1 ==============================
2 LLVM Language Reference Manual
3 ==============================
12 This document is a reference manual for the LLVM assembly language. LLVM
13 is a Static Single Assignment (SSA) based representation that provides
14 type safety, low-level operations, flexibility, and the capability of
15 representing 'all' high-level languages cleanly. It is the common code
16 representation used throughout all phases of the LLVM compilation
22 The LLVM code representation is designed to be used in three different
23 forms: as an in-memory compiler IR, as an on-disk bitcode representation
24 (suitable for fast loading by a Just-In-Time compiler), and as a human
25 readable assembly language representation. This allows LLVM to provide a
26 powerful intermediate representation for efficient compiler
27 transformations and analysis, while providing a natural means to debug
28 and visualize the transformations. The three different forms of LLVM are
29 all equivalent. This document describes the human readable
30 representation and notation.
32 The LLVM representation aims to be light-weight and low-level while
33 being expressive, typed, and extensible at the same time. It aims to be
34 a "universal IR" of sorts, by being at a low enough level that
35 high-level ideas may be cleanly mapped to it (similar to how
36 microprocessors are "universal IR's", allowing many source languages to
37 be mapped to them). By providing type information, LLVM can be used as
38 the target of optimizations: for example, through pointer analysis, it
39 can be proven that a C automatic variable is never accessed outside of
40 the current function, allowing it to be promoted to a simple SSA value
41 instead of a memory location.
48 It is important to note that this document describes 'well formed' LLVM
49 assembly language. There is a difference between what the parser accepts
50 and what is considered 'well formed'. For example, the following
51 instruction is syntactically okay, but not well formed:
57 because the definition of ``%x`` does not dominate all of its uses. The
58 LLVM infrastructure provides a verification pass that may be used to
59 verify that an LLVM module is well formed. This pass is automatically
60 run by the parser after parsing input assembly and by the optimizer
61 before it outputs bitcode. The violations pointed out by the verifier
62 pass indicate bugs in transformation passes or input to the parser.
69 LLVM identifiers come in two basic types: global and local. Global
70 identifiers (functions, global variables) begin with the ``'@'``
71 character. Local identifiers (register names, types) begin with the
72 ``'%'`` character. Additionally, there are three different formats for
73 identifiers, for different purposes:
75 #. Named values are represented as a string of characters with their
76 prefix. For example, ``%foo``, ``@DivisionByZero``,
77 ``%a.really.long.identifier``. The actual regular expression used is
78 '``[%@][-a-zA-Z$._][-a-zA-Z$._0-9]*``'. Identifiers that require other
79 characters in their names can be surrounded with quotes. Special
80 characters may be escaped using ``"\xx"`` where ``xx`` is the ASCII
81 code for the character in hexadecimal. In this way, any character can
82 be used in a name value, even quotes themselves. The ``"\01"`` prefix
83 can be used on global variables to suppress mangling.
84 #. Unnamed values are represented as an unsigned numeric value with
85 their prefix. For example, ``%12``, ``@2``, ``%44``.
86 #. Constants, which are described in the section Constants_ below.
88 LLVM requires that values start with a prefix for two reasons: Compilers
89 don't need to worry about name clashes with reserved words, and the set
90 of reserved words may be expanded in the future without penalty.
91 Additionally, unnamed identifiers allow a compiler to quickly come up
92 with a temporary variable without having to avoid symbol table
95 Reserved words in LLVM are very similar to reserved words in other
96 languages. There are keywords for different opcodes ('``add``',
97 '``bitcast``', '``ret``', etc...), for primitive type names ('``void``',
98 '``i32``', etc...), and others. These reserved words cannot conflict
99 with variable names, because none of them start with a prefix character
100 (``'%'`` or ``'@'``).
102 Here is an example of LLVM code to multiply the integer variable
109 %result = mul i32 %X, 8
111 After strength reduction:
115 %result = shl i32 %X, 3
121 %0 = add i32 %X, %X ; yields i32:%0
122 %1 = add i32 %0, %0 ; yields i32:%1
123 %result = add i32 %1, %1
125 This last way of multiplying ``%X`` by 8 illustrates several important
126 lexical features of LLVM:
128 #. Comments are delimited with a '``;``' and go until the end of line.
129 #. Unnamed temporaries are created when the result of a computation is
130 not assigned to a named value.
131 #. Unnamed temporaries are numbered sequentially (using a per-function
132 incrementing counter, starting with 0). Note that basic blocks and unnamed
133 function parameters are included in this numbering. For example, if the
134 entry basic block is not given a label name and all function parameters are
135 named, then it will get number 0.
137 It also shows a convention that we follow in this document. When
138 demonstrating instructions, we will follow an instruction with a comment
139 that defines the type and name of value produced.
147 LLVM programs are composed of ``Module``'s, each of which is a
148 translation unit of the input programs. Each module consists of
149 functions, global variables, and symbol table entries. Modules may be
150 combined together with the LLVM linker, which merges function (and
151 global variable) definitions, resolves forward declarations, and merges
152 symbol table entries. Here is an example of the "hello world" module:
156 ; Declare the string constant as a global constant.
157 @.str = private unnamed_addr constant [13 x i8] c"hello world\0A\00"
159 ; External declaration of the puts function
160 declare i32 @puts(i8* nocapture) nounwind
162 ; Definition of main function
163 define i32 @main() { ; i32()*
164 ; Convert [13 x i8]* to i8*...
165 %cast210 = getelementptr [13 x i8], [13 x i8]* @.str, i64 0, i64 0
167 ; Call puts function to write out the string to stdout.
168 call i32 @puts(i8* %cast210)
173 !0 = !{i32 42, null, !"string"}
176 This example is made up of a :ref:`global variable <globalvars>` named
177 "``.str``", an external declaration of the "``puts``" function, a
178 :ref:`function definition <functionstructure>` for "``main``" and
179 :ref:`named metadata <namedmetadatastructure>` "``foo``".
181 In general, a module is made up of a list of global values (where both
182 functions and global variables are global values). Global values are
183 represented by a pointer to a memory location (in this case, a pointer
184 to an array of char, and a pointer to a function), and have one of the
185 following :ref:`linkage types <linkage>`.
192 All Global Variables and Functions have one of the following types of
196 Global values with "``private``" linkage are only directly
197 accessible by objects in the current module. In particular, linking
198 code into a module with 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.
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.
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.
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"
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
244 .. _linkage_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
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.
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.
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.
275 It is illegal for a function *declaration* to have any linkage type
276 other than ``external`` or ``extern_weak``.
283 LLVM :ref:`functions <functionstructure>`, :ref:`calls <i_call>` and
284 :ref:`invokes <i_invoke>` can all have an optional calling convention
285 specified for the call. The calling convention of any pair of dynamic
286 caller/callee must match, or the behavior of the program is undefined.
287 The following calling conventions are supported by LLVM, and more may be
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
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
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.
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
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.
371 - On X86-64 the callee preserves all general purpose registers, except for
372 R11. R11 can be used as a scratch register. Floating-point registers
373 (XMMs/YMMs) are not preserved and need to be saved by the caller.
375 The idea behind this convention is to support calls to runtime functions
376 that have a hot path and a cold path. The hot path is usually a small piece
377 of code that doesn't 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.
387 This calling convention will be used by a future version of the ObjectiveC
388 runtime and should therefore still be considered experimental at this time.
389 Although this convention was created to optimize certain runtime calls to
390 the ObjectiveC runtime, it is not limited to this runtime and might be used
391 by other runtimes in the future too. The current implementation only
392 supports X86-64, but the intention is to support more architectures in the
394 "``preserve_allcc``" - The `PreserveAll` calling convention
395 This calling convention attempts to make the code in the caller even less
396 intrusive than the `PreserveMost` calling convention. This calling
397 convention also behaves identical to the `C` calling convention on how
398 arguments and return values are passed, but it uses a different set of
399 caller/callee-saved registers. This removes the burden of saving and
400 recovering a large register set before and after the call in the caller. 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.
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).
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.
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
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).
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.
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`.
433 - On X86-64 the callee preserves all general purpose registers, except for
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.
444 More calling conventions can be added/defined on an as-needed basis, to
445 support Pascal conventions or any other well-known target-independent
448 .. _visibilitystyles:
453 All Global Variables and Functions have one of the following visibility
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.
475 A symbol with ``internal`` or ``private`` linkage must have ``default``
483 All Global Variables, Functions and Aliases can have one of the following
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.
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.
501 Thread Local Storage Models
502 ---------------------------
504 A variable may be defined as ``thread_local``, which means that it will
505 not be shared by threads (each thread will have a separated copy of the
506 variable). Not all targets support thread-local variables. Optionally, a
507 TLS model may be specified:
510 For variables that are only used within the current shared library.
512 For variables in modules that will not be loaded dynamically.
514 For variables defined in the executable and only used within it.
516 If no explicit model is given, the "general dynamic" model is used.
518 The models correspond to the ELF TLS models; see `ELF Handling For
519 Thread-Local Storage <http://people.redhat.com/drepper/tls.pdf>`_ for
520 more information on under which circumstances the different models may
521 be used. The target may choose a different TLS model if the specified
522 model is not supported, or if a better choice of model can be made.
524 A model can also be specified in an alias, but then it only governs how
525 the alias is accessed. It will not have any effect in the aliasee.
527 For platforms without linker support of ELF TLS model, the -femulated-tls
528 flag can be used to generate GCC compatible emulated TLS code.
530 .. _runtime_preemption_model:
532 Runtime Preemption Specifiers
533 -----------------------------
535 Global variables, functions and aliases may have an optional runtime preemption
536 specifier. If a preemption specifier isn't given explicitly, then a
537 symbol is assumed to be ``dso_preemptable``.
540 Indicates that the function or variable may be replaced by a symbol from
541 outside the linkage unit at runtime.
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.
553 LLVM IR allows you to specify both "identified" and "literal" :ref:`structure
554 types <t_struct>`. Literal types are uniqued structurally, but identified types
555 are never uniqued. An :ref:`opaque structural type <t_opaque>` can also be used
556 to forward declare a type that is not yet available.
558 An example of an identified structure specification is:
562 %mytype = type { %mytype*, i32 }
564 Prior to the LLVM 3.0 release, identified types were structurally uniqued. Only
565 literal types are uniqued in recent versions of LLVM.
569 Non-Integral Pointer Type
570 -------------------------
572 Note: non-integral pointer types are a work in progress, and they should be
573 considered experimental at this time.
575 LLVM IR optionally allows the frontend to denote pointers in certain address
576 spaces as "non-integral" via the :ref:`datalayout string<langref_datalayout>`.
577 Non-integral pointer types represent pointers that have an *unspecified* bitwise
578 representation; that is, the integral representation may be target dependent or
579 unstable (not backed by a fixed integer).
581 ``inttoptr`` instructions converting integers to non-integral pointer types are
582 ill-typed, and so are ``ptrtoint`` instructions converting values of
583 non-integral pointer types to integers. Vector versions of said instructions
584 are ill-typed as well.
591 Global variables define regions of memory allocated at compilation time
594 Global variable definitions must be initialized.
596 Global variables in other translation units can also be declared, in which
597 case they don't have an initializer.
599 Either global variable definitions or declarations may have an explicit section
600 to be placed in and may have an optional explicit alignment specified. If there
601 is a mismatch between the explicit or inferred section information for the
602 variable declaration and its definition the resulting behavior is undefined.
604 A variable may be defined as a global ``constant``, which indicates that
605 the contents of the variable will **never** be modified (enabling better
606 optimization, allowing the global data to be placed in the read-only
607 section of an executable, etc). Note that variables that need runtime
608 initialization cannot be marked ``constant`` as there is a store to the
611 LLVM explicitly allows *declarations* of global variables to be marked
612 constant, even if the final definition of the global is not. This
613 capability can be used to enable slightly better optimization of the
614 program, but requires the language definition to guarantee that
615 optimizations based on the 'constantness' are valid for the translation
616 units that do not include the definition.
618 As SSA values, global variables define pointer values that are in scope
619 (i.e. they dominate) all basic blocks in the program. Global variables
620 always define a pointer to their "content" type because they describe a
621 region of memory, and all memory objects in LLVM are accessed through
624 Global variables can be marked with ``unnamed_addr`` which indicates
625 that the address is not significant, only the content. Constants marked
626 like this can be merged with other constants if they have the same
627 initializer. Note that a constant with significant address *can* be
628 merged with a ``unnamed_addr`` constant, the result being a constant
629 whose address is significant.
631 If the ``local_unnamed_addr`` attribute is given, the address is known to
632 not be significant within the module.
634 A global variable may be declared to reside in a target-specific
635 numbered address space. For targets that support them, address spaces
636 may affect how optimizations are performed and/or what target
637 instructions are used to access the variable. The default address space
638 is zero. The address space qualifier must precede any other attributes.
640 LLVM allows an explicit section to be specified for globals. If the
641 target supports it, it will emit globals to the section specified.
642 Additionally, the global can placed in a comdat if the target has the necessary
645 External declarations may have an explicit section specified. Section
646 information is retained in LLVM IR for targets that make use of this
647 information. Attaching section information to an external declaration is an
648 assertion that its definition is located in the specified section. If the
649 definition is located in a different section, the behavior is undefined.
651 By default, global initializers are optimized by assuming that global
652 variables defined within the module are not modified from their
653 initial values before the start of the global initializer. This is
654 true even for variables potentially accessible from outside the
655 module, including those with external linkage or appearing in
656 ``@llvm.used`` or dllexported variables. This assumption may be suppressed
657 by marking the variable with ``externally_initialized``.
659 An explicit alignment may be specified for a global, which must be a
660 power of 2. If not present, or if the alignment is set to zero, the
661 alignment of the global is set by the target to whatever it feels
662 convenient. If an explicit alignment is specified, the global is forced
663 to have exactly that alignment. Targets and optimizers are not allowed
664 to over-align the global if the global has an assigned section. In this
665 case, the extra alignment could be observable: for example, code could
666 assume that the globals are densely packed in their section and try to
667 iterate over them as an array, alignment padding would break this
668 iteration. The maximum alignment is ``1 << 29``.
670 Globals can also have a :ref:`DLL storage class <dllstorageclass>`,
671 an optional :ref:`runtime preemption specifier <runtime_preemption_model>`,
672 an optional :ref:`global attributes <glattrs>` and
673 an optional list of attached :ref:`metadata <metadata>`.
675 Variables and aliases can have a
676 :ref:`Thread Local Storage Model <tls_model>`.
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)*
688 For example, the following defines a global in a numbered address space
689 with an initializer, section, and alignment:
693 @G = addrspace(5) constant float 1.0, section "foo", align 4
695 The following example just declares a global variable
699 @G = external global i32
701 The following example defines a thread-local global with the
702 ``initialexec`` TLS model:
706 @G = thread_local(initialexec) global i32 0, align 4
708 .. _functionstructure:
713 LLVM function definitions consist of the "``define``" keyword, an
714 optional :ref:`linkage type <linkage>`, an optional :ref:`runtime preemption
715 specifier <runtime_preemption_model>`, an optional :ref:`visibility
716 style <visibility>`, an optional :ref:`DLL storage class <dllstorageclass>`,
717 an optional :ref:`calling convention <callingconv>`,
718 an optional ``unnamed_addr`` attribute, a return type, an optional
719 :ref:`parameter attribute <paramattrs>` for the return type, a function
720 name, a (possibly empty) argument list (each with optional :ref:`parameter
721 attributes <paramattrs>`), optional :ref:`function attributes <fnattrs>`,
722 an optional section, an optional alignment,
723 an optional :ref:`comdat <langref_comdats>`,
724 an optional :ref:`garbage collector name <gc>`, an optional :ref:`prefix <prefixdata>`,
725 an optional :ref:`prologue <prologuedata>`,
726 an optional :ref:`personality <personalityfn>`,
727 an optional list of attached :ref:`metadata <metadata>`,
728 an opening curly brace, a list of basic blocks, and a closing curly brace.
730 LLVM function declarations consist of the "``declare``" keyword, an
731 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility style
732 <visibility>`, an optional :ref:`DLL storage class <dllstorageclass>`, an
733 optional :ref:`calling convention <callingconv>`, an optional ``unnamed_addr``
734 or ``local_unnamed_addr`` attribute, a return type, an optional :ref:`parameter
735 attribute <paramattrs>` for the return type, a function name, a possibly
736 empty list of arguments, an optional alignment, an optional :ref:`garbage
737 collector name <gc>`, an optional :ref:`prefix <prefixdata>`, and an optional
738 :ref:`prologue <prologuedata>`.
740 A function definition contains a list of basic blocks, forming the CFG (Control
741 Flow 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,
743 and ends with a :ref:`terminator <terminators>` instruction (such as a branch or
744 function return). If an explicit label is not provided, a block is assigned an
745 implicit numbered label, using the next value from the same counter as used for
746 unnamed temporaries (:ref:`see above<identifiers>`). For example, if a function
747 entry block does not have an explicit label, it will be assigned label "%0",
748 then the first unnamed temporary in that block will be "%1", etc.
750 The first basic block in a function is special in two ways: it is
751 immediately executed on entrance to the function, and it is not allowed
752 to have predecessor basic blocks (i.e. there can not be any branches to
753 the entry block of a function). Because the block can have no
754 predecessors, it also cannot have any :ref:`PHI nodes <i_phi>`.
756 LLVM allows an explicit section to be specified for functions. If the
757 target supports it, it will emit functions to the section specified.
758 Additionally, the function can be placed in a COMDAT.
760 An explicit alignment may be specified for a function. If not present,
761 or if the alignment is set to zero, the alignment of the function is set
762 by the target to whatever it feels convenient. If an explicit alignment
763 is specified, the function is forced to have at least that much
764 alignment. All alignments must be a power of 2.
766 If the ``unnamed_addr`` attribute is given, the address is known to not
767 be significant and two identical functions can be merged.
769 If the ``local_unnamed_addr`` attribute is given, the address is known to
770 not be significant within the module.
774 define [linkage] [PreemptionSpecifier] [visibility] [DLLStorageClass]
776 <ResultType> @<FunctionName> ([argument list])
777 [(unnamed_addr|local_unnamed_addr)] [fn Attrs] [section "name"]
778 [comdat [($name)]] [align N] [gc] [prefix Constant]
779 [prologue Constant] [personality Constant] (!name !N)* { ... }
781 The argument list is a comma separated sequence of arguments where each
782 argument is of the following form:
786 <type> [parameter Attrs] [name]
794 Aliases, unlike function or variables, don't create any new data. They
795 are just a new symbol and metadata for an existing position.
797 Aliases have a name and an aliasee that is either a global value or a
800 Aliases may have an optional :ref:`linkage type <linkage>`, an optional
801 :ref:`runtime preemption specifier <runtime_preemption_model>`, an optional
802 :ref:`visibility style <visibility>`, an optional :ref:`DLL storage class
803 <dllstorageclass>` and an optional :ref:`tls model <tls_model>`.
807 @<Name> = [Linkage] [PreemptionSpecifier] [Visibility] [DLLStorageClass] [ThreadLocal] [(unnamed_addr|local_unnamed_addr)] alias <AliaseeTy>, <AliaseeTy>* @<Aliasee>
809 The linkage must be one of ``private``, ``internal``, ``linkonce``, ``weak``,
810 ``linkonce_odr``, ``weak_odr``, ``external``. Note that some system linkers
811 might not correctly handle dropping a weak symbol that is aliased.
813 Aliases that are not ``unnamed_addr`` are guaranteed to have the same address as
814 the aliasee expression. ``unnamed_addr`` ones are only guaranteed to point
817 If the ``local_unnamed_addr`` attribute is given, the address is known to
818 not be significant within the module.
820 Since aliases are only a second name, some restrictions apply, of which
821 some can only be checked when producing an object file:
823 * The expression defining the aliasee must be computable at assembly
824 time. Since it is just a name, no relocations can be used.
826 * No alias in the expression can be weak as the possibility of the
827 intermediate alias being overridden cannot be represented in an
830 * No global value in the expression can be a declaration, since that
831 would require a relocation, which is not possible.
838 IFuncs, like as aliases, don't create any new data or func. They are just a new
839 symbol that dynamic linker resolves at runtime by calling a resolver function.
841 IFuncs have a name and a resolver that is a function called by dynamic linker
842 that returns address of another function associated with the name.
844 IFunc may have an optional :ref:`linkage type <linkage>` and an optional
845 :ref:`visibility style <visibility>`.
849 @<Name> = [Linkage] [Visibility] ifunc <IFuncTy>, <ResolverTy>* @<Resolver>
857 Comdat IR provides access to COFF and ELF object file COMDAT functionality.
859 Comdats have a name which represents the COMDAT key. All global objects that
860 specify this key will only end up in the final object file if the linker chooses
861 that key over some other key. Aliases are placed in the same COMDAT that their
862 aliasee computes to, if any.
864 Comdats have a selection kind to provide input on how the linker should
865 choose between keys in two different object files.
869 $<Name> = comdat SelectionKind
871 The selection kind must be one of the following:
874 The linker may choose any COMDAT key, the choice is arbitrary.
876 The linker may choose any COMDAT key but the sections must contain the
879 The linker will choose the section containing the largest COMDAT key.
881 The linker requires that only section with this COMDAT key exist.
883 The linker may choose any COMDAT key but the sections must contain the
886 Note that the Mach-O platform doesn't support COMDATs, and ELF and WebAssembly
887 only support ``any`` as a selection kind.
889 Here is an example of a COMDAT group where a function will only be selected if
890 the COMDAT key's section is the largest:
894 $foo = comdat largest
895 @foo = global i32 2, comdat($foo)
897 define void @bar() comdat($foo) {
901 As a syntactic sugar the ``$name`` can be omitted if the name is the same as
907 @foo = global i32 2, comdat
910 In a COFF object file, this will create a COMDAT section with selection kind
911 ``IMAGE_COMDAT_SELECT_LARGEST`` containing the contents of the ``@foo`` symbol
912 and another COMDAT section with selection kind
913 ``IMAGE_COMDAT_SELECT_ASSOCIATIVE`` which is associated with the first COMDAT
914 section and contains the contents of the ``@bar`` symbol.
916 There are some restrictions on the properties of the global object.
917 It, or an alias to it, must have the same name as the COMDAT group when
919 The contents and size of this object may be used during link-time to determine
920 which COMDAT groups get selected depending on the selection kind.
921 Because the name of the object must match the name of the COMDAT group, the
922 linkage of the global object must not be local; local symbols can get renamed
923 if a collision occurs in the symbol table.
925 The combined use of COMDATS and section attributes may yield surprising results.
932 @g1 = global i32 42, section "sec", comdat($foo)
933 @g2 = global i32 42, section "sec", comdat($bar)
935 From the object file perspective, this requires the creation of two sections
936 with the same name. This is necessary because both globals belong to different
937 COMDAT groups and COMDATs, at the object file level, are represented by
940 Note that certain IR constructs like global variables and functions may
941 create COMDATs in the object file in addition to any which are specified using
942 COMDAT IR. This arises when the code generator is configured to emit globals
943 in individual sections (e.g. when `-data-sections` or `-function-sections`
944 is supplied to `llc`).
946 .. _namedmetadatastructure:
951 Named metadata is a collection of metadata. :ref:`Metadata
952 nodes <metadata>` (but not metadata strings) are the only valid
953 operands for a named metadata.
955 #. Named metadata are represented as a string of characters with the
956 metadata prefix. The rules for metadata names are the same as for
957 identifiers, but quoted names are not allowed. ``"\xx"`` type escapes
958 are still valid, which allows any character to be part of a name.
962 ; Some unnamed metadata nodes, which are referenced by the named metadata.
967 !name = !{!0, !1, !2}
974 The return type and each parameter of a function type may have a set of
975 *parameter attributes* associated with them. Parameter attributes are
976 used to communicate additional information about the result or
977 parameters of a function. Parameter attributes are considered to be part
978 of the function, not of the function type, so functions with different
979 parameter attributes can have the same function type.
981 Parameter attributes are simple keywords that follow the type specified.
982 If multiple parameter attributes are needed, they are space separated.
987 declare i32 @printf(i8* noalias nocapture, ...)
988 declare i32 @atoi(i8 zeroext)
989 declare signext i8 @returns_signed_char()
991 Note that any attributes for the function result (``nounwind``,
992 ``readonly``) come immediately after the argument list.
994 Currently, only the following parameter attributes are defined:
997 This indicates to the code generator that the parameter or return
998 value should be zero-extended to the extent required by the target's
999 ABI by the caller (for a parameter) or the callee (for a return value).
1001 This indicates to the code generator that the parameter or return
1002 value should be sign-extended to the extent required by the target's
1003 ABI (which is usually 32-bits) by the caller (for a parameter) or
1004 the callee (for a return value).
1006 This indicates that this parameter or return value should be treated
1007 in a special target-dependent fashion while emitting code for
1008 a function call or return (usually, by putting it in a register as
1009 opposed to memory, though some targets use it to distinguish between
1010 two different kinds of registers). Use of this attribute is
1013 This indicates that the pointer parameter should really be passed by
1014 value to the function. The attribute implies that a hidden copy of
1015 the pointee is made between the caller and the callee, so the callee
1016 is unable to modify the value in the caller. This attribute is only
1017 valid on LLVM pointer arguments. It is generally used to pass
1018 structs and arrays by value, but is also valid on pointers to
1019 scalars. The copy is considered to belong to the caller not the
1020 callee (for example, ``readonly`` functions should not write to
1021 ``byval`` parameters). This is not a valid attribute for return
1024 The byval attribute also supports specifying an alignment with the
1025 align attribute. It indicates the alignment of the stack slot to
1026 form and the known alignment of the pointer specified to the call
1027 site. If the alignment is not specified, then the code generator
1028 makes a target-specific assumption.
1034 The ``inalloca`` argument attribute allows the caller to take the
1035 address of outgoing stack arguments. An ``inalloca`` argument must
1036 be a pointer to stack memory produced by an ``alloca`` instruction.
1037 The alloca, or argument allocation, must also be tagged with the
1038 inalloca keyword. Only the last argument may have the ``inalloca``
1039 attribute, and that argument is guaranteed to be passed in memory.
1041 An argument allocation may be used by a call at most once because
1042 the call may deallocate it. The ``inalloca`` attribute cannot be
1043 used in conjunction with other attributes that affect argument
1044 storage, like ``inreg``, ``nest``, ``sret``, or ``byval``. The
1045 ``inalloca`` attribute also disables LLVM's implicit lowering of
1046 large aggregate return values, which means that frontend authors
1047 must lower them with ``sret`` pointers.
1049 When the call site is reached, the argument allocation must have
1050 been the most recent stack allocation that is still live, or the
1051 results are undefined. It is possible to allocate additional stack
1052 space after an argument allocation and before its call site, but it
1053 must be cleared off with :ref:`llvm.stackrestore
1054 <int_stackrestore>`.
1056 See :doc:`InAlloca` for more information on how to use this
1060 This indicates that the pointer parameter specifies the address of a
1061 structure that is the return value of the function in the source
1062 program. This pointer must be guaranteed by the caller to be valid:
1063 loads and stores to the structure may be assumed by the callee not
1064 to trap and to be properly aligned. This is not a valid attribute
1070 This indicates that the pointer value may be assumed by the optimizer to
1071 have the specified alignment.
1073 Note that this attribute has additional semantics when combined with the
1074 ``byval`` attribute.
1079 This indicates that objects accessed via pointer values
1080 :ref:`based <pointeraliasing>` on the argument or return value are not also
1081 accessed, during the execution of the function, via pointer values not
1082 *based* on the argument or return value. The attribute on a return value
1083 also has additional semantics described below. The caller shares the
1084 responsibility with the callee for ensuring that these requirements are met.
1085 For further details, please see the discussion of the NoAlias response in
1086 :ref:`alias analysis <Must, May, or No>`.
1088 Note that this definition of ``noalias`` is intentionally similar
1089 to the definition of ``restrict`` in C99 for function arguments.
1091 For function return values, C99's ``restrict`` is not meaningful,
1092 while LLVM's ``noalias`` is. Furthermore, the semantics of the ``noalias``
1093 attribute on return values are stronger than the semantics of the attribute
1094 when used on function arguments. On function return values, the ``noalias``
1095 attribute indicates that the function acts like a system memory allocation
1096 function, returning a pointer to allocated storage disjoint from the
1097 storage for any other object accessible to the caller.
1100 This indicates that the callee does not make any copies of the
1101 pointer that outlive the callee itself. This is not a valid
1102 attribute for return values. Addresses used in volatile operations
1103 are considered to be captured.
1108 This indicates that the pointer parameter can be excised using the
1109 :ref:`trampoline intrinsics <int_trampoline>`. This is not a valid
1110 attribute for return values and can only be applied to one parameter.
1113 This indicates that the function always returns the argument as its return
1114 value. This is a hint to the optimizer and code generator used when
1115 generating the caller, allowing value propagation, tail call optimization,
1116 and omission of register saves and restores in some cases; it is not
1117 checked or enforced when generating the callee. The parameter and the
1118 function return type must be valid operands for the
1119 :ref:`bitcast instruction <i_bitcast>`. This is not a valid attribute for
1120 return values and can only be applied to one parameter.
1123 This indicates that the parameter or return pointer is not null. This
1124 attribute may only be applied to pointer typed parameters. This is not
1125 checked or enforced by LLVM, the caller must ensure that the pointer
1126 passed in is non-null, or the callee must ensure that the returned pointer
1129 ``dereferenceable(<n>)``
1130 This indicates that the parameter or return pointer is dereferenceable. This
1131 attribute may only be applied to pointer typed parameters. A pointer that
1132 is dereferenceable can be loaded from speculatively without a risk of
1133 trapping. The number of bytes known to be dereferenceable must be provided
1134 in parentheses. It is legal for the number of bytes to be less than the
1135 size of the pointee type. The ``nonnull`` attribute does not imply
1136 dereferenceability (consider a pointer to one element past the end of an
1137 array), however ``dereferenceable(<n>)`` does imply ``nonnull`` in
1138 ``addrspace(0)`` (which is the default address space).
1140 ``dereferenceable_or_null(<n>)``
1141 This indicates that the parameter or return value isn't both
1142 non-null and non-dereferenceable (up to ``<n>`` bytes) at the same
1143 time. All non-null pointers tagged with
1144 ``dereferenceable_or_null(<n>)`` are ``dereferenceable(<n>)``.
1145 For address space 0 ``dereferenceable_or_null(<n>)`` implies that
1146 a pointer is exactly one of ``dereferenceable(<n>)`` or ``null``,
1147 and in other address spaces ``dereferenceable_or_null(<n>)``
1148 implies that a pointer is at least one of ``dereferenceable(<n>)``
1149 or ``null`` (i.e. it may be both ``null`` and
1150 ``dereferenceable(<n>)``). This attribute may only be applied to
1151 pointer typed parameters.
1154 This indicates that the parameter is the self/context parameter. This is not
1155 a valid attribute for return values and can only be applied to one
1159 This attribute is motivated to model and optimize Swift error handling. It
1160 can be applied to a parameter with pointer to pointer type or a
1161 pointer-sized alloca. At the call site, the actual argument that corresponds
1162 to a ``swifterror`` parameter has to come from a ``swifterror`` alloca or
1163 the ``swifterror`` parameter of the caller. A ``swifterror`` value (either
1164 the parameter or the alloca) can only be loaded and stored from, or used as
1165 a ``swifterror`` argument. This is not a valid attribute for return values
1166 and can only be applied to one parameter.
1168 These constraints allow the calling convention to optimize access to
1169 ``swifterror`` variables by associating them with a specific register at
1170 call boundaries rather than placing them in memory. Since this does change
1171 the calling convention, a function which uses the ``swifterror`` attribute
1172 on a parameter is not ABI-compatible with one which does not.
