13533"""""""""
13534
13535The '``llvm.experimental.vector.reduce.fmul.*``' intrinsics do a floating-point
13536``MUL`` reduction of a vector, returning the result as a scalar. The return type
13537matches the element-type of the vector input.
13538
13539If the intrinsic call has fast-math flags, then the reduction will not preserve
13540the associativity of an equivalent scalarized counterpart. If it does not have
13541fast-math flags, then the reduction will be *ordered*, implying that the
13542operation respects the associativity of a scalarized reduction.
13543
13544
13545Arguments:
13546""""""""""
13547The first argument to this intrinsic is a scalar accumulator value, which is
13548only used when there are no fast-math flags attached. This argument may be undef
13549when fast-math flags are used.
13550
13551The second argument must be a vector of floating-point values.
13552
13553Examples:
13554"""""""""
13555
13556.. code-block:: llvm
13557
13558      %fast = call fast float @llvm.experimental.vector.reduce.fmul.f32.v4f32(float undef, <4 x float> %input) ; fast reduction
13559      %ord = call float @llvm.experimental.vector.reduce.fmul.f32.v4f32(float %acc, <4 x float> %input) ; ordered reduction
13560
13561'``llvm.experimental.vector.reduce.and.*``' Intrinsic
13562^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13563
13564Syntax:
13565"""""""
13566
13567::
13568
13569      declare i32 @llvm.experimental.vector.reduce.and.i32.v4i32(<4 x i32> %a)
13570
13571Overview:
13572"""""""""
13573
13574The '``llvm.experimental.vector.reduce.and.*``' intrinsics do a bitwise ``AND``
13575reduction of a vector, returning the result as a scalar. The return type matches
13576the element-type of the vector input.
13577
13578Arguments:
13579""""""""""
13580The argument to this intrinsic must be a vector of integer values.
13581
13582'``llvm.experimental.vector.reduce.or.*``' Intrinsic
13583^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13584
13585Syntax:
13586"""""""
13587
13588::
13589
13590      declare i32 @llvm.experimental.vector.reduce.or.i32.v4i32(<4 x i32> %a)
13591
13592Overview:
13593"""""""""
13594
13595The '``llvm.experimental.vector.reduce.or.*``' intrinsics do a bitwise ``OR`` reduction
13596of a vector, returning the result as a scalar. The return type matches the
13597element-type of the vector input.
13598
13599Arguments:
13600""""""""""
13601The argument to this intrinsic must be a vector of integer values.
13602
13603'``llvm.experimental.vector.reduce.xor.*``' Intrinsic
13604^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13605
13606Syntax:
13607"""""""
13608
13609::
13610
13611      declare i32 @llvm.experimental.vector.reduce.xor.i32.v4i32(<4 x i32> %a)
13612
13613Overview:
13614"""""""""
13615
13616The '``llvm.experimental.vector.reduce.xor.*``' intrinsics do a bitwise ``XOR``
13617reduction of a vector, returning the result as a scalar. The return type matches
13618the element-type of the vector input.
13619
13620Arguments:
13621""""""""""
13622The argument to this intrinsic must be a vector of integer values.
13623
13624'``llvm.experimental.vector.reduce.smax.*``' Intrinsic
13625^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13626
13627Syntax:
13628"""""""
13629
13630::
13631
13632      declare i32 @llvm.experimental.vector.reduce.smax.i32.v4i32(<4 x i32> %a)
13633
13634Overview:
13635"""""""""
13636
13637The '``llvm.experimental.vector.reduce.smax.*``' intrinsics do a signed integer
13638``MAX`` reduction of a vector, returning the result as a scalar. The return type
13639matches the element-type of the vector input.
13640
13641Arguments:
13642""""""""""
13643The argument to this intrinsic must be a vector of integer values.
13644
13645'``llvm.experimental.vector.reduce.smin.*``' Intrinsic
13646^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13647
13648Syntax:
13649"""""""
13650
13651::
13652
13653      declare i32 @llvm.experimental.vector.reduce.smin.i32.v4i32(<4 x i32> %a)
13654
13655Overview:
13656"""""""""
13657
13658The '``llvm.experimental.vector.reduce.smin.*``' intrinsics do a signed integer
13659``MIN`` reduction of a vector, returning the result as a scalar. The return type
13660matches the element-type of the vector input.
13661
13662Arguments:
13663""""""""""
13664The argument to this intrinsic must be a vector of integer values.
13665
13666'``llvm.experimental.vector.reduce.umax.*``' Intrinsic
13667^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13668
13669Syntax:
13670"""""""
13671
13672::
13673
13674      declare i32 @llvm.experimental.vector.reduce.umax.i32.v4i32(<4 x i32> %a)
13675
13676Overview:
13677"""""""""
13678
13679The '``llvm.experimental.vector.reduce.umax.*``' intrinsics do an unsigned
13680integer ``MAX`` reduction of a vector, returning the result as a scalar. The
13681return type matches the element-type of the vector input.
13682
13683Arguments:
13684""""""""""
13685The argument to this intrinsic must be a vector of integer values.
13686
13687'``llvm.experimental.vector.reduce.umin.*``' Intrinsic
13688^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13689
13690Syntax:
13691"""""""
13692
13693::
13694
13695      declare i32 @llvm.experimental.vector.reduce.umin.i32.v4i32(<4 x i32> %a)
13696
13697Overview:
13698"""""""""
13699
13700The '``llvm.experimental.vector.reduce.umin.*``' intrinsics do an unsigned
13701integer ``MIN`` reduction of a vector, returning the result as a scalar. The
13702return type matches the element-type of the vector input.
13703
13704Arguments:
13705""""""""""
13706The argument to this intrinsic must be a vector of integer values.
13707
13708'``llvm.experimental.vector.reduce.fmax.*``' Intrinsic
13709^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13710
13711Syntax:
13712"""""""
13713
13714::
13715
13716      declare float @llvm.experimental.vector.reduce.fmax.f32.v4f32(<4 x float> %a)
13717      declare double @llvm.experimental.vector.reduce.fmax.f64.v2f64(<2 x double> %a)
13718
13719Overview:
13720"""""""""
13721
13722The '``llvm.experimental.vector.reduce.fmax.*``' intrinsics do a floating-point
13723``MAX`` reduction of a vector, returning the result as a scalar. The return type
13724matches the element-type of the vector input.
13725
13726If the intrinsic call has the ``nnan`` fast-math flag then the operation can
13727assume that NaNs are not present in the input vector.
13728
13729Arguments:
13730""""""""""
13731The argument to this intrinsic must be a vector of floating-point values.
13732
13733'``llvm.experimental.vector.reduce.fmin.*``' Intrinsic
13734^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13735
13736Syntax:
13737"""""""
13738
13739::
13740
13741      declare float @llvm.experimental.vector.reduce.fmin.f32.v4f32(<4 x float> %a)
13742      declare double @llvm.experimental.vector.reduce.fmin.f64.v2f64(<2 x double> %a)
13743
13744Overview:
13745"""""""""
13746
13747The '``llvm.experimental.vector.reduce.fmin.*``' intrinsics do a floating-point
13748``MIN`` reduction of a vector, returning the result as a scalar. The return type
13749matches the element-type of the vector input.
13750
13751If the intrinsic call has the ``nnan`` fast-math flag then the operation can
13752assume that NaNs are not present in the input vector.
13753
13754Arguments:
13755""""""""""
13756The argument to this intrinsic must be a vector of floating-point values.
13757
13758Half Precision Floating-Point Intrinsics
13759----------------------------------------
13760
13761For most target platforms, half precision floating-point is a
13762storage-only format. This means that it is a dense encoding (in memory)
13763but does not support computation in the format.
13764
13765This means that code must first load the half-precision floating-point
13766value as an i16, then convert it to float with
13767:ref:`llvm.convert.from.fp16 <int_convert_from_fp16>`. Computation can
13768then be performed on the float value (including extending to double
13769etc). To store the value back to memory, it is first converted to float
13770if needed, then converted to i16 with
13771:ref:`llvm.convert.to.fp16 <int_convert_to_fp16>`, then storing as an
13772i16 value.
13773
13774.. _int_convert_to_fp16:
13775
13776'``llvm.convert.to.fp16``' Intrinsic
13777^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13778
13779Syntax:
13780"""""""
13781
13782::
13783
13784      declare i16 @llvm.convert.to.fp16.f32(float %a)
13785      declare i16 @llvm.convert.to.fp16.f64(double %a)
13786
13787Overview:
13788"""""""""
13789
13790The '``llvm.convert.to.fp16``' intrinsic function performs a conversion from a
13791conventional floating-point type to half precision floating-point format.
13792
13793Arguments:
13794""""""""""
13795
13796The intrinsic function contains single argument - the value to be
13797converted.
13798
13799Semantics:
13800""""""""""
13801
13802The '``llvm.convert.to.fp16``' intrinsic function performs a conversion from a
13803conventional floating-point format to half precision floating-point format. The
13804return value is an ``i16`` which contains the converted number.
13805
13806Examples:
13807"""""""""
13808
13809.. code-block:: llvm
13810
13811      %res = call i16 @llvm.convert.to.fp16.f32(float %a)
13812      store i16 %res, i16* @x, align 2
13813
13814.. _int_convert_from_fp16:
13815
13816'``llvm.convert.from.fp16``' Intrinsic
13817^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13818
13819Syntax:
13820"""""""
13821
13822::
13823
13824      declare float @llvm.convert.from.fp16.f32(i16 %a)
13825      declare double @llvm.convert.from.fp16.f64(i16 %a)
13826
13827Overview:
13828"""""""""
13829
13830The '``llvm.convert.from.fp16``' intrinsic function performs a
13831conversion from half precision floating-point format to single precision
13832floating-point format.
13833
13834Arguments:
13835""""""""""
13836
13837The intrinsic function contains single argument - the value to be
13838converted.
13839
13840Semantics:
13841""""""""""
13842
13843The '``llvm.convert.from.fp16``' intrinsic function performs a
13844conversion from half single precision floating-point format to single
13845precision floating-point format. The input half-float value is
13846represented by an ``i16`` value.
13847
13848Examples:
13849"""""""""
13850
13851.. code-block:: llvm
13852
13853      %a = load i16, i16* @x, align 2
13854      %res = call float @llvm.convert.from.fp16(i16 %a)
13855
13856.. _dbg_intrinsics:
13857
13858Debugger Intrinsics
13859-------------------
13860
13861The LLVM debugger intrinsics (which all start with ``llvm.dbg.``
13862prefix), are described in the `LLVM Source Level
13863Debugging <SourceLevelDebugging.html#format-common-intrinsics>`_
13864document.
13865
13866Exception Handling Intrinsics
13867-----------------------------
13868
13869The LLVM exception handling intrinsics (which all start with
13870``llvm.eh.`` prefix), are described in the `LLVM Exception
13871Handling <ExceptionHandling.html#format-common-intrinsics>`_ document.
13872
13873.. _int_trampoline:
13874
13875Trampoline Intrinsics
13876---------------------
13877
13878These intrinsics make it possible to excise one parameter, marked with
13879the :ref:`nest <nest>` attribute, from a function. The result is a
13880callable function pointer lacking the nest parameter - the caller does
13881not need to provide a value for it. Instead, the value to use is stored
13882in advance in a "trampoline", a block of memory usually allocated on the
13883stack, which also contains code to splice the nest value into the
13884argument list. This is used to implement the GCC nested function address
13885extension.
13886
13887For example, if the function is ``i32 f(i8* nest %c, i32 %x, i32 %y)``
13888then the resulting function pointer has signature ``i32 (i32, i32)*``.
13889It can be created as follows:
13890
13891.. code-block:: llvm
13892
13893      %tramp = alloca [10 x i8], align 4 ; size and alignment only correct for X86
13894      %tramp1 = getelementptr [10 x i8], [10 x i8]* %tramp, i32 0, i32 0
13895      call i8* @llvm.init.trampoline(i8* %tramp1, i8* bitcast (i32 (i8*, i32, i32)* @f to i8*), i8* %nval)
13896      %p = call i8* @llvm.adjust.trampoline(i8* %tramp1)
13897      %fp = bitcast i8* %p to i32 (i32, i32)*
13898
13899The call ``%val = call i32 %fp(i32 %x, i32 %y)`` is then equivalent to
13900``%val = call i32 %f(i8* %nval, i32 %x, i32 %y)``.
13901
13902.. _int_it:
13903
13904'``llvm.init.trampoline``' Intrinsic
13905^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13906
13907Syntax:
13908"""""""
13909
13910::
13911
13912      declare void @llvm.init.trampoline(i8* <tramp>, i8* <func>, i8* <nval>)
13913
13914Overview:
13915"""""""""
13916
13917This fills the memory pointed to by ``tramp`` with executable code,
13918turning it into a trampoline.
13919
13920Arguments:
13921""""""""""
13922
13923The ``llvm.init.trampoline`` intrinsic takes three arguments, all
13924pointers. The ``tramp`` argument must point to a sufficiently large and
13925sufficiently aligned block of memory; this memory is written to by the
13926intrinsic. Note that the size and the alignment are target-specific -
13927LLVM currently provides no portable way of determining them, so a
13928front-end that generates this intrinsic needs to have some
13929target-specific knowledge. The ``func`` argument must hold a function
13930bitcast to an ``i8*``.
13931
13932Semantics:
13933""""""""""
13934
13935The block of memory pointed to by ``tramp`` is filled with target
13936dependent code, turning it into a function. Then ``tramp`` needs to be
13937passed to :ref:`llvm.adjust.trampoline <int_at>` to get a pointer which can
13938be :ref:`bitcast (to a new function) and called <int_trampoline>`. The new
13939function's signature is the same as that of ``func`` with any arguments
13940marked with the ``nest`` attribute removed. At most one such ``nest``
13941argument is allowed, and it must be of pointer type. Calling the new
13942function is equivalent to calling ``func`` with the same argument list,
13943but with ``nval`` used for the missing ``nest`` argument. If, after
13944calling ``llvm.init.trampoline``, the memory pointed to by ``tramp`` is
13945modified, then the effect of any later call to the returned function
13946pointer is undefined.
13947
13948.. _int_at:
13949
13950'``llvm.adjust.trampoline``' Intrinsic
13951^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13952
13953Syntax:
13954"""""""
13955
13956::
13957
13958      declare i8* @llvm.adjust.trampoline(i8* <tramp>)
13959
13960Overview:
13961"""""""""
13962
13963This performs any required machine-specific adjustment to the address of
13964a trampoline (passed as ``tramp``).
13965
13966Arguments:
13967""""""""""
13968
13969``tramp`` must point to a block of memory which already has trampoline
13970code filled in by a previous call to
13971:ref:`llvm.init.trampoline <int_it>`.
13972
13973Semantics:
13974""""""""""
13975
13976On some architectures the address of the code to be executed needs to be
13977different than the address where the trampoline is actually stored. This
13978intrinsic returns the executable address corresponding to ``tramp``
13979after performing the required machine specific adjustments. The pointer
13980returned can then be :ref:`bitcast and executed <int_trampoline>`.
13981
13982.. _int_mload_mstore:
13983
13984Masked Vector Load and Store Intrinsics
13985---------------------------------------
13986
13987LLVM 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.
13988
13989.. _int_mload:
13990
13991'``llvm.masked.load.*``' Intrinsics
13992^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13993
13994Syntax:
13995"""""""
13996This is an overloaded intrinsic. The loaded data is a vector of any integer, floating-point or pointer data type.
13997
13998::
13999
14000      declare <16 x float>  @llvm.masked.load.v16f32.p0v16f32 (<16 x float>* <ptr>, i32 <alignment>, <16 x i1> <mask>, <16 x float> <passthru>)
14001      declare <2 x double>  @llvm.masked.load.v2f64.p0v2f64  (<2 x double>* <ptr>, i32 <alignment>, <2 x i1>  <mask>, <2 x double> <passthru>)
14002      ;; The data is a vector of pointers to double
14003      declare <8 x double*> @llvm.masked.load.v8p0f64.p0v8p0f64    (<8 x double*>* <ptr>, i32 <alignment>, <8 x i1> <mask>, <8 x double*> <passthru>)
14004      ;; The data is a vector of function pointers
14005      declare <8 x i32 ()*> @llvm.masked.load.v8p0f_i32f.p0v8p0f_i32f (<8 x i32 ()*>* <ptr>, i32 <alignment>, <8 x i1> <mask>, <8 x i32 ()*> <passthru>)
14006
14007Overview:
14008"""""""""
14009
14010Reads 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.
14011
14012
14013Arguments:
14014""""""""""
14015
14016The 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.
14017
14018
14019Semantics:
14020""""""""""
14021
14022The '``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.
14023The 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.
14024
14025
14026::
14027
14028       %res = call <16 x float> @llvm.masked.load.v16f32.p0v16f32 (<16 x float>* %ptr, i32 4, <16 x i1>%mask, <16 x float> %passthru)
14029
14030       ;; The result of the two following instructions is identical aside from potential memory access exception
14031       %loadlal = load <16 x float>, <16 x float>* %ptr, align 4
14032       %res = select <16 x i1> %mask, <16 x float> %loadlal, <16 x float> %passthru
14033
14034.. _int_mstore:
14035
14036'``llvm.masked.store.*``' Intrinsics
14037^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
14038
14039Syntax:
14040"""""""
14041This is an overloaded intrinsic. The data stored in memory is a vector of any integer, floating-point or pointer data type.
14042
14043::
14044
14045       declare void @llvm.masked.store.v8i32.p0v8i32  (<8  x i32>   <value>, <8  x i32>*   <ptr>, i32 <alignment>,  <8  x i1> <mask>)
14046       declare void @llvm.masked.store.v16f32.p0v16f32 (<16 x float> <value>, <16 x float>* <ptr>, i32 <alignment>,  <16 x i1> <mask>)
14047       ;; The data is a vector of pointers to double
14048       declare void @llvm.masked.store.v8p0f64.p0v8p0f64    (<8 x double*> <value>, <8 x double*>* <ptr>, i32 <alignment>, <8 x i1> <mask>)
14049       ;; The data is a vector of function pointers
14050       declare void @llvm.masked.store.v4p0f_i32f.p0v4p0f_i32f (<4 x i32 ()*> <value>, <4 x i32 ()*>* <ptr>, i32 <alignment>, <4 x i1> <mask>)
14051
14052Overview:
14053"""""""""
14054
14055Writes 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.
14056
14057Arguments:
14058""""""""""
14059
14060The 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.
14061
14062
14063Semantics:
14064""""""""""
14065
14066The '``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.
14067The 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.
14068
14069::
14070
14071       call void @llvm.masked.store.v16f32.p0v16f32(<16 x float> %value, <16 x float>* %ptr, i32 4,  <16 x i1> %mask)
14072
14073       ;; The result of the following instructions is identical aside from potential data races and memory access exceptions
14074       %oldval = load <16 x float>, <16 x float>* %ptr, align 4
14075       %res = select <16 x i1> %mask, <16 x float> %value, <16 x float> %oldval
14076       store <16 x float> %res, <16 x float>* %ptr, align 4
14077
14078
14079Masked Vector Gather and Scatter Intrinsics
14080-------------------------------------------
14081
14082LLVM 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.
14083
14084.. _int_mgather:
14085
14086'``llvm.masked.gather.*``' Intrinsics
14087^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
14088
14089Syntax:
14090"""""""
14091This 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.
14092
14093::
14094
14095      declare <16 x float> @llvm.masked.gather.v16f32.v16p0f32   (<16 x float*> <ptrs>, i32 <alignment>, <16 x i1> <mask>, <16 x float> <passthru>)
14096      declare <2 x double> @llvm.masked.gather.v2f64.v2p1f64     (<2 x double addrspace(1)*> <ptrs>, i32 <alignment>, <2 x i1>  <mask>, <2 x double> <passthru>)
14097      declare <8 x float*> @llvm.masked.gather.v8p0f32.v8p0p0f32 (<8 x float**> <ptrs>, i32 <alignment>, <8 x i1>  <mask>, <8 x float*> <passthru>)
14098
14099Overview:
14100"""""""""
14101
14102Reads 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.
14103
14104
14105Arguments:
14106""""""""""
14107
14108The 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.
14109
14110
14111Semantics:
14112""""""""""
14113
14114The '``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.
14115The 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.
14116
14117
14118::
14119
14120       %res = call <4 x double> @llvm.masked.gather.v4f64.v4p0f64 (<4 x double*> %ptrs, i32 8, <4 x i1> <i1 true, i1 true, i1 true, i1 true>, <4 x double> undef)
14121
14122       ;; The gather with all-true mask is equivalent to the following instruction sequence
14123       %ptr0 = extractelement <4 x double*> %ptrs, i32 0
14124       %ptr1 = extractelement <4 x double*> %ptrs, i32 1
14125       %ptr2 = extractelement <4 x double*> %ptrs, i32 2
14126       %ptr3 = extractelement <4 x double*> %ptrs, i32 3
14127
14128       %val0 = load double, double* %ptr0, align 8
14129       %val1 = load double, double* %ptr1, align 8
14130       %val2 = load double, double* %ptr2, align 8
14131       %val3 = load double, double* %ptr3, align 8
14132
14133       %vec0    = insertelement <4 x double>undef, %val0, 0
14134       %vec01   = insertelement <4 x double>%vec0, %val1, 1
14135       %vec012  = insertelement <4 x double>%vec01, %val2, 2
14136       %vec0123 = insertelement <4 x double>%vec012, %val3, 3
14137
14138.. _int_mscatter:
14139
14140'``llvm.masked.scatter.*``' Intrinsics
14141^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
14142
14143Syntax:
14144"""""""
14145This 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.
14146
14147::
14148
14149       declare void @llvm.masked.scatter.v8i32.v8p0i32     (<8 x i32>     <value>, <8 x i32*>     <ptrs>, i32 <alignment>, <8 x i1>  <mask>)
14150       declare void @llvm.masked.scatter.v16f32.v16p1f32   (<16 x float>  <value>, <16 x float addrspace(1)*>  <ptrs>, i32 <alignment>, <16 x i1> <mask>)
14151       declare void @llvm.masked.scatter.v4p0f64.v4p0p0f64 (<4 x double*> <value>, <4 x double**> <ptrs>, i32 <alignment>, <4 x i1>  <mask>)
14152
14153Overview:
14154"""""""""
14155
14156Writes 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.
14157
14158Arguments:
14159""""""""""
14160
14161The 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.
14162
14163
14164Semantics:
14165""""""""""
14166
14167The '``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.
14168
14169::
14170
14171       ;; This instruction unconditionally stores data vector in multiple addresses
14172       call @llvm.masked.scatter.v8i32.v8p0i32 (<8 x i32> %value, <8 x i32*> %ptrs, i32 4,  <8 x i1>  <true, true, .. true>)
14173
14174       ;; It is equivalent to a list of scalar stores
14175       %val0 = extractelement <8 x i32> %value, i32 0
14176       %val1 = extractelement <8 x i32> %value, i32 1
14177       ..
14178       %val7 = extractelement <8 x i32> %value, i32 7
14179       %ptr0 = extractelement <8 x i32*> %ptrs, i32 0
14180       %ptr1 = extractelement <8 x i32*> %ptrs, i32 1
14181       ..
14182       %ptr7 = extractelement <8 x i32*> %ptrs, i32 7
14183       ;; Note: the order of the following stores is important when they overlap:
14184       store i32 %val0, i32* %ptr0, align 4
14185       store i32 %val1, i32* %ptr1, align 4
14186       ..
14187       store i32 %val7, i32* %ptr7, align 4
14188
14189
14190Masked Vector Expanding Load and Compressing Store Intrinsics
14191-------------------------------------------------------------
14192
14193LLVM provides intrinsics for expanding load and compressing store operations. Data selected from a vector according to a mask is stored in consecutive memory addresses (compressed store), and vice-versa (expanding load). These operations effective map to "if (cond.i) a[j++] = v.i" and "if (cond.i) v.i = a[j++]" patterns, respectively. Note that when the mask starts with '1' bits followed by '0' bits, these operations are identical to :ref:`llvm.masked.store <int_mstore>` and :ref:`llvm.masked.load <int_mload>`.
14194
14195.. _int_expandload:
14196
14197'``llvm.masked.expandload.*``' Intrinsics
14198^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
14199
14200Syntax:
14201"""""""
14202This is an overloaded intrinsic. Several values of integer, floating point or pointer data type are loaded from consecutive memory addresses and stored into the elements of a vector according to the mask.
14203
14204::
14205
14206      declare <16 x float>  @llvm.masked.expandload.v16f32 (float* <ptr>, <16 x i1> <mask>, <16 x float> <passthru>)
14207      declare <2 x i64>     @llvm.masked.expandload.v2i64 (i64* <ptr>, <2 x i1>  <mask>, <2 x i64> <passthru>)
14208
14209Overview:
14210"""""""""
14211
14212Reads a number of scalar values sequentially from memory location provided in '``ptr``' and spreads them in a vector. The '``mask``' holds a bit for each vector lane. The number of elements read from memory is equal to the number of '1' bits in the mask. The loaded elements are positioned in the destination vector according to the sequence of '1' and '0' bits in the mask. E.g., if the mask vector is '10010001', "explandload" reads 3 values from memory addresses ptr, ptr+1, ptr+2 and places them in lanes 0, 3 and 7 accordingly. The masked-off lanes are filled by elements from the corresponding lanes of the '``passthru``' operand.
14213
14214
14215Arguments:
14216""""""""""
14217
14218The first operand is the base pointer for the load. It has the same underlying type as the element of the returned vector. The second operand, mask, is a vector of boolean values with the same number of elements as the return type. The third is a pass-through value that is used to fill the masked-off lanes of the result. The return type and the type of the '``passthru``' operand have the same vector type.
14219
14220Semantics:
14221""""""""""
14222
14223The '``llvm.masked.expandload``' intrinsic is designed for reading multiple scalar values from adjacent memory addresses into possibly non-adjacent vector lanes. It is useful for targets that support vector expanding loads and allows vectorizing loop with cross-iteration dependency like in the following example:
14224
14225.. code-block:: c
14226
14227    // In this loop we load from B and spread the elements into array A.
14228    double *A, B; int *C;
14229    for (int i = 0; i < size; ++i) {
14230      if (C[i] != 0)
14231        A[i] = B[j++];
14232    }
14233
14234
14235.. code-block:: llvm
14236
14237    ; Load several elements from array B and expand them in a vector.
14238    ; The number of loaded elements is equal to the number of '1' elements in the Mask.
14239    %Tmp = call <8 x double> @llvm.masked.expandload.v8f64(double* %Bptr, <8 x i1> %Mask, <8 x double> undef)
14240    ; Store the result in A
14241    call void @llvm.masked.store.v8f64.p0v8f64(<8 x double> %Tmp, <8 x double>* %Aptr, i32 8, <8 x i1> %Mask)
14242
14243    ; %Bptr should be increased on each iteration according to the number of '1' elements in the Mask.
14244    %MaskI = bitcast <8 x i1> %Mask to i8
14245    %MaskIPopcnt = call i8 @llvm.ctpop.i8(i8 %MaskI)
14246    %MaskI64 = zext i8 %MaskIPopcnt to i64
14247    %BNextInd = add i64 %BInd, %MaskI64
14248
14249
14250Other targets may support this intrinsic differently, for example, by lowering it into a sequence of conditional scalar load operations and shuffles.
14251If all mask elements are '1', the intrinsic behavior is equivalent to the regular unmasked vector load.
14252
14253.. _int_compressstore:
14254
14255'``llvm.masked.compressstore.*``' Intrinsics
14256^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
14257
14258Syntax:
14259"""""""
14260This is an overloaded intrinsic. A number of scalar values of integer, floating point or pointer data type are collected from an input vector and stored into adjacent memory addresses. A mask defines which elements to collect from the vector.