1174 These constraints also allow LLVM to assume that a ``swifterror`` argument
1175 does not alias any other memory visible within a function and that a
1176 ``swifterror`` alloca passed as an argument does not escape.
1180 Garbage Collector Strategy Names
1181 --------------------------------
1183 Each function may specify a garbage collector strategy name, which is simply a
1186 .. code-block:: llvm
1188 define void @f() gc "name" { ... }
1190 The supported values of *name* includes those :ref:`built in to LLVM
1191 <builtin-gc-strategies>` and any provided by loaded plugins. Specifying a GC
1192 strategy will cause the compiler to alter its output in order to support the
1193 named garbage collection algorithm. Note that LLVM itself does not contain a
1194 garbage collector, this functionality is restricted to generating machine code
1195 which can interoperate with a collector provided externally.
1202 Prefix data is data associated with a function which the code
1203 generator will emit immediately before the function's entrypoint.
1204 The purpose of this feature is to allow frontends to associate
1205 language-specific runtime metadata with specific functions and make it
1206 available through the function pointer while still allowing the
1207 function pointer to be called.
1209 To access the data for a given function, a program may bitcast the
1210 function pointer to a pointer to the constant's type and dereference
1211 index -1. This implies that the IR symbol points just past the end of
1212 the prefix data. For instance, take the example of a function annotated
1213 with a single ``i32``,
1215 .. code-block:: llvm
1217 define void @f() prefix i32 123 { ... }
1219 The prefix data can be referenced as,
1221 .. code-block:: llvm
1223 %0 = bitcast void* () @f to i32*
1224 %a = getelementptr inbounds i32, i32* %0, i32 -1
1225 %b = load i32, i32* %a
1227 Prefix data is laid out as if it were an initializer for a global variable
1228 of the prefix data's type. The function will be placed such that the
1229 beginning of the prefix data is aligned. This means that if the size
1230 of the prefix data is not a multiple of the alignment size, the
1231 function's entrypoint will not be aligned. If alignment of the
1232 function's entrypoint is desired, padding must be added to the prefix
1235 A function may have prefix data but no body. This has similar semantics
1236 to the ``available_externally`` linkage in that the data may be used by the
1237 optimizers but will not be emitted in the object file.
1244 The ``prologue`` attribute allows arbitrary code (encoded as bytes) to
1245 be inserted prior to the function body. This can be used for enabling
1246 function hot-patching and instrumentation.
1248 To maintain the semantics of ordinary function calls, the prologue data must
1249 have a particular format. Specifically, it must begin with a sequence of
1250 bytes which decode to a sequence of machine instructions, valid for the
1251 module's target, which transfer control to the point immediately succeeding
1252 the prologue data, without performing any other visible action. This allows
1253 the inliner and other passes to reason about the semantics of the function
1254 definition without needing to reason about the prologue data. Obviously this
1255 makes the format of the prologue data highly target dependent.
1257 A trivial example of valid prologue data for the x86 architecture is ``i8 144``,
1258 which encodes the ``nop`` instruction:
1260 .. code-block:: text
1262 define void @f() prologue i8 144 { ... }
1264 Generally prologue data can be formed by encoding a relative branch instruction
1265 which skips the metadata, as in this example of valid prologue data for the
1266 x86_64 architecture, where the first two bytes encode ``jmp .+10``:
1268 .. code-block:: text
1270 %0 = type <{ i8, i8, i8* }>
1272 define void @f() prologue %0 <{ i8 235, i8 8, i8* @md}> { ... }
1274 A function may have prologue data but no body. This has similar semantics
1275 to the ``available_externally`` linkage in that the data may be used by the
1276 optimizers but will not be emitted in the object file.
1280 Personality Function
1281 --------------------
1283 The ``personality`` attribute permits functions to specify what function
1284 to use for exception handling.
1291 Attribute groups are groups of attributes that are referenced by objects within
1292 the IR. They are important for keeping ``.ll`` files readable, because a lot of
1293 functions will use the same set of attributes. In the degenerative case of a
1294 ``.ll`` file that corresponds to a single ``.c`` file, the single attribute
1295 group will capture the important command line flags used to build that file.
1297 An attribute group is a module-level object. To use an attribute group, an
1298 object references the attribute group's ID (e.g. ``#37``). An object may refer
1299 to more than one attribute group. In that situation, the attributes from the
1300 different groups are merged.
1302 Here is an example of attribute groups for a function that should always be
1303 inlined, has a stack alignment of 4, and which shouldn't use SSE instructions:
1305 .. code-block:: llvm
1307 ; Target-independent attributes:
1308 attributes #0 = { alwaysinline alignstack=4 }
1310 ; Target-dependent attributes:
1311 attributes #1 = { "no-sse" }
1313 ; Function @f has attributes: alwaysinline, alignstack=4, and "no-sse".
1314 define void @f() #0 #1 { ... }
1321 Function attributes are set to communicate additional information about
1322 a function. Function attributes are considered to be part of the
1323 function, not of the function type, so functions with different function
1324 attributes can have the same function type.
1326 Function attributes are simple keywords that follow the type specified.
1327 If multiple attributes are needed, they are space separated. For
1330 .. code-block:: llvm
1332 define void @f() noinline { ... }
1333 define void @f() alwaysinline { ... }
1334 define void @f() alwaysinline optsize { ... }
1335 define void @f() optsize { ... }
1338 This attribute indicates that, when emitting the prologue and
1339 epilogue, the backend should forcibly align the stack pointer.
1340 Specify the desired alignment, which must be a power of two, in
1342 ``allocsize(<EltSizeParam>[, <NumEltsParam>])``
1343 This attribute indicates that the annotated function will always return at
1344 least a given number of bytes (or null). Its arguments are zero-indexed
1345 parameter numbers; if one argument is provided, then it's assumed that at
1346 least ``CallSite.Args[EltSizeParam]`` bytes will be available at the
1347 returned pointer. If two are provided, then it's assumed that
1348 ``CallSite.Args[EltSizeParam] * CallSite.Args[NumEltsParam]`` bytes are
1349 available. The referenced parameters must be integer types. No assumptions
1350 are made about the contents of the returned block of memory.
1352 This attribute indicates that the inliner should attempt to inline
1353 this function into callers whenever possible, ignoring any active
1354 inlining size threshold for this caller.
1356 This indicates that the callee function at a call site should be
1357 recognized as a built-in function, even though the function's declaration
1358 uses the ``nobuiltin`` attribute. This is only valid at call sites for
1359 direct calls to functions that are declared with the ``nobuiltin``
1362 This attribute indicates that this function is rarely called. When
1363 computing edge weights, basic blocks post-dominated by a cold
1364 function call are also considered to be cold; and, thus, given low
1367 In some parallel execution models, there exist operations that cannot be
1368 made control-dependent on any additional values. We call such operations
1369 ``convergent``, and mark them with this attribute.
1371 The ``convergent`` attribute may appear on functions or call/invoke
1372 instructions. When it appears on a function, it indicates that calls to
1373 this function should not be made control-dependent on additional values.
1374 For example, the intrinsic ``llvm.nvvm.barrier0`` is ``convergent``, so
1375 calls to this intrinsic cannot be made control-dependent on additional
1378 When it appears on a call/invoke, the ``convergent`` attribute indicates
1379 that we should treat the call as though we're calling a convergent
1380 function. This is particularly useful on indirect calls; without this we
1381 may treat such calls as though the target is non-convergent.
1383 The optimizer may remove the ``convergent`` attribute on functions when it
1384 can prove that the function does not execute any convergent operations.
1385 Similarly, the optimizer may remove ``convergent`` on calls/invokes when it
1386 can prove that the call/invoke cannot call a convergent function.
1387 ``inaccessiblememonly``
1388 This attribute indicates that the function may only access memory that
1389 is not accessible by the module being compiled. This is a weaker form
1391 ``inaccessiblemem_or_argmemonly``
1392 This attribute indicates that the function may only access memory that is
1393 either not accessible by the module being compiled, or is pointed to
1394 by its pointer arguments. This is a weaker form of ``argmemonly``
1396 This attribute indicates that the source code contained a hint that
1397 inlining this function is desirable (such as the "inline" keyword in
1398 C/C++). It is just a hint; it imposes no requirements on the
1401 This attribute indicates that the function should be added to a
1402 jump-instruction table at code-generation time, and that all address-taken
1403 references to this function should be replaced with a reference to the
1404 appropriate jump-instruction-table function pointer. Note that this creates
1405 a new pointer for the original function, which means that code that depends
1406 on function-pointer identity can break. So, any function annotated with
1407 ``jumptable`` must also be ``unnamed_addr``.
1409 This attribute suggests that optimization passes and code generator
1410 passes make choices that keep the code size of this function as small
1411 as possible and perform optimizations that may sacrifice runtime
1412 performance in order to minimize the size of the generated code.
1414 This attribute disables prologue / epilogue emission for the
1415 function. This can have very system-specific consequences.
1417 When this attribute is set to true, the jump tables and lookup tables that
1418 can be generated from a switch case lowering are disabled.
1420 This indicates that the callee function at a call site is not recognized as
1421 a built-in function. LLVM will retain the original call and not replace it
1422 with equivalent code based on the semantics of the built-in function, unless
1423 the call site uses the ``builtin`` attribute. This is valid at call sites
1424 and on function declarations and definitions.
1426 This attribute indicates that calls to the function cannot be
1427 duplicated. A call to a ``noduplicate`` function may be moved
1428 within its parent function, but may not be duplicated within
1429 its parent function.
1431 A function containing a ``noduplicate`` call may still
1432 be an inlining candidate, provided that the call is not
1433 duplicated by inlining. That implies that the function has
1434 internal linkage and only has one call site, so the original
1435 call is dead after inlining.
1437 This attributes disables implicit floating point instructions.
1439 This attribute indicates that the inliner should never inline this
1440 function in any situation. This attribute may not be used together
1441 with the ``alwaysinline`` attribute.
1443 This attribute suppresses lazy symbol binding for the function. This
1444 may make calls to the function faster, at the cost of extra program
1445 startup time if the function is not called during program startup.
1447 This attribute indicates that the code generator should not use a
1448 red zone, even if the target-specific ABI normally permits it.
1450 This function attribute indicates that the function never returns
1451 normally. This produces undefined behavior at runtime if the
1452 function ever does dynamically return.
1454 This function attribute indicates that the function does not call itself
1455 either directly or indirectly down any possible call path. This produces
1456 undefined behavior at runtime if the function ever does recurse.
1458 This function attribute indicates that the function never raises an
1459 exception. If the function does raise an exception, its runtime
1460 behavior is undefined. However, functions marked nounwind may still
1461 trap or generate asynchronous exceptions. Exception handling schemes
1462 that are recognized by LLVM to handle asynchronous exceptions, such
1463 as SEH, will still provide their implementation defined semantics.
1465 This function attribute indicates that most optimization passes will skip
1466 this function, with the exception of interprocedural optimization passes.
1467 Code generation defaults to the "fast" instruction selector.
1468 This attribute cannot be used together with the ``alwaysinline``
1469 attribute; this attribute is also incompatible
1470 with the ``minsize`` attribute and the ``optsize`` attribute.
1472 This attribute requires the ``noinline`` attribute to be specified on
1473 the function as well, so the function is never inlined into any caller.
1474 Only functions with the ``alwaysinline`` attribute are valid
1475 candidates for inlining into the body of this function.
1477 This attribute suggests that optimization passes and code generator
1478 passes make choices that keep the code size of this function low,
1479 and otherwise do optimizations specifically to reduce code size as
1480 long as they do not significantly impact runtime performance.
1481 ``"patchable-function"``
1482 This attribute tells the code generator that the code
1483 generated for this function needs to follow certain conventions that
1484 make it possible for a runtime function to patch over it later.
1485 The exact effect of this attribute depends on its string value,
1486 for which there currently is one legal possibility:
1488 * ``"prologue-short-redirect"`` - This style of patchable
1489 function is intended to support patching a function prologue to
1490 redirect control away from the function in a thread safe
1491 manner. It guarantees that the first instruction of the
1492 function will be large enough to accommodate a short jump
1493 instruction, and will be sufficiently aligned to allow being
1494 fully changed via an atomic compare-and-swap instruction.
1495 While the first requirement can be satisfied by inserting large
1496 enough NOP, LLVM can and will try to re-purpose an existing
1497 instruction (i.e. one that would have to be emitted anyway) as
1498 the patchable instruction larger than a short jump.
1500 ``"prologue-short-redirect"`` is currently only supported on
1503 This attribute by itself does not imply restrictions on
1504 inter-procedural optimizations. All of the semantic effects the
1505 patching may have to be separately conveyed via the linkage type.
1507 This attribute indicates that the function will trigger a guard region
1508 in the end of the stack. It ensures that accesses to the stack must be
1509 no further apart than the size of the guard region to a previous
1510 access of the stack. It takes one required string value, the name of
1511 the stack probing function that will be called.
1513 If a function that has a ``"probe-stack"`` attribute is inlined into
1514 a function with another ``"probe-stack"`` attribute, the resulting
1515 function has the ``"probe-stack"`` attribute of the caller. If a
1516 function that has a ``"probe-stack"`` attribute is inlined into a
1517 function that has no ``"probe-stack"`` attribute at all, the resulting
1518 function has the ``"probe-stack"`` attribute of the callee.
1520 On a function, this attribute indicates that the function computes its
1521 result (or decides to unwind an exception) based strictly on its arguments,
1522 without dereferencing any pointer arguments or otherwise accessing
1523 any mutable state (e.g. memory, control registers, etc) visible to
1524 caller functions. It does not write through any pointer arguments
1525 (including ``byval`` arguments) and never changes any state visible
1526 to callers. This means while it cannot unwind exceptions by calling
1527 the ``C++`` exception throwing methods (since they write to memory), there may
1528 be non-``C++`` mechanisms that throw exceptions without writing to LLVM
1531 On an argument, this attribute indicates that the function does not
1532 dereference that pointer argument, even though it may read or write the
1533 memory that the pointer points to if accessed through other pointers.
1535 On a function, this attribute indicates that the function does not write
1536 through any pointer arguments (including ``byval`` arguments) or otherwise
1537 modify any state (e.g. memory, control registers, etc) visible to
1538 caller functions. It may dereference pointer arguments and read
1539 state that may be set in the caller. A readonly function always
1540 returns the same value (or unwinds an exception identically) when
1541 called with the same set of arguments and global state. This means while it
1542 cannot unwind exceptions by calling the ``C++`` exception throwing methods
1543 (since they write to memory), there may be non-``C++`` mechanisms that throw
1544 exceptions without writing to LLVM visible memory.
1546 On an argument, this attribute indicates that the function does not write
1547 through this pointer argument, even though it may write to the memory that
1548 the pointer points to.
1549 ``"stack-probe-size"``
1550 This attribute controls the behavior of stack probes: either
1551 the ``"probe-stack"`` attribute, or ABI-required stack probes, if any.
1552 It defines the size of the guard region. It ensures that if the function
1553 may use more stack space than the size of the guard region, stack probing
1554 sequence will be emitted. It takes one required integer value, which
1557 If a function that has a ``"stack-probe-size"`` attribute is inlined into
1558 a function with another ``"stack-probe-size"`` attribute, the resulting
1559 function has the ``"stack-probe-size"`` attribute that has the lower
1560 numeric value. If a function that has a ``"stack-probe-size"`` attribute is
1561 inlined into a function that has no ``"stack-probe-size"`` attribute
1562 at all, the resulting function has the ``"stack-probe-size"`` attribute
1564 ``"no-stack-arg-probe"``
1565 This attribute disables ABI-required stack probes, if any.
1567 On a function, this attribute indicates that the function may write to but
1568 does not read from memory.
1570 On an argument, this attribute indicates that the function may write to but
1571 does not read through this pointer argument (even though it may read from
1572 the memory that the pointer points to).
1574 This attribute indicates that the only memory accesses inside function are
1575 loads and stores from objects pointed to by its pointer-typed arguments,
1576 with arbitrary offsets. Or in other words, all memory operations in the
1577 function can refer to memory only using pointers based on its function
1579 Note that ``argmemonly`` can be used together with ``readonly`` attribute
1580 in order to specify that function reads only from its arguments.
1582 This attribute indicates that this function can return twice. The C
1583 ``setjmp`` is an example of such a function. The compiler disables
1584 some optimizations (like tail calls) in the caller of these
1587 This attribute indicates that
1588 `SafeStack <http://clang.llvm.org/docs/SafeStack.html>`_
1589 protection is enabled for this function.
1591 If a function that has a ``safestack`` attribute is inlined into a
1592 function that doesn't have a ``safestack`` attribute or which has an
1593 ``ssp``, ``sspstrong`` or ``sspreq`` attribute, then the resulting
1594 function will have a ``safestack`` attribute.
1595 ``sanitize_address``
1596 This attribute indicates that AddressSanitizer checks
1597 (dynamic address safety analysis) are enabled for this function.
1599 This attribute indicates that MemorySanitizer checks (dynamic detection
1600 of accesses to uninitialized memory) are enabled for this function.
1602 This attribute indicates that ThreadSanitizer checks
1603 (dynamic thread safety analysis) are enabled for this function.
1604 ``sanitize_hwaddress``
1605 This attribute indicates that HWAddressSanitizer checks
1606 (dynamic address safety analysis based on tagged pointers) are enabled for
1609 This function attribute indicates that the function does not have any
1610 effects besides calculating its result and does not have undefined behavior.
1611 Note that ``speculatable`` is not enough to conclude that along any
1612 particular execution path the number of calls to this function will not be
1613 externally observable. This attribute is only valid on functions
1614 and declarations, not on individual call sites. If a function is
1615 incorrectly marked as speculatable and really does exhibit
1616 undefined behavior, the undefined behavior may be observed even
1617 if the call site is dead code.
1620 This attribute indicates that the function should emit a stack
1621 smashing protector. It is in the form of a "canary" --- a random value
1622 placed on the stack before the local variables that's checked upon
1623 return from the function to see if it has been overwritten. A
1624 heuristic is used to determine if a function needs stack protectors
1625 or not. The heuristic used will enable protectors for functions with:
1627 - Character arrays larger than ``ssp-buffer-size`` (default 8).
1628 - Aggregates containing character arrays larger than ``ssp-buffer-size``.
1629 - Calls to alloca() with variable sizes or constant sizes greater than
1630 ``ssp-buffer-size``.
1632 Variables that are identified as requiring a protector will be arranged
1633 on the stack such that they are adjacent to the stack protector guard.
1635 If a function that has an ``ssp`` attribute is inlined into a
1636 function that doesn't have an ``ssp`` attribute, then the resulting
1637 function will have an ``ssp`` attribute.
1639 This attribute indicates that the function should *always* emit a
1640 stack smashing protector. This overrides the ``ssp`` function
1643 Variables that are identified as requiring a protector will be arranged
1644 on the stack such that they are adjacent to the stack protector guard.
1645 The specific layout rules are:
1647 #. Large arrays and structures containing large arrays
1648 (``>= ssp-buffer-size``) are closest to the stack protector.
1649 #. Small arrays and structures containing small arrays
1650 (``< ssp-buffer-size``) are 2nd closest to the protector.
1651 #. Variables that have had their address taken are 3rd closest to the
1654 If a function that has an ``sspreq`` attribute is inlined into a
1655 function that doesn't have an ``sspreq`` attribute or which has an
1656 ``ssp`` or ``sspstrong`` attribute, then the resulting function will have
1657 an ``sspreq`` attribute.
1659 This attribute indicates that the function should emit a stack smashing
1660 protector. This attribute causes a strong heuristic to be used when
1661 determining if a function needs stack protectors. The strong heuristic
1662 will enable protectors for functions with:
1664 - Arrays of any size and type
1665 - Aggregates containing an array of any size and type.
1666 - Calls to alloca().
1667 - Local variables that have had their address taken.
1669 Variables that are identified as requiring a protector will be arranged
1670 on the stack such that they are adjacent to the stack protector guard.
1671 The specific layout rules are:
1673 #. Large arrays and structures containing large arrays
1674 (``>= ssp-buffer-size``) are closest to the stack protector.
1675 #. Small arrays and structures containing small arrays
1676 (``< ssp-buffer-size``) are 2nd closest to the protector.
1677 #. Variables that have had their address taken are 3rd closest to the
1680 This overrides the ``ssp`` function attribute.
1682 If a function that has an ``sspstrong`` attribute is inlined into a
1683 function that doesn't have an ``sspstrong`` attribute, then the
1684 resulting function will have an ``sspstrong`` attribute.
1686 This attribute indicates that the function was called from a scope that
1687 requires strict floating point semantics. LLVM will not attempt any
1688 optimizations that require assumptions about the floating point rounding
1689 mode or that might alter the state of floating point status flags that
1690 might otherwise be set or cleared by calling this function.
1692 This attribute indicates that the function will delegate to some other
1693 function with a tail call. The prototype of a thunk should not be used for
1694 optimization purposes. The caller is expected to cast the thunk prototype to
1695 match the thunk target prototype.
1697 This attribute indicates that the ABI being targeted requires that
1698 an unwind table entry be produced for this function even if we can
1699 show that no exceptions passes by it. This is normally the case for
1700 the ELF x86-64 abi, but it can be disabled for some compilation
1703 This attribute indicates that no control-flow check will be perfomed on
1704 the attributed entity. It disables -fcf-protection=<> for a specific
1705 entity to fine grain the HW control flow protection mechanism. The flag
1706 is target independant and currently appertains to a function or function
1714 Attributes may be set to communicate additional information about a global variable.
1715 Unlike :ref:`function attributes <fnattrs>`, attributes on a global variable
1716 are grouped into a single :ref:`attribute group <attrgrp>`.
1723 Operand bundles are tagged sets of SSA values that can be associated
1724 with certain LLVM instructions (currently only ``call`` s and
1725 ``invoke`` s). In a way they are like metadata, but dropping them is
1726 incorrect and will change program semantics.
1730 operand bundle set ::= '[' operand bundle (, operand bundle )* ']'
1731 operand bundle ::= tag '(' [ bundle operand ] (, bundle operand )* ')'
1732 bundle operand ::= SSA value
1733 tag ::= string constant
1735 Operand bundles are **not** part of a function's signature, and a
1736 given function may be called from multiple places with different kinds
1737 of operand bundles. This reflects the fact that the operand bundles
1738 are conceptually a part of the ``call`` (or ``invoke``), not the
1739 callee being dispatched to.
1741 Operand bundles are a generic mechanism intended to support
1742 runtime-introspection-like functionality for managed languages. While
1743 the exact semantics of an operand bundle depend on the bundle tag,
1744 there are certain limitations to how much the presence of an operand
1745 bundle can influence the semantics of a program. These restrictions
1746 are described as the semantics of an "unknown" operand bundle. As
1747 long as the behavior of an operand bundle is describable within these
1748 restrictions, LLVM does not need to have special knowledge of the
1749 operand bundle to not miscompile programs containing it.
1751 - The bundle operands for an unknown operand bundle escape in unknown
1752 ways before control is transferred to the callee or invokee.
1753 - Calls and invokes with operand bundles have unknown read / write
1754 effect on the heap on entry and exit (even if the call target is
1755 ``readnone`` or ``readonly``), unless they're overridden with
1756 callsite specific attributes.
1757 - An operand bundle at a call site cannot change the implementation
1758 of the called function. Inter-procedural optimizations work as
1759 usual as long as they take into account the first two properties.
1761 More specific types of operand bundles are described below.
1763 .. _deopt_opbundles:
1765 Deoptimization Operand Bundles
1766 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1768 Deoptimization operand bundles are characterized by the ``"deopt"``
1769 operand bundle tag. These operand bundles represent an alternate
1770 "safe" continuation for the call site they're attached to, and can be
1771 used by a suitable runtime to deoptimize the compiled frame at the
1772 specified call site. There can be at most one ``"deopt"`` operand
1773 bundle attached to a call site. Exact details of deoptimization is
1774 out of scope for the language reference, but it usually involves
1775 rewriting a compiled frame into a set of interpreted frames.
1777 From the compiler's perspective, deoptimization operand bundles make
1778 the call sites they're attached to at least ``readonly``. They read
1779 through all of their pointer typed operands (even if they're not
1780 otherwise escaped) and the entire visible heap. Deoptimization
1781 operand bundles do not capture their operands except during
1782 deoptimization, in which case control will not be returned to the
1785 The inliner knows how to inline through calls that have deoptimization
1786 operand bundles. Just like inlining through a normal call site
1787 involves composing the normal and exceptional continuations, inlining
1788 through a call site with a deoptimization operand bundle needs to
1789 appropriately compose the "safe" deoptimization continuation. The
1790 inliner does this by prepending the parent's deoptimization
1791 continuation to every deoptimization continuation in the inlined body.
1792 E.g. inlining ``@f`` into ``@g`` in the following example
1794 .. code-block:: llvm
1797 call void @x() ;; no deopt state
1798 call void @y() [ "deopt"(i32 10) ]
1799 call void @y() [ "deopt"(i32 10), "unknown"(i8* null) ]
1804 call void @f() [ "deopt"(i32 20) ]
1810 .. code-block:: llvm
1813 call void @x() ;; still no deopt state
1814 call void @y() [ "deopt"(i32 20, i32 10) ]
1815 call void @y() [ "deopt"(i32 20, i32 10), "unknown"(i8* null) ]
1819 It is the frontend's responsibility to structure or encode the
1820 deoptimization state in a way that syntactically prepending the
1821 caller's deoptimization state to the callee's deoptimization state is
1822 semantically equivalent to composing the caller's deoptimization
1823 continuation after the callee's deoptimization continuation.
1827 Funclet Operand Bundles
1828 ^^^^^^^^^^^^^^^^^^^^^^^
1830 Funclet operand bundles are characterized by the ``"funclet"``
1831 operand bundle tag. These operand bundles indicate that a call site
1832 is within a particular funclet. There can be at most one
1833 ``"funclet"`` operand bundle attached to a call site and it must have
1834 exactly one bundle operand.
1836 If any funclet EH pads have been "entered" but not "exited" (per the
1837 `description in the EH doc\ <ExceptionHandling.html#wineh-constraints>`_),
1838 it is undefined behavior to execute a ``call`` or ``invoke`` which:
1840 * does not have a ``"funclet"`` bundle and is not a ``call`` to a nounwind
1842 * has a ``"funclet"`` bundle whose operand is not the most-recently-entered
1843 not-yet-exited funclet EH pad.
1845 Similarly, if no funclet EH pads have been entered-but-not-yet-exited,
1846 executing a ``call`` or ``invoke`` with a ``"funclet"`` bundle is undefined behavior.
1848 GC Transition Operand Bundles
1849 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1851 GC transition operand bundles are characterized by the
1852 ``"gc-transition"`` operand bundle tag. These operand bundles mark a
1853 call as a transition between a function with one GC strategy to a
1854 function with a different GC strategy. If coordinating the transition
1855 between GC strategies requires additional code generation at the call
1856 site, these bundles may contain any values that are needed by the
1857 generated code. For more details, see :ref:`GC Transitions
1858 <gc_transition_args>`.
1862 Module-Level Inline Assembly
1863 ----------------------------
1865 Modules may contain "module-level inline asm" blocks, which corresponds
1866 to the GCC "file scope inline asm" blocks. These blocks are internally
1867 concatenated by LLVM and treated as a single unit, but may be separated
1868 in the ``.ll`` file if desired. The syntax is very simple:
1870 .. code-block:: llvm
1872 module asm "inline asm code goes here"
1873 module asm "more can go here"
1875 The strings can contain any character by escaping non-printable
1876 characters. The escape sequence used is simply "\\xx" where "xx" is the
1877 two digit hex code for the number.
1879 Note that the assembly string *must* be parseable by LLVM's integrated assembler
1880 (unless it is disabled), even when emitting a ``.s`` file.
1882 .. _langref_datalayout:
1887 A module may specify a target specific data layout string that specifies
1888 how data is to be laid out in memory. The syntax for the data layout is
1891 .. code-block:: llvm
1893 target datalayout = "layout specification"
1895 The *layout specification* consists of a list of specifications
1896 separated by the minus sign character ('-'). Each specification starts
1897 with a letter and may include other information after the letter to
1898 define some aspect of the data layout. The specifications accepted are
1902 Specifies that the target lays out data in big-endian form. That is,
1903 the bits with the most significance have the lowest address
1906 Specifies that the target lays out data in little-endian form. That
1907 is, the bits with the least significance have the lowest address
1910 Specifies the natural alignment of the stack in bits. Alignment
1911 promotion of stack variables is limited to the natural stack
1912 alignment to avoid dynamic stack realignment. The stack alignment
1913 must be a multiple of 8-bits. If omitted, the natural stack
1914 alignment defaults to "unspecified", which does not prevent any
1915 alignment promotions.
1916 ``P<address space>``
1917 Specifies the address space that corresponds to program memory.
1918 Harvard architectures can use this to specify what space LLVM
1919 should place things such as functions into. If omitted, the
1920 program memory space defaults to the default address space of 0,
1921 which corresponds to a Von Neumann architecture that has code
1922 and data in the same space.
1923 ``A<address space>``
1924 Specifies the address space of objects created by '``alloca``'.
1925 Defaults to the default address space of 0.
1926 ``p[n]:<size>:<abi>:<pref>:<idx>``
1927 This specifies the *size* of a pointer and its ``<abi>`` and
1928 ``<pref>``\erred alignments for address space ``n``. The fourth parameter
1929 ``<idx>`` is a size of index that used for address calculation. If not
1930 specified, the default index size is equal to the pointer size. All sizes
1931 are in bits. The address space, ``n``, is optional, and if not specified,
1932 denotes the default address space 0. The value of ``n`` must be
1933 in the range [1,2^23).
1934 ``i<size>:<abi>:<pref>``
1935 This specifies the alignment for an integer type of a given bit
1936 ``<size>``. The value of ``<size>`` must be in the range [1,2^23).
1937 ``v<size>:<abi>:<pref>``
1938 This specifies the alignment for a vector type of a given bit
1940 ``f<size>:<abi>:<pref>``
1941 This specifies the alignment for a floating point type of a given bit
1942 ``<size>``. Only values of ``<size>`` that are supported by the target
1943 will work. 32 (float) and 64 (double) are supported on all targets; 80
1944 or 128 (different flavors of long double) are also supported on some
1947 This specifies the alignment for an object of aggregate type.
1949 If present, specifies that llvm names are mangled in the output. Symbols
1950 prefixed with the mangling escape character ``\01`` are passed through
1951 directly to the assembler without the escape character. The mangling style
1954 * ``e``: ELF mangling: Private symbols get a ``.L`` prefix.
1955 * ``m``: Mips mangling: Private symbols get a ``$`` prefix.
1956 * ``o``: Mach-O mangling: Private symbols get ``L`` prefix. Other
1957 symbols get a ``_`` prefix.
1958 * ``x``: Windows x86 COFF mangling: Private symbols get the usual prefix.
1959 Regular C symbols get a ``_`` prefix. Functions with ``__stdcall``,
1960 ``__fastcall``, and ``__vectorcall`` have custom mangling that appends
1961 ``@N`` where N is the number of bytes used to pass parameters. C++ symbols
1962 starting with ``?`` are not mangled in any way.
1963 * ``w``: Windows COFF mangling: Similar to ``x``, except that normal C
1964 symbols do not receive a ``_`` prefix.
1965 ``n<size1>:<size2>:<size3>...``
1966 This specifies a set of native integer widths for the target CPU in
1967 bits. For example, it might contain ``n32`` for 32-bit PowerPC,
1968 ``n32:64`` for PowerPC 64, or ``n8:16:32:64`` for X86-64. Elements of
1969 this set are considered to support most general arithmetic operations
1971 ``ni:<address space0>:<address space1>:<address space2>...``
1972 This specifies pointer types with the specified address spaces
1973 as :ref:`Non-Integral Pointer Type <nointptrtype>` s. The ``0``
1974 address space cannot be specified as non-integral.