14261
14262::
14263
14264      declare void @llvm.masked.compressstore.v8i32  (<8  x i32>   <value>, i32*   <ptr>, <8  x i1> <mask>)
14265      declare void @llvm.masked.compressstore.v16f32 (<16 x float> <value>, float* <ptr>, <16 x i1> <mask>)
14266
14267Overview:
14268"""""""""
14269
14270Selects elements from input vector '``value``' according to the '``mask``'. All selected elements are written into adjacent memory addresses starting at address '`ptr`', from lower to higher. The mask holds a bit for each vector lane, and is used to select elements to be stored. The number of elements to be stored is equal to the number of active bits in the mask.
14271
14272Arguments:
14273""""""""""
14274
14275The first operand is the input vector, from which elements are collected and written to memory. The second operand is the base pointer for the store, it has the same underlying type as the element of the input vector operand. The third operand is the mask, a vector of boolean values. The mask and the input vector must have the same number of vector elements.
14276
14277
14278Semantics:
14279""""""""""
14280
14281The '``llvm.masked.compressstore``' intrinsic is designed for compressing data in memory. It allows to collect elements from possibly non-adjacent lanes of a vector and store them contiguously in memory in one IR operation. It is useful for targets that support compressing store operations and allows vectorizing loops with cross-iteration dependences like in the following example:
14282
14283.. code-block:: c
14284
14285    // In this loop we load elements from A and store them consecutively in B
14286    double *A, B; int *C;
14287    for (int i = 0; i < size; ++i) {
14288      if (C[i] != 0)
14289        B[j++] = A[i]
14290    }
14291
14292
14293.. code-block:: llvm
14294
14295    ; Load elements from A.
14296    %Tmp = call <8 x double> @llvm.masked.load.v8f64.p0v8f64(<8 x double>* %Aptr, i32 8, <8 x i1> %Mask, <8 x double> undef)
14297    ; Store all selected elements consecutively in array B
14298    call <void> @llvm.masked.compressstore.v8f64(<8 x double> %Tmp, double* %Bptr, <8 x i1> %Mask)
14299
14300    ; %Bptr should be increased on each iteration according to the number of '1' elements in the Mask.
14301    %MaskI = bitcast <8 x i1> %Mask to i8
14302    %MaskIPopcnt = call i8 @llvm.ctpop.i8(i8 %MaskI)
14303    %MaskI64 = zext i8 %MaskIPopcnt to i64
14304    %BNextInd = add i64 %BInd, %MaskI64
14305
14306
14307Other targets may support this intrinsic differently, for example, by lowering it into a sequence of branches that guard scalar store operations.
14308
14309
14310Memory Use Markers
14311------------------
14312
14313This class of intrinsics provides information about the lifetime of
14314memory objects and ranges where variables are immutable.
14315
14316.. _int_lifestart:
14317
14318'``llvm.lifetime.start``' Intrinsic
14319^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
14320
14321Syntax:
14322"""""""
14323
14324::
14325
14326      declare void @llvm.lifetime.start(i64 <size>, i8* nocapture <ptr>)
14327
14328Overview:
14329"""""""""
14330
14331The '``llvm.lifetime.start``' intrinsic specifies the start of a memory
14332object's lifetime.
14333
14334Arguments:
14335""""""""""
14336
14337The first argument is a constant integer representing the size of the
14338object, or -1 if it is variable sized. The second argument is a pointer
14339to the object.
14340
14341Semantics:
14342""""""""""
14343
14344This intrinsic indicates that before this point in the code, the value
14345of the memory pointed to by ``ptr`` is dead. This means that it is known
14346to never be used and has an undefined value. A load from the pointer
14347that precedes this intrinsic can be replaced with ``'undef'``.
14348
14349.. _int_lifeend:
14350
14351'``llvm.lifetime.end``' Intrinsic
14352^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
14353
14354Syntax:
14355"""""""
14356
14357::
14358
14359      declare void @llvm.lifetime.end(i64 <size>, i8* nocapture <ptr>)
14360
14361Overview:
14362"""""""""
14363
14364The '``llvm.lifetime.end``' intrinsic specifies the end of a memory
14365object's lifetime.
14366
14367Arguments:
14368""""""""""
14369
14370The first argument is a constant integer representing the size of the
14371object, or -1 if it is variable sized. The second argument is a pointer
14372to the object.
14373
14374Semantics:
14375""""""""""
14376
14377This intrinsic indicates that after this point in the code, the value of
14378the memory pointed to by ``ptr`` is dead. This means that it is known to
14379never be used and has an undefined value. Any stores into the memory
14380object following this intrinsic may be removed as dead.
14381
14382'``llvm.invariant.start``' Intrinsic
14383^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
14384
14385Syntax:
14386"""""""
14387This is an overloaded intrinsic. The memory object can belong to any address space.
14388
14389::
14390
14391      declare {}* @llvm.invariant.start.p0i8(i64 <size>, i8* nocapture <ptr>)
14392
14393Overview:
14394"""""""""
14395
14396The '``llvm.invariant.start``' intrinsic specifies that the contents of
14397a memory object will not change.
14398
14399Arguments:
14400""""""""""
14401
14402The first argument is a constant integer representing the size of the
14403object, or -1 if it is variable sized. The second argument is a pointer
14404to the object.
14405
14406Semantics:
14407""""""""""
14408
14409This intrinsic indicates that until an ``llvm.invariant.end`` that uses
14410the return value, the referenced memory location is constant and
14411unchanging.
14412
14413'``llvm.invariant.end``' Intrinsic
14414^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
14415
14416Syntax:
14417"""""""
14418This is an overloaded intrinsic. The memory object can belong to any address space.
14419
14420::
14421
14422      declare void @llvm.invariant.end.p0i8({}* <start>, i64 <size>, i8* nocapture <ptr>)
14423
14424Overview:
14425"""""""""
14426
14427The '``llvm.invariant.end``' intrinsic specifies that the contents of a
14428memory object are mutable.
14429
14430Arguments:
14431""""""""""
14432
14433The first argument is the matching ``llvm.invariant.start`` intrinsic.
14434The second argument is a constant integer representing the size of the
14435object, or -1 if it is variable sized and the third argument is a
14436pointer to the object.
14437
14438Semantics:
14439""""""""""
14440
14441This intrinsic indicates that the memory is mutable again.
14442
14443'``llvm.launder.invariant.group``' Intrinsic
14444^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
14445
14446Syntax:
14447"""""""
14448This is an overloaded intrinsic. The memory object can belong to any address
14449space. The returned pointer must belong to the same address space as the
14450argument.
14451
14452::
14453
14454      declare i8* @llvm.launder.invariant.group.p0i8(i8* <ptr>)
14455
14456Overview:
14457"""""""""
14458
14459The '``llvm.launder.invariant.group``' intrinsic can be used when an invariant
14460established by ``invariant.group`` metadata no longer holds, to obtain a new
14461pointer value that carries fresh invariant group information. It is an
14462experimental intrinsic, which means that its semantics might change in the
14463future.
14464
14465
14466Arguments:
14467""""""""""
14468
14469The ``llvm.launder.invariant.group`` takes only one argument, which is a pointer
14470to the memory.
14471
14472Semantics:
14473""""""""""
14474
14475Returns another pointer that aliases its argument but which is considered different
14476for the purposes of ``load``/``store`` ``invariant.group`` metadata.
14477It does not read any accessible memory and the execution can be speculated.
14478
14479'``llvm.strip.invariant.group``' Intrinsic
14480^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
14481
14482Syntax:
14483"""""""
14484This is an overloaded intrinsic. The memory object can belong to any address
14485space. The returned pointer must belong to the same address space as the
14486argument.
14487
14488::
14489
14490      declare i8* @llvm.strip.invariant.group.p0i8(i8* <ptr>)
14491
14492Overview:
14493"""""""""
14494
14495The '``llvm.strip.invariant.group``' intrinsic can be used when an invariant
14496established by ``invariant.group`` metadata no longer holds, to obtain a new pointer
14497value that does not carry the invariant information. It is an experimental
14498intrinsic, which means that its semantics might change in the future.
14499
14500
14501Arguments:
14502""""""""""
14503
14504The ``llvm.strip.invariant.group`` takes only one argument, which is a pointer
14505to the memory.
14506
14507Semantics:
14508""""""""""
14509
14510Returns another pointer that aliases its argument but which has no associated
14511``invariant.group`` metadata.
14512It does not read any memory and can be speculated.
14513
14514
14515
14516.. _constrainedfp:
14517
14518Constrained Floating-Point Intrinsics
14519-------------------------------------
14520
14521These intrinsics are used to provide special handling of floating-point
14522operations when specific rounding mode or floating-point exception behavior is
14523required.  By default, LLVM optimization passes assume that the rounding mode is
14524round-to-nearest and that floating-point exceptions will not be monitored.
14525Constrained FP intrinsics are used to support non-default rounding modes and
14526accurately preserve exception behavior without compromising LLVM's ability to
14527optimize FP code when the default behavior is used.
14528
14529Each of these intrinsics corresponds to a normal floating-point operation.  The
14530first two arguments and the return value are the same as the corresponding FP
14531operation.
14532
14533The third argument is a metadata argument specifying the rounding mode to be
14534assumed. This argument must be one of the following strings:
14535
14536::
14537
14538      "round.dynamic"
14539      "round.tonearest"
14540      "round.downward"
14541      "round.upward"
14542      "round.towardzero"
14543
14544If this argument is "round.dynamic" optimization passes must assume that the
14545rounding mode is unknown and may change at runtime.  No transformations that
14546depend on rounding mode may be performed in this case.
14547
14548The other possible values for the rounding mode argument correspond to the
14549similarly named IEEE rounding modes.  If the argument is any of these values
14550optimization passes may perform transformations as long as they are consistent
14551with the specified rounding mode.
14552
14553For example, 'x-0'->'x' is not a valid transformation if the rounding mode is
14554"round.downward" or "round.dynamic" because if the value of 'x' is +0 then
14555'x-0' should evaluate to '-0' when rounding downward.  However, this
14556transformation is legal for all other rounding modes.
14557
14558For values other than "round.dynamic" optimization passes may assume that the
14559actual runtime rounding mode (as defined in a target-specific manner) matches
14560the specified rounding mode, but this is not guaranteed.  Using a specific
14561non-dynamic rounding mode which does not match the actual rounding mode at
14562runtime results in undefined behavior.
14563
14564The fourth argument to the constrained floating-point intrinsics specifies the
14565required exception behavior.  This argument must be one of the following
14566strings:
14567
14568::
14569
14570      "fpexcept.ignore"
14571      "fpexcept.maytrap"
14572      "fpexcept.strict"
14573
14574If this argument is "fpexcept.ignore" optimization passes may assume that the
14575exception status flags will not be read and that floating-point exceptions will
14576be masked.  This allows transformations to be performed that may change the
14577exception semantics of the original code.  For example, FP operations may be
14578speculatively executed in this case whereas they must not be for either of the
14579other possible values of this argument.
14580
14581If the exception behavior argument is "fpexcept.maytrap" optimization passes
14582must avoid transformations that may raise exceptions that would not have been
14583raised by the original code (such as speculatively executing FP operations), but
14584passes are not required to preserve all exceptions that are implied by the
14585original code.  For example, exceptions may be potentially hidden by constant
14586folding.
14587
14588If the exception behavior argument is "fpexcept.strict" all transformations must
14589strictly preserve the floating-point exception semantics of the original code.
14590Any FP exception that would have been raised by the original code must be raised
14591by the transformed code, and the transformed code must not raise any FP
14592exceptions that would not have been raised by the original code.  This is the
14593exception behavior argument that will be used if the code being compiled reads
14594the FP exception status flags, but this mode can also be used with code that
14595unmasks FP exceptions.
14596
14597The number and order of floating-point exceptions is NOT guaranteed.  For
14598example, a series of FP operations that each may raise exceptions may be
14599vectorized into a single instruction that raises each unique exception a single
14600time.
14601
14602
14603'``llvm.experimental.constrained.fadd``' Intrinsic
14604^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
14605
14606Syntax:
14607"""""""
14608
14609::
14610
14611      declare <type>
14612      @llvm.experimental.constrained.fadd(<type> <op1>, <type> <op2>,
14613                                          metadata <rounding mode>,
14614                                          metadata <exception behavior>)
14615
14616Overview:
14617"""""""""
14618
14619The '``llvm.experimental.constrained.fadd``' intrinsic returns the sum of its
14620two operands.
14621
14622
14623Arguments:
14624""""""""""
14625
14626The first two arguments to the '``llvm.experimental.constrained.fadd``'
14627intrinsic must be :ref:`floating-point <t_floating>` or :ref:`vector <t_vector>`
14628of floating-point values. Both arguments must have identical types.
14629
14630The third and fourth arguments specify the rounding mode and exception
14631behavior as described above.
14632
14633Semantics:
14634""""""""""
14635
14636The value produced is the floating-point sum of the two value operands and has
14637the same type as the operands.
14638
14639
14640'``llvm.experimental.constrained.fsub``' Intrinsic
14641^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
14642
14643Syntax:
14644"""""""
14645
14646::
14647
14648      declare <type>
14649      @llvm.experimental.constrained.fsub(<type> <op1>, <type> <op2>,
14650                                          metadata <rounding mode>,
14651                                          metadata <exception behavior>)
14652
14653Overview:
14654"""""""""
14655
14656The '``llvm.experimental.constrained.fsub``' intrinsic returns the difference
14657of its two operands.
14658
14659
14660Arguments:
14661""""""""""
14662
14663The first two arguments to the '``llvm.experimental.constrained.fsub``'
14664intrinsic must be :ref:`floating-point <t_floating>` or :ref:`vector <t_vector>`
14665of floating-point values. Both arguments must have identical types.