1976 On every specification that takes a ``<abi>:<pref>``, specifying the
1977 ``<pref>`` alignment is optional. If omitted, the preceding ``:``
1978 should be omitted too and ``<pref>`` will be equal to ``<abi>``.
1980 When constructing the data layout for a given target, LLVM starts with a
1981 default set of specifications which are then (possibly) overridden by
1982 the specifications in the ``datalayout`` keyword. The default
1983 specifications are given in this list:
1985 - ``E`` - big endian
1986 - ``p:64:64:64`` - 64-bit pointers with 64-bit alignment.
1987 - ``p[n]:64:64:64`` - Other address spaces are assumed to be the
1988 same as the default address space.
1989 - ``S0`` - natural stack alignment is unspecified
1990 - ``i1:8:8`` - i1 is 8-bit (byte) aligned
1991 - ``i8:8:8`` - i8 is 8-bit (byte) aligned
1992 - ``i16:16:16`` - i16 is 16-bit aligned
1993 - ``i32:32:32`` - i32 is 32-bit aligned
1994 - ``i64:32:64`` - i64 has ABI alignment of 32-bits but preferred
1995 alignment of 64-bits
1996 - ``f16:16:16`` - half is 16-bit aligned
1997 - ``f32:32:32`` - float is 32-bit aligned
1998 - ``f64:64:64`` - double is 64-bit aligned
1999 - ``f128:128:128`` - quad is 128-bit aligned
2000 - ``v64:64:64`` - 64-bit vector is 64-bit aligned
2001 - ``v128:128:128`` - 128-bit vector is 128-bit aligned
2002 - ``a:0:64`` - aggregates are 64-bit aligned
2004 When LLVM is determining the alignment for a given type, it uses the
2007 #. If the type sought is an exact match for one of the specifications,
2008 that specification is used.
2009 #. If no match is found, and the type sought is an integer type, then
2010 the smallest integer type that is larger than the bitwidth of the
2011 sought type is used. If none of the specifications are larger than
2012 the bitwidth then the largest integer type is used. For example,
2013 given the default specifications above, the i7 type will use the
2014 alignment of i8 (next largest) while both i65 and i256 will use the
2015 alignment of i64 (largest specified).
2016 #. If no match is found, and the type sought is a vector type, then the
2017 largest vector type that is smaller than the sought vector type will
2018 be used as a fall back. This happens because <128 x double> can be
2019 implemented in terms of 64 <2 x double>, for example.
2021 The function of the data layout string may not be what you expect.
2022 Notably, this is not a specification from the frontend of what alignment
2023 the code generator should use.
2025 Instead, if specified, the target data layout is required to match what
2026 the ultimate *code generator* expects. This string is used by the
2027 mid-level optimizers to improve code, and this only works if it matches
2028 what the ultimate code generator uses. There is no way to generate IR
2029 that does not embed this target-specific detail into the IR. If you
2030 don't specify the string, the default specifications will be used to
2031 generate a Data Layout and the optimization phases will operate
2032 accordingly and introduce target specificity into the IR with respect to
2033 these default specifications.
2040 A module may specify a target triple string that describes the target
2041 host. The syntax for the target triple is simply:
2043 .. code-block:: llvm
2045 target triple = "x86_64-apple-macosx10.7.0"
2047 The *target triple* string consists of a series of identifiers delimited
2048 by the minus sign character ('-'). The canonical forms are:
2052 ARCHITECTURE-VENDOR-OPERATING_SYSTEM
2053 ARCHITECTURE-VENDOR-OPERATING_SYSTEM-ENVIRONMENT
2055 This information is passed along to the backend so that it generates
2056 code for the proper architecture. It's possible to override this on the
2057 command line with the ``-mtriple`` command line option.
2059 .. _pointeraliasing:
2061 Pointer Aliasing Rules
2062 ----------------------
2064 Any memory access must be done through a pointer value associated with
2065 an address range of the memory access, otherwise the behavior is
2066 undefined. Pointer values are associated with address ranges according
2067 to the following rules:
2069 - A pointer value is associated with the addresses associated with any
2070 value it is *based* on.
2071 - An address of a global variable is associated with the address range
2072 of the variable's storage.
2073 - The result value of an allocation instruction is associated with the
2074 address range of the allocated storage.
2075 - A null pointer in the default address-space is associated with no
2077 - An integer constant other than zero or a pointer value returned from
2078 a function not defined within LLVM may be associated with address
2079 ranges allocated through mechanisms other than those provided by
2080 LLVM. Such ranges shall not overlap with any ranges of addresses
2081 allocated by mechanisms provided by LLVM.
2083 A pointer value is *based* on another pointer value according to the
2086 - A pointer value formed from a scalar ``getelementptr`` operation is *based* on
2087 the pointer-typed operand of the ``getelementptr``.
2088 - The pointer in lane *l* of the result of a vector ``getelementptr`` operation
2089 is *based* on the pointer in lane *l* of the vector-of-pointers-typed operand
2090 of the ``getelementptr``.
2091 - The result value of a ``bitcast`` is *based* on the operand of the
2093 - A pointer value formed by an ``inttoptr`` is *based* on all pointer
2094 values that contribute (directly or indirectly) to the computation of
2095 the pointer's value.
2096 - The "*based* on" relationship is transitive.
2098 Note that this definition of *"based"* is intentionally similar to the
2099 definition of *"based"* in C99, though it is slightly weaker.
2101 LLVM IR does not associate types with memory. The result type of a
2102 ``load`` merely indicates the size and alignment of the memory from
2103 which to load, as well as the interpretation of the value. The first
2104 operand type of a ``store`` similarly only indicates the size and
2105 alignment of the store.
2107 Consequently, type-based alias analysis, aka TBAA, aka
2108 ``-fstrict-aliasing``, is not applicable to general unadorned LLVM IR.
2109 :ref:`Metadata <metadata>` may be used to encode additional information
2110 which specialized optimization passes may use to implement type-based
2115 Volatile Memory Accesses
2116 ------------------------
2118 Certain memory accesses, such as :ref:`load <i_load>`'s,
2119 :ref:`store <i_store>`'s, and :ref:`llvm.memcpy <int_memcpy>`'s may be
2120 marked ``volatile``. The optimizers must not change the number of
2121 volatile operations or change their order of execution relative to other
2122 volatile operations. The optimizers *may* change the order of volatile
2123 operations relative to non-volatile operations. This is not Java's
2124 "volatile" and has no cross-thread synchronization behavior.
2126 IR-level volatile loads and stores cannot safely be optimized into
2127 llvm.memcpy or llvm.memmove intrinsics even when those intrinsics are
2128 flagged volatile. Likewise, the backend should never split or merge
2129 target-legal volatile load/store instructions.
2131 .. admonition:: Rationale
2133 Platforms may rely on volatile loads and stores of natively supported
2134 data width to be executed as single instruction. For example, in C
2135 this holds for an l-value of volatile primitive type with native
2136 hardware support, but not necessarily for aggregate types. The
2137 frontend upholds these expectations, which are intentionally
2138 unspecified in the IR. The rules above ensure that IR transformations
2139 do not violate the frontend's contract with the language.
2143 Memory Model for Concurrent Operations
2144 --------------------------------------
2146 The LLVM IR does not define any way to start parallel threads of
2147 execution or to register signal handlers. Nonetheless, there are
2148 platform-specific ways to create them, and we define LLVM IR's behavior
2149 in their presence. This model is inspired by the C++0x memory model.
2151 For a more informal introduction to this model, see the :doc:`Atomics`.
2153 We define a *happens-before* partial order as the least partial order
2156 - Is a superset of single-thread program order, and
2157 - When a *synchronizes-with* ``b``, includes an edge from ``a`` to
2158 ``b``. *Synchronizes-with* pairs are introduced by platform-specific
2159 techniques, like pthread locks, thread creation, thread joining,
2160 etc., and by atomic instructions. (See also :ref:`Atomic Memory Ordering
2161 Constraints <ordering>`).
2163 Note that program order does not introduce *happens-before* edges
2164 between a thread and signals executing inside that thread.
2166 Every (defined) read operation (load instructions, memcpy, atomic
2167 loads/read-modify-writes, etc.) R reads a series of bytes written by
2168 (defined) write operations (store instructions, atomic
2169 stores/read-modify-writes, memcpy, etc.). For the purposes of this
2170 section, initialized globals are considered to have a write of the
2171 initializer which is atomic and happens before any other read or write
2172 of the memory in question. For each byte of a read R, R\ :sub:`byte`
2173 may see any write to the same byte, except:
2175 - If write\ :sub:`1` happens before write\ :sub:`2`, and
2176 write\ :sub:`2` happens before R\ :sub:`byte`, then
2177 R\ :sub:`byte` does not see write\ :sub:`1`.
2178 - If R\ :sub:`byte` happens before write\ :sub:`3`, then
2179 R\ :sub:`byte` does not see write\ :sub:`3`.
2181 Given that definition, R\ :sub:`byte` is defined as follows:
2183 - If R is volatile, the result is target-dependent. (Volatile is
2184 supposed to give guarantees which can support ``sig_atomic_t`` in
2185 C/C++, and may be used for accesses to addresses that do not behave
2186 like normal memory. It does not generally provide cross-thread
2188 - Otherwise, if there is no write to the same byte that happens before
2189 R\ :sub:`byte`, R\ :sub:`byte` returns ``undef`` for that byte.
2190 - Otherwise, if R\ :sub:`byte` may see exactly one write,
2191 R\ :sub:`byte` returns the value written by that write.
2192 - Otherwise, if R is atomic, and all the writes R\ :sub:`byte` may
2193 see are atomic, it chooses one of the values written. See the :ref:`Atomic
2194 Memory Ordering Constraints <ordering>` section for additional
2195 constraints on how the choice is made.
2196 - Otherwise R\ :sub:`byte` returns ``undef``.
2198 R returns the value composed of the series of bytes it read. This
2199 implies that some bytes within the value may be ``undef`` **without**
2200 the entire value being ``undef``. Note that this only defines the
2201 semantics of the operation; it doesn't mean that targets will emit more
2202 than one instruction to read the series of bytes.
2204 Note that in cases where none of the atomic intrinsics are used, this
2205 model places only one restriction on IR transformations on top of what
2206 is required for single-threaded execution: introducing a store to a byte
2207 which might not otherwise be stored is not allowed in general.
2208 (Specifically, in the case where another thread might write to and read
2209 from an address, introducing a store can change a load that may see
2210 exactly one write into a load that may see multiple writes.)
2214 Atomic Memory Ordering Constraints
2215 ----------------------------------
2217 Atomic instructions (:ref:`cmpxchg <i_cmpxchg>`,
2218 :ref:`atomicrmw <i_atomicrmw>`, :ref:`fence <i_fence>`,
2219 :ref:`atomic load <i_load>`, and :ref:`atomic store <i_store>`) take
2220 ordering parameters that determine which other atomic instructions on
2221 the same address they *synchronize with*. These semantics are borrowed
2222 from Java and C++0x, but are somewhat more colloquial. If these
2223 descriptions aren't precise enough, check those specs (see spec
2224 references in the :doc:`atomics guide <Atomics>`).
2225 :ref:`fence <i_fence>` instructions treat these orderings somewhat
2226 differently since they don't take an address. See that instruction's
2227 documentation for details.
2229 For a simpler introduction to the ordering constraints, see the
2233 The set of values that can be read is governed by the happens-before
2234 partial order. A value cannot be read unless some operation wrote
2235 it. This is intended to provide a guarantee strong enough to model
2236 Java's non-volatile shared variables. This ordering cannot be
2237 specified for read-modify-write operations; it is not strong enough
2238 to make them atomic in any interesting way.
2240 In addition to the guarantees of ``unordered``, there is a single
2241 total order for modifications by ``monotonic`` operations on each
2242 address. All modification orders must be compatible with the
2243 happens-before order. There is no guarantee that the modification
2244 orders can be combined to a global total order for the whole program
2245 (and this often will not be possible). The read in an atomic
2246 read-modify-write operation (:ref:`cmpxchg <i_cmpxchg>` and
2247 :ref:`atomicrmw <i_atomicrmw>`) reads the value in the modification
2248 order immediately before the value it writes. If one atomic read
2249 happens before another atomic read of the same address, the later
2250 read must see the same value or a later value in the address's
2251 modification order. This disallows reordering of ``monotonic`` (or
2252 stronger) operations on the same address. If an address is written
2253 ``monotonic``-ally by one thread, and other threads ``monotonic``-ally
2254 read that address repeatedly, the other threads must eventually see
2255 the write. This corresponds to the C++0x/C1x
2256 ``memory_order_relaxed``.
2258 In addition to the guarantees of ``monotonic``, a
2259 *synchronizes-with* edge may be formed with a ``release`` operation.
2260 This is intended to model C++'s ``memory_order_acquire``.
2262 In addition to the guarantees of ``monotonic``, if this operation
2263 writes a value which is subsequently read by an ``acquire``
2264 operation, it *synchronizes-with* that operation. (This isn't a
2265 complete description; see the C++0x definition of a release
2266 sequence.) This corresponds to the C++0x/C1x
2267 ``memory_order_release``.
2268 ``acq_rel`` (acquire+release)
2269 Acts as both an ``acquire`` and ``release`` operation on its
2270 address. This corresponds to the C++0x/C1x ``memory_order_acq_rel``.
2271 ``seq_cst`` (sequentially consistent)
2272 In addition to the guarantees of ``acq_rel`` (``acquire`` for an
2273 operation that only reads, ``release`` for an operation that only
2274 writes), there is a global total order on all
2275 sequentially-consistent operations on all addresses, which is
2276 consistent with the *happens-before* partial order and with the
2277 modification orders of all the affected addresses. Each
2278 sequentially-consistent read sees the last preceding write to the
2279 same address in this global order. This corresponds to the C++0x/C1x
2280 ``memory_order_seq_cst`` and Java volatile.
2284 If an atomic operation is marked ``syncscope("singlethread")``, it only
2285 *synchronizes with* and only participates in the seq\_cst total orderings of
2286 other operations running in the same thread (for example, in signal handlers).
2288 If an atomic operation is marked ``syncscope("<target-scope>")``, where
2289 ``<target-scope>`` is a target specific synchronization scope, then it is target
2290 dependent if it *synchronizes with* and participates in the seq\_cst total
2291 orderings of other operations.
2293 Otherwise, an atomic operation that is not marked ``syncscope("singlethread")``
2294 or ``syncscope("<target-scope>")`` *synchronizes with* and participates in the
2295 seq\_cst total orderings of other operations that are not marked
2296 ``syncscope("singlethread")`` or ``syncscope("<target-scope>")``.
2300 Floating-Point Environment
2301 --------------------------
2303 The default LLVM floating-point environment assumes that floating-point
2304 instructions do not have side effects. Results assume the round-to-nearest
2305 rounding mode. No floating-point exception state is maintained in this
2306 environment. Therefore, there is no attempt to create or preserve invalid
2307 operation (SNaN) or division-by-zero exceptions in these examples:
2309 .. code-block:: llvm
2311 %A = fdiv 0x7ff0000000000001, %X ; 64-bit SNaN hex value
2317 The benefit of this exception-free assumption is that floating-point
2318 operations may be speculated freely without any other fast-math relaxations
2319 to the floating-point model.
2321 Code that requires different behavior than this should use the
2322 :ref:`Constrained Floating-Point Intrinsics <constrainedfp>`.
2329 LLVM IR floating-point operations (:ref:`fadd <i_fadd>`,
2330 :ref:`fsub <i_fsub>`, :ref:`fmul <i_fmul>`, :ref:`fdiv <i_fdiv>`,
2331 :ref:`frem <i_frem>`, :ref:`fcmp <i_fcmp>`) and :ref:`call <i_call>`
2332 may use the following flags to enable otherwise unsafe
2333 floating-point transformations.
2336 No NaNs - Allow optimizations to assume the arguments and result are not
2337 NaN. Such optimizations are required to retain defined behavior over
2338 NaNs, but the value of the result is undefined.
2341 No Infs - Allow optimizations to assume the arguments and result are not
2342 +/-Inf. Such optimizations are required to retain defined behavior over
2343 +/-Inf, but the value of the result is undefined.
2346 No Signed Zeros - Allow optimizations to treat the sign of a zero
2347 argument or result as insignificant.
2350 Allow Reciprocal - Allow optimizations to use the reciprocal of an
2351 argument rather than perform division.
2354 Allow floating-point contraction (e.g. fusing a multiply followed by an
2355 addition into a fused multiply-and-add).
2358 Approximate functions - Allow substitution of approximate calculations for
2359 functions (sin, log, sqrt, etc). See floating-point intrinsic definitions
2360 for places where this can apply to LLVM's intrinsic math functions.
2363 Allow reassociation transformations for floating-point instructions.
2364 This may dramatically change results in floating point.
2367 This flag implies all of the others.
2371 Use-list Order Directives
2372 -------------------------
2374 Use-list directives encode the in-memory order of each use-list, allowing the
2375 order to be recreated. ``<order-indexes>`` is a comma-separated list of
2376 indexes that are assigned to the referenced value's uses. The referenced
2377 value's use-list is immediately sorted by these indexes.
2379 Use-list directives may appear at function scope or global scope. They are not
2380 instructions, and have no effect on the semantics of the IR. When they're at
2381 function scope, they must appear after the terminator of the final basic block.
2383 If basic blocks have their address taken via ``blockaddress()`` expressions,
2384 ``uselistorder_bb`` can be used to reorder their use-lists from outside their
2391 uselistorder <ty> <value>, { <order-indexes> }
2392 uselistorder_bb @function, %block { <order-indexes> }
2398 define void @foo(i32 %arg1, i32 %arg2) {
2400 ; ... instructions ...
2402 ; ... instructions ...
2404 ; At function scope.
2405 uselistorder i32 %arg1, { 1, 0, 2 }
2406 uselistorder label %bb, { 1, 0 }
2410 uselistorder i32* @global, { 1, 2, 0 }
2411 uselistorder i32 7, { 1, 0 }
2412 uselistorder i32 (i32) @bar, { 1, 0 }
2413 uselistorder_bb @foo, %bb, { 5, 1, 3, 2, 0, 4 }
2415 .. _source_filename:
2420 The *source filename* string is set to the original module identifier,
2421 which will be the name of the compiled source file when compiling from
2422 source through the clang front end, for example. It is then preserved through
2425 This is currently necessary to generate a consistent unique global
2426 identifier for local functions used in profile data, which prepends the
2427 source file name to the local function name.
2429 The syntax for the source file name is simply:
2431 .. code-block:: text
2433 source_filename = "/path/to/source.c"
2440 The LLVM type system is one of the most important features of the
2441 intermediate representation. Being typed enables a number of
2442 optimizations to be performed on the intermediate representation
2443 directly, without having to do extra analyses on the side before the
2444 transformation. A strong type system makes it easier to read the
2445 generated code and enables novel analyses and transformations that are
2446 not feasible to perform on normal three address code representations.
2456 The void type does not represent any value and has no size.
2474 The function type can be thought of as a function signature. It consists of a
2475 return type and a list of formal parameter types. The return type of a function
2476 type is a void type or first class type --- except for :ref:`label <t_label>`
2477 and :ref:`metadata <t_metadata>` types.
2483 <returntype> (<parameter list>)
2485 ...where '``<parameter list>``' is a comma-separated list of type
2486 specifiers. Optionally, the parameter list may include a type ``...``, which
2487 indicates that the function takes a variable number of arguments. Variable
2488 argument functions can access their arguments with the :ref:`variable argument
2489 handling intrinsic <int_varargs>` functions. '``<returntype>``' is any type
2490 except :ref:`label <t_label>` and :ref:`metadata <t_metadata>`.
2494 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2495 | ``i32 (i32)`` | function taking an ``i32``, returning an ``i32`` |
2496 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2497 | ``float (i16, i32 *) *`` | :ref:`Pointer <t_pointer>` to a function that takes an ``i16`` and a :ref:`pointer <t_pointer>` to ``i32``, returning ``float``. |
2498 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2499 | ``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. |
2500 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2501 | ``{i32, i32} (i32)`` | A function taking an ``i32``, returning a :ref:`structure <t_struct>` containing two ``i32`` values |
2502 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2509 The :ref:`first class <t_firstclass>` types are perhaps the most important.
2510 Values of these types are the only ones which can be produced by
2518 These are the types that are valid in registers from CodeGen's perspective.
2527 The integer type is a very simple type that simply specifies an
2528 arbitrary bit width for the integer type desired. Any bit width from 1
2529 bit to 2\ :sup:`23`\ -1 (about 8 million) can be specified.
2537 The number of bits the integer will occupy is specified by the ``N``
2543 +----------------+------------------------------------------------+
2544 | ``i1`` | a single-bit integer. |
2545 +----------------+------------------------------------------------+
2546 | ``i32`` | a 32-bit integer. |
2547 +----------------+------------------------------------------------+
2548 | ``i1942652`` | a really big integer of over 1 million bits. |
2549 +----------------+------------------------------------------------+
2553 Floating Point Types
2554 """"""""""""""""""""
2563 - 16-bit floating point value
2566 - 32-bit floating point value
2569 - 64-bit floating point value
2572 - 128-bit floating point value (112-bit mantissa)
2575 - 80-bit floating point value (X87)
2578 - 128-bit floating point value (two 64-bits)
2585 The x86_mmx type represents a value held in an MMX register on an x86
2586 machine. The operations allowed on it are quite limited: parameters and
2587 return values, load and store, and bitcast. User-specified MMX
2588 instructions are represented as intrinsic or asm calls with arguments
2589 and/or results of this type. There are no arrays, vectors or constants
2606 The pointer type is used to specify memory locations. Pointers are
2607 commonly used to reference objects in memory.
2609 Pointer types may have an optional address space attribute defining the
2610 numbered address space where the pointed-to object resides. The default
2611 address space is number zero. The semantics of non-zero address spaces
2612 are target-specific.
2614 Note that LLVM does not permit pointers to void (``void*``) nor does it
2615 permit pointers to labels (``label*``). Use ``i8*`` instead.
2625 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2626 | ``[4 x i32]*`` | A :ref:`pointer <t_pointer>` to :ref:`array <t_array>` of four ``i32`` values. |
2627 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2628 | ``i32 (i32*) *`` | A :ref:`pointer <t_pointer>` to a :ref:`function <t_function>` that takes an ``i32*``, returning an ``i32``. |
2629 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2630 | ``i32 addrspace(5)*`` | A :ref:`pointer <t_pointer>` to an ``i32`` value that resides in address space #5. |
2631 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2640 A vector type is a simple derived type that represents a vector of
2641 elements. Vector types are used when multiple primitive data are
2642 operated in parallel using a single instruction (SIMD). A vector type
2643 requires a size (number of elements) and an underlying primitive data
2644 type. Vector types are considered :ref:`first class <t_firstclass>`.
2650 < <# elements> x <elementtype> >
2652 The number of elements is a constant integer value larger than 0;
2653 elementtype may be any integer, floating point or pointer type. Vectors
2654 of size zero are not allowed.
2658 +-------------------+--------------------------------------------------+
2659 | ``<4 x i32>`` | Vector of 4 32-bit integer values. |
2660 +-------------------+--------------------------------------------------+
2661 | ``<8 x float>`` | Vector of 8 32-bit floating-point values. |
2662 +-------------------+--------------------------------------------------+
2663 | ``<2 x i64>`` | Vector of 2 64-bit integer values. |
2664 +-------------------+--------------------------------------------------+
2665 | ``<4 x i64*>`` | Vector of 4 pointers to 64-bit integer values. |
2666 +-------------------+--------------------------------------------------+
2675 The label type represents code labels.
2690 The token type is used when a value is associated with an instruction
2691 but all uses of the value must not attempt to introspect or obscure it.
2692 As such, it is not appropriate to have a :ref:`phi <i_phi>` or
2693 :ref:`select <i_select>` of type token.
2710 The metadata type represents embedded metadata. No derived types may be
2711 created from metadata except for :ref:`function <t_function>` arguments.
2724 Aggregate Types are a subset of derived types that can contain multiple
2725 member types. :ref:`Arrays <t_array>` and :ref:`structs <t_struct>` are
2726 aggregate types. :ref:`Vectors <t_vector>` are not considered to be
2736 The array type is a very simple derived type that arranges elements
2737 sequentially in memory. The array type requires a size (number of
2738 elements) and an underlying data type.
2744 [<# elements> x <elementtype>]
2746 The number of elements is a constant integer value; ``elementtype`` may
2747 be any type with a size.
2751 +------------------+--------------------------------------+
2752 | ``[40 x i32]`` | Array of 40 32-bit integer values. |
2753 +------------------+--------------------------------------+
2754 | ``[41 x i32]`` | Array of 41 32-bit integer values. |
2755 +------------------+--------------------------------------+
2756 | ``[4 x i8]`` | Array of 4 8-bit integer values. |
2757 +------------------+--------------------------------------+
2759 Here are some examples of multidimensional arrays:
2761 +-----------------------------+----------------------------------------------------------+
2762 | ``[3 x [4 x i32]]`` | 3x4 array of 32-bit integer values. |
2763 +-----------------------------+----------------------------------------------------------+
2764 | ``[12 x [10 x float]]`` | 12x10 array of single precision floating point values. |
2765 +-----------------------------+----------------------------------------------------------+
2766 | ``[2 x [3 x [4 x i16]]]`` | 2x3x4 array of 16-bit integer values. |
2767 +-----------------------------+----------------------------------------------------------+
2769 There is no restriction on indexing beyond the end of the array implied
2770 by a static type (though there are restrictions on indexing beyond the
2771 bounds of an allocated object in some cases). This means that
2772 single-dimension 'variable sized array' addressing can be implemented in
2773 LLVM with a zero length array type. An implementation of 'pascal style
2774 arrays' in LLVM could use the type "``{ i32, [0 x float]}``", for
2784 The structure type is used to represent a collection of data members
2785 together in memory. The elements of a structure may be any type that has
2788 Structures in memory are accessed using '``load``' and '``store``' by
2789 getting a pointer to a field with the '``getelementptr``' instruction.
2790 Structures in registers are accessed using the '``extractvalue``' and
2791 '``insertvalue``' instructions.
2793 Structures may optionally be "packed" structures, which indicate that
2794 the alignment of the struct is one byte, and that there is no padding
2795 between the elements. In non-packed structs, padding between field types
2796 is inserted as defined by the DataLayout string in the module, which is
2797 required to match what the underlying code generator expects.
2799 Structures can either be "literal" or "identified". A literal structure
2800 is defined inline with other types (e.g. ``{i32, i32}*``) whereas
2801 identified types are always defined at the top level with a name.
2802 Literal types are uniqued by their contents and can never be recursive
2803 or opaque since there is no way to write one. Identified types can be
2804 recursive, can be opaqued, and are never uniqued.
2810 %T1 = type { <type list> } ; Identified normal struct type
2811 %T2 = type <{ <type list> }> ; Identified packed struct type
2815 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2816 | ``{ i32, i32, i32 }`` | A triple of three ``i32`` values |
2817 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2818 | ``{ 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``. |
2819 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2820 | ``<{ i8, i32 }>`` | A packed struct known to be 5 bytes in size. |
2821 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2825 Opaque Structure Types
2826 """"""""""""""""""""""
2830 Opaque structure types are used to represent named structure types that
2831 do not have a body specified. This corresponds (for example) to the C
2832 notion of a forward declared structure.
2843 +--------------+-------------------+
2844 | ``opaque`` | An opaque type. |
2845 +--------------+-------------------+
2852 LLVM has several different basic types of constants. This section
2853 describes them all and their syntax.
2858 **Boolean constants**
2859 The two strings '``true``' and '``false``' are both valid constants
2861 **Integer constants**
2862 Standard integers (such as '4') are constants of the
2863 :ref:`integer <t_integer>` type. Negative numbers may be used with
2865 **Floating point constants**
2866 Floating point constants use standard decimal notation (e.g.
2867 123.421), exponential notation (e.g. 1.23421e+2), or a more precise
2868 hexadecimal notation (see below). The assembler requires the exact
2869 decimal value of a floating-point constant. For example, the
2870 assembler accepts 1.25 but rejects 1.3 because 1.3 is a repeating
2871 decimal in binary. Floating point constants must have a :ref:`floating
2872 point <t_floating>` type.
2873 **Null pointer constants**
2874 The identifier '``null``' is recognized as a null pointer constant
2875 and must be of :ref:`pointer type <t_pointer>`.
2877 The identifier '``none``' is recognized as an empty token constant
2878 and must be of :ref:`token type <t_token>`.
2880 The one non-intuitive notation for constants is the hexadecimal form of
2881 floating point constants. For example, the form
2882 '``double 0x432ff973cafa8000``' is equivalent to (but harder to read
2883 than) '``double 4.5e+15``'. The only time hexadecimal floating point
2884 constants are required (and the only time that they are generated by the
2885 disassembler) is when a floating point constant must be emitted but it
2886 cannot be represented as a decimal floating point number in a reasonable
2887 number of digits. For example, NaN's, infinities, and other special
2888 values are represented in their IEEE hexadecimal format so that assembly
2889 and disassembly do not cause any bits to change in the constants.
2891 When using the hexadecimal form, constants of types half, float, and
2892 double are represented using the 16-digit form shown above (which
2893 matches the IEEE754 representation for double); half and float values
2894 must, however, be exactly representable as IEEE 754 half and single
2895 precision, respectively. Hexadecimal format is always used for long
2896 double, and there are three forms of long double. The 80-bit format used
2897 by x86 is represented as ``0xK`` followed by 20 hexadecimal digits. The
2898 128-bit format used by PowerPC (two adjacent doubles) is represented by
2899 ``0xM`` followed by 32 hexadecimal digits. The IEEE 128-bit format is
2900 represented by ``0xL`` followed by 32 hexadecimal digits. Long doubles
2901 will only work if they match the long double format on your target.
2902 The IEEE 16-bit format (half precision) is represented by ``0xH``
2903 followed by 4 hexadecimal digits. All hexadecimal formats are big-endian
2904 (sign bit at the left).
2906 There are no constants of type x86_mmx.
2908 .. _complexconstants:
2913 Complex constants are a (potentially recursive) combination of simple
2914 constants and smaller complex constants.
2916 **Structure constants**
2917 Structure constants are represented with notation similar to
2918 structure type definitions (a comma separated list of elements,
2919 surrounded by braces (``{}``)). For example:
2920 "``{ i32 4, float 17.0, i32* @G }``", where "``@G``" is declared as
2921 "``@G = external global i32``". Structure constants must have
2922 :ref:`structure type <t_struct>`, and the number and types of elements
2923 must match those specified by the type.
2925 Array constants are represented with notation similar to array type
2926 definitions (a comma separated list of elements, surrounded by
2927 square brackets (``[]``)). For example:
2928 "``[ i32 42, i32 11, i32 74 ]``". Array constants must have
2929 :ref:`array type <t_array>`, and the number and types of elements must
2930 match those specified by the type. As a special case, character array
2931 constants may also be represented as a double-quoted string using the ``c``
2932 prefix. For example: "``c"Hello World\0A\00"``".
2933 **Vector constants**
2934 Vector constants are represented with notation similar to vector
2935 type definitions (a comma separated list of elements, surrounded by
2936 less-than/greater-than's (``<>``)). For example:
2937 "``< i32 42, i32 11, i32 74, i32 100 >``". Vector constants
2938 must have :ref:`vector type <t_vector>`, and the number and types of
2939 elements must match those specified by the type.
2940 **Zero initialization**
2941 The string '``zeroinitializer``' can be used to zero initialize a
2942 value to zero of *any* type, including scalar and
2943 :ref:`aggregate <t_aggregate>` types. This is often used to avoid
2944 having to print large zero initializers (e.g. for large arrays) and
2945 is always exactly equivalent to using explicit zero initializers.
2947 A metadata node is a constant tuple without types. For example:
2948 "``!{!0, !{!2, !0}, !"test"}``". Metadata can reference constant values,
2949 for example: "``!{!0, i32 0, i8* @global, i64 (i64)* @function, !"str"}``".
2950 Unlike other typed constants that are meant to be interpreted as part of
2951 the instruction stream, metadata is a place to attach additional
2952 information such as debug info.
2954 Global Variable and Function Addresses
2955 --------------------------------------
2957 The addresses of :ref:`global variables <globalvars>` and
2958 :ref:`functions <functionstructure>` are always implicitly valid
2959 (link-time) constants. These constants are explicitly referenced when
2960 the :ref:`identifier for the global <identifiers>` is used and always have
2961 :ref:`pointer <t_pointer>` type. For example, the following is a legal LLVM
2964 .. code-block:: llvm
2968 @Z = global [2 x i32*] [ i32* @X, i32* @Y ]
2975 The string '``undef``' can be used anywhere a constant is expected, and
2976 indicates that the user of the value may receive an unspecified
2977 bit-pattern. Undefined values may be of any type (other than '``label``'
2978 or '``void``') and be used anywhere a constant is permitted.
2980 Undefined values are useful because they indicate to the compiler that
2981 the program is well defined no matter what value is used. This gives the
2982 compiler more freedom to optimize. Here are some examples of
2983 (potentially surprising) transformations that are valid (in pseudo IR):
2985 .. code-block:: llvm
2995 This is safe because all of the output bits are affected by the undef
2996 bits. Any output bit can have a zero or one depending on the input bits.
2998 .. code-block:: llvm
3006 %A = %X ;; By choosing undef as 0
3007 %B = %X ;; By choosing undef as -1
3012 These logical operations have bits that are not always affected by the
3013 input. For example, if ``%X`` has a zero bit, then the output of the
3014 '``and``' operation will always be a zero for that bit, no matter what
3015 the corresponding bit from the '``undef``' is. As such, it is unsafe to
3016 optimize or assume that the result of the '``and``' is '``undef``'.
3017 However, it is safe to assume that all bits of the '``undef``' could be
3018 0, and optimize the '``and``' to 0. Likewise, it is safe to assume that
3019 all the bits of the '``undef``' operand to the '``or``' could be set,
3020 allowing the '``or``' to be folded to -1.
3022 .. code-block:: llvm
3024 %A = select undef, %X, %Y
3025 %B = select undef, 42, %Y
3026 %C = select %X, %Y, undef
3036 This set of examples shows that undefined '``select``' (and conditional
3037 branch) conditions can go *either way*, but they have to come from one
3038 of the two operands. In the ``%A`` example, if ``%X`` and ``%Y`` were
3039 both known to have a clear low bit, then ``%A`` would have to have a
3040 cleared low bit. However, in the ``%C`` example, the optimizer is
3041 allowed to assume that the '``undef``' operand could be the same as
3042 ``%Y``, allowing the whole '``select``' to be eliminated.
3044 .. code-block:: text
3046 %A = xor undef, undef
3063 This example points out that two '``undef``' operands are not
3064 necessarily the same. This can be surprising to people (and also matches
3065 C semantics) where they assume that "``X^X``" is always zero, even if
3066 ``X`` is undefined. This isn't true for a number of reasons, but the
3067 short answer is that an '``undef``' "variable" can arbitrarily change
3068 its value over its "live range". This is true because the variable
3069 doesn't actually *have a live range*. Instead, the value is logically
3070 read from arbitrary registers that happen to be around when needed, so
3071 the value is not necessarily consistent over time. In fact, ``%A`` and
3072 ``%C`` need to have the same semantics or the core LLVM "replace all
3073 uses with" concept would not hold.
3075 .. code-block:: llvm
3083 These examples show the crucial difference between an *undefined value*
3084 and *undefined behavior*. An undefined value (like '``undef``') is
3085 allowed to have an arbitrary bit-pattern. This means that the ``%A``
3086 operation can be constant folded to '``0``', because the '``undef``'
3087 could be zero, and zero divided by any value is zero.
3088 However, in the second example, we can make a more aggressive
3089 assumption: because the ``undef`` is allowed to be an arbitrary value,
3090 we are allowed to assume that it could be zero. Since a divide by zero
3091 has *undefined behavior*, we are allowed to assume that the operation
3092 does not execute at all. This allows us to delete the divide and all
3093 code after it. Because the undefined operation "can't happen", the
3094 optimizer can assume that it occurs in dead code.
3096 .. code-block:: text
3098 a: store undef -> %X
3099 b: store %X -> undef
3104 A store *of* an undefined value can be assumed to not have any effect;
3105 we can assume that the value is overwritten with bits that happen to
3106 match what was already there. However, a store *to* an undefined
3107 location could clobber arbitrary memory, therefore, it has undefined
3115 Poison values are similar to :ref:`undef values <undefvalues>`, however
3116 they also represent the fact that an instruction or constant expression
3117 that cannot evoke side effects has nevertheless detected a condition
3118 that results in undefined behavior.
3120 There is currently no way of representing a poison value in the IR; they
3121 only exist when produced by operations such as :ref:`add <i_add>` with
3124 Poison value behavior is defined in terms of value *dependence*:
3126 - Values other than :ref:`phi <i_phi>` nodes depend on their operands.
3127 - :ref:`Phi <i_phi>` nodes depend on the operand corresponding to
3128 their dynamic predecessor basic block.
3129 - Function arguments depend on the corresponding actual argument values
3130 in the dynamic callers of their functions.
3131 - :ref:`Call <i_call>` instructions depend on the :ref:`ret <i_ret>`
3132 instructions that dynamically transfer control back to them.
3133 - :ref:`Invoke <i_invoke>` instructions depend on the
3134 :ref:`ret <i_ret>`, :ref:`resume <i_resume>`, or exception-throwing
3135 call instructions that dynamically transfer control back to them.
3136 - Non-volatile loads and stores depend on the most recent stores to all
3137 of the referenced memory addresses, following the order in the IR
3138 (including loads and stores implied by intrinsics such as
3139 :ref:`@llvm.memcpy <int_memcpy>`.)
3140 - An instruction with externally visible side effects depends on the
3141 most recent preceding instruction with externally visible side
3142 effects, following the order in the IR. (This includes :ref:`volatile
3143 operations <volatile>`.)
3144 - An instruction *control-depends* on a :ref:`terminator
3145 instruction <terminators>` if the terminator instruction has
3146 multiple successors and the instruction is always executed when
3147 control transfers to one of the successors, and may not be executed
3148 when control is transferred to another.
3149 - Additionally, an instruction also *control-depends* on a terminator
3150 instruction if the set of instructions it otherwise depends on would
3151 be different if the terminator had transferred control to a different
3153 - Dependence is transitive.
3155 Poison values have the same behavior as :ref:`undef values <undefvalues>`,
3156 with the additional effect that any instruction that has a *dependence*
3157 on a poison value has undefined behavior.
3159 Here are some examples:
3161 .. code-block:: llvm
3164 %poison = sub nuw i32 0, 1 ; Results in a poison value.
3165 %still_poison = and i32 %poison, 0 ; 0, but also poison.
3166 %poison_yet_again = getelementptr i32, i32* @h, i32 %still_poison
3167 store i32 0, i32* %poison_yet_again ; memory at @h[0] is poisoned
3169 store i32 %poison, i32* @g ; Poison value stored to memory.
3170 %poison2 = load i32, i32* @g ; Poison value loaded back from memory.
3172 store volatile i32 %poison, i32* @g ; External observation; undefined behavior.
3174 %narrowaddr = bitcast i32* @g to i16*
3175 %wideaddr = bitcast i32* @g to i64*
3176 %poison3 = load i16, i16* %narrowaddr ; Returns a poison value.
3177 %poison4 = load i64, i64* %wideaddr ; Returns a poison value.
3179 %cmp = icmp slt i32 %poison, 0 ; Returns a poison value.
3180 br i1 %cmp, label %true, label %end ; Branch to either destination.
3183 store volatile i32 0, i32* @g ; This is control-dependent on %cmp, so
3184 ; it has undefined behavior.
3188 %p = phi i32 [ 0, %entry ], [ 1, %true ]
3189 ; Both edges into this PHI are
3190 ; control-dependent on %cmp, so this
3191 ; always results in a poison value.
3193 store volatile i32 0, i32* @g ; This would depend on the store in %true
3194 ; if %cmp is true, or the store in %entry
3195 ; otherwise, so this is undefined behavior.
3197 br i1 %cmp, label %second_true, label %second_end
3198 ; The same branch again, but this time the
3199 ; true block doesn't have side effects.
3206 store volatile i32 0, i32* @g ; This time, the instruction always depends
3207 ; on the store in %end. Also, it is
3208 ; control-equivalent to %end, so this is
3209 ; well-defined (ignoring earlier undefined
3210 ; behavior in this example).
3214 Addresses of Basic Blocks
3215 -------------------------
3217 ``blockaddress(@function, %block)``
3219 The '``blockaddress``' constant computes the address of the specified
3220 basic block in the specified function, and always has an ``i8*`` type.
3221 Taking the address of the entry block is illegal.
3223 This value only has defined behavior when used as an operand to the
3224 ':ref:`indirectbr <i_indirectbr>`' instruction, or for comparisons
3225 against null. Pointer equality tests between labels addresses results in
3226 undefined behavior --- though, again, comparison against null is ok, and
3227 no label is equal to the null pointer. This may be passed around as an
3228 opaque pointer sized value as long as the bits are not inspected. This
3229 allows ``ptrtoint`` and arithmetic to be performed on these values so
3230 long as the original value is reconstituted before the ``indirectbr``
3233 Finally, some targets may provide defined semantics when using the value
3234 as the operand to an inline assembly, but that is target specific.
3238 Constant Expressions
3239 --------------------
3241 Constant expressions are used to allow expressions involving other
3242 constants to be used as constants. Constant expressions may be of any
3243 :ref:`first class <t_firstclass>` type and may involve any LLVM operation
3244 that does not have side effects (e.g. load and call are not supported).
3245 The following is the syntax for constant expressions:
3247 ``trunc (CST to TYPE)``
3248 Perform the :ref:`trunc operation <i_trunc>` on constants.
3249 ``zext (CST to TYPE)``
3250 Perform the :ref:`zext operation <i_zext>` on constants.
3251 ``sext (CST to TYPE)``
3252 Perform the :ref:`sext operation <i_sext>` on constants.
3253 ``fptrunc (CST to TYPE)``
3254 Truncate a floating point constant to another floating point type.
3255 The size of CST must be larger than the size of TYPE. Both types
3256 must be floating point.
3257 ``fpext (CST to TYPE)``
3258 Floating point extend a constant to another type. The size of CST
3259 must be smaller or equal to the size of TYPE. Both types must be
3261 ``fptoui (CST to TYPE)``
3262 Convert a floating point constant to the corresponding unsigned
3263 integer constant. TYPE must be a scalar or vector integer type. CST
3264 must be of scalar or vector floating point type. Both CST and TYPE
3265 must be scalars, or vectors of the same number of elements. If the
3266 value won't fit in the integer type, the results are undefined.
3267 ``fptosi (CST to TYPE)``
3268 Convert a floating point constant to the corresponding signed
3269 integer constant. TYPE must be a scalar or vector integer type. CST
3270 must be of scalar or vector floating point type. Both CST and TYPE
3271 must be scalars, or vectors of the same number of elements. If the
3272 value won't fit in the integer type, the results are undefined.
3273 ``uitofp (CST to TYPE)``
3274 Convert an unsigned integer constant to the corresponding floating
3275 point constant. TYPE must be a scalar or vector floating point type.
3276 CST must be of scalar or vector integer type. Both CST and TYPE must
3277 be scalars, or vectors of the same number of elements. If the value
3278 won't fit in the floating point type, the results are undefined.
3279 ``sitofp (CST to TYPE)``
3280 Convert a signed integer constant to the corresponding floating
3281 point constant. TYPE must be a scalar or vector floating point type.
3282 CST must be of scalar or vector integer type. Both CST and TYPE must
3283 be scalars, or vectors of the same number of elements. If the value
3284 won't fit in the floating point type, the results are undefined.
3285 ``ptrtoint (CST to TYPE)``
3286 Perform the :ref:`ptrtoint operation <i_ptrtoint>` on constants.
3287 ``inttoptr (CST to TYPE)``
3288 Perform the :ref:`inttoptr operation <i_inttoptr>` on constants.
3289 This one is *really* dangerous!
3290 ``bitcast (CST to TYPE)``
3291 Convert a constant, CST, to another TYPE.
3292 The constraints of the operands are the same as those for the
3293 :ref:`bitcast instruction <i_bitcast>`.
3294 ``addrspacecast (CST to TYPE)``
3295 Convert a constant pointer or constant vector of pointer, CST, to another
3296 TYPE in a different address space. The constraints of the operands are the
3297 same as those for the :ref:`addrspacecast instruction <i_addrspacecast>`.
3298 ``getelementptr (TY, CSTPTR, IDX0, IDX1, ...)``, ``getelementptr inbounds (TY, CSTPTR, IDX0, IDX1, ...)``
3299 Perform the :ref:`getelementptr operation <i_getelementptr>` on
3300 constants. As with the :ref:`getelementptr <i_getelementptr>`
3301 instruction, the index list may have one or more indexes, which are
3302 required to make sense for the type of "pointer to TY".
3303 ``select (COND, VAL1, VAL2)``
3304 Perform the :ref:`select operation <i_select>` on constants.
3305 ``icmp COND (VAL1, VAL2)``
3306 Perform the :ref:`icmp operation <i_icmp>` on constants.
3307 ``fcmp COND (VAL1, VAL2)``
3308 Perform the :ref:`fcmp operation <i_fcmp>` on constants.
3309 ``extractelement (VAL, IDX)``
3310 Perform the :ref:`extractelement operation <i_extractelement>` on
3312 ``insertelement (VAL, ELT, IDX)``
3313 Perform the :ref:`insertelement operation <i_insertelement>` on
3315 ``shufflevector (VEC1, VEC2, IDXMASK)``
3316 Perform the :ref:`shufflevector operation <i_shufflevector>` on
3318 ``extractvalue (VAL, IDX0, IDX1, ...)``
3319 Perform the :ref:`extractvalue operation <i_extractvalue>` on
3320 constants. The index list is interpreted in a similar manner as
3321 indices in a ':ref:`getelementptr <i_getelementptr>`' operation. At
3322 least one index value must be specified.
3323 ``insertvalue (VAL, ELT, IDX0, IDX1, ...)``
3324 Perform the :ref:`insertvalue operation <i_insertvalue>` on constants.
3325 The index list is interpreted in a similar manner as indices in a
3326 ':ref:`getelementptr <i_getelementptr>`' operation. At least one index
3327 value must be specified.
3328 ``OPCODE (LHS, RHS)``
3329 Perform the specified operation of the LHS and RHS constants. OPCODE
3330 may be any of the :ref:`binary <binaryops>` or :ref:`bitwise
3331 binary <bitwiseops>` operations. The constraints on operands are
3332 the same as those for the corresponding instruction (e.g. no bitwise
3333 operations on floating point values are allowed).
3340 Inline Assembler Expressions
3341 ----------------------------
3343 LLVM supports inline assembler expressions (as opposed to :ref:`Module-Level
3344 Inline Assembly <moduleasm>`) through the use of a special value. This value
3345 represents the inline assembler as a template string (containing the
3346 instructions to emit), a list of operand constraints (stored as a string), a
3347 flag that indicates whether or not the inline asm expression has side effects,
3348 and a flag indicating whether the function containing the asm needs to align its
3349 stack conservatively.
3351 The template string supports argument substitution of the operands using "``$``"
3352 followed by a number, to indicate substitution of the given register/memory
3353 location, as specified by the constraint string. "``${NUM:MODIFIER}``" may also
3354 be used, where ``MODIFIER`` is a target-specific annotation for how to print the
3355 operand (See :ref:`inline-asm-modifiers`).
3357 A literal "``$``" may be included by using "``$$``" in the template. To include
3358 other special characters into the output, the usual "``\XX``" escapes may be
3359 used, just as in other strings. Note that after template substitution, the
3360 resulting assembly string is parsed by LLVM's integrated assembler unless it is
3361 disabled -- even when emitting a ``.s`` file -- and thus must contain assembly
3362 syntax known to LLVM.
3364 LLVM also supports a few more substitions useful for writing inline assembly:
3366 - ``${:uid}``: Expands to a decimal integer unique to this inline assembly blob.
3367 This substitution is useful when declaring a local label. Many standard
3368 compiler optimizations, such as inlining, may duplicate an inline asm blob.
3369 Adding a blob-unique identifier ensures that the two labels will not conflict
3370 during assembly. This is used to implement `GCC's %= special format
3371 string <https://gcc.gnu.org/onlinedocs/gcc/Extended-Asm.html>`_.
3372 - ``${:comment}``: Expands to the comment character of the current target's
3373 assembly dialect. This is usually ``#``, but many targets use other strings,
3374 such as ``;``, ``//``, or ``!``.
3375 - ``${:private}``: Expands to the assembler private label prefix. Labels with
3376 this prefix will not appear in the symbol table of the assembled object.
3377 Typically the prefix is ``L``, but targets may use other strings. ``.L`` is
3380 LLVM's support for inline asm is modeled closely on the requirements of Clang's
3381 GCC-compatible inline-asm support. Thus, the feature-set and the constraint and
3382 modifier codes listed here are similar or identical to those in GCC's inline asm
3383 support. However, to be clear, the syntax of the template and constraint strings
3384 described here is *not* the same as the syntax accepted by GCC and Clang, and,
3385 while most constraint letters are passed through as-is by Clang, some get
3386 translated to other codes when converting from the C source to the LLVM
3389 An example inline assembler expression is:
3391 .. code-block:: llvm
3393 i32 (i32) asm "bswap $0", "=r,r"
3395 Inline assembler expressions may **only** be used as the callee operand
3396 of a :ref:`call <i_call>` or an :ref:`invoke <i_invoke>` instruction.
3397 Thus, typically we have:
3399 .. code-block:: llvm
3401 %X = call i32 asm "bswap $0", "=r,r"(i32 %Y)
3403 Inline asms with side effects not visible in the constraint list must be
3404 marked as having side effects. This is done through the use of the
3405 '``sideeffect``' keyword, like so:
3407 .. code-block:: llvm
3409 call void asm sideeffect "eieio", ""()
3411 In some cases inline asms will contain code that will not work unless
3412 the stack is aligned in some way, such as calls or SSE instructions on
3413 x86, yet will not contain code that does that alignment within the asm.
3414 The compiler should make conservative assumptions about what the asm
3415 might contain and should generate its usual stack alignment code in the
3416 prologue if the '``alignstack``' keyword is present:
3418 .. code-block:: llvm
3420 call void asm alignstack "eieio", ""()
3422 Inline asms also support using non-standard assembly dialects. The
3423 assumed dialect is ATT. When the '``inteldialect``' keyword is present,
3424 the inline asm is using the Intel dialect. Currently, ATT and Intel are
3425 the only supported dialects. An example is:
3427 .. code-block:: llvm
3429 call void asm inteldialect "eieio", ""()
3431 If multiple keywords appear the '``sideeffect``' keyword must come
3432 first, the '``alignstack``' keyword second and the '``inteldialect``'
3435 Inline Asm Constraint String
3436 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3438 The constraint list is a comma-separated string, each element containing one or
3439 more constraint codes.
3441 For each element in the constraint list an appropriate register or memory
3442 operand will be chosen, and it will be made available to assembly template
3443 string expansion as ``$0`` for the first constraint in the list, ``$1`` for the
3446 There are three different types of constraints, which are distinguished by a
3447 prefix symbol in front of the constraint code: Output, Input, and Clobber. The
3448 constraints must always be given in that order: outputs first, then inputs, then
3449 clobbers. They cannot be intermingled.
3451 There are also three different categories of constraint codes:
3453 - Register constraint. This is either a register class, or a fixed physical
3454 register. This kind of constraint will allocate a register, and if necessary,
3455 bitcast the argument or result to the appropriate type.
3456 - Memory constraint. This kind of constraint is for use with an instruction
3457 taking a memory operand. Different constraints allow for different addressing
3458 modes used by the target.
3459 - Immediate value constraint. This kind of constraint is for an integer or other
3460 immediate value which can be rendered directly into an instruction. The
3461 various target-specific constraints allow the selection of a value in the
3462 proper range for the instruction you wish to use it with.
3467 Output constraints are specified by an "``=``" prefix (e.g. "``=r``"). This
3468 indicates that the assembly will write to this operand, and the operand will
3469 then be made available as a return value of the ``asm`` expression. Output
3470 constraints do not consume an argument from the call instruction. (Except, see
3471 below about indirect outputs).
3473 Normally, it is expected that no output locations are written to by the assembly
3474 expression until *all* of the inputs have been read. As such, LLVM may assign
3475 the same register to an output and an input. If this is not safe (e.g. if the
3476 assembly contains two instructions, where the first writes to one output, and
3477 the second reads an input and writes to a second output), then the "``&``"
3478 modifier must be used (e.g. "``=&r``") to specify that the output is an
3479 "early-clobber" output. Marking an output as "early-clobber" ensures that LLVM
3480 will not use the same register for any inputs (other than an input tied to this
3486 Input constraints do not have a prefix -- just the constraint codes. Each input
3487 constraint will consume one argument from the call instruction. It is not
3488 permitted for the asm to write to any input register or memory location (unless
3489 that input is tied to an output). Note also that multiple inputs may all be
3490 assigned to the same register, if LLVM can determine that they necessarily all
3491 contain the same value.
3493 Instead of providing a Constraint Code, input constraints may also "tie"
3494 themselves to an output constraint, by providing an integer as the constraint
3495 string. Tied inputs still consume an argument from the call instruction, and
3496 take up a position in the asm template numbering as is usual -- they will simply
3497 be constrained to always use the same register as the output they've been tied
3498 to. For example, a constraint string of "``=r,0``" says to assign a register for
3499 output, and use that register as an input as well (it being the 0'th
3502 It is permitted to tie an input to an "early-clobber" output. In that case, no
3503 *other* input may share the same register as the input tied to the early-clobber
3504 (even when the other input has the same value).
3506 You may only tie an input to an output which has a register constraint, not a
3507 memory constraint. Only a single input may be tied to an output.
3509 There is also an "interesting" feature which deserves a bit of explanation: if a
3510 register class constraint allocates a register which is too small for the value
3511 type operand provided as input, the input value will be split into multiple
3512 registers, and all of them passed to the inline asm.
3514 However, this feature is often not as useful as you might think.
3516 Firstly, the registers are *not* guaranteed to be consecutive. So, on those
3517 architectures that have instructions which operate on multiple consecutive
3518 instructions, this is not an appropriate way to support them. (e.g. the 32-bit
3519 SparcV8 has a 64-bit load, which instruction takes a single 32-bit register. The
3520 hardware then loads into both the named register, and the next register. This
3521 feature of inline asm would not be useful to support that.)
3523 A few of the targets provide a template string modifier allowing explicit access
3524 to the second register of a two-register operand (e.g. MIPS ``L``, ``M``, and
3525 ``D``). On such an architecture, you can actually access the second allocated
3526 register (yet, still, not any subsequent ones). But, in that case, you're still
3527 probably better off simply splitting the value into two separate operands, for
3528 clarity. (e.g. see the description of the ``A`` constraint on X86, which,
3529 despite existing only for use with this feature, is not really a good idea to
3532 Indirect inputs and outputs
3533 """""""""""""""""""""""""""
3535 Indirect output or input constraints can be specified by the "``*``" modifier
3536 (which goes after the "``=``" in case of an output). This indicates that the asm
3537 will write to or read from the contents of an *address* provided as an input
3538 argument. (Note that in this way, indirect outputs act more like an *input* than
3539 an output: just like an input, they consume an argument of the call expression,
3540 rather than producing a return value. An indirect output constraint is an
3541 "output" only in that the asm is expected to write to the contents of the input
3542 memory location, instead of just read from it).
3544 This is most typically used for memory constraint, e.g. "``=*m``", to pass the
3545 address of a variable as a value.
3547 It is also possible to use an indirect *register* constraint, but only on output
3548 (e.g. "``=*r``"). This will cause LLVM to allocate a register for an output
3549 value normally, and then, separately emit a store to the address provided as
3550 input, after the provided inline asm. (It's not clear what value this
3551 functionality provides, compared to writing the store explicitly after the asm
3552 statement, and it can only produce worse code, since it bypasses many
3553 optimization passes. I would recommend not using it.)
3559 A clobber constraint is indicated by a "``~``" prefix. A clobber does not
3560 consume an input operand, nor generate an output. Clobbers cannot use any of the
3561 general constraint code letters -- they may use only explicit register
3562 constraints, e.g. "``~{eax}``". The one exception is that a clobber string of
3563 "``~{memory}``" indicates that the assembly writes to arbitrary undeclared
3564 memory locations -- not only the memory pointed to by a declared indirect
3567 Note that clobbering named registers that are also present in output
3568 constraints is not legal.
3573 After a potential prefix comes constraint code, or codes.
3575 A Constraint Code is either a single letter (e.g. "``r``"), a "``^``" character
3576 followed by two letters (e.g. "``^wc``"), or "``{``" register-name "``}``"
3579 The one and two letter constraint codes are typically chosen to be the same as
3580 GCC's constraint codes.
3582 A single constraint may include one or more than constraint code in it, leaving
3583 it up to LLVM to choose which one to use. This is included mainly for
3584 compatibility with the translation of GCC inline asm coming from clang.
3586 There are two ways to specify alternatives, and either or both may be used in an
3587 inline asm constraint list:
3589 1) Append the codes to each other, making a constraint code set. E.g. "``im``"
3590 or "``{eax}m``". This means "choose any of the options in the set". The
3591 choice of constraint is made independently for each constraint in the
3594 2) Use "``|``" between constraint code sets, creating alternatives. Every
3595 constraint in the constraint list must have the same number of alternative
3596 sets. With this syntax, the same alternative in *all* of the items in the
3597 constraint list will be chosen together.
3599 Putting those together, you might have a two operand constraint string like
3600 ``"rm|r,ri|rm"``. This indicates that if operand 0 is ``r`` or ``m``, then
3601 operand 1 may be one of ``r`` or ``i``. If operand 0 is ``r``, then operand 1
3602 may be one of ``r`` or ``m``. But, operand 0 and 1 cannot both be of type m.
3604 However, the use of either of the alternatives features is *NOT* recommended, as
3605 LLVM is not able to make an intelligent choice about which one to use. (At the
3606 point it currently needs to choose, not enough information is available to do so
3607 in a smart way.) Thus, it simply tries to make a choice that's most likely to
3608 compile, not one that will be optimal performance. (e.g., given "``rm``", it'll
3609 always choose to use memory, not registers). And, if given multiple registers,
3610 or multiple register classes, it will simply choose the first one. (In fact, it
3611 doesn't currently even ensure explicitly specified physical registers are
3612 unique, so specifying multiple physical registers as alternatives, like
3613 ``{r11}{r12},{r11}{r12}``, will assign r11 to both operands, not at all what was
3616 Supported Constraint Code List
3617 """"""""""""""""""""""""""""""
3619 The constraint codes are, in general, expected to behave the same way they do in
3620 GCC. LLVM's support is often implemented on an 'as-needed' basis, to support C
3621 inline asm code which was supported by GCC. A mismatch in behavior between LLVM
3622 and GCC likely indicates a bug in LLVM.
3624 Some constraint codes are typically supported by all targets:
3626 - ``r``: A register in the target's general purpose register class.
3627 - ``m``: A memory address operand. It is target-specific what addressing modes
3628 are supported, typical examples are register, or register + register offset,
3629 or register + immediate offset (of some target-specific size).
3630 - ``i``: An integer constant (of target-specific width). Allows either a simple
3631 immediate, or a relocatable value.
3632 - ``n``: An integer constant -- *not* including relocatable values.
3633 - ``s``: An integer constant, but allowing *only* relocatable values.
3634 - ``X``: Allows an operand of any kind, no constraint whatsoever. Typically
3635 useful to pass a label for an asm branch or call.
3637 .. FIXME: but that surely isn't actually okay to jump out of an asm
3638 block without telling llvm about the control transfer???)
3640 - ``{register-name}``: Requires exactly the named physical register.
3642 Other constraints are target-specific:
3646 - ``z``: An immediate integer 0. Outputs ``WZR`` or ``XZR``, as appropriate.
3647 - ``I``: An immediate integer valid for an ``ADD`` or ``SUB`` instruction,
3648 i.e. 0 to 4095 with optional shift by 12.
3649 - ``J``: An immediate integer that, when negated, is valid for an ``ADD`` or
3650 ``SUB`` instruction, i.e. -1 to -4095 with optional left shift by 12.
3651 - ``K``: An immediate integer that is valid for the 'bitmask immediate 32' of a
3652 logical instruction like ``AND``, ``EOR``, or ``ORR`` with a 32-bit register.
3653 - ``L``: An immediate integer that is valid for the 'bitmask immediate 64' of a
3654 logical instruction like ``AND``, ``EOR``, or ``ORR`` with a 64-bit register.
3655 - ``M``: An immediate integer for use with the ``MOV`` assembly alias on a
3656 32-bit register. This is a superset of ``K``: in addition to the bitmask
3657 immediate, also allows immediate integers which can be loaded with a single
3658 ``MOVZ`` or ``MOVL`` instruction.
3659 - ``N``: An immediate integer for use with the ``MOV`` assembly alias on a
3660 64-bit register. This is a superset of ``L``.
3661 - ``Q``: Memory address operand must be in a single register (no
3662 offsets). (However, LLVM currently does this for the ``m`` constraint as
3664 - ``r``: A 32 or 64-bit integer register (W* or X*).
3665 - ``w``: A 32, 64, or 128-bit floating-point/SIMD register.
3666 - ``x``: A lower 128-bit floating-point/SIMD register (``V0`` to ``V15``).
3670 - ``r``: A 32 or 64-bit integer register.
3671 - ``[0-9]v``: The 32-bit VGPR register, number 0-9.
3672 - ``[0-9]s``: The 32-bit SGPR register, number 0-9.
3677 - ``Q``, ``Um``, ``Un``, ``Uq``, ``Us``, ``Ut``, ``Uv``, ``Uy``: Memory address
3678 operand. Treated the same as operand ``m``, at the moment.
3680 ARM and ARM's Thumb2 mode:
3682 - ``j``: An immediate integer between 0 and 65535 (valid for ``MOVW``)
3683 - ``I``: An immediate integer valid for a data-processing instruction.
3684 - ``J``: An immediate integer between -4095 and 4095.
3685 - ``K``: An immediate integer whose bitwise inverse is valid for a
3686 data-processing instruction. (Can be used with template modifier "``B``" to
3687 print the inverted value).
3688 - ``L``: An immediate integer whose negation is valid for a data-processing
3689 instruction. (Can be used with template modifier "``n``" to print the negated
3691 - ``M``: A power of two or a integer between 0 and 32.
3692 - ``N``: Invalid immediate constraint.
3693 - ``O``: Invalid immediate constraint.
3694 - ``r``: A general-purpose 32-bit integer register (``r0-r15``).
3695 - ``l``: In Thumb2 mode, low 32-bit GPR registers (``r0-r7``). In ARM mode, same
3697 - ``h``: In Thumb2 mode, a high 32-bit GPR register (``r8-r15``). In ARM mode,
3699 - ``w``: A 32, 64, or 128-bit floating-point/SIMD register: ``s0-s31``,
3700 ``d0-d31``, or ``q0-q15``.
3701 - ``x``: A 32, 64, or 128-bit floating-point/SIMD register: ``s0-s15``,
3702 ``d0-d7``, or ``q0-q3``.