14666
14667The third and fourth arguments specify the rounding mode and exception
14668behavior as described above.
14669
14670Semantics:
14671""""""""""
14672
14673The value produced is the floating-point difference of the two value operands
14674and has the same type as the operands.
14675
14676
14677'``llvm.experimental.constrained.fmul``' Intrinsic
14678^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
14679
14680Syntax:
14681"""""""
14682
14683::
14684
14685      declare <type>
14686      @llvm.experimental.constrained.fmul(<type> <op1>, <type> <op2>,
14687                                          metadata <rounding mode>,
14688                                          metadata <exception behavior>)
14689
14690Overview:
14691"""""""""
14692
14693The '``llvm.experimental.constrained.fmul``' intrinsic returns the product of
14694its two operands.
14695
14696
14697Arguments:
14698""""""""""
14699
14700The first two arguments to the '``llvm.experimental.constrained.fmul``'
14701intrinsic must be :ref:`floating-point <t_floating>` or :ref:`vector <t_vector>`
14702of floating-point values. Both arguments must have identical types.
14703
14704The third and fourth arguments specify the rounding mode and exception
14705behavior as described above.
14706
14707Semantics:
14708""""""""""
14709
14710The value produced is the floating-point product of the two value operands and
14711has the same type as the operands.
14712
14713
14714'``llvm.experimental.constrained.fdiv``' Intrinsic
14715^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
14716
14717Syntax:
14718"""""""
14719
14720::
14721
14722      declare <type>
14723      @llvm.experimental.constrained.fdiv(<type> <op1>, <type> <op2>,
14724                                          metadata <rounding mode>,
14725                                          metadata <exception behavior>)
14726
14727Overview:
14728"""""""""
14729
14730The '``llvm.experimental.constrained.fdiv``' intrinsic returns the quotient of
14731its two operands.
14732
14733
14734Arguments:
14735""""""""""
14736
14737The first two arguments to the '``llvm.experimental.constrained.fdiv``'
14738intrinsic must be :ref:`floating-point <t_floating>` or :ref:`vector <t_vector>`
14739of floating-point values. Both arguments must have identical types.
14740
14741The third and fourth arguments specify the rounding mode and exception
14742behavior as described above.
14743
14744Semantics:
14745""""""""""
14746
14747The value produced is the floating-point quotient of the two value operands and
14748has the same type as the operands.
14749
14750
14751'``llvm.experimental.constrained.frem``' Intrinsic
14752^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
14753
14754Syntax:
14755"""""""
14756
14757::
14758
14759      declare <type>
14760      @llvm.experimental.constrained.frem(<type> <op1>, <type> <op2>,
14761                                          metadata <rounding mode>,
14762                                          metadata <exception behavior>)
14763
14764Overview:
14765"""""""""
14766
14767The '``llvm.experimental.constrained.frem``' intrinsic returns the remainder
14768from the division of its two operands.
14769
14770
14771Arguments:
14772""""""""""
14773
14774The first two arguments to the '``llvm.experimental.constrained.frem``'
14775intrinsic must be :ref:`floating-point <t_floating>` or :ref:`vector <t_vector>`
14776of floating-point values. Both arguments must have identical types.
14777
14778The third and fourth arguments specify the rounding mode and exception
14779behavior as described above.  The rounding mode argument has no effect, since
14780the result of frem is never rounded, but the argument is included for
14781consistency with the other constrained floating-point intrinsics.
14782
14783Semantics:
14784""""""""""
14785
14786The value produced is the floating-point remainder from the division of the two
14787value operands and has the same type as the operands.  The remainder has the
14788same sign as the dividend.
14789
14790'``llvm.experimental.constrained.fma``' Intrinsic
14791^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
14792
14793Syntax:
14794"""""""
14795
14796::
14797
14798      declare <type>
14799      @llvm.experimental.constrained.fma(<type> <op1>, <type> <op2>, <type> <op3>,
14800                                          metadata <rounding mode>,
14801                                          metadata <exception behavior>)
14802
14803Overview:
14804"""""""""
14805
14806The '``llvm.experimental.constrained.fma``' intrinsic returns the result of a
14807fused-multiply-add operation on its operands.
14808
14809Arguments:
14810""""""""""
14811
14812The first three arguments to the '``llvm.experimental.constrained.fma``'
14813intrinsic must be :ref:`floating-point <t_floating>` or :ref:`vector
14814<t_vector>` of floating-point values. All arguments must have identical types.
14815
14816The fourth and fifth arguments specify the rounding mode and exception behavior
14817as described above.
14818
14819Semantics:
14820""""""""""
14821
14822The result produced is the product of the first two operands added to the third
14823operand computed with infinite precision, and then rounded to the target
14824precision.
14825
14826Constrained libm-equivalent Intrinsics
14827--------------------------------------
14828
14829In addition to the basic floating-point operations for which constrained
14830intrinsics are described above, there are constrained versions of various
14831operations which provide equivalent behavior to a corresponding libm function.
14832These intrinsics allow the precise behavior of these operations with respect to
14833rounding mode and exception behavior to be controlled.
14834
14835As with the basic constrained floating-point intrinsics, the rounding mode
14836and exception behavior arguments only control the behavior of the optimizer.
14837They do not change the runtime floating-point environment.
14838
14839
14840'``llvm.experimental.constrained.sqrt``' Intrinsic
14841^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
14842
14843Syntax:
14844"""""""
14845
14846::
14847
14848      declare <type>
14849      @llvm.experimental.constrained.sqrt(<type> <op1>,
14850                                          metadata <rounding mode>,
14851                                          metadata <exception behavior>)
14852
14853Overview:
14854"""""""""
14855
14856The '``llvm.experimental.constrained.sqrt``' intrinsic returns the square root
14857of the specified value, returning the same value as the libm '``sqrt``'
14858functions would, but without setting ``errno``.
14859
14860Arguments:
14861""""""""""
14862
14863The first argument and the return type are floating-point numbers of the same
14864type.
14865
14866The second and third arguments specify the rounding mode and exception
14867behavior as described above.
14868
14869Semantics:
14870""""""""""
14871
14872This function returns the nonnegative square root of the specified value.
14873If the value is less than negative zero, a floating-point exception occurs
14874and the return value is architecture specific.
14875
14876
14877'``llvm.experimental.constrained.pow``' Intrinsic
14878^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
14879
14880Syntax:
14881"""""""
14882
14883::
14884
14885      declare <type>
14886      @llvm.experimental.constrained.pow(<type> <op1>, <type> <op2>,
14887                                         metadata <rounding mode>,
14888                                         metadata <exception behavior>)
14889
14890Overview:
14891"""""""""
14892
14893The '``llvm.experimental.constrained.pow``' intrinsic returns the first operand
14894raised to the (positive or negative) power specified by the second operand.
14895
14896Arguments:
14897""""""""""
14898
14899The first two arguments and the return value are floating-point numbers of the
14900same type.  The second argument specifies the power to which the first argument
14901should be raised.
14902
14903The third and fourth arguments specify the rounding mode and exception
14904behavior as described above.
14905
14906Semantics:
14907""""""""""
14908
14909This function returns the first value raised to the second power,
14910returning the same values as the libm ``pow`` functions would, and
14911handles error conditions in the same way.
14912
14913
14914'``llvm.experimental.constrained.powi``' Intrinsic
14915^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
14916
14917Syntax:
14918"""""""
14919
14920::
14921
14922      declare <type>
14923      @llvm.experimental.constrained.powi(<type> <op1>, i32 <op2>,
14924                                          metadata <rounding mode>,
14925                                          metadata <exception behavior>)
14926
14927Overview:
14928"""""""""
14929
14930The '``llvm.experimental.constrained.powi``' intrinsic returns the first operand
14931raised to the (positive or negative) power specified by the second operand. The
14932order of evaluation of multiplications is not defined. When a vector of
14933floating-point type is used, the second argument remains a scalar integer value.
14934
14935
14936Arguments:
14937""""""""""
14938
14939The first argument and the return value are floating-point numbers of the same
14940type.  The second argument is a 32-bit signed integer specifying the power to
14941which the first argument should be raised.
14942
14943The third and fourth arguments specify the rounding mode and exception
14944behavior as described above.
14945
14946Semantics:
14947""""""""""
14948
14949This function returns the first value raised to the second power with an
14950unspecified sequence of rounding operations.
14951
14952
14953'``llvm.experimental.constrained.sin``' Intrinsic
14954^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
14955
14956Syntax:
14957"""""""
14958
14959::
14960
14961      declare <type>
14962      @llvm.experimental.constrained.sin(<type> <op1>,
14963                                         metadata <rounding mode>,
14964                                         metadata <exception behavior>)
14965
14966Overview:
14967"""""""""
14968
14969The '``llvm.experimental.constrained.sin``' intrinsic returns the sine of the
14970first operand.
14971
14972Arguments:
14973""""""""""
14974
14975The first argument and the return type are floating-point numbers of the same
14976type.
14977
14978The second and third arguments specify the rounding mode and exception
14979behavior as described above.
14980
14981Semantics:
14982""""""""""
14983
14984This function returns the sine of the specified operand, returning the
14985same values as the libm ``sin`` functions would, and handles error
14986conditions in the same way.
14987
14988
14989'``llvm.experimental.constrained.cos``' Intrinsic
14990^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
14991
14992Syntax:
14993"""""""
14994
14995::
14996
14997      declare <type>
14998      @llvm.experimental.constrained.cos(<type> <op1>,
14999                                         metadata <rounding mode>,
15000                                         metadata <exception behavior>)
15001
15002Overview:
15003"""""""""
15004
15005The '``llvm.experimental.constrained.cos``' intrinsic returns the cosine of the
15006first operand.
15007
15008Arguments:
15009""""""""""
15010
15011The first argument and the return type are floating-point numbers of the same
15012type.
15013
15014The second and third arguments specify the rounding mode and exception
15015behavior as described above.
15016
15017Semantics:
15018""""""""""
15019
15020This function returns the cosine of the specified operand, returning the
15021same values as the libm ``cos`` functions would, and handles error
15022conditions in the same way.
15023
15024
15025'``llvm.experimental.constrained.exp``' Intrinsic
15026^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
15027
15028Syntax:
15029"""""""
15030
15031::
15032
15033      declare <type>
15034      @llvm.experimental.constrained.exp(<type> <op1>,
15035                                         metadata <rounding mode>,
15036                                         metadata <exception behavior>)
15037
15038Overview:
15039"""""""""
15040
15041The '``llvm.experimental.constrained.exp``' intrinsic computes the base-e
15042exponential of the specified value.
15043
15044Arguments:
15045""""""""""
15046
15047The first argument and the return value are floating-point numbers of the same
15048type.
15049
15050The second and third arguments specify the rounding mode and exception
15051behavior as described above.
15052
15053Semantics:
15054""""""""""
15055
15056This function returns the same values as the libm ``exp`` functions
15057would, and handles error conditions in the same way.
15058
15059
15060'``llvm.experimental.constrained.exp2``' Intrinsic
15061^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
15062
15063Syntax:
15064"""""""
15065
15066::
15067
15068      declare <type>
15069      @llvm.experimental.constrained.exp2(<type> <op1>,
15070                                          metadata <rounding mode>,
15071                                          metadata <exception behavior>)
15072
15073Overview:
15074"""""""""
15075
15076The '``llvm.experimental.constrained.exp2``' intrinsic computes the base-2
15077exponential of the specified value.
15078
15079
15080Arguments:
15081""""""""""
15082
15083The first argument and the return value are floating-point numbers of the same
15084type.
15085
15086The second and third arguments specify the rounding mode and exception
15087behavior as described above.
15088
15089Semantics:
15090""""""""""
15091
15092This function returns the same values as the libm ``exp2`` functions
15093would, and handles error conditions in the same way.
15094
15095
15096'``llvm.experimental.constrained.log``' Intrinsic
15097^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
15098
15099Syntax:
15100"""""""
15101
15102::
15103
15104      declare <type>
15105      @llvm.experimental.constrained.log(<type> <op1>,
15106                                         metadata <rounding mode>,
15107                                         metadata <exception behavior>)
15108
15109Overview:
15110"""""""""
15111
15112The '``llvm.experimental.constrained.log``' intrinsic computes the base-e
15113logarithm of the specified value.
15114
15115Arguments:
15116""""""""""
15117
15118The first argument and the return value are floating-point numbers of the same
15119type.
15120
15121The second and third arguments specify the rounding mode and exception
15122behavior as described above.
15123
15124
15125Semantics:
15126""""""""""
15127
15128This function returns the same values as the libm ``log`` functions
15129would, and handles error conditions in the same way.
15130
15131
15132'``llvm.experimental.constrained.log10``' Intrinsic
15133^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
15134
15135Syntax:
15136"""""""
15137
15138::
15139
15140      declare <type>
15141      @llvm.experimental.constrained.log10(<type> <op1>,
15142                                           metadata <rounding mode>,
15143                                           metadata <exception behavior>)
15144
15145Overview:
15146"""""""""
15147
15148The '``llvm.experimental.constrained.log10``' intrinsic computes the base-10
15149logarithm of the specified value.
15150
15151Arguments:
15152""""""""""
15153
15154The first argument and the return value are floating-point numbers of the same
15155type.
15156
15157The second and third arguments specify the rounding mode and exception
15158behavior as described above.
15159
15160Semantics:
15161""""""""""
15162
15163This function returns the same values as the libm ``log10`` functions
15164would, and handles error conditions in the same way.
15165
15166
15167'``llvm.experimental.constrained.log2``' Intrinsic
15168^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
15169
15170Syntax:
15171"""""""
15172
15173::
15174
15175      declare <type>
15176      @llvm.experimental.constrained.log2(<type> <op1>,
15177                                          metadata <rounding mode>,
15178                                          metadata <exception behavior>)
15179
15180Overview:
15181"""""""""
15182
15183The '``llvm.experimental.constrained.log2``' intrinsic computes the base-2
15184logarithm of the specified value.