3703 - ``t``: A low floating-point/SIMD register: ``s0-s31``, ``d0-d16``, or
3708 - ``I``: An immediate integer between 0 and 255.
3709 - ``J``: An immediate integer between -255 and -1.
3710 - ``K``: An immediate integer between 0 and 255, with optional left-shift by
3712 - ``L``: An immediate integer between -7 and 7.
3713 - ``M``: An immediate integer which is a multiple of 4 between 0 and 1020.
3714 - ``N``: An immediate integer between 0 and 31.
3715 - ``O``: An immediate integer which is a multiple of 4 between -508 and 508.
3716 - ``r``: A low 32-bit GPR register (``r0-r7``).
3717 - ``l``: A low 32-bit GPR register (``r0-r7``).
3718 - ``h``: A high GPR register (``r0-r7``).
3719 - ``w``: A 32, 64, or 128-bit floating-point/SIMD register: ``s0-s31``,
3720 ``d0-d31``, or ``q0-q15``.
3721 - ``x``: A 32, 64, or 128-bit floating-point/SIMD register: ``s0-s15``,
3722 ``d0-d7``, or ``q0-q3``.
3723 - ``t``: A low floating-point/SIMD register: ``s0-s31``, ``d0-d16``, or
3729 - ``o``, ``v``: A memory address operand, treated the same as constraint ``m``,
3731 - ``r``: A 32 or 64-bit register.
3735 - ``r``: An 8 or 16-bit register.
3739 - ``I``: An immediate signed 16-bit integer.
3740 - ``J``: An immediate integer zero.
3741 - ``K``: An immediate unsigned 16-bit integer.
3742 - ``L``: An immediate 32-bit integer, where the lower 16 bits are 0.
3743 - ``N``: An immediate integer between -65535 and -1.
3744 - ``O``: An immediate signed 15-bit integer.
3745 - ``P``: An immediate integer between 1 and 65535.
3746 - ``m``: A memory address operand. In MIPS-SE mode, allows a base address
3747 register plus 16-bit immediate offset. In MIPS mode, just a base register.
3748 - ``R``: A memory address operand. In MIPS-SE mode, allows a base address
3749 register plus a 9-bit signed offset. In MIPS mode, the same as constraint
3751 - ``ZC``: A memory address operand, suitable for use in a ``pref``, ``ll``, or
3752 ``sc`` instruction on the given subtarget (details vary).
3753 - ``r``, ``d``, ``y``: A 32 or 64-bit GPR register.
3754 - ``f``: A 32 or 64-bit FPU register (``F0-F31``), or a 128-bit MSA register
3755 (``W0-W31``). In the case of MSA registers, it is recommended to use the ``w``
3756 argument modifier for compatibility with GCC.
3757 - ``c``: A 32-bit or 64-bit GPR register suitable for indirect jump (always
3759 - ``l``: The ``lo`` register, 32 or 64-bit.
3764 - ``b``: A 1-bit integer register.
3765 - ``c`` or ``h``: A 16-bit integer register.
3766 - ``r``: A 32-bit integer register.
3767 - ``l`` or ``N``: A 64-bit integer register.
3768 - ``f``: A 32-bit float register.
3769 - ``d``: A 64-bit float register.
3774 - ``I``: An immediate signed 16-bit integer.
3775 - ``J``: An immediate unsigned 16-bit integer, shifted left 16 bits.
3776 - ``K``: An immediate unsigned 16-bit integer.
3777 - ``L``: An immediate signed 16-bit integer, shifted left 16 bits.
3778 - ``M``: An immediate integer greater than 31.
3779 - ``N``: An immediate integer that is an exact power of 2.
3780 - ``O``: The immediate integer constant 0.
3781 - ``P``: An immediate integer constant whose negation is a signed 16-bit
3783 - ``es``, ``o``, ``Q``, ``Z``, ``Zy``: A memory address operand, currently
3784 treated the same as ``m``.
3785 - ``r``: A 32 or 64-bit integer register.
3786 - ``b``: A 32 or 64-bit integer register, excluding ``R0`` (that is:
3788 - ``f``: A 32 or 64-bit float register (``F0-F31``), or when QPX is enabled, a
3789 128 or 256-bit QPX register (``Q0-Q31``; aliases the ``F`` registers).
3790 - ``v``: For ``4 x f32`` or ``4 x f64`` types, when QPX is enabled, a
3791 128 or 256-bit QPX register (``Q0-Q31``), otherwise a 128-bit
3792 altivec vector register (``V0-V31``).
3794 .. FIXME: is this a bug that v accepts QPX registers? I think this
3795 is supposed to only use the altivec vector registers?
3797 - ``y``: Condition register (``CR0-CR7``).
3798 - ``wc``: An individual CR bit in a CR register.
3799 - ``wa``, ``wd``, ``wf``: Any 128-bit VSX vector register, from the full VSX
3800 register set (overlapping both the floating-point and vector register files).
3801 - ``ws``: A 32 or 64-bit floating point register, from the full VSX register
3806 - ``I``: An immediate 13-bit signed integer.
3807 - ``r``: A 32-bit integer register.
3808 - ``f``: Any floating-point register on SparcV8, or a floating point
3809 register in the "low" half of the registers on SparcV9.
3810 - ``e``: Any floating point register. (Same as ``f`` on SparcV8.)
3814 - ``I``: An immediate unsigned 8-bit integer.
3815 - ``J``: An immediate unsigned 12-bit integer.
3816 - ``K``: An immediate signed 16-bit integer.
3817 - ``L``: An immediate signed 20-bit integer.
3818 - ``M``: An immediate integer 0x7fffffff.
3819 - ``Q``: A memory address operand with a base address and a 12-bit immediate
3820 unsigned displacement.
3821 - ``R``: A memory address operand with a base address, a 12-bit immediate
3822 unsigned displacement, and an index register.
3823 - ``S``: A memory address operand with a base address and a 20-bit immediate
3824 signed displacement.
3825 - ``T``: A memory address operand with a base address, a 20-bit immediate
3826 signed displacement, and an index register.
3827 - ``r`` or ``d``: A 32, 64, or 128-bit integer register.
3828 - ``a``: A 32, 64, or 128-bit integer address register (excludes R0, which in an
3829 address context evaluates as zero).
3830 - ``h``: A 32-bit value in the high part of a 64bit data register
3832 - ``f``: A 32, 64, or 128-bit floating point register.
3836 - ``I``: An immediate integer between 0 and 31.
3837 - ``J``: An immediate integer between 0 and 64.
3838 - ``K``: An immediate signed 8-bit integer.
3839 - ``L``: An immediate integer, 0xff or 0xffff or (in 64-bit mode only)
3841 - ``M``: An immediate integer between 0 and 3.
3842 - ``N``: An immediate unsigned 8-bit integer.
3843 - ``O``: An immediate integer between 0 and 127.
3844 - ``e``: An immediate 32-bit signed integer.
3845 - ``Z``: An immediate 32-bit unsigned integer.
3846 - ``o``, ``v``: Treated the same as ``m``, at the moment.
3847 - ``q``: An 8, 16, 32, or 64-bit register which can be accessed as an 8-bit
3848 ``l`` integer register. On X86-32, this is the ``a``, ``b``, ``c``, and ``d``
3849 registers, and on X86-64, it is all of the integer registers.
3850 - ``Q``: An 8, 16, 32, or 64-bit register which can be accessed as an 8-bit
3851 ``h`` integer register. This is the ``a``, ``b``, ``c``, and ``d`` registers.
3852 - ``r`` or ``l``: An 8, 16, 32, or 64-bit integer register.
3853 - ``R``: An 8, 16, 32, or 64-bit "legacy" integer register -- one which has
3854 existed since i386, and can be accessed without the REX prefix.
3855 - ``f``: A 32, 64, or 80-bit '387 FPU stack pseudo-register.
3856 - ``y``: A 64-bit MMX register, if MMX is enabled.
3857 - ``x``: If SSE is enabled: a 32 or 64-bit scalar operand, or 128-bit vector
3858 operand in a SSE register. If AVX is also enabled, can also be a 256-bit
3859 vector operand in an AVX register. If AVX-512 is also enabled, can also be a
3860 512-bit vector operand in an AVX512 register, Otherwise, an error.
3861 - ``Y``: The same as ``x``, if *SSE2* is enabled, otherwise an error.
3862 - ``A``: Special case: allocates EAX first, then EDX, for a single operand (in
3863 32-bit mode, a 64-bit integer operand will get split into two registers). It
3864 is not recommended to use this constraint, as in 64-bit mode, the 64-bit
3865 operand will get allocated only to RAX -- if two 32-bit operands are needed,
3866 you're better off splitting it yourself, before passing it to the asm
3871 - ``r``: A 32-bit integer register.
3874 .. _inline-asm-modifiers:
3876 Asm template argument modifiers
3877 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3879 In the asm template string, modifiers can be used on the operand reference, like
3882 The modifiers are, in general, expected to behave the same way they do in
3883 GCC. LLVM's support is often implemented on an 'as-needed' basis, to support C
3884 inline asm code which was supported by GCC. A mismatch in behavior between LLVM
3885 and GCC likely indicates a bug in LLVM.
3889 - ``c``: Print an immediate integer constant unadorned, without
3890 the target-specific immediate punctuation (e.g. no ``$`` prefix).
3891 - ``n``: Negate and print immediate integer constant unadorned, without the
3892 target-specific immediate punctuation (e.g. no ``$`` prefix).
3893 - ``l``: Print as an unadorned label, without the target-specific label
3894 punctuation (e.g. no ``$`` prefix).
3898 - ``w``: Print a GPR register with a ``w*`` name instead of ``x*`` name. E.g.,
3899 instead of ``x30``, print ``w30``.
3900 - ``x``: Print a GPR register with a ``x*`` name. (this is the default, anyhow).
3901 - ``b``, ``h``, ``s``, ``d``, ``q``: Print a floating-point/SIMD register with a
3902 ``b*``, ``h*``, ``s*``, ``d*``, or ``q*`` name, rather than the default of
3911 - ``a``: Print an operand as an address (with ``[`` and ``]`` surrounding a
3915 - ``y``: Print a VFP single-precision register as an indexed double (e.g. print
3916 as ``d4[1]`` instead of ``s9``)
3917 - ``B``: Bitwise invert and print an immediate integer constant without ``#``
3919 - ``L``: Print the low 16-bits of an immediate integer constant.
3920 - ``M``: Print as a register set suitable for ldm/stm. Also prints *all*
3921 register operands subsequent to the specified one (!), so use carefully.
3922 - ``Q``: Print the low-order register of a register-pair, or the low-order
3923 register of a two-register operand.
3924 - ``R``: Print the high-order register of a register-pair, or the high-order
3925 register of a two-register operand.
3926 - ``H``: Print the second register of a register-pair. (On a big-endian system,
3927 ``H`` is equivalent to ``Q``, and on little-endian system, ``H`` is equivalent
3930 .. FIXME: H doesn't currently support printing the second register
3931 of a two-register operand.
3933 - ``e``: Print the low doubleword register of a NEON quad register.
3934 - ``f``: Print the high doubleword register of a NEON quad register.
3935 - ``m``: Print the base register of a memory operand without the ``[`` and ``]``
3940 - ``L``: Print the second register of a two-register operand. Requires that it
3941 has been allocated consecutively to the first.
3943 .. FIXME: why is it restricted to consecutive ones? And there's
3944 nothing that ensures that happens, is there?
3946 - ``I``: Print the letter 'i' if the operand is an integer constant, otherwise
3947 nothing. Used to print 'addi' vs 'add' instructions.
3951 No additional modifiers.
3955 - ``X``: Print an immediate integer as hexadecimal
3956 - ``x``: Print the low 16 bits of an immediate integer as hexadecimal.
3957 - ``d``: Print an immediate integer as decimal.
3958 - ``m``: Subtract one and print an immediate integer as decimal.
3959 - ``z``: Print $0 if an immediate zero, otherwise print normally.
3960 - ``L``: Print the low-order register of a two-register operand, or prints the
3961 address of the low-order word of a double-word memory operand.
3963 .. FIXME: L seems to be missing memory operand support.
3965 - ``M``: Print the high-order register of a two-register operand, or prints the
3966 address of the high-order word of a double-word memory operand.
3968 .. FIXME: M seems to be missing memory operand support.
3970 - ``D``: Print the second register of a two-register operand, or prints the
3971 second word of a double-word memory operand. (On a big-endian system, ``D`` is
3972 equivalent to ``L``, and on little-endian system, ``D`` is equivalent to
3974 - ``w``: No effect. Provided for compatibility with GCC which requires this
3975 modifier in order to print MSA registers (``W0-W31``) with the ``f``
3984 - ``L``: Print the second register of a two-register operand. Requires that it
3985 has been allocated consecutively to the first.
3987 .. FIXME: why is it restricted to consecutive ones? And there's
3988 nothing that ensures that happens, is there?
3990 - ``I``: Print the letter 'i' if the operand is an integer constant, otherwise
3991 nothing. Used to print 'addi' vs 'add' instructions.
3992 - ``y``: For a memory operand, prints formatter for a two-register X-form
3993 instruction. (Currently always prints ``r0,OPERAND``).
3994 - ``U``: Prints 'u' if the memory operand is an update form, and nothing
3995 otherwise. (NOTE: LLVM does not support update form, so this will currently
3996 always print nothing)
3997 - ``X``: Prints 'x' if the memory operand is an indexed form. (NOTE: LLVM does
3998 not support indexed form, so this will currently always print nothing)
4006 SystemZ implements only ``n``, and does *not* support any of the other
4007 target-independent modifiers.
4011 - ``c``: Print an unadorned integer or symbol name. (The latter is
4012 target-specific behavior for this typically target-independent modifier).
4013 - ``A``: Print a register name with a '``*``' before it.
4014 - ``b``: Print an 8-bit register name (e.g. ``al``); do nothing on a memory
4016 - ``h``: Print the upper 8-bit register name (e.g. ``ah``); do nothing on a
4018 - ``w``: Print the 16-bit register name (e.g. ``ax``); do nothing on a memory
4020 - ``k``: Print the 32-bit register name (e.g. ``eax``); do nothing on a memory
4022 - ``q``: Print the 64-bit register name (e.g. ``rax``), if 64-bit registers are
4023 available, otherwise the 32-bit register name; do nothing on a memory operand.
4024 - ``n``: Negate and print an unadorned integer, or, for operands other than an
4025 immediate integer (e.g. a relocatable symbol expression), print a '-' before
4026 the operand. (The behavior for relocatable symbol expressions is a
4027 target-specific behavior for this typically target-independent modifier)
4028 - ``H``: Print a memory reference with additional offset +8.
4029 - ``P``: Print a memory reference or operand for use as the argument of a call
4030 instruction. (E.g. omit ``(rip)``, even though it's PC-relative.)
4034 No additional modifiers.
4040 The call instructions that wrap inline asm nodes may have a
4041 "``!srcloc``" MDNode attached to it that contains a list of constant
4042 integers. If present, the code generator will use the integer as the
4043 location cookie value when report errors through the ``LLVMContext``
4044 error reporting mechanisms. This allows a front-end to correlate backend
4045 errors that occur with inline asm back to the source code that produced
4048 .. code-block:: llvm
4050 call void asm sideeffect "something bad", ""(), !srcloc !42
4052 !42 = !{ i32 1234567 }
4054 It is up to the front-end to make sense of the magic numbers it places
4055 in the IR. If the MDNode contains multiple constants, the code generator
4056 will use the one that corresponds to the line of the asm that the error
4064 LLVM IR allows metadata to be attached to instructions in the program
4065 that can convey extra information about the code to the optimizers and
4066 code generator. One example application of metadata is source-level
4067 debug information. There are two metadata primitives: strings and nodes.
4069 Metadata does not have a type, and is not a value. If referenced from a
4070 ``call`` instruction, it uses the ``metadata`` type.
4072 All metadata are identified in syntax by a exclamation point ('``!``').
4074 .. _metadata-string:
4076 Metadata Nodes and Metadata Strings
4077 -----------------------------------
4079 A metadata string is a string surrounded by double quotes. It can
4080 contain any character by escaping non-printable characters with
4081 "``\xx``" where "``xx``" is the two digit hex code. For example:
4084 Metadata nodes are represented with notation similar to structure
4085 constants (a comma separated list of elements, surrounded by braces and
4086 preceded by an exclamation point). Metadata nodes can have any values as
4087 their operand. For example:
4089 .. code-block:: llvm
4091 !{ !"test\00", i32 10}
4093 Metadata nodes that aren't uniqued use the ``distinct`` keyword. For example:
4095 .. code-block:: text
4097 !0 = distinct !{!"test\00", i32 10}
4099 ``distinct`` nodes are useful when nodes shouldn't be merged based on their
4100 content. They can also occur when transformations cause uniquing collisions
4101 when metadata operands change.
4103 A :ref:`named metadata <namedmetadatastructure>` is a collection of
4104 metadata nodes, which can be looked up in the module symbol table. For
4107 .. code-block:: llvm
4111 Metadata can be used as function arguments. Here the ``llvm.dbg.value``
4112 intrinsic is using three metadata arguments:
4114 .. code-block:: llvm
4116 call void @llvm.dbg.value(metadata !24, metadata !25, metadata !26)
4118 Metadata can be attached to an instruction. Here metadata ``!21`` is attached
4119 to the ``add`` instruction using the ``!dbg`` identifier:
4121 .. code-block:: llvm
4123 %indvar.next = add i64 %indvar, 1, !dbg !21
4125 Metadata can also be attached to a function or a global variable. Here metadata
4126 ``!22`` is attached to the ``f1`` and ``f2 functions, and the globals ``g1``
4127 and ``g2`` using the ``!dbg`` identifier:
4129 .. code-block:: llvm
4131 declare !dbg !22 void @f1()
4132 define void @f2() !dbg !22 {
4136 @g1 = global i32 0, !dbg !22
4137 @g2 = external global i32, !dbg !22
4139 A transformation is required to drop any metadata attachment that it does not
4140 know or know it can't preserve. Currently there is an exception for metadata
4141 attachment to globals for ``!type`` and ``!absolute_symbol`` which can't be
4142 unconditionally dropped unless the global is itself deleted.
4144 Metadata attached to a module using named metadata may not be dropped, with
4145 the exception of debug metadata (named metadata with the name ``!llvm.dbg.*``).
4147 More information about specific metadata nodes recognized by the
4148 optimizers and code generator is found below.
4150 .. _specialized-metadata:
4152 Specialized Metadata Nodes
4153 ^^^^^^^^^^^^^^^^^^^^^^^^^^
4155 Specialized metadata nodes are custom data structures in metadata (as opposed
4156 to generic tuples). Their fields are labelled, and can be specified in any
4159 These aren't inherently debug info centric, but currently all the specialized
4160 metadata nodes are related to debug info.
4167 ``DICompileUnit`` nodes represent a compile unit. The ``enums:``,
4168 ``retainedTypes:``, ``globals:``, ``imports:`` and ``macros:`` fields are tuples
4169 containing the debug info to be emitted along with the compile unit, regardless
4170 of code optimizations (some nodes are only emitted if there are references to
4171 them from instructions). The ``debugInfoForProfiling:`` field is a boolean
4172 indicating whether or not line-table discriminators are updated to provide
4173 more-accurate debug info for profiling results.
4175 .. code-block:: text
4177 !0 = !DICompileUnit(language: DW_LANG_C99, file: !1, producer: "clang",
4178 isOptimized: true, flags: "-O2", runtimeVersion: 2,
4179 splitDebugFilename: "abc.debug", emissionKind: FullDebug,
4180 enums: !2, retainedTypes: !3, globals: !4, imports: !5,
4181 macros: !6, dwoId: 0x0abcd)
4183 Compile unit descriptors provide the root scope for objects declared in a
4184 specific compilation unit. File descriptors are defined using this scope. These
4185 descriptors are collected by a named metadata node ``!llvm.dbg.cu``. They keep
4186 track of global variables, type information, and imported entities (declarations
4194 ``DIFile`` nodes represent files. The ``filename:`` can include slashes.
4196 .. code-block:: none
4198 !0 = !DIFile(filename: "path/to/file", directory: "/path/to/dir",
4199 checksumkind: CSK_MD5,
4200 checksum: "000102030405060708090a0b0c0d0e0f")
4202 Files are sometimes used in ``scope:`` fields, and are the only valid target
4203 for ``file:`` fields.
4204 Valid values for ``checksumkind:`` field are: {CSK_None, CSK_MD5, CSK_SHA1}
4211 ``DIBasicType`` nodes represent primitive types, such as ``int``, ``bool`` and
4212 ``float``. ``tag:`` defaults to ``DW_TAG_base_type``.
4214 .. code-block:: text
4216 !0 = !DIBasicType(name: "unsigned char", size: 8, align: 8,
4217 encoding: DW_ATE_unsigned_char)
4218 !1 = !DIBasicType(tag: DW_TAG_unspecified_type, name: "decltype(nullptr)")
4220 The ``encoding:`` describes the details of the type. Usually it's one of the
4223 .. code-block:: text
4229 DW_ATE_signed_char = 6
4231 DW_ATE_unsigned_char = 8
4233 .. _DISubroutineType:
4238 ``DISubroutineType`` nodes represent subroutine types. Their ``types:`` field
4239 refers to a tuple; the first operand is the return type, while the rest are the
4240 types of the formal arguments in order. If the first operand is ``null``, that
4241 represents a function with no return value (such as ``void foo() {}`` in C++).
4243 .. code-block:: text
4245 !0 = !BasicType(name: "int", size: 32, align: 32, DW_ATE_signed)
4246 !1 = !BasicType(name: "char", size: 8, align: 8, DW_ATE_signed_char)
4247 !2 = !DISubroutineType(types: !{null, !0, !1}) ; void (int, char)
4254 ``DIDerivedType`` nodes represent types derived from other types, such as
4257 .. code-block:: text
4259 !0 = !DIBasicType(name: "unsigned char", size: 8, align: 8,
4260 encoding: DW_ATE_unsigned_char)
4261 !1 = !DIDerivedType(tag: DW_TAG_pointer_type, baseType: !0, size: 32,
4264 The following ``tag:`` values are valid:
4266 .. code-block:: text
4269 DW_TAG_pointer_type = 15
4270 DW_TAG_reference_type = 16
4272 DW_TAG_inheritance = 28
4273 DW_TAG_ptr_to_member_type = 31
4274 DW_TAG_const_type = 38
4276 DW_TAG_volatile_type = 53
4277 DW_TAG_restrict_type = 55
4278 DW_TAG_atomic_type = 71
4280 .. _DIDerivedTypeMember:
4282 ``DW_TAG_member`` is used to define a member of a :ref:`composite type
4283 <DICompositeType>`. The type of the member is the ``baseType:``. The
4284 ``offset:`` is the member's bit offset. If the composite type has an ODR
4285 ``identifier:`` and does not set ``flags: DIFwdDecl``, then the member is
4286 uniqued based only on its ``name:`` and ``scope:``.
4288 ``DW_TAG_inheritance`` and ``DW_TAG_friend`` are used in the ``elements:``
4289 field of :ref:`composite types <DICompositeType>` to describe parents and
4292 ``DW_TAG_typedef`` is used to provide a name for the ``baseType:``.
4294 ``DW_TAG_pointer_type``, ``DW_TAG_reference_type``, ``DW_TAG_const_type``,
4295 ``DW_TAG_volatile_type``, ``DW_TAG_restrict_type`` and ``DW_TAG_atomic_type``
4296 are used to qualify the ``baseType:``.
4298 Note that the ``void *`` type is expressed as a type derived from NULL.
4300 .. _DICompositeType:
4305 ``DICompositeType`` nodes represent types composed of other types, like
4306 structures and unions. ``elements:`` points to a tuple of the composed types.
4308 If the source language supports ODR, the ``identifier:`` field gives the unique
4309 identifier used for type merging between modules. When specified,
4310 :ref:`subprogram declarations <DISubprogramDeclaration>` and :ref:`member
4311 derived types <DIDerivedTypeMember>` that reference the ODR-type in their
4312 ``scope:`` change uniquing rules.
4314 For a given ``identifier:``, there should only be a single composite type that
4315 does not have ``flags: DIFlagFwdDecl`` set. LLVM tools that link modules
4316 together will unique such definitions at parse time via the ``identifier:``
4317 field, even if the nodes are ``distinct``.
4319 .. code-block:: text
4321 !0 = !DIEnumerator(name: "SixKind", value: 7)
4322 !1 = !DIEnumerator(name: "SevenKind", value: 7)
4323 !2 = !DIEnumerator(name: "NegEightKind", value: -8)
4324 !3 = !DICompositeType(tag: DW_TAG_enumeration_type, name: "Enum", file: !12,
4325 line: 2, size: 32, align: 32, identifier: "_M4Enum",
4326 elements: !{!0, !1, !2})
4328 The following ``tag:`` values are valid:
4330 .. code-block:: text
4332 DW_TAG_array_type = 1
4333 DW_TAG_class_type = 2
4334 DW_TAG_enumeration_type = 4
4335 DW_TAG_structure_type = 19
4336 DW_TAG_union_type = 23
4338 For ``DW_TAG_array_type``, the ``elements:`` should be :ref:`subrange
4339 descriptors <DISubrange>`, each representing the range of subscripts at that
4340 level of indexing. The ``DIFlagVector`` flag to ``flags:`` indicates that an
4341 array type is a native packed vector.
4343 For ``DW_TAG_enumeration_type``, the ``elements:`` should be :ref:`enumerator
4344 descriptors <DIEnumerator>`, each representing the definition of an enumeration
4345 value for the set. All enumeration type descriptors are collected in the
4346 ``enums:`` field of the :ref:`compile unit <DICompileUnit>`.
4348 For ``DW_TAG_structure_type``, ``DW_TAG_class_type``, and
4349 ``DW_TAG_union_type``, the ``elements:`` should be :ref:`derived types
4350 <DIDerivedType>` with ``tag: DW_TAG_member``, ``tag: DW_TAG_inheritance``, or
4351 ``tag: DW_TAG_friend``; or :ref:`subprograms <DISubprogram>` with
4352 ``isDefinition: false``.
4359 ``DISubrange`` nodes are the elements for ``DW_TAG_array_type`` variants of
4360 :ref:`DICompositeType`.
4362 - ``count: -1`` indicates an empty array.
4363 - ``count: !9`` describes the count with a :ref:`DILocalVariable`.
4364 - ``count: !11`` describes the count with a :ref:`DIGlobalVariable`.
4366 .. code-block:: llvm
4368 !0 = !DISubrange(count: 5, lowerBound: 0) ; array counting from 0
4369 !1 = !DISubrange(count: 5, lowerBound: 1) ; array counting from 1
4370 !2 = !DISubrange(count: -1) ; empty array.
4372 ; Scopes used in rest of example
4373 !6 = !DIFile(filename: "vla.c", directory: "/path/to/file")
4374 !7 = distinct !DICompileUnit(language: DW_LANG_C99, ...
4375 !8 = distinct !DISubprogram(name: "foo", scope: !7, file: !6, line: 5, ...
4377 ; Use of local variable as count value
4378 !9 = !DIBasicType(name: "int", size: 32, encoding: DW_ATE_signed)
4379 !10 = !DILocalVariable(name: "count", scope: !8, file: !6, line: 42, type: !9)
4380 !11 = !DISubrange(count !10, lowerBound: 0)
4382 ; Use of global variable as count value
4383 !12 = !DIGlobalVariable(name: "count", scope: !8, file: !6, line: 22, type: !9)
4384 !13 = !DISubrange(count !12, lowerBound: 0)
4391 ``DIEnumerator`` nodes are the elements for ``DW_TAG_enumeration_type``
4392 variants of :ref:`DICompositeType`.
4394 .. code-block:: llvm
4396 !0 = !DIEnumerator(name: "SixKind", value: 7)
4397 !1 = !DIEnumerator(name: "SevenKind", value: 7)
4398 !2 = !DIEnumerator(name: "NegEightKind", value: -8)
4400 DITemplateTypeParameter
4401 """""""""""""""""""""""
4403 ``DITemplateTypeParameter`` nodes represent type parameters to generic source
4404 language constructs. They are used (optionally) in :ref:`DICompositeType` and
4405 :ref:`DISubprogram` ``templateParams:`` fields.
4407 .. code-block:: llvm
4409 !0 = !DITemplateTypeParameter(name: "Ty", type: !1)
4411 DITemplateValueParameter
4412 """"""""""""""""""""""""
4414 ``DITemplateValueParameter`` nodes represent value parameters to generic source
4415 language constructs. ``tag:`` defaults to ``DW_TAG_template_value_parameter``,
4416 but if specified can also be set to ``DW_TAG_GNU_template_template_param`` or
4417 ``DW_TAG_GNU_template_param_pack``. They are used (optionally) in
4418 :ref:`DICompositeType` and :ref:`DISubprogram` ``templateParams:`` fields.
4420 .. code-block:: llvm
4422 !0 = !DITemplateValueParameter(name: "Ty", type: !1, value: i32 7)
4427 ``DINamespace`` nodes represent namespaces in the source language.
4429 .. code-block:: llvm
4431 !0 = !DINamespace(name: "myawesomeproject", scope: !1, file: !2, line: 7)
4433 .. _DIGlobalVariable:
4438 ``DIGlobalVariable`` nodes represent global variables in the source language.
4440 .. code-block:: llvm
4442 !0 = !DIGlobalVariable(name: "foo", linkageName: "foo", scope: !1,
4443 file: !2, line: 7, type: !3, isLocal: true,
4444 isDefinition: false, variable: i32* @foo,
4447 All global variables should be referenced by the `globals:` field of a
4448 :ref:`compile unit <DICompileUnit>`.
4455 ``DISubprogram`` nodes represent functions from the source language. A
4456 ``DISubprogram`` may be attached to a function definition using ``!dbg``
4457 metadata. The ``variables:`` field points at :ref:`variables <DILocalVariable>`
4458 that must be retained, even if their IR counterparts are optimized out of
4459 the IR. The ``type:`` field must point at an :ref:`DISubroutineType`.
4461 .. _DISubprogramDeclaration:
4463 When ``isDefinition: false``, subprograms describe a declaration in the type
4464 tree as opposed to a definition of a function. If the scope is a composite
4465 type with an ODR ``identifier:`` and that does not set ``flags: DIFwdDecl``,
4466 then the subprogram declaration is uniqued based only on its ``linkageName:``
4469 .. code-block:: text
4471 define void @_Z3foov() !dbg !0 {
4475 !0 = distinct !DISubprogram(name: "foo", linkageName: "_Zfoov", scope: !1,
4476 file: !2, line: 7, type: !3, isLocal: true,
4477 isDefinition: true, scopeLine: 8,
4479 virtuality: DW_VIRTUALITY_pure_virtual,
4480 virtualIndex: 10, flags: DIFlagPrototyped,
4481 isOptimized: true, unit: !5, templateParams: !6,
4482 declaration: !7, variables: !8, thrownTypes: !9)
4489 ``DILexicalBlock`` nodes describe nested blocks within a :ref:`subprogram
4490 <DISubprogram>`. The line number and column numbers are used to distinguish
4491 two lexical blocks at same depth. They are valid targets for ``scope:``
4494 .. code-block:: text
4496 !0 = distinct !DILexicalBlock(scope: !1, file: !2, line: 7, column: 35)
4498 Usually lexical blocks are ``distinct`` to prevent node merging based on
4501 .. _DILexicalBlockFile:
4506 ``DILexicalBlockFile`` nodes are used to discriminate between sections of a
4507 :ref:`lexical block <DILexicalBlock>`. The ``file:`` field can be changed to
4508 indicate textual inclusion, or the ``discriminator:`` field can be used to
4509 discriminate between control flow within a single block in the source language.
4511 .. code-block:: llvm
4513 !0 = !DILexicalBlock(scope: !3, file: !4, line: 7, column: 35)
4514 !1 = !DILexicalBlockFile(scope: !0, file: !4, discriminator: 0)
4515 !2 = !DILexicalBlockFile(scope: !0, file: !4, discriminator: 1)
4522 ``DILocation`` nodes represent source debug locations. The ``scope:`` field is
4523 mandatory, and points at an :ref:`DILexicalBlockFile`, an
4524 :ref:`DILexicalBlock`, or an :ref:`DISubprogram`.
4526 .. code-block:: llvm
4528 !0 = !DILocation(line: 2900, column: 42, scope: !1, inlinedAt: !2)
4530 .. _DILocalVariable:
4535 ``DILocalVariable`` nodes represent local variables in the source language. If
4536 the ``arg:`` field is set to non-zero, then this variable is a subprogram
4537 parameter, and it will be included in the ``variables:`` field of its
4538 :ref:`DISubprogram`.
4540 .. code-block:: text
4542 !0 = !DILocalVariable(name: "this", arg: 1, scope: !3, file: !2, line: 7,
4543 type: !3, flags: DIFlagArtificial)
4544 !1 = !DILocalVariable(name: "x", arg: 2, scope: !4, file: !2, line: 7,
4546 !2 = !DILocalVariable(name: "y", scope: !5, file: !2, line: 7, type: !3)
4551 ``DIExpression`` nodes represent expressions that are inspired by the DWARF
4552 expression language. They are used in :ref:`debug intrinsics<dbg_intrinsics>`
4553 (such as ``llvm.dbg.declare`` and ``llvm.dbg.value``) to describe how the
4554 referenced LLVM variable relates to the source language variable.
4556 The current supported vocabulary is limited:
4558 - ``DW_OP_deref`` dereferences the top of the expression stack.
4559 - ``DW_OP_plus`` pops the last two entries from the expression stack, adds
4560 them together and appends the result to the expression stack.
4561 - ``DW_OP_minus`` pops the last two entries from the expression stack, subtracts
4562 the last entry from the second last entry and appends the result to the
4564 - ``DW_OP_plus_uconst, 93`` adds ``93`` to the working expression.
4565 - ``DW_OP_LLVM_fragment, 16, 8`` specifies the offset and size (``16`` and ``8``
4566 here, respectively) of the variable fragment from the working expression. Note
4567 that contrary to DW_OP_bit_piece, the offset is describing the location
4568 within the described source variable.
4569 - ``DW_OP_swap`` swaps top two stack entries.
4570 - ``DW_OP_xderef`` provides extended dereference mechanism. The entry at the top
4571 of the stack is treated as an address. The second stack entry is treated as an
4572 address space identifier.
4573 - ``DW_OP_stack_value`` marks a constant value.
4575 DWARF specifies three kinds of simple location descriptions: Register, memory,
4576 and implicit location descriptions. Register and memory location descriptions
4577 describe the *location* of a source variable (in the sense that a debugger might
4578 modify its value), whereas implicit locations describe merely the *value* of a
4579 source variable. DIExpressions also follow this model: A DIExpression that
4580 doesn't have a trailing ``DW_OP_stack_value`` will describe an *address* when
4581 combined with a concrete location.
4583 .. code-block:: text
4585 !0 = !DIExpression(DW_OP_deref)
4586 !1 = !DIExpression(DW_OP_plus_uconst, 3)
4587 !1 = !DIExpression(DW_OP_constu, 3, DW_OP_plus)
4588 !2 = !DIExpression(DW_OP_bit_piece, 3, 7)
4589 !3 = !DIExpression(DW_OP_deref, DW_OP_constu, 3, DW_OP_plus, DW_OP_LLVM_fragment, 3, 7)
4590 !4 = !DIExpression(DW_OP_constu, 2, DW_OP_swap, DW_OP_xderef)
4591 !5 = !DIExpression(DW_OP_constu, 42, DW_OP_stack_value)
4596 ``DIObjCProperty`` nodes represent Objective-C property nodes.
4598 .. code-block:: llvm
4600 !3 = !DIObjCProperty(name: "foo", file: !1, line: 7, setter: "setFoo",
4601 getter: "getFoo", attributes: 7, type: !2)
4606 ``DIImportedEntity`` nodes represent entities (such as modules) imported into a
4609 .. code-block:: text
4611 !2 = !DIImportedEntity(tag: DW_TAG_imported_module, name: "foo", scope: !0,
4612 entity: !1, line: 7)
4617 ``DIMacro`` nodes represent definition or undefinition of a macro identifiers.
4618 The ``name:`` field is the macro identifier, followed by macro parameters when
4619 defining a function-like macro, and the ``value`` field is the token-string
4620 used to expand the macro identifier.
4622 .. code-block:: text
4624 !2 = !DIMacro(macinfo: DW_MACINFO_define, line: 7, name: "foo(x)",
4626 !3 = !DIMacro(macinfo: DW_MACINFO_undef, line: 30, name: "foo")
4631 ``DIMacroFile`` nodes represent inclusion of source files.
4632 The ``nodes:`` field is a list of ``DIMacro`` and ``DIMacroFile`` nodes that
4633 appear in the included source file.
4635 .. code-block:: text
4637 !2 = !DIMacroFile(macinfo: DW_MACINFO_start_file, line: 7, file: !2,
4643 In LLVM IR, memory does not have types, so LLVM's own type system is not
4644 suitable for doing type based alias analysis (TBAA). Instead, metadata is
4645 added to the IR to describe a type system of a higher level language. This
4646 can be used to implement C/C++ strict type aliasing rules, but it can also
4647 be used to implement custom alias analysis behavior for other languages.
4649 This description of LLVM's TBAA system is broken into two parts:
4650 :ref:`Semantics<tbaa_node_semantics>` talks about high level issues, and
4651 :ref:`Representation<tbaa_node_representation>` talks about the metadata
4652 encoding of various entities.
4654 It is always possible to trace any TBAA node to a "root" TBAA node (details
4655 in the :ref:`Representation<tbaa_node_representation>` section). TBAA
4656 nodes with different roots have an unknown aliasing relationship, and LLVM
4657 conservatively infers ``MayAlias`` between them. The rules mentioned in
4658 this section only pertain to TBAA nodes living under the same root.
4660 .. _tbaa_node_semantics:
4665 The TBAA metadata system, referred to as "struct path TBAA" (not to be
4666 confused with ``tbaa.struct``), consists of the following high level
4667 concepts: *Type Descriptors*, further subdivided into scalar type
4668 descriptors and struct type descriptors; and *Access Tags*.
4670 **Type descriptors** describe the type system of the higher level language
4671 being compiled. **Scalar type descriptors** describe types that do not
4672 contain other types. Each scalar type has a parent type, which must also
4673 be a scalar type or the TBAA root. Via this parent relation, scalar types
4674 within a TBAA root form a tree. **Struct type descriptors** denote types
4675 that contain a sequence of other type descriptors, at known offsets. These
4676 contained type descriptors can either be struct type descriptors themselves
4677 or scalar type descriptors.
4679 **Access tags** are metadata nodes attached to load and store instructions.
4680 Access tags use type descriptors to describe the *location* being accessed
4681 in terms of the type system of the higher level language. Access tags are
4682 tuples consisting of a base type, an access type and an offset. The base
4683 type is a scalar type descriptor or a struct type descriptor, the access
4684 type is a scalar type descriptor, and the offset is a constant integer.
4686 The access tag ``(BaseTy, AccessTy, Offset)`` can describe one of two
4689 * If ``BaseTy`` is a struct type, the tag describes a memory access (load
4690 or store) of a value of type ``AccessTy`` contained in the struct type
4691 ``BaseTy`` at offset ``Offset``.
4693 * If ``BaseTy`` is a scalar type, ``Offset`` must be 0 and ``BaseTy`` and
4694 ``AccessTy`` must be the same; and the access tag describes a scalar
4695 access with scalar type ``AccessTy``.
4697 We first define an ``ImmediateParent`` relation on ``(BaseTy, Offset)``
4700 * If ``BaseTy`` is a scalar type then ``ImmediateParent(BaseTy, 0)`` is
4701 ``(ParentTy, 0)`` where ``ParentTy`` is the parent of the scalar type as
4702 described in the TBAA metadata. ``ImmediateParent(BaseTy, Offset)`` is
4703 undefined if ``Offset`` is non-zero.
4705 * If ``BaseTy`` is a struct type then ``ImmediateParent(BaseTy, Offset)``
4706 is ``(NewTy, NewOffset)`` where ``NewTy`` is the type contained in
4707 ``BaseTy`` at offset ``Offset`` and ``NewOffset`` is ``Offset`` adjusted
4708 to be relative within that inner type.
4710 A memory access with an access tag ``(BaseTy1, AccessTy1, Offset1)``
4711 aliases a memory access with an access tag ``(BaseTy2, AccessTy2,
4712 Offset2)`` if either ``(BaseTy1, Offset1)`` is reachable from ``(Base2,
4713 Offset2)`` via the ``Parent`` relation or vice versa.
4715 As a concrete example, the type descriptor graph for the following program
4721 float f; // offset 4
4725 float f; // offset 0
4726 double d; // offset 4
4727 struct Inner inner_a; // offset 12
4730 void f(struct Outer* outer, struct Inner* inner, float* f, int* i, char* c) {
4731 outer->f = 0; // tag0: (OuterStructTy, FloatScalarTy, 0)
4732 outer->inner_a.i = 0; // tag1: (OuterStructTy, IntScalarTy, 12)
4733 outer->inner_a.f = 0.0; // tag2: (OuterStructTy, IntScalarTy, 16)
4734 *f = 0.0; // tag3: (FloatScalarTy, FloatScalarTy, 0)
4737 is (note that in C and C++, ``char`` can be used to access any arbitrary
4740 .. code-block:: text
4743 CharScalarTy = ("char", Root, 0)
4744 FloatScalarTy = ("float", CharScalarTy, 0)
4745 DoubleScalarTy = ("double", CharScalarTy, 0)
4746 IntScalarTy = ("int", CharScalarTy, 0)
4747 InnerStructTy = {"Inner" (IntScalarTy, 0), (FloatScalarTy, 4)}
4748 OuterStructTy = {"Outer", (FloatScalarTy, 0), (DoubleScalarTy, 4),
4749 (InnerStructTy, 12)}
4752 with (e.g.) ``ImmediateParent(OuterStructTy, 12)`` = ``(InnerStructTy,
4753 0)``, ``ImmediateParent(InnerStructTy, 0)`` = ``(IntScalarTy, 0)``, and
4754 ``ImmediateParent(IntScalarTy, 0)`` = ``(CharScalarTy, 0)``.
4756 .. _tbaa_node_representation:
4761 The root node of a TBAA type hierarchy is an ``MDNode`` with 0 operands or
4762 with exactly one ``MDString`` operand.
4764 Scalar type descriptors are represented as an ``MDNode`` s with two
4765 operands. The first operand is an ``MDString`` denoting the name of the
4766 struct type. LLVM does not assign meaning to the value of this operand, it
4767 only cares about it being an ``MDString``. The second operand is an
4768 ``MDNode`` which points to the parent for said scalar type descriptor,
4769 which is either another scalar type descriptor or the TBAA root. Scalar
4770 type descriptors can have an optional third argument, but that must be the
4771 constant integer zero.
4773 Struct type descriptors are represented as ``MDNode`` s with an odd number
4774 of operands greater than 1. The first operand is an ``MDString`` denoting
4775 the name of the struct type. Like in scalar type descriptors the actual
4776 value of this name operand is irrelevant to LLVM. After the name operand,
4777 the struct type descriptors have a sequence of alternating ``MDNode`` and
4778 ``ConstantInt`` operands. With N starting from 1, the 2N - 1 th operand,
4779 an ``MDNode``, denotes a contained field, and the 2N th operand, a
4780 ``ConstantInt``, is the offset of the said contained field. The offsets
4781 must be in non-decreasing order.
4783 Access tags are represented as ``MDNode`` s with either 3 or 4 operands.
4784 The first operand is an ``MDNode`` pointing to the node representing the
4785 base type. The second operand is an ``MDNode`` pointing to the node
4786 representing the access type. The third operand is a ``ConstantInt`` that
4787 states the offset of the access. If a fourth field is present, it must be
4788 a ``ConstantInt`` valued at 0 or 1. If it is 1 then the access tag states
4789 that the location being accessed is "constant" (meaning
4790 ``pointsToConstantMemory`` should return true; see `other useful
4791 AliasAnalysis methods <AliasAnalysis.html#OtherItfs>`_). The TBAA root of
4792 the access type and the base type of an access tag must be the same, and
4793 that is the TBAA root of the access tag.
4795 '``tbaa.struct``' Metadata
4796 ^^^^^^^^^^^^^^^^^^^^^^^^^^
4798 The :ref:`llvm.memcpy <int_memcpy>` is often used to implement
4799 aggregate assignment operations in C and similar languages, however it
4800 is defined to copy a contiguous region of memory, which is more than
4801 strictly necessary for aggregate types which contain holes due to
4802 padding. Also, it doesn't contain any TBAA information about the fields
4805 ``!tbaa.struct`` metadata can describe which memory subregions in a
4806 memcpy are padding and what the TBAA tags of the struct are.
4808 The current metadata format is very simple. ``!tbaa.struct`` metadata
4809 nodes are a list of operands which are in conceptual groups of three.
4810 For each group of three, the first operand gives the byte offset of a
4811 field in bytes, the second gives its size in bytes, and the third gives
4814 .. code-block:: llvm
4816 !4 = !{ i64 0, i64 4, !1, i64 8, i64 4, !2 }
4818 This describes a struct with two fields. The first is at offset 0 bytes
4819 with size 4 bytes, and has tbaa tag !1. The second is at offset 8 bytes
4820 and has size 4 bytes and has tbaa tag !2.
4822 Note that the fields need not be contiguous. In this example, there is a
4823 4 byte gap between the two fields. This gap represents padding which
4824 does not carry useful data and need not be preserved.
4826 '``noalias``' and '``alias.scope``' Metadata
4827 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4829 ``noalias`` and ``alias.scope`` metadata provide the ability to specify generic
4830 noalias memory-access sets. This means that some collection of memory access
4831 instructions (loads, stores, memory-accessing calls, etc.) that carry
4832 ``noalias`` metadata can specifically be specified not to alias with some other
4833 collection of memory access instructions that carry ``alias.scope`` metadata.
4834 Each type of metadata specifies a list of scopes where each scope has an id and
4837 When evaluating an aliasing query, if for some domain, the set
4838 of scopes with that domain in one instruction's ``alias.scope`` list is a
4839 subset of (or equal to) the set of scopes for that domain in another
4840 instruction's ``noalias`` list, then the two memory accesses are assumed not to
4843 Because scopes in one domain don't affect scopes in other domains, separate
4844 domains can be used to compose multiple independent noalias sets. This is
4845 used for example during inlining. As the noalias function parameters are
4846 turned into noalias scope metadata, a new domain is used every time the
4847 function is inlined.
4849 The metadata identifying each domain is itself a list containing one or two
4850 entries. The first entry is the name of the domain. Note that if the name is a
4851 string then it can be combined across functions and translation units. A
4852 self-reference can be used to create globally unique domain names. A
4853 descriptive string may optionally be provided as a second list entry.
4855 The metadata identifying each scope is also itself a list containing two or
4856 three entries. The first entry is the name of the scope. Note that if the name
4857 is a string then it can be combined across functions and translation units. A
4858 self-reference can be used to create globally unique scope names. A metadata
4859 reference to the scope's domain is the second entry. A descriptive string may
4860 optionally be provided as a third list entry.
4864 .. code-block:: llvm
4866 ; Two scope domains:
4870 ; Some scopes in these domains:
4876 !5 = !{!4} ; A list containing only scope !4
4880 ; These two instructions don't alias:
4881 %0 = load float, float* %c, align 4, !alias.scope !5
4882 store float %0, float* %arrayidx.i, align 4, !noalias !5
4884 ; These two instructions also don't alias (for domain !1, the set of scopes
4885 ; in the !alias.scope equals that in the !noalias list):
4886 %2 = load float, float* %c, align 4, !alias.scope !5
4887 store float %2, float* %arrayidx.i2, align 4, !noalias !6
4889 ; These two instructions may alias (for domain !0, the set of scopes in
4890 ; the !noalias list is not a superset of, or equal to, the scopes in the
4891 ; !alias.scope list):
4892 %2 = load float, float* %c, align 4, !alias.scope !6
4893 store float %0, float* %arrayidx.i, align 4, !noalias !7
4895 '``fpmath``' Metadata
4896 ^^^^^^^^^^^^^^^^^^^^^
4898 ``fpmath`` metadata may be attached to any instruction of floating point
4899 type. It can be used to express the maximum acceptable error in the
4900 result of that instruction, in ULPs, thus potentially allowing the
4901 compiler to use a more efficient but less accurate method of computing
4902 it. ULP is defined as follows:
4904 If ``x`` is a real number that lies between two finite consecutive
4905 floating-point numbers ``a`` and ``b``, without being equal to one
4906 of them, then ``ulp(x) = |b - a|``, otherwise ``ulp(x)`` is the
4907 distance between the two non-equal finite floating-point numbers
4908 nearest ``x``. Moreover, ``ulp(NaN)`` is ``NaN``.
4910 The metadata node shall consist of a single positive float type number
4911 representing the maximum relative error, for example:
4913 .. code-block:: llvm
4915 !0 = !{ float 2.5 } ; maximum acceptable inaccuracy is 2.5 ULPs
4919 '``range``' Metadata
4920 ^^^^^^^^^^^^^^^^^^^^
4922 ``range`` metadata may be attached only to ``load``, ``call`` and ``invoke`` of
4923 integer types. It expresses the possible ranges the loaded value or the value
4924 returned by the called function at this call site is in. The ranges are
4925 represented with a flattened list of integers. The loaded value or the value
4926 returned is known to be in the union of the ranges defined by each consecutive
4927 pair. Each pair has the following properties:
4929 - The type must match the type loaded by the instruction.
4930 - The pair ``a,b`` represents the range ``[a,b)``.
4931 - Both ``a`` and ``b`` are constants.
4932 - The range is allowed to wrap.
4933 - The range should not represent the full or empty set. That is,
4936 In addition, the pairs must be in signed order of the lower bound and
4937 they must be non-contiguous.
4941 .. code-block:: llvm
4943 %a = load i8, i8* %x, align 1, !range !0 ; Can only be 0 or 1
4944 %b = load i8, i8* %y, align 1, !range !1 ; Can only be 255 (-1), 0 or 1
4945 %c = call i8 @foo(), !range !2 ; Can only be 0, 1, 3, 4 or 5
4946 %d = invoke i8 @bar() to label %cont
4947 unwind label %lpad, !range !3 ; Can only be -2, -1, 3, 4 or 5
4949 !0 = !{ i8 0, i8 2 }
4950 !1 = !{ i8 255, i8 2 }
4951 !2 = !{ i8 0, i8 2, i8 3, i8 6 }
4952 !3 = !{ i8 -2, i8 0, i8 3, i8 6 }
4954 '``absolute_symbol``' Metadata
4955 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4957 ``absolute_symbol`` metadata may be attached to a global variable
4958 declaration. It marks the declaration as a reference to an absolute symbol,
4959 which causes the backend to use absolute relocations for the symbol even
4960 in position independent code, and expresses the possible ranges that the
4961 global variable's *address* (not its value) is in, in the same format as
4962 ``range`` metadata, with the extension that the pair ``all-ones,all-ones``
4963 may be used to represent the full set.
4965 Example (assuming 64-bit pointers):
4967 .. code-block:: llvm
4969 @a = external global i8, !absolute_symbol !0 ; Absolute symbol in range [0,256)
4970 @b = external global i8, !absolute_symbol !1 ; Absolute symbol in range [0,2^64)
4973 !0 = !{ i64 0, i64 256 }
4974 !1 = !{ i64 -1, i64 -1 }
4976 '``callees``' Metadata
4977 ^^^^^^^^^^^^^^^^^^^^^^
4979 ``callees`` metadata may be attached to indirect call sites. If ``callees``
4980 metadata is attached to a call site, and any callee is not among the set of
4981 functions provided by the metadata, the behavior is undefined. The intent of
4982 this metadata is to facilitate optimizations such as indirect-call promotion.
4983 For example, in the code below, the call instruction may only target the
4984 ``add`` or ``sub`` functions:
4986 .. code-block:: llvm
4988 %result = call i64 %binop(i64 %x, i64 %y), !callees !0
4991 !0 = !{i64 (i64, i64)* @add, i64 (i64, i64)* @sub}
4993 '``unpredictable``' Metadata
4994 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4996 ``unpredictable`` metadata may be attached to any branch or switch
4997 instruction. It can be used to express the unpredictability of control
4998 flow. Similar to the llvm.expect intrinsic, it may be used to alter
4999 optimizations related to compare and branch instructions. The metadata
5000 is treated as a boolean value; if it exists, it signals that the branch
5001 or switch that it is attached to is completely unpredictable.
5006 It is sometimes useful to attach information to loop constructs. Currently,
5007 loop metadata is implemented as metadata attached to the branch instruction
5008 in the loop latch block. This type of metadata refer to a metadata node that is
5009 guaranteed to be separate for each loop. The loop identifier metadata is
5010 specified with the name ``llvm.loop``.
5012 The loop identifier metadata is implemented using a metadata that refers to
5013 itself to avoid merging it with any other identifier metadata, e.g.,
5014 during module linkage or function inlining. That is, each loop should refer
5015 to their own identification metadata even if they reside in separate functions.
5016 The following example contains loop identifier metadata for two separate loop
5019 .. code-block:: llvm
5024 The loop identifier metadata can be used to specify additional
5025 per-loop metadata. Any operands after the first operand can be treated
5026 as user-defined metadata. For example the ``llvm.loop.unroll.count``
5027 suggests an unroll factor to the loop unroller:
5029 .. code-block:: llvm
5031 br i1 %exitcond, label %._crit_edge, label %.lr.ph, !llvm.loop !0
5034 !1 = !{!"llvm.loop.unroll.count", i32 4}
5036 '``llvm.loop.vectorize``' and '``llvm.loop.interleave``'
5037 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5039 Metadata prefixed with ``llvm.loop.vectorize`` or ``llvm.loop.interleave`` are
5040 used to control per-loop vectorization and interleaving parameters such as
5041 vectorization width and interleave count. These metadata should be used in
5042 conjunction with ``llvm.loop`` loop identification metadata. The
5043 ``llvm.loop.vectorize`` and ``llvm.loop.interleave`` metadata are only
5044 optimization hints and the optimizer will only interleave and vectorize loops if
5045 it believes it is safe to do so. The ``llvm.mem.parallel_loop_access`` metadata
5046 which contains information about loop-carried memory dependencies can be helpful
5047 in determining the safety of these transformations.
5049 '``llvm.loop.interleave.count``' Metadata
5050 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5052 This metadata suggests an interleave count to the loop interleaver.
5053 The first operand is the string ``llvm.loop.interleave.count`` and the
5054 second operand is an integer specifying the interleave count. For
5057 .. code-block:: llvm
5059 !0 = !{!"llvm.loop.interleave.count", i32 4}
5061 Note that setting ``llvm.loop.interleave.count`` to 1 disables interleaving
5062 multiple iterations of the loop. If ``llvm.loop.interleave.count`` is set to 0
5063 then the interleave count will be determined automatically.
5065 '``llvm.loop.vectorize.enable``' Metadata
5066 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5068 This metadata selectively enables or disables vectorization for the loop. The
5069 first operand is the string ``llvm.loop.vectorize.enable`` and the second operand
5070 is a bit. If the bit operand value is 1 vectorization is enabled. A value of
5071 0 disables vectorization:
5073 .. code-block:: llvm
5075 !0 = !{!"llvm.loop.vectorize.enable", i1 0}
5076 !1 = !{!"llvm.loop.vectorize.enable", i1 1}
5078 '``llvm.loop.vectorize.width``' Metadata
5079 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5081 This metadata sets the target width of the vectorizer. The first
5082 operand is the string ``llvm.loop.vectorize.width`` and the second
5083 operand is an integer specifying the width. For example:
5085 .. code-block:: llvm
5087 !0 = !{!"llvm.loop.vectorize.width", i32 4}
5089 Note that setting ``llvm.loop.vectorize.width`` to 1 disables
5090 vectorization of the loop. If ``llvm.loop.vectorize.width`` is set to
5091 0 or if the loop does not have this metadata the width will be
5092 determined automatically.
5094 '``llvm.loop.unroll``'
5095 ^^^^^^^^^^^^^^^^^^^^^^
5097 Metadata prefixed with ``llvm.loop.unroll`` are loop unrolling
5098 optimization hints such as the unroll factor. ``llvm.loop.unroll``
5099 metadata should be used in conjunction with ``llvm.loop`` loop
5100 identification metadata. The ``llvm.loop.unroll`` metadata are only
5101 optimization hints and the unrolling will only be performed if the
5102 optimizer believes it is safe to do so.
5104 '``llvm.loop.unroll.count``' Metadata
5105 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5107 This metadata suggests an unroll factor to the loop unroller. The
5108 first operand is the string ``llvm.loop.unroll.count`` and the second
5109 operand is a positive integer specifying the unroll factor. For
5112 .. code-block:: llvm
5114 !0 = !{!"llvm.loop.unroll.count", i32 4}
5116 If the trip count of the loop is less than the unroll count the loop
5117 will be partially unrolled.
5119 '``llvm.loop.unroll.disable``' Metadata
5120 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5122 This metadata disables loop unrolling. The metadata has a single operand
5123 which is the string ``llvm.loop.unroll.disable``. For example:
5125 .. code-block:: llvm
5127 !0 = !{!"llvm.loop.unroll.disable"}
5129 '``llvm.loop.unroll.runtime.disable``' Metadata
5130 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5132 This metadata disables runtime loop unrolling. The metadata has a single
5133 operand which is the string ``llvm.loop.unroll.runtime.disable``. For example:
5135 .. code-block:: llvm
5137 !0 = !{!"llvm.loop.unroll.runtime.disable"}
5139 '``llvm.loop.unroll.enable``' Metadata
5140 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5142 This metadata suggests that the loop should be fully unrolled if the trip count
5143 is known at compile time and partially unrolled if the trip count is not known
5144 at compile time. The metadata has a single operand which is the string
5145 ``llvm.loop.unroll.enable``. For example:
5147 .. code-block:: llvm
5149 !0 = !{!"llvm.loop.unroll.enable"}
5151 '``llvm.loop.unroll.full``' Metadata
5152 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5154 This metadata suggests that the loop should be unrolled fully. The
5155 metadata has a single operand which is the string ``llvm.loop.unroll.full``.
5158 .. code-block:: llvm
5160 !0 = !{!"llvm.loop.unroll.full"}
5162 '``llvm.loop.licm_versioning.disable``' Metadata
5163 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5165 This metadata indicates that the loop should not be versioned for the purpose
5166 of enabling loop-invariant code motion (LICM). The metadata has a single operand
5167 which is the string ``llvm.loop.licm_versioning.disable``. For example:
5169 .. code-block:: llvm
5171 !0 = !{!"llvm.loop.licm_versioning.disable"}
5173 '``llvm.loop.distribute.enable``' Metadata
5174 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5176 Loop distribution allows splitting a loop into multiple loops. Currently,
5177 this is only performed if the entire loop cannot be vectorized due to unsafe
5178 memory dependencies. The transformation will attempt to isolate the unsafe
5179 dependencies into their own loop.
5181 This metadata can be used to selectively enable or disable distribution of the
5182 loop. The first operand is the string ``llvm.loop.distribute.enable`` and the
5183 second operand is a bit. If the bit operand value is 1 distribution is
5184 enabled. A value of 0 disables distribution:
5186 .. code-block:: llvm
5188 !0 = !{!"llvm.loop.distribute.enable", i1 0}
5189 !1 = !{!"llvm.loop.distribute.enable", i1 1}
5191 This metadata should be used in conjunction with ``llvm.loop`` loop
5192 identification metadata.
5197 Metadata types used to annotate memory accesses with information helpful
5198 for optimizations are prefixed with ``llvm.mem``.
5200 '``llvm.mem.parallel_loop_access``' Metadata
5201 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5203 The ``llvm.mem.parallel_loop_access`` metadata refers to a loop identifier,
5204 or metadata containing a list of loop identifiers for nested loops.
5205 The metadata is attached to memory accessing instructions and denotes that
5206 no loop carried memory dependence exist between it and other instructions denoted
5207 with the same loop identifier. The metadata on memory reads also implies that
5208 if conversion (i.e. speculative execution within a loop iteration) is safe.
5210 Precisely, given two instructions ``m1`` and ``m2`` that both have the
5211 ``llvm.mem.parallel_loop_access`` metadata, with ``L1`` and ``L2`` being the
5212 set of loops associated with that metadata, respectively, then there is no loop
5213 carried dependence between ``m1`` and ``m2`` for loops in both ``L1`` and
5216 As a special case, if all memory accessing instructions in a loop have
5217 ``llvm.mem.parallel_loop_access`` metadata that refers to that loop, then the
5218 loop has no loop carried memory dependences and is considered to be a parallel
5221 Note that if not all memory access instructions have such metadata referring to
5222 the loop, then the loop is considered not being trivially parallel. Additional
5223 memory dependence analysis is required to make that determination. As a fail
5224 safe mechanism, this causes loops that were originally parallel to be considered
5225 sequential (if optimization passes that are unaware of the parallel semantics
5226 insert new memory instructions into the loop body).
5228 Example of a loop that is considered parallel due to its correct use of
5229 both ``llvm.loop`` and ``llvm.mem.parallel_loop_access``
5230 metadata types that refer to the same loop identifier metadata.
5232 .. code-block:: llvm
5236 %val0 = load i32, i32* %arrayidx, !llvm.mem.parallel_loop_access !0
5238 store i32 %val0, i32* %arrayidx1, !llvm.mem.parallel_loop_access !0
5240 br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !0
5246 It is also possible to have nested parallel loops. In that case the
5247 memory accesses refer to a list of loop identifier metadata nodes instead of
5248 the loop identifier metadata node directly:
5250 .. code-block:: llvm
5254 %val1 = load i32, i32* %arrayidx3, !llvm.mem.parallel_loop_access !2
5256 br label %inner.for.body
5260 %val0 = load i32, i32* %arrayidx1, !llvm.mem.parallel_loop_access !0
5262 store i32 %val0, i32* %arrayidx2, !llvm.mem.parallel_loop_access !0
5264 br i1 %exitcond, label %inner.for.end, label %inner.for.body, !llvm.loop !1
5268 store i32 %val1, i32* %arrayidx4, !llvm.mem.parallel_loop_access !2
5270 br i1 %exitcond, label %outer.for.end, label %outer.for.body, !llvm.loop !2
5272 outer.for.end: ; preds = %for.body
5274 !0 = !{!1, !2} ; a list of loop identifiers
5275 !1 = !{!1} ; an identifier for the inner loop
5276 !2 = !{!2} ; an identifier for the outer loop
5278 '``irr_loop``' Metadata
5279 ^^^^^^^^^^^^^^^^^^^^^^^
5281 ``irr_loop`` metadata may be attached to the terminator instruction of a basic
5282 block that's an irreducible loop header (note that an irreducible loop has more
5283 than once header basic blocks.) If ``irr_loop`` metadata is attached to the
5284 terminator instruction of a basic block that is not really an irreducible loop
5285 header, the behavior is undefined. The intent of this metadata is to improve the
5286 accuracy of the block frequency propagation. For example, in the code below, the
5287 block ``header0`` may have a loop header weight (relative to the other headers of
5288 the irreducible loop) of 100:
5290 .. code-block:: llvm
5294 br i1 %cmp, label %t1, label %t2, !irr_loop !0
5297 !0 = !{"loop_header_weight", i64 100}
5299 Irreducible loop header weights are typically based on profile data.
5301 '``invariant.group``' Metadata
5302 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5304 The ``invariant.group`` metadata may be attached to ``load``/``store`` instructions.