15185
15186Arguments:
15187""""""""""
15188
15189The first argument and the return value are floating-point numbers of the same
15190type.
15191
15192The second and third arguments specify the rounding mode and exception
15193behavior as described above.
15194
15195Semantics:
15196""""""""""
15197
15198This function returns the same values as the libm ``log2`` functions
15199would, and handles error conditions in the same way.
15200
15201
15202'``llvm.experimental.constrained.rint``' Intrinsic
15203^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
15204
15205Syntax:
15206"""""""
15207
15208::
15209
15210      declare <type>
15211      @llvm.experimental.constrained.rint(<type> <op1>,
15212                                          metadata <rounding mode>,
15213                                          metadata <exception behavior>)
15214
15215Overview:
15216"""""""""
15217
15218The '``llvm.experimental.constrained.rint``' intrinsic returns the first
15219operand rounded to the nearest integer. It may raise an inexact floating-point
15220exception if the operand is not an integer.
15221
15222Arguments:
15223""""""""""
15224
15225The first argument and the return value are floating-point numbers of the same
15226type.
15227
15228The second and third arguments specify the rounding mode and exception
15229behavior as described above.
15230
15231Semantics:
15232""""""""""
15233
15234This function returns the same values as the libm ``rint`` functions
15235would, and handles error conditions in the same way.  The rounding mode is
15236described, not determined, by the rounding mode argument.  The actual rounding
15237mode is determined by the runtime floating-point environment.  The rounding
15238mode argument is only intended as information to the compiler.
15239
15240
15241'``llvm.experimental.constrained.nearbyint``' Intrinsic
15242^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
15243
15244Syntax:
15245"""""""
15246
15247::
15248
15249      declare <type>
15250      @llvm.experimental.constrained.nearbyint(<type> <op1>,
15251                                               metadata <rounding mode>,
15252                                               metadata <exception behavior>)
15253
15254Overview:
15255"""""""""
15256
15257The '``llvm.experimental.constrained.nearbyint``' intrinsic returns the first
15258operand rounded to the nearest integer. It will not raise an inexact
15259floating-point exception if the operand is not an integer.
15260
15261
15262Arguments:
15263""""""""""
15264
15265The first argument and the return value are floating-point numbers of the same
15266type.
15267
15268The second and third arguments specify the rounding mode and exception
15269behavior as described above.
15270
15271Semantics:
15272""""""""""
15273
15274This function returns the same values as the libm ``nearbyint`` functions
15275would, and handles error conditions in the same way.  The rounding mode is
15276described, not determined, by the rounding mode argument.  The actual rounding
15277mode is determined by the runtime floating-point environment.  The rounding
15278mode argument is only intended as information to the compiler.
15279
15280
15281'``llvm.experimental.constrained.maxnum``' Intrinsic
15282^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
15283
15284Syntax:
15285"""""""
15286
15287::
15288
15289      declare <type>
15290      @llvm.experimental.constrained.maxnum(<type> <op1>, <type> <op2>
15291                                            metadata <rounding mode>,
15292                                            metadata <exception behavior>)
15293
15294Overview:
15295"""""""""
15296
15297The '``llvm.experimental.constrained.maxnum``' intrinsic returns the maximum
15298of the two arguments.
15299
15300Arguments:
15301""""""""""
15302
15303The first two arguments and the return value are floating-point numbers
15304of the same type.
15305
15306The third and forth arguments specify the rounding mode and exception
15307behavior as described above.
15308
15309Semantics:
15310""""""""""
15311
15312This function follows the IEEE-754 semantics for maxNum. The rounding mode is
15313described, not determined, by the rounding mode argument. The actual rounding
15314mode is determined by the runtime floating-point environment. The rounding
15315mode argument is only intended as information to the compiler.
15316
15317
15318'``llvm.experimental.constrained.minnum``' Intrinsic
15319^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
15320
15321Syntax:
15322"""""""
15323
15324::
15325
15326      declare <type>
15327      @llvm.experimental.constrained.minnum(<type> <op1>, <type> <op2>
15328                                            metadata <rounding mode>,
15329                                            metadata <exception behavior>)
15330
15331Overview:
15332"""""""""
15333
15334The '``llvm.experimental.constrained.minnum``' intrinsic returns the minimum
15335of the two arguments.
15336
15337Arguments:
15338""""""""""
15339
15340The first two arguments and the return value are floating-point numbers
15341of the same type.
15342
15343The third and forth arguments specify the rounding mode and exception
15344behavior as described above.
15345
15346Semantics:
15347""""""""""
15348
15349This function follows the IEEE-754 semantics for minNum. The rounding mode is
15350described, not determined, by the rounding mode argument. The actual rounding
15351mode is determined by the runtime floating-point environment. The rounding
15352mode argument is only intended as information to the compiler.
15353
15354
15355'``llvm.experimental.constrained.ceil``' Intrinsic
15356^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
15357
15358Syntax:
15359"""""""
15360
15361::
15362
15363      declare <type>
15364      @llvm.experimental.constrained.ceil(<type> <op1>,
15365                                          metadata <rounding mode>,
15366                                          metadata <exception behavior>)
15367
15368Overview:
15369"""""""""
15370
15371The '``llvm.experimental.constrained.ceil``' intrinsic returns the ceiling of the
15372first operand.
15373
15374Arguments:
15375""""""""""
15376
15377The first argument and the return value are floating-point numbers of the same
15378type.
15379
15380The second and third arguments specify the rounding mode and exception
15381behavior as described above. The rounding mode is currently unused for this
15382intrinsic.
15383
15384Semantics:
15385""""""""""
15386
15387This function returns the same values as the libm ``ceil`` functions
15388would and handles error conditions in the same way.
15389
15390
15391'``llvm.experimental.constrained.floor``' Intrinsic
15392^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
15393
15394Syntax:
15395"""""""
15396
15397::
15398
15399      declare <type>
15400      @llvm.experimental.constrained.floor(<type> <op1>,
15401                                           metadata <rounding mode>,
15402                                           metadata <exception behavior>)
15403
15404Overview:
15405"""""""""
15406
15407The '``llvm.experimental.constrained.floor``' intrinsic returns the floor of the
15408first operand.
15409
15410Arguments:
15411""""""""""
15412
15413The first argument and the return value are floating-point numbers of the same
15414type.
15415
15416The second and third arguments specify the rounding mode and exception
15417behavior as described above. The rounding mode is currently unused for this
15418intrinsic.
15419
15420Semantics:
15421""""""""""
15422
15423This function returns the same values as the libm ``floor`` functions
15424would and handles error conditions in the same way.
15425
15426
15427'``llvm.experimental.constrained.round``' Intrinsic
15428^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
15429
15430Syntax:
15431"""""""
15432
15433::
15434
15435      declare <type>
15436      @llvm.experimental.constrained.round(<type> <op1>,
15437                                           metadata <rounding mode>,
15438                                           metadata <exception behavior>)
15439
15440Overview:
15441"""""""""
15442
15443The '``llvm.experimental.constrained.round``' intrinsic returns the first
15444operand rounded to the nearest integer.
15445
15446Arguments:
15447""""""""""
15448
15449The first argument and the return value are floating-point numbers of the same
15450type.
15451
15452The second and third arguments specify the rounding mode and exception
15453behavior as described above. The rounding mode is currently unused for this
15454intrinsic.
15455
15456Semantics:
15457""""""""""
15458
15459This function returns the same values as the libm ``round`` functions
15460would and handles error conditions in the same way.
15461
15462
15463'``llvm.experimental.constrained.trunc``' Intrinsic
15464^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
15465
15466Syntax:
15467"""""""
15468
15469::
15470
15471      declare <type>
15472      @llvm.experimental.constrained.trunc(<type> <op1>,
15473                                           metadata <truncing mode>,
15474                                           metadata <exception behavior>)
15475
15476Overview:
15477"""""""""
15478
15479The '``llvm.experimental.constrained.trunc``' intrinsic returns the first
15480operand rounded to the nearest integer not larger in magnitude than the
15481operand.
15482
15483Arguments:
15484""""""""""
15485
15486The first argument and the return value are floating-point numbers of the same
15487type.
15488
15489The second and third arguments specify the truncing mode and exception
15490behavior as described above. The truncing mode is currently unused for this
15491intrinsic.
15492
15493Semantics:
15494""""""""""
15495
15496This function returns the same values as the libm ``trunc`` functions
15497would and handles error conditions in the same way.
15498
15499
15500General Intrinsics
15501------------------
15502
15503This class of intrinsics is designed to be generic and has no specific
15504purpose.
15505
15506'``llvm.var.annotation``' Intrinsic
15507^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
15508
15509Syntax:
15510"""""""
15511
15512::
15513
15514      declare void @llvm.var.annotation(i8* <val>, i8* <str>, i8* <str>, i32  <int>)
15515
15516Overview:
15517"""""""""
15518
15519The '``llvm.var.annotation``' intrinsic.
15520
15521Arguments:
15522""""""""""
15523
15524The first argument is a pointer to a value, the second is a pointer to a
15525global string, the third is a pointer to a global string which is the
15526source file name, and the last argument is the line number.
15527
15528Semantics:
15529""""""""""
15530
15531This intrinsic allows annotation of local variables with arbitrary
15532strings. This can be useful for special purpose optimizations that want
15533to look for these annotations. These have no other defined use; they are
15534ignored by code generation and optimization.
15535
15536'``llvm.ptr.annotation.*``' Intrinsic
15537^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
15538
15539Syntax:
15540"""""""
15541
15542This is an overloaded intrinsic. You can use '``llvm.ptr.annotation``' on a
15543pointer to an integer of any width. *NOTE* you must specify an address space for
15544the pointer. The identifier for the default address space is the integer
15545'``0``'.
15546
15547::
15548
15549      declare i8*   @llvm.ptr.annotation.p<address space>i8(i8* <val>, i8* <str>, i8* <str>, i32  <int>)
15550      declare i16*  @llvm.ptr.annotation.p<address space>i16(i16* <val>, i8* <str>, i8* <str>, i32  <int>)
15551      declare i32*  @llvm.ptr.annotation.p<address space>i32(i32* <val>, i8* <str>, i8* <str>, i32  <int>)
15552      declare i64*  @llvm.ptr.annotation.p<address space>i64(i64* <val>, i8* <str>, i8* <str>, i32  <int>)
15553      declare i256* @llvm.ptr.annotation.p<address space>i256(i256* <val>, i8* <str>, i8* <str>, i32  <int>)
15554
15555Overview:
15556"""""""""
15557
15558The '``llvm.ptr.annotation``' intrinsic.
15559
15560Arguments:
15561""""""""""
15562
15563The first argument is a pointer to an integer value of arbitrary bitwidth
15564(result of some expression), the second is a pointer to a global string, the
15565third is a pointer to a global string which is the source file name, and the
15566last argument is the line number. It returns the value of the first argument.
15567
15568Semantics:
15569""""""""""
15570
15571This intrinsic allows annotation of a pointer to an integer with arbitrary
15572strings. This can be useful for special purpose optimizations that want to look
15573for these annotations. These have no other defined use; they are ignored by code
15574generation and optimization.
15575
15576'``llvm.annotation.*``' Intrinsic
15577^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
15578
15579Syntax:
15580"""""""
15581
15582This is an overloaded intrinsic. You can use '``llvm.annotation``' on
15583any integer bit width.
15584
15585::
15586
15587      declare i8 @llvm.annotation.i8(i8 <val>, i8* <str>, i8* <str>, i32  <int>)
15588      declare i16 @llvm.annotation.i16(i16 <val>, i8* <str>, i8* <str>, i32  <int>)
15589      declare i32 @llvm.annotation.i32(i32 <val>, i8* <str>, i8* <str>, i32  <int>)
15590      declare i64 @llvm.annotation.i64(i64 <val>, i8* <str>, i8* <str>, i32  <int>)
15591      declare i256 @llvm.annotation.i256(i256 <val>, i8* <str>, i8* <str>, i32  <int>)
15592
15593Overview:
15594"""""""""
15595
15596The '``llvm.annotation``' intrinsic.
15597
15598Arguments:
15599""""""""""
15600
15601The first argument is an integer value (result of some expression), the
15602second is a pointer to a global string, the third is a pointer to a
15603global string which is the source file name, and the last argument is
15604the line number. It returns the value of the first argument.
15605
15606Semantics:
15607""""""""""
15608
15609This intrinsic allows annotations to be put on arbitrary expressions
15610with arbitrary strings. This can be useful for special purpose
15611optimizations that want to look for these annotations. These have no
15612other defined use; they are ignored by code generation and optimization.
15613
15614'``llvm.codeview.annotation``' Intrinsic
15615^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
15616
15617Syntax:
15618"""""""
15619
15620This annotation emits a label at its program point and an associated
15621``S_ANNOTATION`` codeview record with some additional string metadata. This is
15622used to implement MSVC's ``__annotation`` intrinsic. It is marked
15623``noduplicate``, so calls to this intrinsic prevent inlining and should be
15624considered expensive.
15625
15626::
15627
15628      declare void @llvm.codeview.annotation(metadata)
15629
15630Arguments:
15631""""""""""
15632
15633The argument should be an MDTuple containing any number of MDStrings.
15634
15635'``llvm.trap``' Intrinsic
15636^^^^^^^^^^^^^^^^^^^^^^^^^
15637
15638Syntax:
15639"""""""
15640
15641::
15642
15643      declare void @llvm.trap() cold noreturn nounwind
15644
15645Overview:
15646"""""""""
15647
15648The '``llvm.trap``' intrinsic.
15649
15650Arguments:
15651""""""""""
15652
15653None.
15654
15655Semantics:
15656""""""""""
15657
15658This intrinsic is lowered to the target dependent trap instruction. If
15659the target does not have a trap instruction, this intrinsic will be
15660lowered to a call of the ``abort()`` function.
15661
15662'``llvm.debugtrap``' Intrinsic
15663^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
15664
15665Syntax:
15666"""""""
15667
15668::
15669
15670      declare void @llvm.debugtrap() nounwind
15671
15672Overview:
15673"""""""""
15674
15675The '``llvm.debugtrap``' intrinsic.
15676
15677Arguments:
15678""""""""""
15679
15680None.
15681
15682Semantics:
15683""""""""""
15684
15685This intrinsic is lowered to code which is intended to cause an
15686execution trap with the intention of requesting the attention of a
15687debugger.
15688
15689'``llvm.stackprotector``' Intrinsic
15690^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
15691
15692Syntax:
15693"""""""
15694
15695::
15696
15697      declare void @llvm.stackprotector(i8* <guard>, i8** <slot>)
15698
15699Overview:
15700"""""""""
15701
15702The ``llvm.stackprotector`` intrinsic takes the ``guard`` and stores it
15703onto the stack at ``slot``. The stack slot is adjusted to ensure that it
15704is placed on the stack before local variables.
15705
15706Arguments:
15707""""""""""
15708
15709The ``llvm.stackprotector`` intrinsic requires two pointer arguments.
15710The first argument is the value loaded from the stack guard
15711``@__stack_chk_guard``. The second variable is an ``alloca`` that has
15712enough space to hold the value of the guard.
15713
15714Semantics:
15715""""""""""
15716
15717This intrinsic causes the prologue/epilogue inserter to force the position of
15718the ``AllocaInst`` stack slot to be before local variables on the stack. This is
15719to ensure that if a local variable on the stack is overwritten, it will destroy
15720the value of the guard. When the function exits, the guard on the stack is
15721checked against the original guard by ``llvm.stackprotectorcheck``. If they are
15722different, then ``llvm.stackprotectorcheck`` causes the program to abort by
15723calling the ``__stack_chk_fail()`` function.
15724
15725'``llvm.stackguard``' Intrinsic
15726^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
15727
15728Syntax:
15729"""""""
15730
15731::
15732
15733      declare i8* @llvm.stackguard()
15734
15735Overview:
15736"""""""""
15737
15738The ``llvm.stackguard`` intrinsic returns the system stack guard value.
15739
15740It should not be generated by frontends, since it is only for internal usage.
15741The reason why we create this intrinsic is that we still support IR form Stack
15742Protector in FastISel.
15743
15744Arguments:
15745""""""""""
15746
15747None.
15748
15749Semantics:
15750""""""""""
15751
15752On some platforms, the value returned by this intrinsic remains unchanged
15753between loads in the same thread. On other platforms, it returns the same
15754global variable value, if any, e.g. ``@__stack_chk_guard``.
15755
15756Currently some platforms have IR-level customized stack guard loading (e.g.
15757X86 Linux) that is not handled by ``llvm.stackguard()``, while they should be
15758in the future.
15759
15760'``llvm.objectsize``' Intrinsic
15761^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
15762
15763Syntax:
15764"""""""
15765
15766::
15767
15768      declare i32 @llvm.objectsize.i32(i8* <object>, i1 <min>, i1 <nullunknown>, i1 <dynamic>)
15769      declare i64 @llvm.objectsize.i64(i8* <object>, i1 <min>, i1 <nullunknown>, i1 <dynamic>)
15770
15771Overview:
15772"""""""""
15773
15774The ``llvm.objectsize`` intrinsic is designed to provide information to the
15775optimizer to determine whether a) an operation (like memcpy) will overflow a
15776buffer that corresponds to an object, or b) that a runtime check for overflow
15777isn't necessary. An object in this context means an allocation of a specific
15778class, structure, array, or other object.
15779
15780Arguments:
15781""""""""""
15782
15783The ``llvm.objectsize`` intrinsic takes four arguments. The first argument is a
15784pointer to or into the ``object``. The second argument determines whether
15785``llvm.objectsize`` returns 0 (if true) or -1 (if false) when the object size is
15786unknown. The third argument controls how ``llvm.objectsize`` acts when ``null``
15787in address space 0 is used as its pointer argument. If it's ``false``,
15788``llvm.objectsize`` reports 0 bytes available when given ``null``. Otherwise, if
15789the ``null`` is in a non-zero address space or if ``true`` is given for the
15790third argument of ``llvm.objectsize``, we assume its size is unknown. The fourth
15791argument to ``llvm.objectsize`` determines if the value should be evaluated at
15792runtime.
15793
15794The second, third, and fourth arguments only accept constants.
15795
15796Semantics:
15797""""""""""
15798
15799The ``llvm.objectsize`` intrinsic is lowered to a value representing the size of
15800the object concerned. If the size cannot be determined, ``llvm.objectsize``
15801returns ``i32/i64 -1 or 0`` (depending on the ``min`` argument).
15802
15803'``llvm.expect``' Intrinsic
15804^^^^^^^^^^^^^^^^^^^^^^^^^^^
15805
15806Syntax:
15807"""""""
15808
15809This is an overloaded intrinsic. You can use ``llvm.expect`` on any
15810integer bit width.
15811
15812::
15813
15814      declare i1 @llvm.expect.i1(i1 <val>, i1 <expected_val>)
15815      declare i32 @llvm.expect.i32(i32 <val>, i32 <expected_val>)
15816      declare i64 @llvm.expect.i64(i64 <val>, i64 <expected_val>)
15817
15818Overview:
15819"""""""""
15820
15821The ``llvm.expect`` intrinsic provides information about expected (the
15822most probable) value of ``val``, which can be used by optimizers.
15823
15824Arguments:
15825""""""""""
15826
15827The ``llvm.expect`` intrinsic takes two arguments. The first argument is
15828a value. The second argument is an expected value.
15829
15830Semantics:
15831""""""""""
15832
15833This intrinsic is lowered to the ``val``.
15834
15835.. _int_assume:
15836
15837'``llvm.assume``' Intrinsic
15838^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
15839
15840Syntax:
15841"""""""
15842
15843::
15844
15845      declare void @llvm.assume(i1 %cond)
15846
15847Overview:
15848"""""""""
15849
15850The ``llvm.assume`` allows the optimizer to assume that the provided
15851condition is true. This information can then be used in simplifying other parts
15852of the code.
15853
15854Arguments:
15855""""""""""
15856
15857The condition which the optimizer may assume is always true.
15858
15859Semantics:
15860""""""""""
15861
15862The intrinsic allows the optimizer to assume that the provided condition is
15863always true whenever the control flow reaches the intrinsic call. No code is
15864generated for this intrinsic, and instructions that contribute only to the
15865provided condition are not used for code generation. If the condition is
15866violated during execution, the behavior is undefined.
15867
15868Note that the optimizer might limit the transformations performed on values
15869used by the ``llvm.assume`` intrinsic in order to preserve the instructions
15870only used to form the intrinsic's input argument. This might prove undesirable
15871if the extra information provided by the ``llvm.assume`` intrinsic does not cause
15872sufficient overall improvement in code quality. For this reason,
15873``llvm.assume`` should not be used to document basic mathematical invariants
15874that the optimizer can otherwise deduce or facts that are of little use to the
15875optimizer.
15876
15877.. _int_ssa_copy:
15878
15879'``llvm.ssa_copy``' Intrinsic
15880^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
15881
15882Syntax:
15883"""""""
15884
15885::
15886
15887      declare type @llvm.ssa_copy(type %operand) returned(1) readnone
15888
15889Arguments:
15890""""""""""
15891
15892The first argument is an operand which is used as the returned value.
15893
15894Overview:
15895""""""""""
15896
15897The ``llvm.ssa_copy`` intrinsic can be used to attach information to
15898operations by copying them and giving them new names.  For example,
15899the PredicateInfo utility uses it to build Extended SSA form, and
15900attach various forms of information to operands that dominate specific
15901uses.  It is not meant for general use, only for building temporary
15902renaming forms that require value splits at certain points.
15903
15904.. _type.test:
15905
15906'``llvm.type.test``' Intrinsic
15907^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
15908
15909Syntax:
15910"""""""
15911
15912::
15913
15914      declare i1 @llvm.type.test(i8* %ptr, metadata %type) nounwind readnone
15915
15916
15917Arguments:
15918""""""""""
15919
15920The first argument is a pointer to be tested. The second argument is a
15921metadata object representing a :doc:`type identifier <TypeMetadata>`.
15922
15923Overview:
15924"""""""""
15925
15926The ``llvm.type.test`` intrinsic tests whether the given pointer is associated
15927with the given type identifier.
15928
15929'``llvm.type.checked.load``' Intrinsic
15930^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
15931
15932Syntax:
15933"""""""
15934
15935::
15936
15937      declare {i8*, i1} @llvm.type.checked.load(i8* %ptr, i32 %offset, metadata %type) argmemonly nounwind readonly
15938
15939
15940Arguments:
15941""""""""""
15942
15943The first argument is a pointer from which to load a function pointer. The
15944second argument is the byte offset from which to load the function pointer. The
15945third argument is a metadata object representing a :doc:`type identifier
15946<TypeMetadata>`.
15947
15948Overview:
15949"""""""""
15950
15951The ``llvm.type.checked.load`` intrinsic safely loads a function pointer from a
15952virtual table pointer using type metadata. This intrinsic is used to implement
15953control flow integrity in conjunction with virtual call optimization. The
15954virtual call optimization pass will optimize away ``llvm.type.checked.load``
15955intrinsics associated with devirtualized calls, thereby removing the type
15956check in cases where it is not needed to enforce the control flow integrity
15957constraint.
15958
15959If the given pointer is associated with a type metadata identifier, this
15960function returns true as the second element of its return value. (Note that
15961the function may also return true if the given pointer is not associated
15962with a type metadata identifier.) If the function's return value's second
15963element is true, the following rules apply to the first element:
15964
15965- If the given pointer is associated with the given type metadata identifier,
15966  it is the function pointer loaded from the given byte offset from the given
15967  pointer.
15968
15969- If the given pointer is not associated with the given type metadata
15970  identifier, it is one of the following (the choice of which is unspecified):
15971
15972  1. The function pointer that would have been loaded from an arbitrarily chosen
15973     (through an unspecified mechanism) pointer associated with the type
15974     metadata.
15975
15976  2. If the function has a non-void return type, a pointer to a function that
15977     returns an unspecified value without causing side effects.
15978
15979If the function's return value's second element is false, the value of the
15980first element is undefined.
15981
15982
15983'``llvm.donothing``' Intrinsic
15984^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
15985
15986Syntax:
15987"""""""
15988
15989::
15990
15991      declare void @llvm.donothing() nounwind readnone
15992
15993Overview:
15994"""""""""
15995
15996The ``llvm.donothing`` intrinsic doesn't perform any operation. It's one of only
15997three intrinsics (besides ``llvm.experimental.patchpoint`` and
15998``llvm.experimental.gc.statepoint``) that can be called with an invoke
15999instruction.
16000
16001Arguments:
16002""""""""""
16003
16004None.
16005
16006Semantics:
16007""""""""""
16008
16009This intrinsic does nothing, and it's removed by optimizers and ignored
16010by codegen.
16011
16012'``llvm.experimental.deoptimize``' Intrinsic
16013^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
16014
16015Syntax:
16016"""""""
16017
16018::
16019
16020      declare type @llvm.experimental.deoptimize(...) [ "deopt"(...) ]
16021
16022Overview:
16023"""""""""
16024
16025This intrinsic, together with :ref:`deoptimization operand bundles
16026<deopt_opbundles>`, allow frontends to express transfer of control and
16027frame-local state from the currently executing (typically more specialized,
16028hence faster) version of a function into another (typically more generic, hence
16029slower) version.
16030
16031In languages with a fully integrated managed runtime like Java and JavaScript
16032this intrinsic can be used to implement "uncommon trap" or "side exit" like
16033functionality.  In unmanaged languages like C and C++, this intrinsic can be
16034used to represent the slow paths of specialized functions.
16035
16036
16037Arguments:
16038""""""""""
16039
16040The intrinsic takes an arbitrary number of arguments, whose meaning is
16041decided by the :ref:`lowering strategy<deoptimize_lowering>`.
16042
16043Semantics:
16044""""""""""
16045
16046The ``@llvm.experimental.deoptimize`` intrinsic executes an attached
16047deoptimization continuation (denoted using a :ref:`deoptimization
16048operand bundle <deopt_opbundles>`) and returns the value returned by
16049the deoptimization continuation.  Defining the semantic properties of
16050the continuation itself is out of scope of the language reference --
16051as far as LLVM is concerned, the deoptimization continuation can
16052invoke arbitrary side effects, including reading from and writing to
16053the entire heap.
16054
16055Deoptimization continuations expressed using ``"deopt"`` operand bundles always
16056continue execution to the end of the physical frame containing them, so all
16057calls to ``@llvm.experimental.deoptimize`` must be in "tail position":
16058
16059   - ``@llvm.experimental.deoptimize`` cannot be invoked.
16060   - The call must immediately precede a :ref:`ret <i_ret>` instruction.
16061   - The ``ret`` instruction must return the value produced by the
16062     ``@llvm.experimental.deoptimize`` call if there is one, or void.
16063
16064Note that the above restrictions imply that the return type for a call to
16065``@llvm.experimental.deoptimize`` will match the return type of its immediate
16066caller.
16067
16068The inliner composes the ``"deopt"`` continuations of the caller into the
16069``"deopt"`` continuations present in the inlinee, and also updates calls to this
16070intrinsic to return directly from the frame of the function it inlined into.
16071
16072All declarations of ``@llvm.experimental.deoptimize`` must share the
16073same calling convention.
16074
16075.. _deoptimize_lowering:
16076
16077Lowering:
16078"""""""""
16079
16080Calls to ``@llvm.experimental.deoptimize`` are lowered to calls to the
16081symbol ``__llvm_deoptimize`` (it is the frontend's responsibility to
16082ensure that this symbol is defined).  The call arguments to
16083``@llvm.experimental.deoptimize`` are lowered as if they were formal
16084arguments of the specified types, and not as varargs.