5305 The existence of the ``invariant.group`` metadata on the instruction tells
5306 the optimizer that every ``load`` and ``store`` to the same pointer operand
5307 within the same invariant group can be assumed to load or store the same
5308 value (but see the ``llvm.invariant.group.barrier`` intrinsic which affects
5309 when two pointers are considered the same). Pointers returned by bitcast or
5310 getelementptr with only zero indices are considered the same.
5314 .. code-block:: llvm
5316 @unknownPtr = external global i8
5319 store i8 42, i8* %ptr, !invariant.group !0
5320 call void @foo(i8* %ptr)
5322 %a = load i8, i8* %ptr, !invariant.group !0 ; Can assume that value under %ptr didn't change
5323 call void @foo(i8* %ptr)
5324 %b = load i8, i8* %ptr, !invariant.group !1 ; Can't assume anything, because group changed
5326 %newPtr = call i8* @getPointer(i8* %ptr)
5327 %c = load i8, i8* %newPtr, !invariant.group !0 ; Can't assume anything, because we only have information about %ptr
5329 %unknownValue = load i8, i8* @unknownPtr
5330 store i8 %unknownValue, i8* %ptr, !invariant.group !0 ; Can assume that %unknownValue == 42
5332 call void @foo(i8* %ptr)
5333 %newPtr2 = call i8* @llvm.invariant.group.barrier(i8* %ptr)
5334 %d = load i8, i8* %newPtr2, !invariant.group !0 ; Can't step through invariant.group.barrier to get value of %ptr
5337 declare void @foo(i8*)
5338 declare i8* @getPointer(i8*)
5339 declare i8* @llvm.invariant.group.barrier(i8*)
5341 !0 = !{!"magic ptr"}
5342 !1 = !{!"other ptr"}
5344 The invariant.group metadata must be dropped when replacing one pointer by
5345 another based on aliasing information. This is because invariant.group is tied
5346 to the SSA value of the pointer operand.
5348 .. code-block:: llvm
5350 %v = load i8, i8* %x, !invariant.group !0
5351 ; if %x mustalias %y then we can replace the above instruction with
5352 %v = load i8, i8* %y
5358 See :doc:`TypeMetadata`.
5360 '``associated``' Metadata
5361 ^^^^^^^^^^^^^^^^^^^^^^^^^
5363 The ``associated`` metadata may be attached to a global object
5364 declaration with a single argument that references another global object.
5366 This metadata prevents discarding of the global object in linker GC
5367 unless the referenced object is also discarded. The linker support for
5368 this feature is spotty. For best compatibility, globals carrying this
5371 - Be in a comdat with the referenced global.
5372 - Be in @llvm.compiler.used.
5373 - Have an explicit section with a name which is a valid C identifier.
5375 It does not have any effect on non-ELF targets.
5379 .. code-block:: text
5382 @a = global i32 1, comdat $a
5383 @b = internal global i32 2, comdat $a, section "abc", !associated !0
5390 The ``prof`` metadata is used to record profile data in the IR.
5391 The first operand of the metadata node indicates the profile metadata
5392 type. There are currently 3 types:
5393 :ref:`branch_weights<prof_node_branch_weights>`,
5394 :ref:`function_entry_count<prof_node_function_entry_count>`, and
5395 :ref:`VP<prof_node_VP>`.
5397 .. _prof_node_branch_weights:
5402 Branch weight metadata attached to a branch, select, switch or call instruction
5403 represents the likeliness of the associated branch being taken.
5404 For more information, see :doc:`BranchWeightMetadata`.
5406 .. _prof_node_function_entry_count:
5408 function_entry_count
5409 """"""""""""""""""""
5411 Function entry count metadata can be attached to function definitions
5412 to record the number of times the function is called. Used with BFI
5413 information, it is also used to derive the basic block profile count.
5414 For more information, see :doc:`BranchWeightMetadata`.
5421 VP (value profile) metadata can be attached to instructions that have
5422 value profile information. Currently this is indirect calls (where it
5423 records the hottest callees) and calls to memory intrinsics such as memcpy,
5424 memmove, and memset (where it records the hottest byte lengths).
5426 Each VP metadata node contains "VP" string, then a uint32_t value for the value
5427 profiling kind, a uint64_t value for the total number of times the instruction
5428 is executed, followed by uint64_t value and execution count pairs.
5429 The value profiling kind is 0 for indirect call targets and 1 for memory
5430 operations. For indirect call targets, each profile value is a hash
5431 of the callee function name, and for memory operations each value is the
5434 Note that the value counts do not need to add up to the total count
5435 listed in the third operand (in practice only the top hottest values
5436 are tracked and reported).
5438 Indirect call example:
5440 .. code-block:: llvm
5442 call void %f(), !prof !1
5443 !1 = !{!"VP", i32 0, i64 1600, i64 7651369219802541373, i64 1030, i64 -4377547752858689819, i64 410}
5445 Note that the VP type is 0 (the second operand), which indicates this is
5446 an indirect call value profile data. The third operand indicates that the
5447 indirect call executed 1600 times. The 4th and 6th operands give the
5448 hashes of the 2 hottest target functions' names (this is the same hash used
5449 to represent function names in the profile database), and the 5th and 7th
5450 operands give the execution count that each of the respective prior target
5451 functions was called.
5453 Module Flags Metadata
5454 =====================
5456 Information about the module as a whole is difficult to convey to LLVM's
5457 subsystems. The LLVM IR isn't sufficient to transmit this information.
5458 The ``llvm.module.flags`` named metadata exists in order to facilitate
5459 this. These flags are in the form of key / value pairs --- much like a
5460 dictionary --- making it easy for any subsystem who cares about a flag to
5463 The ``llvm.module.flags`` metadata contains a list of metadata triplets.
5464 Each triplet has the following form:
5466 - The first element is a *behavior* flag, which specifies the behavior
5467 when two (or more) modules are merged together, and it encounters two
5468 (or more) metadata with the same ID. The supported behaviors are
5470 - The second element is a metadata string that is a unique ID for the
5471 metadata. Each module may only have one flag entry for each unique ID (not
5472 including entries with the **Require** behavior).
5473 - The third element is the value of the flag.
5475 When two (or more) modules are merged together, the resulting
5476 ``llvm.module.flags`` metadata is the union of the modules' flags. That is, for
5477 each unique metadata ID string, there will be exactly one entry in the merged
5478 modules ``llvm.module.flags`` metadata table, and the value for that entry will
5479 be determined by the merge behavior flag, as described below. The only exception
5480 is that entries with the *Require* behavior are always preserved.
5482 The following behaviors are supported:
5493 Emits an error if two values disagree, otherwise the resulting value
5494 is that of the operands.
5498 Emits a warning if two values disagree. The result value will be the
5499 operand for the flag from the first module being linked.
5503 Adds a requirement that another module flag be present and have a
5504 specified value after linking is performed. The value must be a
5505 metadata pair, where the first element of the pair is the ID of the
5506 module flag to be restricted, and the second element of the pair is
5507 the value the module flag should be restricted to. This behavior can
5508 be used to restrict the allowable results (via triggering of an
5509 error) of linking IDs with the **Override** behavior.
5513 Uses the specified value, regardless of the behavior or value of the
5514 other module. If both modules specify **Override**, but the values
5515 differ, an error will be emitted.
5519 Appends the two values, which are required to be metadata nodes.
5523 Appends the two values, which are required to be metadata
5524 nodes. However, duplicate entries in the second list are dropped
5525 during the append operation.
5529 Takes the max of the two values, which are required to be integers.
5531 It is an error for a particular unique flag ID to have multiple behaviors,
5532 except in the case of **Require** (which adds restrictions on another metadata
5533 value) or **Override**.
5535 An example of module flags:
5537 .. code-block:: llvm
5539 !0 = !{ i32 1, !"foo", i32 1 }
5540 !1 = !{ i32 4, !"bar", i32 37 }
5541 !2 = !{ i32 2, !"qux", i32 42 }
5542 !3 = !{ i32 3, !"qux",
5547 !llvm.module.flags = !{ !0, !1, !2, !3 }
5549 - Metadata ``!0`` has the ID ``!"foo"`` and the value '1'. The behavior
5550 if two or more ``!"foo"`` flags are seen is to emit an error if their
5551 values are not equal.
5553 - Metadata ``!1`` has the ID ``!"bar"`` and the value '37'. The
5554 behavior if two or more ``!"bar"`` flags are seen is to use the value
5557 - Metadata ``!2`` has the ID ``!"qux"`` and the value '42'. The
5558 behavior if two or more ``!"qux"`` flags are seen is to emit a
5559 warning if their values are not equal.
5561 - Metadata ``!3`` has the ID ``!"qux"`` and the value:
5567 The behavior is to emit an error if the ``llvm.module.flags`` does not
5568 contain a flag with the ID ``!"foo"`` that has the value '1' after linking is
5571 Objective-C Garbage Collection Module Flags Metadata
5572 ----------------------------------------------------
5574 On the Mach-O platform, Objective-C stores metadata about garbage
5575 collection in a special section called "image info". The metadata
5576 consists of a version number and a bitmask specifying what types of
5577 garbage collection are supported (if any) by the file. If two or more
5578 modules are linked together their garbage collection metadata needs to
5579 be merged rather than appended together.
5581 The Objective-C garbage collection module flags metadata consists of the
5582 following key-value pairs:
5591 * - ``Objective-C Version``
5592 - **[Required]** --- The Objective-C ABI version. Valid values are 1 and 2.
5594 * - ``Objective-C Image Info Version``
5595 - **[Required]** --- The version of the image info section. Currently
5598 * - ``Objective-C Image Info Section``
5599 - **[Required]** --- The section to place the metadata. Valid values are
5600 ``"__OBJC, __image_info, regular"`` for Objective-C ABI version 1, and
5601 ``"__DATA,__objc_imageinfo, regular, no_dead_strip"`` for
5602 Objective-C ABI version 2.
5604 * - ``Objective-C Garbage Collection``
5605 - **[Required]** --- Specifies whether garbage collection is supported or
5606 not. Valid values are 0, for no garbage collection, and 2, for garbage
5607 collection supported.
5609 * - ``Objective-C GC Only``
5610 - **[Optional]** --- Specifies that only garbage collection is supported.
5611 If present, its value must be 6. This flag requires that the
5612 ``Objective-C Garbage Collection`` flag have the value 2.
5614 Some important flag interactions:
5616 - If a module with ``Objective-C Garbage Collection`` set to 0 is
5617 merged with a module with ``Objective-C Garbage Collection`` set to
5618 2, then the resulting module has the
5619 ``Objective-C Garbage Collection`` flag set to 0.
5620 - A module with ``Objective-C Garbage Collection`` set to 0 cannot be
5621 merged with a module with ``Objective-C GC Only`` set to 6.
5623 C type width Module Flags Metadata
5624 ----------------------------------
5626 The ARM backend emits a section into each generated object file describing the
5627 options that it was compiled with (in a compiler-independent way) to prevent
5628 linking incompatible objects, and to allow automatic library selection. Some
5629 of these options are not visible at the IR level, namely wchar_t width and enum
5632 To pass this information to the backend, these options are encoded in module
5633 flags metadata, using the following key-value pairs:
5643 - * 0 --- sizeof(wchar_t) == 4
5644 * 1 --- sizeof(wchar_t) == 2
5647 - * 0 --- Enums are at least as large as an ``int``.
5648 * 1 --- Enums are stored in the smallest integer type which can
5649 represent all of its values.
5651 For example, the following metadata section specifies that the module was
5652 compiled with a ``wchar_t`` width of 4 bytes, and the underlying type of an
5653 enum is the smallest type which can represent all of its values::
5655 !llvm.module.flags = !{!0, !1}
5656 !0 = !{i32 1, !"short_wchar", i32 1}
5657 !1 = !{i32 1, !"short_enum", i32 0}
5659 Automatic Linker Flags Named Metadata
5660 =====================================
5662 Some targets support embedding flags to the linker inside individual object
5663 files. Typically this is used in conjunction with language extensions which
5664 allow source files to explicitly declare the libraries they depend on, and have
5665 these automatically be transmitted to the linker via object files.
5667 These flags are encoded in the IR using named metadata with the name
5668 ``!llvm.linker.options``. Each operand is expected to be a metadata node
5669 which should be a list of other metadata nodes, each of which should be a
5670 list of metadata strings defining linker options.
5672 For example, the following metadata section specifies two separate sets of
5673 linker options, presumably to link against ``libz`` and the ``Cocoa``
5677 !1 = !{ !"-framework", !"Cocoa" } } }
5678 !llvm.linker.options = !{ !0, !1 }
5680 The metadata encoding as lists of lists of options, as opposed to a collapsed
5681 list of options, is chosen so that the IR encoding can use multiple option
5682 strings to specify e.g., a single library, while still having that specifier be
5683 preserved as an atomic element that can be recognized by a target specific
5684 assembly writer or object file emitter.
5686 Each individual option is required to be either a valid option for the target's
5687 linker, or an option that is reserved by the target specific assembly writer or
5688 object file emitter. No other aspect of these options is defined by the IR.
5690 .. _intrinsicglobalvariables:
5692 Intrinsic Global Variables
5693 ==========================
5695 LLVM has a number of "magic" global variables that contain data that
5696 affect code generation or other IR semantics. These are documented here.
5697 All globals of this sort should have a section specified as
5698 "``llvm.metadata``". This section and all globals that start with
5699 "``llvm.``" are reserved for use by LLVM.
5703 The '``llvm.used``' Global Variable
5704 -----------------------------------
5706 The ``@llvm.used`` global is an array which has
5707 :ref:`appending linkage <linkage_appending>`. This array contains a list of
5708 pointers to named global variables, functions and aliases which may optionally
5709 have a pointer cast formed of bitcast or getelementptr. For example, a legal
5712 .. code-block:: llvm
5717 @llvm.used = appending global [2 x i8*] [
5719 i8* bitcast (i32* @Y to i8*)
5720 ], section "llvm.metadata"
5722 If a symbol appears in the ``@llvm.used`` list, then the compiler, assembler,
5723 and linker are required to treat the symbol as if there is a reference to the
5724 symbol that it cannot see (which is why they have to be named). For example, if
5725 a variable has internal linkage and no references other than that from the
5726 ``@llvm.used`` list, it cannot be deleted. This is commonly used to represent
5727 references from inline asms and other things the compiler cannot "see", and
5728 corresponds to "``attribute((used))``" in GNU C.
5730 On some targets, the code generator must emit a directive to the
5731 assembler or object file to prevent the assembler and linker from
5732 molesting the symbol.
5734 .. _gv_llvmcompilerused:
5736 The '``llvm.compiler.used``' Global Variable
5737 --------------------------------------------
5739 The ``@llvm.compiler.used`` directive is the same as the ``@llvm.used``
5740 directive, except that it only prevents the compiler from touching the
5741 symbol. On targets that support it, this allows an intelligent linker to
5742 optimize references to the symbol without being impeded as it would be
5745 This is a rare construct that should only be used in rare circumstances,
5746 and should not be exposed to source languages.
5748 .. _gv_llvmglobalctors:
5750 The '``llvm.global_ctors``' Global Variable
5751 -------------------------------------------
5753 .. code-block:: llvm
5755 %0 = type { i32, void ()*, i8* }
5756 @llvm.global_ctors = appending global [1 x %0] [%0 { i32 65535, void ()* @ctor, i8* @data }]
5758 The ``@llvm.global_ctors`` array contains a list of constructor
5759 functions, priorities, and an optional associated global or function.
5760 The functions referenced by this array will be called in ascending order
5761 of priority (i.e. lowest first) when the module is loaded. The order of
5762 functions with the same priority is not defined.
5764 If the third field is present, non-null, and points to a global variable
5765 or function, the initializer function will only run if the associated
5766 data from the current module is not discarded.
5768 .. _llvmglobaldtors:
5770 The '``llvm.global_dtors``' Global Variable
5771 -------------------------------------------
5773 .. code-block:: llvm
5775 %0 = type { i32, void ()*, i8* }
5776 @llvm.global_dtors = appending global [1 x %0] [%0 { i32 65535, void ()* @dtor, i8* @data }]
5778 The ``@llvm.global_dtors`` array contains a list of destructor
5779 functions, priorities, and an optional associated global or function.
5780 The functions referenced by this array will be called in descending
5781 order of priority (i.e. highest first) when the module is unloaded. The
5782 order of functions with the same priority is not defined.
5784 If the third field is present, non-null, and points to a global variable
5785 or function, the destructor function will only run if the associated
5786 data from the current module is not discarded.
5788 Instruction Reference
5789 =====================
5791 The LLVM instruction set consists of several different classifications
5792 of instructions: :ref:`terminator instructions <terminators>`, :ref:`binary
5793 instructions <binaryops>`, :ref:`bitwise binary
5794 instructions <bitwiseops>`, :ref:`memory instructions <memoryops>`, and
5795 :ref:`other instructions <otherops>`.
5799 Terminator Instructions
5800 -----------------------
5802 As mentioned :ref:`previously <functionstructure>`, every basic block in a
5803 program ends with a "Terminator" instruction, which indicates which
5804 block should be executed after the current block is finished. These
5805 terminator instructions typically yield a '``void``' value: they produce
5806 control flow, not values (the one exception being the
5807 ':ref:`invoke <i_invoke>`' instruction).
5809 The terminator instructions are: ':ref:`ret <i_ret>`',
5810 ':ref:`br <i_br>`', ':ref:`switch <i_switch>`',
5811 ':ref:`indirectbr <i_indirectbr>`', ':ref:`invoke <i_invoke>`',
5812 ':ref:`resume <i_resume>`', ':ref:`catchswitch <i_catchswitch>`',
5813 ':ref:`catchret <i_catchret>`',
5814 ':ref:`cleanupret <i_cleanupret>`',
5815 and ':ref:`unreachable <i_unreachable>`'.
5819 '``ret``' Instruction
5820 ^^^^^^^^^^^^^^^^^^^^^
5827 ret <type> <value> ; Return a value from a non-void function
5828 ret void ; Return from void function
5833 The '``ret``' instruction is used to return control flow (and optionally
5834 a value) from a function back to the caller.
5836 There are two forms of the '``ret``' instruction: one that returns a
5837 value and then causes control flow, and one that just causes control
5843 The '``ret``' instruction optionally accepts a single argument, the
5844 return value. The type of the return value must be a ':ref:`first
5845 class <t_firstclass>`' type.
5847 A function is not :ref:`well formed <wellformed>` if it it has a non-void
5848 return type and contains a '``ret``' instruction with no return value or
5849 a return value with a type that does not match its type, or if it has a
5850 void return type and contains a '``ret``' instruction with a return
5856 When the '``ret``' instruction is executed, control flow returns back to
5857 the calling function's context. If the caller is a
5858 ":ref:`call <i_call>`" instruction, execution continues at the
5859 instruction after the call. If the caller was an
5860 ":ref:`invoke <i_invoke>`" instruction, execution continues at the
5861 beginning of the "normal" destination block. If the instruction returns
5862 a value, that value shall set the call or invoke instruction's return
5868 .. code-block:: llvm
5870 ret i32 5 ; Return an integer value of 5
5871 ret void ; Return from a void function
5872 ret { i32, i8 } { i32 4, i8 2 } ; Return a struct of values 4 and 2
5876 '``br``' Instruction
5877 ^^^^^^^^^^^^^^^^^^^^
5884 br i1 <cond>, label <iftrue>, label <iffalse>
5885 br label <dest> ; Unconditional branch
5890 The '``br``' instruction is used to cause control flow to transfer to a
5891 different basic block in the current function. There are two forms of
5892 this instruction, corresponding to a conditional branch and an
5893 unconditional branch.
5898 The conditional branch form of the '``br``' instruction takes a single
5899 '``i1``' value and two '``label``' values. The unconditional form of the
5900 '``br``' instruction takes a single '``label``' value as a target.
5905 Upon execution of a conditional '``br``' instruction, the '``i1``'
5906 argument is evaluated. If the value is ``true``, control flows to the
5907 '``iftrue``' ``label`` argument. If "cond" is ``false``, control flows
5908 to the '``iffalse``' ``label`` argument.
5913 .. code-block:: llvm
5916 %cond = icmp eq i32 %a, %b
5917 br i1 %cond, label %IfEqual, label %IfUnequal
5925 '``switch``' Instruction
5926 ^^^^^^^^^^^^^^^^^^^^^^^^
5933 switch <intty> <value>, label <defaultdest> [ <intty> <val>, label <dest> ... ]
5938 The '``switch``' instruction is used to transfer control flow to one of
5939 several different places. It is a generalization of the '``br``'
5940 instruction, allowing a branch to occur to one of many possible
5946 The '``switch``' instruction uses three parameters: an integer
5947 comparison value '``value``', a default '``label``' destination, and an
5948 array of pairs of comparison value constants and '``label``'s. The table
5949 is not allowed to contain duplicate constant entries.
5954 The ``switch`` instruction specifies a table of values and destinations.
5955 When the '``switch``' instruction is executed, this table is searched
5956 for the given value. If the value is found, control flow is transferred
5957 to the corresponding destination; otherwise, control flow is transferred
5958 to the default destination.
5963 Depending on properties of the target machine and the particular
5964 ``switch`` instruction, this instruction may be code generated in
5965 different ways. For example, it could be generated as a series of
5966 chained conditional branches or with a lookup table.
5971 .. code-block:: llvm
5973 ; Emulate a conditional br instruction
5974 %Val = zext i1 %value to i32
5975 switch i32 %Val, label %truedest [ i32 0, label %falsedest ]
5977 ; Emulate an unconditional br instruction
5978 switch i32 0, label %dest [ ]
5980 ; Implement a jump table:
5981 switch i32 %val, label %otherwise [ i32 0, label %onzero
5983 i32 2, label %ontwo ]
5987 '``indirectbr``' Instruction
5988 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5995 indirectbr <somety>* <address>, [ label <dest1>, label <dest2>, ... ]
6000 The '``indirectbr``' instruction implements an indirect branch to a
6001 label within the current function, whose address is specified by
6002 "``address``". Address must be derived from a
6003 :ref:`blockaddress <blockaddress>` constant.
6008 The '``address``' argument is the address of the label to jump to. The
6009 rest of the arguments indicate the full set of possible destinations
6010 that the address may point to. Blocks are allowed to occur multiple
6011 times in the destination list, though this isn't particularly useful.
6013 This destination list is required so that dataflow analysis has an
6014 accurate understanding of the CFG.
6019 Control transfers to the block specified in the address argument. All
6020 possible destination blocks must be listed in the label list, otherwise
6021 this instruction has undefined behavior. This implies that jumps to
6022 labels defined in other functions have undefined behavior as well.
6027 This is typically implemented with a jump through a register.
6032 .. code-block:: llvm
6034 indirectbr i8* %Addr, [ label %bb1, label %bb2, label %bb3 ]
6038 '``invoke``' Instruction
6039 ^^^^^^^^^^^^^^^^^^^^^^^^
6046 <result> = invoke [cconv] [ret attrs] <ty>|<fnty> <fnptrval>(<function args>) [fn attrs]
6047 [operand bundles] to label <normal label> unwind label <exception label>
6052 The '``invoke``' instruction causes control to transfer to a specified
6053 function, with the possibility of control flow transfer to either the
6054 '``normal``' label or the '``exception``' label. If the callee function
6055 returns with the "``ret``" instruction, control flow will return to the
6056 "normal" label. If the callee (or any indirect callees) returns via the
6057 ":ref:`resume <i_resume>`" instruction or other exception handling
6058 mechanism, control is interrupted and continued at the dynamically
6059 nearest "exception" label.
6061 The '``exception``' label is a `landing
6062 pad <ExceptionHandling.html#overview>`_ for the exception. As such,
6063 '``exception``' label is required to have the
6064 ":ref:`landingpad <i_landingpad>`" instruction, which contains the
6065 information about the behavior of the program after unwinding happens,
6066 as its first non-PHI instruction. The restrictions on the
6067 "``landingpad``" instruction's tightly couples it to the "``invoke``"
6068 instruction, so that the important information contained within the
6069 "``landingpad``" instruction can't be lost through normal code motion.
6074 This instruction requires several arguments:
6076 #. The optional "cconv" marker indicates which :ref:`calling
6077 convention <callingconv>` the call should use. If none is
6078 specified, the call defaults to using C calling conventions.
6079 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
6080 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
6082 #. '``ty``': the type of the call instruction itself which is also the
6083 type of the return value. Functions that return no value are marked
6085 #. '``fnty``': shall be the signature of the function being invoked. The
6086 argument types must match the types implied by this signature. This
6087 type can be omitted if the function is not varargs.
6088 #. '``fnptrval``': An LLVM value containing a pointer to a function to
6089 be invoked. In most cases, this is a direct function invocation, but
6090 indirect ``invoke``'s are just as possible, calling an arbitrary pointer
6092 #. '``function args``': argument list whose types match the function
6093 signature argument types and parameter attributes. All arguments must
6094 be of :ref:`first class <t_firstclass>` type. If the function signature
6095 indicates the function accepts a variable number of arguments, the
6096 extra arguments can be specified.
6097 #. '``normal label``': the label reached when the called function
6098 executes a '``ret``' instruction.
6099 #. '``exception label``': the label reached when a callee returns via
6100 the :ref:`resume <i_resume>` instruction or other exception handling
6102 #. The optional :ref:`function attributes <fnattrs>` list.
6103 #. The optional :ref:`operand bundles <opbundles>` list.
6108 This instruction is designed to operate as a standard '``call``'
6109 instruction in most regards. The primary difference is that it
6110 establishes an association with a label, which is used by the runtime
6111 library to unwind the stack.
6113 This instruction is used in languages with destructors to ensure that
6114 proper cleanup is performed in the case of either a ``longjmp`` or a
6115 thrown exception. Additionally, this is important for implementation of
6116 '``catch``' clauses in high-level languages that support them.
6118 For the purposes of the SSA form, the definition of the value returned
6119 by the '``invoke``' instruction is deemed to occur on the edge from the
6120 current block to the "normal" label. If the callee unwinds then no
6121 return value is available.
6126 .. code-block:: llvm
6128 %retval = invoke i32 @Test(i32 15) to label %Continue
6129 unwind label %TestCleanup ; i32:retval set
6130 %retval = invoke coldcc i32 %Testfnptr(i32 15) to label %Continue
6131 unwind label %TestCleanup ; i32:retval set
6135 '``resume``' Instruction
6136 ^^^^^^^^^^^^^^^^^^^^^^^^
6143 resume <type> <value>
6148 The '``resume``' instruction is a terminator instruction that has no
6154 The '``resume``' instruction requires one argument, which must have the
6155 same type as the result of any '``landingpad``' instruction in the same
6161 The '``resume``' instruction resumes propagation of an existing
6162 (in-flight) exception whose unwinding was interrupted with a
6163 :ref:`landingpad <i_landingpad>` instruction.
6168 .. code-block:: llvm
6170 resume { i8*, i32 } %exn
6174 '``catchswitch``' Instruction
6175 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6182 <resultval> = catchswitch within <parent> [ label <handler1>, label <handler2>, ... ] unwind to caller
6183 <resultval> = catchswitch within <parent> [ label <handler1>, label <handler2>, ... ] unwind label <default>
6188 The '``catchswitch``' instruction is used by `LLVM's exception handling system
6189 <ExceptionHandling.html#overview>`_ to describe the set of possible catch handlers
6190 that may be executed by the :ref:`EH personality routine <personalityfn>`.
6195 The ``parent`` argument is the token of the funclet that contains the
6196 ``catchswitch`` instruction. If the ``catchswitch`` is not inside a funclet,
6197 this operand may be the token ``none``.
6199 The ``default`` argument is the label of another basic block beginning with
6200 either a ``cleanuppad`` or ``catchswitch`` instruction. This unwind destination
6201 must be a legal target with respect to the ``parent`` links, as described in
6202 the `exception handling documentation\ <ExceptionHandling.html#wineh-constraints>`_.
6204 The ``handlers`` are a nonempty list of successor blocks that each begin with a
6205 :ref:`catchpad <i_catchpad>` instruction.
6210 Executing this instruction transfers control to one of the successors in
6211 ``handlers``, if appropriate, or continues to unwind via the unwind label if
6214 The ``catchswitch`` is both a terminator and a "pad" instruction, meaning that
6215 it must be both the first non-phi instruction and last instruction in the basic
6216 block. Therefore, it must be the only non-phi instruction in the block.
6221 .. code-block:: text
6224 %cs1 = catchswitch within none [label %handler0, label %handler1] unwind to caller
6226 %cs2 = catchswitch within %parenthandler [label %handler0] unwind label %cleanup
6230 '``catchret``' Instruction
6231 ^^^^^^^^^^^^^^^^^^^^^^^^^^
6238 catchret from <token> to label <normal>
6243 The '``catchret``' instruction is a terminator instruction that has a
6250 The first argument to a '``catchret``' indicates which ``catchpad`` it
6251 exits. It must be a :ref:`catchpad <i_catchpad>`.
6252 The second argument to a '``catchret``' specifies where control will
6258 The '``catchret``' instruction ends an existing (in-flight) exception whose
6259 unwinding was interrupted with a :ref:`catchpad <i_catchpad>` instruction. The
6260 :ref:`personality function <personalityfn>` gets a chance to execute arbitrary
6261 code to, for example, destroy the active exception. Control then transfers to
6264 The ``token`` argument must be a token produced by a ``catchpad`` instruction.
6265 If the specified ``catchpad`` is not the most-recently-entered not-yet-exited
6266 funclet pad (as described in the `EH documentation\ <ExceptionHandling.html#wineh-constraints>`_),
6267 the ``catchret``'s behavior is undefined.
6272 .. code-block:: text
6274 catchret from %catch label %continue
6278 '``cleanupret``' Instruction
6279 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6286 cleanupret from <value> unwind label <continue>
6287 cleanupret from <value> unwind to caller
6292 The '``cleanupret``' instruction is a terminator instruction that has
6293 an optional successor.
6299 The '``cleanupret``' instruction requires one argument, which indicates
6300 which ``cleanuppad`` it exits, and must be a :ref:`cleanuppad <i_cleanuppad>`.
6301 If the specified ``cleanuppad`` is not the most-recently-entered not-yet-exited
6302 funclet pad (as described in the `EH documentation\ <ExceptionHandling.html#wineh-constraints>`_),
6303 the ``cleanupret``'s behavior is undefined.
6305 The '``cleanupret``' instruction also has an optional successor, ``continue``,
6306 which must be the label of another basic block beginning with either a
6307 ``cleanuppad`` or ``catchswitch`` instruction. This unwind destination must
6308 be a legal target with respect to the ``parent`` links, as described in the
6309 `exception handling documentation\ <ExceptionHandling.html#wineh-constraints>`_.
6314 The '``cleanupret``' instruction indicates to the
6315 :ref:`personality function <personalityfn>` that one
6316 :ref:`cleanuppad <i_cleanuppad>` it transferred control to has ended.
6317 It transfers control to ``continue`` or unwinds out of the function.
6322 .. code-block:: text
6324 cleanupret from %cleanup unwind to caller
6325 cleanupret from %cleanup unwind label %continue
6329 '``unreachable``' Instruction
6330 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6342 The '``unreachable``' instruction has no defined semantics. This
6343 instruction is used to inform the optimizer that a particular portion of
6344 the code is not reachable. This can be used to indicate that the code
6345 after a no-return function cannot be reached, and other facts.
6350 The '``unreachable``' instruction has no defined semantics.
6357 Binary operators are used to do most of the computation in a program.
6358 They require two operands of the same type, execute an operation on
6359 them, and produce a single value. The operands might represent multiple
6360 data, as is the case with the :ref:`vector <t_vector>` data type. The
6361 result value has the same type as its operands.