16085
16086
16087'``llvm.experimental.guard``' Intrinsic
16088^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
16089
16090Syntax:
16091"""""""
16092
16093::
16094
16095      declare void @llvm.experimental.guard(i1, ...) [ "deopt"(...) ]
16096
16097Overview:
16098"""""""""
16099
16100This intrinsic, together with :ref:`deoptimization operand bundles
16101<deopt_opbundles>`, allows frontends to express guards or checks on
16102optimistic assumptions made during compilation.  The semantics of
16103``@llvm.experimental.guard`` is defined in terms of
16104``@llvm.experimental.deoptimize`` -- its body is defined to be
16105equivalent to:
16106
16107.. code-block:: text
16108
16109  define void @llvm.experimental.guard(i1 %pred, <args...>) {
16110    %realPred = and i1 %pred, undef
16111    br i1 %realPred, label %continue, label %leave [, !make.implicit !{}]
16112
16113  leave:
16114    call void @llvm.experimental.deoptimize(<args...>) [ "deopt"() ]
16115    ret void
16116
16117  continue:
16118    ret void
16119  }
16120
16121
16122with the optional ``[, !make.implicit !{}]`` present if and only if it
16123is present on the call site.  For more details on ``!make.implicit``,
16124see :doc:`FaultMaps`.
16125
16126In words, ``@llvm.experimental.guard`` executes the attached
16127``"deopt"`` continuation if (but **not** only if) its first argument
16128is ``false``.  Since the optimizer is allowed to replace the ``undef``
16129with an arbitrary value, it can optimize guard to fail "spuriously",
16130i.e. without the original condition being false (hence the "not only
16131if"); and this allows for "check widening" type optimizations.
16132
16133``@llvm.experimental.guard`` cannot be invoked.
16134
16135
16136'``llvm.experimental.widenable.condition``' Intrinsic
16137^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
16138
16139Syntax:
16140"""""""
16141
16142::
16143
16144      declare i1 @llvm.experimental.widenable.condition()
16145
16146Overview:
16147"""""""""
16148
16149This intrinsic represents a "widenable condition" which is
16150boolean expressions with the following property: whether this
16151expression is `true` or `false`, the program is correct and
16152well-defined.
16153
16154Together with :ref:`deoptimization operand bundles <deopt_opbundles>`,
16155``@llvm.experimental.widenable.condition`` allows frontends to
16156express guards or checks on optimistic assumptions made during
16157compilation and represent them as branch instructions on special
16158conditions.
16159
16160While this may appear similar in semantics to `undef`, it is very
16161different in that an invocation produces a particular, singular
16162value. It is also intended to be lowered late, and remain available
16163for specific optimizations and transforms that can benefit from its
16164special properties.
16165
16166Arguments:
16167""""""""""
16168
16169None.
16170
16171Semantics:
16172""""""""""
16173
16174The intrinsic ``@llvm.experimental.widenable.condition()``
16175returns either `true` or `false`. For each evaluation of a call
16176to this intrinsic, the program must be valid and correct both if
16177it returns `true` and if it returns `false`. This allows
16178transformation passes to replace evaluations of this intrinsic
16179with either value whenever one is beneficial.
16180
16181When used in a branch condition, it allows us to choose between
16182two alternative correct solutions for the same problem, like
16183in example below:
16184
16185.. code-block:: text
16186
16187    %cond = call i1 @llvm.experimental.widenable.condition()
16188    br i1 %cond, label %solution_1, label %solution_2
16189
16190  label %fast_path:
16191    ; Apply memory-consuming but fast solution for a task.
16192
16193  label %slow_path:
16194    ; Cheap in memory but slow solution.
16195
16196Whether the result of intrinsic's call is `true` or `false`,
16197it should be correct to pick either solution. We can switch
16198between them by replacing the result of
16199``@llvm.experimental.widenable.condition`` with different
16200`i1` expressions.
16201
16202This is how it can be used to represent guards as widenable branches:
16203
16204.. code-block:: text
16205
16206  block:
16207    ; Unguarded instructions
16208    call void @llvm.experimental.guard(i1 %cond, <args...>) ["deopt"(<deopt_args...>)]
16209    ; Guarded instructions
16210
16211Can be expressed in an alternative equivalent form of explicit branch using
16212``@llvm.experimental.widenable.condition``:
16213
16214.. code-block:: text
16215
16216  block:
16217    ; Unguarded instructions
16218    %widenable_condition = call i1 @llvm.experimental.widenable.condition()
16219    %guard_condition = and i1 %cond, %widenable_condition
16220    br i1 %guard_condition, label %guarded, label %deopt
16221
16222  guarded:
16223    ; Guarded instructions
16224
16225  deopt:
16226    call type @llvm.experimental.deoptimize(<args...>) [ "deopt"(<deopt_args...>) ]
16227
16228So the block `guarded` is only reachable when `%cond` is `true`,
16229and it should be valid to go to the block `deopt` whenever `%cond`
16230is `true` or `false`.
16231
16232``@llvm.experimental.widenable.condition`` will never throw, thus
16233it cannot be invoked.
16234
16235Guard widening:
16236"""""""""""""""
16237
16238When ``@llvm.experimental.widenable.condition()`` is used in
16239condition of a guard represented as explicit branch, it is
16240legal to widen the guard's condition with any additional
16241conditions.
16242
16243Guard widening looks like replacement of
16244
16245.. code-block:: text
16246
16247  %widenable_cond = call i1 @llvm.experimental.widenable.condition()
16248  %guard_cond = and i1 %cond, %widenable_cond
16249  br i1 %guard_cond, label %guarded, label %deopt
16250
16251with
16252
16253.. code-block:: text
16254
16255  %widenable_cond = call i1 @llvm.experimental.widenable.condition()
16256  %new_cond = and i1 %any_other_cond, %widenable_cond
16257  %new_guard_cond = and i1 %cond, %new_cond
16258  br i1 %new_guard_cond, label %guarded, label %deopt
16259
16260for this branch. Here `%any_other_cond` is an arbitrarily chosen
16261well-defined `i1` value. By making guard widening, we may
16262impose stricter conditions on `guarded` block and bail to the
16263deopt when the new condition is not met.
16264
16265Lowering:
16266"""""""""
16267
16268Default lowering strategy is replacing the result of
16269call of ``@llvm.experimental.widenable.condition``  with
16270constant `true`. However it is always correct to replace
16271it with any other `i1` value. Any pass can
16272freely do it if it can benefit from non-default lowering.
16273
16274
16275'``llvm.load.relative``' Intrinsic
16276^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
16277
16278Syntax:
16279"""""""
16280
16281::
16282
16283      declare i8* @llvm.load.relative.iN(i8* %ptr, iN %offset) argmemonly nounwind readonly
16284
16285Overview:
16286"""""""""
16287
16288This intrinsic loads a 32-bit value from the address ``%ptr + %offset``,
16289adds ``%ptr`` to that value and returns it. The constant folder specifically
16290recognizes the form of this intrinsic and the constant initializers it may
16291load from; if a loaded constant initializer is known to have the form
16292``i32 trunc(x - %ptr)``, the intrinsic call is folded to ``x``.
16293
16294LLVM provides that the calculation of such a constant initializer will
16295not overflow at link time under the medium code model if ``x`` is an
16296``unnamed_addr`` function. However, it does not provide this guarantee for
16297a constant initializer folded into a function body. This intrinsic can be
16298used to avoid the possibility of overflows when loading from such a constant.
16299
16300'``llvm.sideeffect``' Intrinsic
16301^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
16302
16303Syntax:
16304"""""""
16305
16306::
16307
16308      declare void @llvm.sideeffect() inaccessiblememonly nounwind
16309
16310Overview:
16311"""""""""
16312
16313The ``llvm.sideeffect`` intrinsic doesn't perform any operation. Optimizers
16314treat it as having side effects, so it can be inserted into a loop to
16315indicate that the loop shouldn't be assumed to terminate (which could
16316potentially lead to the loop being optimized away entirely), even if it's
16317an infinite loop with no other side effects.
16318
16319Arguments:
16320""""""""""
16321
16322None.
16323
16324Semantics:
16325""""""""""
16326
16327This intrinsic actually does nothing, but optimizers must assume that it
16328has externally observable side effects.
16329
16330'``llvm.is.constant.*``' Intrinsic
16331^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
16332
16333Syntax:
16334"""""""
16335
16336This is an overloaded intrinsic. You can use llvm.is.constant with any argument type.
16337
16338::
16339
16340      declare i1 @llvm.is.constant.i32(i32 %operand) nounwind readnone
16341      declare i1 @llvm.is.constant.f32(float %operand) nounwind readnone
16342      declare i1 @llvm.is.constant.TYPENAME(TYPE %operand) nounwind readnone
16343
16344Overview:
16345"""""""""
16346
16347The '``llvm.is.constant``' intrinsic will return true if the argument
16348is known to be a manifest compile-time constant. It is guaranteed to
16349fold to either true or false before generating machine code.
16350
16351Semantics:
16352""""""""""
16353
16354This intrinsic generates no code. If its argument is known to be a
16355manifest compile-time constant value, then the intrinsic will be
16356converted to a constant true value. Otherwise, it will be converted to
16357a constant false value.
16358
16359In particular, note that if the argument is a constant expression
16360which refers to a global (the address of which _is_ a constant, but
16361not manifest during the compile), then the intrinsic evaluates to
16362false.
16363
16364The result also intentionally depends on the result of optimization
16365passes -- e.g., the result can change depending on whether a
16366function gets inlined or not. A function's parameters are
16367obviously not constant. However, a call like
16368``llvm.is.constant.i32(i32 %param)`` *can* return true after the
16369function is inlined, if the value passed to the function parameter was
16370a constant.
16371
16372On the other hand, if constant folding is not run, it will never
16373evaluate to true, even in simple cases.
16374
16375Stack Map Intrinsics
16376--------------------
16377
16378LLVM provides experimental intrinsics to support runtime patching
16379mechanisms commonly desired in dynamic language JITs. These intrinsics
16380are described in :doc:`StackMaps`.
16381
16382Element Wise Atomic Memory Intrinsics
16383-------------------------------------
16384
16385These intrinsics are similar to the standard library memory intrinsics except
16386that they perform memory transfer as a sequence of atomic memory accesses.
16387
16388.. _int_memcpy_element_unordered_atomic:
16389
16390'``llvm.memcpy.element.unordered.atomic``' Intrinsic
16391^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
16392
16393Syntax:
16394"""""""
16395
16396This is an overloaded intrinsic. You can use ``llvm.memcpy.element.unordered.atomic`` on
16397any integer bit width and for different address spaces. Not all targets
16398support all bit widths however.
16399
16400::
16401
16402      declare void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i32(i8* <dest>,
16403                                                                       i8* <src>,
16404                                                                       i32 <len>,
16405                                                                       i32 <element_size>)
16406      declare void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* <dest>,
16407                                                                       i8* <src>,
16408                                                                       i64 <len>,
16409                                                                       i32 <element_size>)
16410
16411Overview:
16412"""""""""
16413
16414The '``llvm.memcpy.element.unordered.atomic.*``' intrinsic is a specialization of the
16415'``llvm.memcpy.*``' intrinsic. It differs in that the ``dest`` and ``src`` are treated
16416as arrays with elements that are exactly ``element_size`` bytes, and the copy between
16417buffers uses a sequence of :ref:`unordered atomic <ordering>` load/store operations
16418that are a positive integer multiple of the ``element_size`` in size.
16419
16420Arguments:
16421""""""""""
16422
16423The first three arguments are the same as they are in the :ref:`@llvm.memcpy <int_memcpy>`
16424intrinsic, with the added constraint that ``len`` is required to be a positive integer
16425multiple of the ``element_size``. If ``len`` is not a positive integer multiple of
16426``element_size``, then the behaviour of the intrinsic is undefined.
16427
16428``element_size`` must be a compile-time constant positive power of two no greater than
16429target-specific atomic access size limit.
16430
16431For each of the input pointers ``align`` parameter attribute must be specified. It
16432must be a power of two no less than the ``element_size``. Caller guarantees that
16433both the source and destination pointers are aligned to that boundary.
16434
16435Semantics:
16436""""""""""
16437
16438The '``llvm.memcpy.element.unordered.atomic.*``' intrinsic copies ``len`` bytes of
16439memory from the source location to the destination location. These locations are not
16440allowed to overlap. The memory copy is performed as a sequence of load/store operations
16441where each access is guaranteed to be a multiple of ``element_size`` bytes wide and
16442aligned at an ``element_size`` boundary.
16443
16444The order of the copy is unspecified. The same value may be read from the source
16445buffer many times, but only one write is issued to the destination buffer per
16446element. It is well defined to have concurrent reads and writes to both source and
16447destination provided those reads and writes are unordered atomic when specified.
16448
16449This intrinsic does not provide any additional ordering guarantees over those
16450provided by a set of unordered loads from the source location and stores to the
16451destination.
16452
16453Lowering:
16454"""""""""
16455
16456In the most general case call to the '``llvm.memcpy.element.unordered.atomic.*``' is
16457lowered to a call to the symbol ``__llvm_memcpy_element_unordered_atomic_*``. Where '*'
16458is replaced with an actual element size.
16459
16460Optimizer is allowed to inline memory copy when it's profitable to do so.
16461
16462'``llvm.memmove.element.unordered.atomic``' Intrinsic
16463^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
16464
16465Syntax:
16466"""""""
16467
16468This is an overloaded intrinsic. You can use
16469``llvm.memmove.element.unordered.atomic`` on any integer bit width and for
16470different address spaces. Not all targets support all bit widths however.
16471
16472::
16473
16474      declare void @llvm.memmove.element.unordered.atomic.p0i8.p0i8.i32(i8* <dest>,
16475                                                                        i8* <src>,
16476                                                                        i32 <len>,
16477                                                                        i32 <element_size>)
16478      declare void @llvm.memmove.element.unordered.atomic.p0i8.p0i8.i64(i8* <dest>,
16479                                                                        i8* <src>,
16480                                                                        i64 <len>,
16481                                                                        i32 <element_size>)
16482
16483Overview:
16484"""""""""
16485
16486The '``llvm.memmove.element.unordered.atomic.*``' intrinsic is a specialization
16487of the '``llvm.memmove.*``' intrinsic. It differs in that the ``dest`` and
16488``src`` are treated as arrays with elements that are exactly ``element_size``
16489bytes, and the copy between buffers uses a sequence of
16490:ref:`unordered atomic <ordering>` load/store operations that are a positive
16491integer multiple of the ``element_size`` in size.
16492
16493Arguments:
16494""""""""""
16495
16496The first three arguments are the same as they are in the
16497:ref:`@llvm.memmove <int_memmove>` intrinsic, with the added constraint that
16498``len`` is required to be a positive integer multiple of the ``element_size``.
16499If ``len`` is not a positive integer multiple of ``element_size``, then the
16500behaviour of the intrinsic is undefined.
16501
16502``element_size`` must be a compile-time constant positive power of two no
16503greater than a target-specific atomic access size limit.
16504
16505For each of the input pointers the ``align`` parameter attribute must be
16506specified. It must be a power of two no less than the ``element_size``. Caller
16507guarantees that both the source and destination pointers are aligned to that
16508boundary.
16509
16510Semantics:
16511""""""""""
16512
16513The '``llvm.memmove.element.unordered.atomic.*``' intrinsic copies ``len`` bytes
16514of memory from the source location to the destination location. These locations
16515are allowed to overlap. The memory copy is performed as a sequence of load/store
16516operations where each access is guaranteed to be a multiple of ``element_size``
16517bytes wide and aligned at an ``element_size`` boundary.
16518
16519The order of the copy is unspecified. The same value may be read from the source
16520buffer many times, but only one write is issued to the destination buffer per
16521element. It is well defined to have concurrent reads and writes to both source
16522and destination provided those reads and writes are unordered atomic when
16523specified.
16524
16525This intrinsic does not provide any additional ordering guarantees over those
16526provided by a set of unordered loads from the source location and stores to the
16527destination.
16528
16529Lowering:
16530"""""""""
16531
16532In the most general case call to the
16533'``llvm.memmove.element.unordered.atomic.*``' is lowered to a call to the symbol
16534``__llvm_memmove_element_unordered_atomic_*``. Where '*' is replaced with an
16535actual element size.
16536
16537The optimizer is allowed to inline the memory copy when it's profitable to do so.
16538
16539.. _int_memset_element_unordered_atomic:
16540
16541'``llvm.memset.element.unordered.atomic``' Intrinsic
16542^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
16543
16544Syntax:
16545"""""""
16546
16547This is an overloaded intrinsic. You can use ``llvm.memset.element.unordered.atomic`` on
16548any integer bit width and for different address spaces. Not all targets
16549support all bit widths however.
16550
16551::
16552
16553      declare void @llvm.memset.element.unordered.atomic.p0i8.i32(i8* <dest>,
16554                                                                  i8 <value>,
16555                                                                  i32 <len>,
16556                                                                  i32 <element_size>)
16557      declare void @llvm.memset.element.unordered.atomic.p0i8.i64(i8* <dest>,
16558                                                                  i8 <value>,
16559                                                                  i64 <len>,
16560                                                                  i32 <element_size>)
16561
16562Overview:
16563"""""""""
16564
16565The '``llvm.memset.element.unordered.atomic.*``' intrinsic is a specialization of the
16566'``llvm.memset.*``' intrinsic. It differs in that the ``dest`` is treated as an array
16567with elements that are exactly ``element_size`` bytes, and the assignment to that array
16568uses uses a sequence of :ref:`unordered atomic <ordering>` store operations
16569that are a positive integer multiple of the ``element_size`` in size.
16570
16571Arguments:
16572""""""""""
16573
16574The first three arguments are the same as they are in the :ref:`@llvm.memset <int_memset>`
16575intrinsic, with the added constraint that ``len`` is required to be a positive integer
16576multiple of the ``element_size``. If ``len`` is not a positive integer multiple of
16577``element_size``, then the behaviour of the intrinsic is undefined.
16578
16579``element_size`` must be a compile-time constant positive power of two no greater than
16580target-specific atomic access size limit.
16581
16582The ``dest`` input pointer must have the ``align`` parameter attribute specified. It
16583must be a power of two no less than the ``element_size``. Caller guarantees that
16584the destination pointer is aligned to that boundary.
16585
16586Semantics:
16587""""""""""
16588
16589The '``llvm.memset.element.unordered.atomic.*``' intrinsic sets the ``len`` bytes of
16590memory starting at the destination location to the given ``value``. The memory is
16591set with a sequence of store operations where each access is guaranteed to be a
16592multiple of ``element_size`` bytes wide and aligned at an ``element_size`` boundary.
16593
16594The order of the assignment is unspecified. Only one write is issued to the
16595destination buffer per element. It is well defined to have concurrent reads and
16596writes to the destination provided those reads and writes are unordered atomic
16597when specified.
16598
16599This intrinsic does not provide any additional ordering guarantees over those
16600provided by a set of unordered stores to the destination.
16601
16602Lowering:
16603"""""""""
16604
16605In the most general case call to the '``llvm.memset.element.unordered.atomic.*``' is
16606lowered to a call to the symbol ``__llvm_memset_element_unordered_atomic_*``. Where '*'
16607is replaced with an actual element size.
16608
16609The optimizer is allowed to inline the memory assignment when it's profitable to do so.
16610
16611Objective-C ARC Runtime Intrinsics
16612----------------------------------
16613
16614LLVM provides intrinsics that lower to Objective-C ARC runtime entry points.
16615LLVM is aware of the semantics of these functions, and optimizes based on that
16616knowledge. You can read more about the details of Objective-C ARC `here
16617<https://clang.llvm.org/docs/AutomaticReferenceCounting.html>`_.
16618
16619'``llvm.objc.autorelease``' Intrinsic
16620^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
16621
16622Syntax:
16623"""""""
16624::
16625
16626      declare i8* @llvm.objc.autorelease(i8*)
16627
16628Lowering:
16629"""""""""
16630
16631Lowers to a call to `objc_autorelease <https://clang.llvm.org/docs/AutomaticReferenceCounting.html#arc-runtime-objc-autorelease>`_.
16632
16633'``llvm.objc.autoreleasePoolPop``' Intrinsic
16634^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
16635
16636Syntax:
16637"""""""
16638::
16639
16640      declare void @llvm.objc.autoreleasePoolPop(i8*)
16641
16642Lowering:
16643"""""""""
16644
16645Lowers to a call to `objc_autoreleasePoolPop <https://clang.llvm.org/docs/AutomaticReferenceCounting.html#void-objc-autoreleasepoolpop-void-pool>`_.
16646
16647'``llvm.objc.autoreleasePoolPush``' Intrinsic
16648^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
16649
16650Syntax:
16651"""""""
16652::
16653
16654      declare i8* @llvm.objc.autoreleasePoolPush()
16655
16656Lowering:
16657"""""""""
16658
16659Lowers to a call to `objc_autoreleasePoolPush <https://clang.llvm.org/docs/AutomaticReferenceCounting.html#void-objc-autoreleasepoolpush-void>`_.
16660
16661'``llvm.objc.autoreleaseReturnValue``' Intrinsic
16662^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
16663
16664Syntax:
16665"""""""
16666::
16667
16668      declare i8* @llvm.objc.autoreleaseReturnValue(i8*)
16669
16670Lowering:
16671"""""""""
16672
16673Lowers to a call to `objc_autoreleaseReturnValue <https://clang.llvm.org/docs/AutomaticReferenceCounting.html#arc-runtime-objc-autoreleasereturnvalue>`_.
16674
16675'``llvm.objc.copyWeak``' Intrinsic
16676^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
16677
16678Syntax:
16679"""""""
16680::
16681
16682      declare void @llvm.objc.copyWeak(i8**, i8**)
16683
16684Lowering:
16685"""""""""
16686
16687Lowers to a call to `objc_copyWeak <https://clang.llvm.org/docs/AutomaticReferenceCounting.html#void-objc-copyweak-id-dest-id-src>`_.
16688
16689'``llvm.objc.destroyWeak``' Intrinsic
16690^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
16691
16692Syntax:
16693"""""""
16694::
16695
16696      declare void @llvm.objc.destroyWeak(i8**)
16697
16698Lowering:
16699"""""""""
16700
16701Lowers to a call to `objc_destroyWeak <https://clang.llvm.org/docs/AutomaticReferenceCounting.html#void-objc-destroyweak-id-object>`_.
16702
16703'``llvm.objc.initWeak``' Intrinsic
16704^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
16705
16706Syntax:
16707"""""""
16708::
16709
16710      declare i8* @llvm.objc.initWeak(i8**, i8*)
16711
16712Lowering:
16713"""""""""
16714
16715Lowers to a call to `objc_initWeak <https://clang.llvm.org/docs/AutomaticReferenceCounting.html#arc-runtime-objc-initweak>`_.
16716
16717'``llvm.objc.loadWeak``' Intrinsic
16718^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
16719
16720Syntax:
16721"""""""
16722::
16723
16724      declare i8* @llvm.objc.loadWeak(i8**)
16725
16726Lowering:
16727"""""""""
16728
16729Lowers to a call to `objc_loadWeak <https://clang.llvm.org/docs/AutomaticReferenceCounting.html#arc-runtime-objc-loadweak>`_.
16730
16731'``llvm.objc.loadWeakRetained``' Intrinsic
16732^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
16733
16734Syntax:
16735"""""""
16736::
16737
16738      declare i8* @llvm.objc.loadWeakRetained(i8**)
16739
16740Lowering:
16741"""""""""
16742
16743Lowers to a call to `objc_loadWeakRetained <https://clang.llvm.org/docs/AutomaticReferenceCounting.html#arc-runtime-objc-loadweakretained>`_.
16744
16745'``llvm.objc.moveWeak``' Intrinsic
16746^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
16747
16748Syntax:
16749"""""""
16750::
16751
16752      declare void @llvm.objc.moveWeak(i8**, i8**)
16753
16754Lowering:
16755"""""""""
16756
16757Lowers to a call to `objc_moveWeak <https://clang.llvm.org/docs/AutomaticReferenceCounting.html#void-objc-moveweak-id-dest-id-src>`_.
16758
16759'``llvm.objc.release``' Intrinsic
16760^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
16761
16762Syntax:
16763"""""""
16764::
16765
16766      declare void @llvm.objc.release(i8*)
16767
16768Lowering:
16769"""""""""
16770
16771Lowers to a call to `objc_release <https://clang.llvm.org/docs/AutomaticReferenceCounting.html#void-objc-release-id-value>`_.
16772
16773'``llvm.objc.retain``' Intrinsic
16774^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
16775
16776Syntax:
16777"""""""
16778::
16779
16780      declare i8* @llvm.objc.retain(i8*)
16781
16782Lowering:
16783"""""""""
16784
16785Lowers to a call to `objc_retain <https://clang.llvm.org/docs/AutomaticReferenceCounting.html#arc-runtime-objc-retain>`_.
16786
16787'``llvm.objc.retainAutorelease``' Intrinsic
16788^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
16789
16790Syntax:
16791"""""""
16792::
16793
16794      declare i8* @llvm.objc.retainAutorelease(i8*)
16795
16796Lowering:
16797"""""""""
16798
16799Lowers to a call to `objc_retainAutorelease <https://clang.llvm.org/docs/AutomaticReferenceCounting.html#arc-runtime-objc-retainautorelease>`_.
16800
16801'``llvm.objc.retainAutoreleaseReturnValue``' Intrinsic
16802^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
16803
16804Syntax:
16805"""""""
16806::
16807
16808      declare i8* @llvm.objc.retainAutoreleaseReturnValue(i8*)
16809
16810Lowering:
16811"""""""""
16812
16813Lowers to a call to `objc_retainAutoreleaseReturnValue <https://clang.llvm.org/docs/AutomaticReferenceCounting.html#arc-runtime-objc-retainautoreleasereturnvalue>`_.
16814
16815'``llvm.objc.retainAutoreleasedReturnValue``' Intrinsic
16816^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
16817
16818Syntax:
16819"""""""
16820::
16821
16822      declare i8* @llvm.objc.retainAutoreleasedReturnValue(i8*)
16823
16824Lowering:
16825"""""""""
16826
16827Lowers to a call to `objc_retainAutoreleasedReturnValue <https://clang.llvm.org/docs/AutomaticReferenceCounting.html#arc-runtime-objc-retainautoreleasedreturnvalue>`_.
16828
16829'``llvm.objc.retainBlock``' Intrinsic
16830^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
16831
16832Syntax:
16833"""""""
16834::
16835
16836      declare i8* @llvm.objc.retainBlock(i8*)
16837
16838Lowering:
16839"""""""""
16840
16841Lowers to a call to `objc_retainBlock <https://clang.llvm.org/docs/AutomaticReferenceCounting.html#arc-runtime-objc-retainblock>`_.
16842
16843'``llvm.objc.storeStrong``' Intrinsic
16844^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
16845
16846Syntax:
16847"""""""
16848::
16849
16850      declare void @llvm.objc.storeStrong(i8**, i8*)
16851
16852Lowering:
16853"""""""""
16854
16855Lowers to a call to `objc_storeStrong <https://clang.llvm.org/docs/AutomaticReferenceCounting.html#void-objc-storestrong-id-object-id-value>`_.
16856
16857'``llvm.objc.storeWeak``' Intrinsic
16858^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
16859
16860Syntax:
16861"""""""
16862::
16863
16864      declare i8* @llvm.objc.storeWeak(i8**, i8*)
16865
16866Lowering:
16867"""""""""
16868
16869Lowers to a call to `objc_storeWeak <https://clang.llvm.org/docs/AutomaticReferenceCounting.html#arc-runtime-objc-storeweak>`_.
16870