6363 There are several different binary operators:
6367 '``add``' Instruction
6368 ^^^^^^^^^^^^^^^^^^^^^
6375 <result> = add <ty> <op1>, <op2> ; yields ty:result
6376 <result> = add nuw <ty> <op1>, <op2> ; yields ty:result
6377 <result> = add nsw <ty> <op1>, <op2> ; yields ty:result
6378 <result> = add nuw nsw <ty> <op1>, <op2> ; yields ty:result
6383 The '``add``' instruction returns the sum of its two operands.
6388 The two arguments to the '``add``' instruction must be
6389 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6390 arguments must have identical types.
6395 The value produced is the integer sum of the two operands.
6397 If the sum has unsigned overflow, the result returned is the
6398 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
6401 Because LLVM integers use a two's complement representation, this
6402 instruction is appropriate for both signed and unsigned integers.
6404 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
6405 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
6406 result value of the ``add`` is a :ref:`poison value <poisonvalues>` if
6407 unsigned and/or signed overflow, respectively, occurs.
6412 .. code-block:: text
6414 <result> = add i32 4, %var ; yields i32:result = 4 + %var
6418 '``fadd``' Instruction
6419 ^^^^^^^^^^^^^^^^^^^^^^
6426 <result> = fadd [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
6431 The '``fadd``' instruction returns the sum of its two operands.
6436 The two arguments to the '``fadd``' instruction must be :ref:`floating
6437 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
6438 Both arguments must have identical types.
6443 The value produced is the floating-point sum of the two operands.
6444 This instruction is assumed to execute in the default :ref:`floating-point
6445 environment <floatenv>`.
6446 This instruction can also take any number of :ref:`fast-math
6447 flags <fastmath>`, which are optimization hints to enable otherwise
6448 unsafe floating-point optimizations:
6453 .. code-block:: text
6455 <result> = fadd float 4.0, %var ; yields float:result = 4.0 + %var
6457 '``sub``' Instruction
6458 ^^^^^^^^^^^^^^^^^^^^^
6465 <result> = sub <ty> <op1>, <op2> ; yields ty:result
6466 <result> = sub nuw <ty> <op1>, <op2> ; yields ty:result
6467 <result> = sub nsw <ty> <op1>, <op2> ; yields ty:result
6468 <result> = sub nuw nsw <ty> <op1>, <op2> ; yields ty:result
6473 The '``sub``' instruction returns the difference of its two operands.
6475 Note that the '``sub``' instruction is used to represent the '``neg``'
6476 instruction present in most other intermediate representations.
6481 The two arguments to the '``sub``' instruction must be
6482 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6483 arguments must have identical types.
6488 The value produced is the integer difference of the two operands.
6490 If the difference has unsigned overflow, the result returned is the
6491 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
6494 Because LLVM integers use a two's complement representation, this
6495 instruction is appropriate for both signed and unsigned integers.
6497 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
6498 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
6499 result value of the ``sub`` is a :ref:`poison value <poisonvalues>` if
6500 unsigned and/or signed overflow, respectively, occurs.
6505 .. code-block:: text
6507 <result> = sub i32 4, %var ; yields i32:result = 4 - %var
6508 <result> = sub i32 0, %val ; yields i32:result = -%var
6512 '``fsub``' Instruction
6513 ^^^^^^^^^^^^^^^^^^^^^^
6520 <result> = fsub [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
6525 The '``fsub``' instruction returns the difference of its two operands.
6527 Note that the '``fsub``' instruction is used to represent the '``fneg``'
6528 instruction present in most other intermediate representations.
6533 The two arguments to the '``fsub``' instruction must be :ref:`floating
6534 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
6535 Both arguments must have identical types.
6540 The value produced is the floating-point difference of the two operands.
6541 This instruction is assumed to execute in the default :ref:`floating-point
6542 environment <floatenv>`.
6543 This instruction can also take any number of :ref:`fast-math
6544 flags <fastmath>`, which are optimization hints to enable otherwise
6545 unsafe floating-point optimizations:
6550 .. code-block:: text
6552 <result> = fsub float 4.0, %var ; yields float:result = 4.0 - %var
6553 <result> = fsub float -0.0, %val ; yields float:result = -%var
6555 '``mul``' Instruction
6556 ^^^^^^^^^^^^^^^^^^^^^
6563 <result> = mul <ty> <op1>, <op2> ; yields ty:result
6564 <result> = mul nuw <ty> <op1>, <op2> ; yields ty:result
6565 <result> = mul nsw <ty> <op1>, <op2> ; yields ty:result
6566 <result> = mul nuw nsw <ty> <op1>, <op2> ; yields ty:result
6571 The '``mul``' instruction returns the product of its two operands.
6576 The two arguments to the '``mul``' instruction must be
6577 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6578 arguments must have identical types.
6583 The value produced is the integer product of the two operands.
6585 If the result of the multiplication has unsigned overflow, the result
6586 returned is the mathematical result modulo 2\ :sup:`n`\ , where n is the
6587 bit width of the result.
6589 Because LLVM integers use a two's complement representation, and the
6590 result is the same width as the operands, this instruction returns the
6591 correct result for both signed and unsigned integers. If a full product
6592 (e.g. ``i32`` * ``i32`` -> ``i64``) is needed, the operands should be
6593 sign-extended or zero-extended as appropriate to the width of the full
6596 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
6597 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
6598 result value of the ``mul`` is a :ref:`poison value <poisonvalues>` if
6599 unsigned and/or signed overflow, respectively, occurs.
6604 .. code-block:: text
6606 <result> = mul i32 4, %var ; yields i32:result = 4 * %var
6610 '``fmul``' Instruction
6611 ^^^^^^^^^^^^^^^^^^^^^^
6618 <result> = fmul [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
6623 The '``fmul``' instruction returns the product of its two operands.
6628 The two arguments to the '``fmul``' instruction must be :ref:`floating
6629 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
6630 Both arguments must have identical types.
6635 The value produced is the floating-point product of the two operands.
6636 This instruction is assumed to execute in the default :ref:`floating-point
6637 environment <floatenv>`.
6638 This instruction can also take any number of :ref:`fast-math
6639 flags <fastmath>`, which are optimization hints to enable otherwise
6640 unsafe floating-point optimizations:
6645 .. code-block:: text
6647 <result> = fmul float 4.0, %var ; yields float:result = 4.0 * %var
6649 '``udiv``' Instruction
6650 ^^^^^^^^^^^^^^^^^^^^^^
6657 <result> = udiv <ty> <op1>, <op2> ; yields ty:result
6658 <result> = udiv exact <ty> <op1>, <op2> ; yields ty:result
6663 The '``udiv``' instruction returns the quotient of its two operands.
6668 The two arguments to the '``udiv``' instruction must be
6669 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6670 arguments must have identical types.
6675 The value produced is the unsigned integer quotient of the two operands.
6677 Note that unsigned integer division and signed integer division are
6678 distinct operations; for signed integer division, use '``sdiv``'.
6680 Division by zero is undefined behavior. For vectors, if any element
6681 of the divisor is zero, the operation has undefined behavior.
6684 If the ``exact`` keyword is present, the result value of the ``udiv`` is
6685 a :ref:`poison value <poisonvalues>` if %op1 is not a multiple of %op2 (as
6686 such, "((a udiv exact b) mul b) == a").
6691 .. code-block:: text
6693 <result> = udiv i32 4, %var ; yields i32:result = 4 / %var
6695 '``sdiv``' Instruction
6696 ^^^^^^^^^^^^^^^^^^^^^^
6703 <result> = sdiv <ty> <op1>, <op2> ; yields ty:result
6704 <result> = sdiv exact <ty> <op1>, <op2> ; yields ty:result
6709 The '``sdiv``' instruction returns the quotient of its two operands.
6714 The two arguments to the '``sdiv``' instruction must be
6715 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6716 arguments must have identical types.
6721 The value produced is the signed integer quotient of the two operands
6722 rounded towards zero.
6724 Note that signed integer division and unsigned integer division are
6725 distinct operations; for unsigned integer division, use '``udiv``'.
6727 Division by zero is undefined behavior. For vectors, if any element
6728 of the divisor is zero, the operation has undefined behavior.
6729 Overflow also leads to undefined behavior; this is a rare case, but can
6730 occur, for example, by doing a 32-bit division of -2147483648 by -1.
6732 If the ``exact`` keyword is present, the result value of the ``sdiv`` is
6733 a :ref:`poison value <poisonvalues>` if the result would be rounded.
6738 .. code-block:: text
6740 <result> = sdiv i32 4, %var ; yields i32:result = 4 / %var
6744 '``fdiv``' Instruction
6745 ^^^^^^^^^^^^^^^^^^^^^^
6752 <result> = fdiv [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
6757 The '``fdiv``' instruction returns the quotient of its two operands.
6762 The two arguments to the '``fdiv``' instruction must be :ref:`floating
6763 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
6764 Both arguments must have identical types.
6769 The value produced is the floating-point quotient of the two operands.
6770 This instruction is assumed to execute in the default :ref:`floating-point
6771 environment <floatenv>`.
6772 This instruction can also take any number of :ref:`fast-math
6773 flags <fastmath>`, which are optimization hints to enable otherwise
6774 unsafe floating-point optimizations:
6779 .. code-block:: text
6781 <result> = fdiv float 4.0, %var ; yields float:result = 4.0 / %var
6783 '``urem``' Instruction
6784 ^^^^^^^^^^^^^^^^^^^^^^
6791 <result> = urem <ty> <op1>, <op2> ; yields ty:result
6796 The '``urem``' instruction returns the remainder from the unsigned
6797 division of its two arguments.
6802 The two arguments to the '``urem``' instruction must be
6803 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6804 arguments must have identical types.
6809 This instruction returns the unsigned integer *remainder* of a division.
6810 This instruction always performs an unsigned division to get the
6813 Note that unsigned integer remainder and signed integer remainder are
6814 distinct operations; for signed integer remainder, use '``srem``'.
6816 Taking the remainder of a division by zero is undefined behavior.
6817 For vectors, if any element of the divisor is zero, the operation has
6823 .. code-block:: text
6825 <result> = urem i32 4, %var ; yields i32:result = 4 % %var
6827 '``srem``' Instruction
6828 ^^^^^^^^^^^^^^^^^^^^^^
6835 <result> = srem <ty> <op1>, <op2> ; yields ty:result
6840 The '``srem``' instruction returns the remainder from the signed
6841 division of its two operands. This instruction can also take
6842 :ref:`vector <t_vector>` versions of the values in which case the elements
6848 The two arguments to the '``srem``' instruction must be
6849 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6850 arguments must have identical types.
6855 This instruction returns the *remainder* of a division (where the result
6856 is either zero or has the same sign as the dividend, ``op1``), not the
6857 *modulo* operator (where the result is either zero or has the same sign
6858 as the divisor, ``op2``) of a value. For more information about the
6859 difference, see `The Math
6860 Forum <http://mathforum.org/dr.math/problems/anne.4.28.99.html>`_. For a
6861 table of how this is implemented in various languages, please see
6863 operation <http://en.wikipedia.org/wiki/Modulo_operation>`_.
6865 Note that signed integer remainder and unsigned integer remainder are
6866 distinct operations; for unsigned integer remainder, use '``urem``'.
6868 Taking the remainder of a division by zero is undefined behavior.
6869 For vectors, if any element of the divisor is zero, the operation has
6871 Overflow also leads to undefined behavior; this is a rare case, but can
6872 occur, for example, by taking the remainder of a 32-bit division of
6873 -2147483648 by -1. (The remainder doesn't actually overflow, but this
6874 rule lets srem be implemented using instructions that return both the
6875 result of the division and the remainder.)
6880 .. code-block:: text
6882 <result> = srem i32 4, %var ; yields i32:result = 4 % %var
6886 '``frem``' Instruction
6887 ^^^^^^^^^^^^^^^^^^^^^^
6894 <result> = frem [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
6899 The '``frem``' instruction returns the remainder from the division of
6905 The two arguments to the '``frem``' instruction must be :ref:`floating
6906 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
6907 Both arguments must have identical types.
6912 The value produced is the floating-point remainder of the two operands.
6913 This is the same output as a libm '``fmod``' function, but without any
6914 possibility of setting ``errno``. The remainder has the same sign as the
6916 This instruction is assumed to execute in the default :ref:`floating-point
6917 environment <floatenv>`.
6918 This instruction can also take any number of :ref:`fast-math
6919 flags <fastmath>`, which are optimization hints to enable otherwise
6920 unsafe floating-point optimizations:
6925 .. code-block:: text
6927 <result> = frem float 4.0, %var ; yields float:result = 4.0 % %var
6931 Bitwise Binary Operations
6932 -------------------------
6934 Bitwise binary operators are used to do various forms of bit-twiddling
6935 in a program. They are generally very efficient instructions and can
6936 commonly be strength reduced from other instructions. They require two
6937 operands of the same type, execute an operation on them, and produce a
6938 single value. The resulting value is the same type as its operands.
6940 '``shl``' Instruction
6941 ^^^^^^^^^^^^^^^^^^^^^
6948 <result> = shl <ty> <op1>, <op2> ; yields ty:result
6949 <result> = shl nuw <ty> <op1>, <op2> ; yields ty:result
6950 <result> = shl nsw <ty> <op1>, <op2> ; yields ty:result
6951 <result> = shl nuw nsw <ty> <op1>, <op2> ; yields ty:result
6956 The '``shl``' instruction returns the first operand shifted to the left
6957 a specified number of bits.
6962 Both arguments to the '``shl``' instruction must be the same
6963 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
6964 '``op2``' is treated as an unsigned value.
6969 The value produced is ``op1`` \* 2\ :sup:`op2` mod 2\ :sup:`n`,
6970 where ``n`` is the width of the result. If ``op2`` is (statically or
6971 dynamically) equal to or larger than the number of bits in
6972 ``op1``, this instruction returns a :ref:`poison value <poisonvalues>`.
6973 If the arguments are vectors, each vector element of ``op1`` is shifted
6974 by the corresponding shift amount in ``op2``.
6976 If the ``nuw`` keyword is present, then the shift produces a poison
6977 value if it shifts out any non-zero bits.
6978 If the ``nsw`` keyword is present, then the shift produces a poison
6979 value it shifts out any bits that disagree with the resultant sign bit.
6984 .. code-block:: text
6986 <result> = shl i32 4, %var ; yields i32: 4 << %var
6987 <result> = shl i32 4, 2 ; yields i32: 16
6988 <result> = shl i32 1, 10 ; yields i32: 1024
6989 <result> = shl i32 1, 32 ; undefined
6990 <result> = shl <2 x i32> < i32 1, i32 1>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 2, i32 4>
6992 '``lshr``' Instruction
6993 ^^^^^^^^^^^^^^^^^^^^^^
7000 <result> = lshr <ty> <op1>, <op2> ; yields ty:result
7001 <result> = lshr exact <ty> <op1>, <op2> ; yields ty:result
7006 The '``lshr``' instruction (logical shift right) returns the first
7007 operand shifted to the right a specified number of bits with zero fill.
7012 Both arguments to the '``lshr``' instruction must be the same
7013 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
7014 '``op2``' is treated as an unsigned value.
7019 This instruction always performs a logical shift right operation. The
7020 most significant bits of the result will be filled with zero bits after
7021 the shift. If ``op2`` is (statically or dynamically) equal to or larger
7022 than the number of bits in ``op1``, this instruction returns a :ref:`poison
7023 value <poisonvalues>`. If the arguments are vectors, each vector element
7024 of ``op1`` is shifted by the corresponding shift amount in ``op2``.
7026 If the ``exact`` keyword is present, the result value of the ``lshr`` is
7027 a poison value if any of the bits shifted out are non-zero.
7032 .. code-block:: text
7034 <result> = lshr i32 4, 1 ; yields i32:result = 2
7035 <result> = lshr i32 4, 2 ; yields i32:result = 1
7036 <result> = lshr i8 4, 3 ; yields i8:result = 0
7037 <result> = lshr i8 -2, 1 ; yields i8:result = 0x7F
7038 <result> = lshr i32 1, 32 ; undefined
7039 <result> = lshr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 0x7FFFFFFF, i32 1>
7041 '``ashr``' Instruction
7042 ^^^^^^^^^^^^^^^^^^^^^^
7049 <result> = ashr <ty> <op1>, <op2> ; yields ty:result
7050 <result> = ashr exact <ty> <op1>, <op2> ; yields ty:result
7055 The '``ashr``' instruction (arithmetic shift right) returns the first
7056 operand shifted to the right a specified number of bits with sign
7062 Both arguments to the '``ashr``' instruction must be the same
7063 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
7064 '``op2``' is treated as an unsigned value.
7069 This instruction always performs an arithmetic shift right operation,
7070 The most significant bits of the result will be filled with the sign bit
7071 of ``op1``. If ``op2`` is (statically or dynamically) equal to or larger
7072 than the number of bits in ``op1``, this instruction returns a :ref:`poison
7073 value <poisonvalues>`. If the arguments are vectors, each vector element
7074 of ``op1`` is shifted by the corresponding shift amount in ``op2``.
7076 If the ``exact`` keyword is present, the result value of the ``ashr`` is
7077 a poison value if any of the bits shifted out are non-zero.
7082 .. code-block:: text
7084 <result> = ashr i32 4, 1 ; yields i32:result = 2
7085 <result> = ashr i32 4, 2 ; yields i32:result = 1
7086 <result> = ashr i8 4, 3 ; yields i8:result = 0
7087 <result> = ashr i8 -2, 1 ; yields i8:result = -1
7088 <result> = ashr i32 1, 32 ; undefined
7089 <result> = ashr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 3> ; yields: result=<2 x i32> < i32 -1, i32 0>
7091 '``and``' Instruction
7092 ^^^^^^^^^^^^^^^^^^^^^
7099 <result> = and <ty> <op1>, <op2> ; yields ty:result
7104 The '``and``' instruction returns the bitwise logical and of its two
7110 The two arguments to the '``and``' instruction must be
7111 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
7112 arguments must have identical types.
7117 The truth table used for the '``and``' instruction is:
7134 .. code-block:: text
7136 <result> = and i32 4, %var ; yields i32:result = 4 & %var
7137 <result> = and i32 15, 40 ; yields i32:result = 8
7138 <result> = and i32 4, 8 ; yields i32:result = 0
7140 '``or``' Instruction
7141 ^^^^^^^^^^^^^^^^^^^^
7148 <result> = or <ty> <op1>, <op2> ; yields ty:result
7153 The '``or``' instruction returns the bitwise logical inclusive or of its
7159 The two arguments to the '``or``' instruction must be
7160 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
7161 arguments must have identical types.
7166 The truth table used for the '``or``' instruction is:
7185 <result> = or i32 4, %var ; yields i32:result = 4 | %var
7186 <result> = or i32 15, 40 ; yields i32:result = 47
7187 <result> = or i32 4, 8 ; yields i32:result = 12
7189 '``xor``' Instruction
7190 ^^^^^^^^^^^^^^^^^^^^^
7197 <result> = xor <ty> <op1>, <op2> ; yields ty:result
7202 The '``xor``' instruction returns the bitwise logical exclusive or of
7203 its two operands. The ``xor`` is used to implement the "one's
7204 complement" operation, which is the "~" operator in C.
7209 The two arguments to the '``xor``' instruction must be
7210 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
7211 arguments must have identical types.
7216 The truth table used for the '``xor``' instruction is:
7233 .. code-block:: text
7235 <result> = xor i32 4, %var ; yields i32:result = 4 ^ %var
7236 <result> = xor i32 15, 40 ; yields i32:result = 39
7237 <result> = xor i32 4, 8 ; yields i32:result = 12
7238 <result> = xor i32 %V, -1 ; yields i32:result = ~%V
7243 LLVM supports several instructions to represent vector operations in a
7244 target-independent manner. These instructions cover the element-access
7245 and vector-specific operations needed to process vectors effectively.
7246 While LLVM does directly support these vector operations, many
7247 sophisticated algorithms will want to use target-specific intrinsics to
7248 take full advantage of a specific target.
7250 .. _i_extractelement:
7252 '``extractelement``' Instruction
7253 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7260 <result> = extractelement <n x <ty>> <val>, <ty2> <idx> ; yields <ty>
7265 The '``extractelement``' instruction extracts a single scalar element
7266 from a vector at a specified index.
7271 The first operand of an '``extractelement``' instruction is a value of
7272 :ref:`vector <t_vector>` type. The second operand is an index indicating
7273 the position from which to extract the element. The index may be a
7274 variable of any integer type.
7279 The result is a scalar of the same type as the element type of ``val``.
7280 Its value is the value at position ``idx`` of ``val``. If ``idx``
7281 exceeds the length of ``val``, the results are undefined.
7286 .. code-block:: text
7288 <result> = extractelement <4 x i32> %vec, i32 0 ; yields i32
7290 .. _i_insertelement:
7292 '``insertelement``' Instruction
7293 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7300 <result> = insertelement <n x <ty>> <val>, <ty> <elt>, <ty2> <idx> ; yields <n x <ty>>
7305 The '``insertelement``' instruction inserts a scalar element into a
7306 vector at a specified index.
7311 The first operand of an '``insertelement``' instruction is a value of
7312 :ref:`vector <t_vector>` type. The second operand is a scalar value whose
7313 type must equal the element type of the first operand. The third operand
7314 is an index indicating the position at which to insert the value. The
7315 index may be a variable of any integer type.
7320 The result is a vector of the same type as ``val``. Its element values
7321 are those of ``val`` except at position ``idx``, where it gets the value
7322 ``elt``. If ``idx`` exceeds the length of ``val``, the results are
7328 .. code-block:: text
7330 <result> = insertelement <4 x i32> %vec, i32 1, i32 0 ; yields <4 x i32>
7332 .. _i_shufflevector:
7334 '``shufflevector``' Instruction
7335 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7342 <result> = shufflevector <n x <ty>> <v1>, <n x <ty>> <v2>, <m x i32> <mask> ; yields <m x <ty>>
7347 The '``shufflevector``' instruction constructs a permutation of elements
7348 from two input vectors, returning a vector with the same element type as
7349 the input and length that is the same as the shuffle mask.
7354 The first two operands of a '``shufflevector``' instruction are vectors
7355 with the same type. The third argument is a shuffle mask whose element
7356 type is always 'i32'. The result of the instruction is a vector whose
7357 length is the same as the shuffle mask and whose element type is the
7358 same as the element type of the first two operands.
7360 The shuffle mask operand is required to be a constant vector with either
7361 constant integer or undef values.
7366 The elements of the two input vectors are numbered from left to right
7367 across both of the vectors. The shuffle mask operand specifies, for each
7368 element of the result vector, which element of the two input vectors the
7369 result element gets. If the shuffle mask is undef, the result vector is
7370 undef. If any element of the mask operand is undef, that element of the
7371 result is undef. If the shuffle mask selects an undef element from one
7372 of the input vectors, the resulting element is undef.
7377 .. code-block:: text
7379 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
7380 <4 x i32> <i32 0, i32 4, i32 1, i32 5> ; yields <4 x i32>
7381 <result> = shufflevector <4 x i32> %v1, <4 x i32> undef,
7382 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32> - Identity shuffle.
7383 <result> = shufflevector <8 x i32> %v1, <8 x i32> undef,
7384 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32>
7385 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
7386 <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7 > ; yields <8 x i32>
7388 Aggregate Operations
7389 --------------------
7391 LLVM supports several instructions for working with
7392 :ref:`aggregate <t_aggregate>` values.
7396 '``extractvalue``' Instruction
7397 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7404 <result> = extractvalue <aggregate type> <val>, <idx>{, <idx>}*
7409 The '``extractvalue``' instruction extracts the value of a member field
7410 from an :ref:`aggregate <t_aggregate>` value.
7415 The first operand of an '``extractvalue``' instruction is a value of
7416 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The other operands are
7417 constant indices to specify which value to extract in a similar manner
7418 as indices in a '``getelementptr``' instruction.
7420 The major differences to ``getelementptr`` indexing are:
7422 - Since the value being indexed is not a pointer, the first index is
7423 omitted and assumed to be zero.
7424 - At least one index must be specified.
7425 - Not only struct indices but also array indices must be in bounds.
7430 The result is the value at the position in the aggregate specified by
7436 .. code-block:: text
7438 <result> = extractvalue {i32, float} %agg, 0 ; yields i32
7442 '``insertvalue``' Instruction
7443 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7450 <result> = insertvalue <aggregate type> <val>, <ty> <elt>, <idx>{, <idx>}* ; yields <aggregate type>
7455 The '``insertvalue``' instruction inserts a value into a member field in
7456 an :ref:`aggregate <t_aggregate>` value.
7461 The first operand of an '``insertvalue``' instruction is a value of
7462 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The second operand is
7463 a first-class value to insert. The following operands are constant
7464 indices indicating the position at which to insert the value in a
7465 similar manner as indices in a '``extractvalue``' instruction. The value
7466 to insert must have the same type as the value identified by the
7472 The result is an aggregate of the same type as ``val``. Its value is
7473 that of ``val`` except that the value at the position specified by the
7474 indices is that of ``elt``.
7479 .. code-block:: llvm
7481 %agg1 = insertvalue {i32, float} undef, i32 1, 0 ; yields {i32 1, float undef}
7482 %agg2 = insertvalue {i32, float} %agg1, float %val, 1 ; yields {i32 1, float %val}
7483 %agg3 = insertvalue {i32, {float}} undef, float %val, 1, 0 ; yields {i32 undef, {float %val}}
7487 Memory Access and Addressing Operations
7488 ---------------------------------------
7490 A key design point of an SSA-based representation is how it represents
7491 memory. In LLVM, no memory locations are in SSA form, which makes things
7492 very simple. This section describes how to read, write, and allocate
7497 '``alloca``' Instruction
7498 ^^^^^^^^^^^^^^^^^^^^^^^^
7505 <result> = alloca [inalloca] <type> [, <ty> <NumElements>] [, align <alignment>] [, addrspace(<num>)] ; yields type addrspace(num)*:result
7510 The '``alloca``' instruction allocates memory on the stack frame of the
7511 currently executing function, to be automatically released when this
7512 function returns to its caller. The object is always allocated in the
7513 address space for allocas indicated in the datalayout.
7518 The '``alloca``' instruction allocates ``sizeof(<type>)*NumElements``
7519 bytes of memory on the runtime stack, returning a pointer of the
7520 appropriate type to the program. If "NumElements" is specified, it is
7521 the number of elements allocated, otherwise "NumElements" is defaulted
7522 to be one. If a constant alignment is specified, the value result of the
7523 allocation is guaranteed to be aligned to at least that boundary. The
7524 alignment may not be greater than ``1 << 29``. If not specified, or if
7525 zero, the target can choose to align the allocation on any convenient
7526 boundary compatible with the type.
7528 '``type``' may be any sized type.
7533 Memory is allocated; a pointer is returned. The operation is undefined
7534 if there is insufficient stack space for the allocation. '``alloca``'d
7535 memory is automatically released when the function returns. The
7536 '``alloca``' instruction is commonly used to represent automatic
7537 variables that must have an address available. When the function returns
7538 (either with the ``ret`` or ``resume`` instructions), the memory is
7539 reclaimed. Allocating zero bytes is legal, but the result is undefined.
7540 The order in which memory is allocated (ie., which way the stack grows)
7546 .. code-block:: llvm
7548 %ptr = alloca i32 ; yields i32*:ptr
7549 %ptr = alloca i32, i32 4 ; yields i32*:ptr
7550 %ptr = alloca i32, i32 4, align 1024 ; yields i32*:ptr
7551 %ptr = alloca i32, align 1024 ; yields i32*:ptr
7555 '``load``' Instruction
7556 ^^^^^^^^^^^^^^^^^^^^^^
7563 <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>]
7564 <result> = load atomic [volatile] <ty>, <ty>* <pointer> [syncscope("<target-scope>")] <ordering>, align <alignment> [, !invariant.group !<index>]
7565 !<index> = !{ i32 1 }
7566 !<deref_bytes_node> = !{i64 <dereferenceable_bytes>}
7567 !<align_node> = !{ i64 <value_alignment> }
7572 The '``load``' instruction is used to read from memory.
7577 The argument to the ``load`` instruction specifies the memory address from which
7578 to load. The type specified must be a :ref:`first class <t_firstclass>` type of
7579 known size (i.e. not containing an :ref:`opaque structural type <t_opaque>`). If
7580 the ``load`` is marked as ``volatile``, then the optimizer is not allowed to
7581 modify the number or order of execution of this ``load`` with other
7582 :ref:`volatile operations <volatile>`.
7584 If the ``load`` is marked as ``atomic``, it takes an extra :ref:`ordering
7585 <ordering>` and optional ``syncscope("<target-scope>")`` argument. The
7586 ``release`` and ``acq_rel`` orderings are not valid on ``load`` instructions.
7587 Atomic loads produce :ref:`defined <memmodel>` results when they may see
7588 multiple atomic stores. The type of the pointee must be an integer, pointer, or
7589 floating-point type whose bit width is a power of two greater than or equal to
7590 eight and less than or equal to a target-specific size limit. ``align`` must be
7591 explicitly specified on atomic loads, and the load has undefined behavior if the
7592 alignment is not set to a value which is at least the size in bytes of the
7593 pointee. ``!nontemporal`` does not have any defined semantics for atomic loads.
7595 The optional constant ``align`` argument specifies the alignment of the
7596 operation (that is, the alignment of the memory address). A value of 0
7597 or an omitted ``align`` argument means that the operation has the ABI
7598 alignment for the target. It is the responsibility of the code emitter
7599 to ensure that the alignment information is correct. Overestimating the
7600 alignment results in undefined behavior. Underestimating the alignment
7601 may produce less efficient code. An alignment of 1 is always safe. The
7602 maximum possible alignment is ``1 << 29``. An alignment value higher
7603 than the size of the loaded type implies memory up to the alignment
7604 value bytes can be safely loaded without trapping in the default
7605 address space. Access of the high bytes can interfere with debugging
7606 tools, so should not be accessed if the function has the
7607 ``sanitize_thread`` or ``sanitize_address`` attributes.
7609 The optional ``!nontemporal`` metadata must reference a single
7610 metadata name ``<index>`` corresponding to a metadata node with one
7611 ``i32`` entry of value 1. The existence of the ``!nontemporal``
7612 metadata on the instruction tells the optimizer and code generator
7613 that this load is not expected to be reused in the cache. The code
7614 generator may select special instructions to save cache bandwidth, such
7615 as the ``MOVNT`` instruction on x86.
7617 The optional ``!invariant.load`` metadata must reference a single
7618 metadata name ``<index>`` corresponding to a metadata node with no
7619 entries. If a load instruction tagged with the ``!invariant.load``
7620 metadata is executed, the optimizer may assume the memory location
7621 referenced by the load contains the same value at all points in the
7622 program where the memory location is known to be dereferenceable.
7624 The optional ``!invariant.group`` metadata must reference a single metadata name
7625 ``<index>`` corresponding to a metadata node. See ``invariant.group`` metadata.
7627 The optional ``!nonnull`` metadata must reference a single
7628 metadata name ``<index>`` corresponding to a metadata node with no
7629 entries. The existence of the ``!nonnull`` metadata on the
7630 instruction tells the optimizer that the value loaded is known to
7631 never be null. This is analogous to the ``nonnull`` attribute
7632 on parameters and return values. This metadata can only be applied
7633 to loads of a pointer type.
7635 The optional ``!dereferenceable`` metadata must reference a single metadata
7636 name ``<deref_bytes_node>`` corresponding to a metadata node with one ``i64``
7637 entry. The existence of the ``!dereferenceable`` metadata on the instruction
7638 tells the optimizer that the value loaded is known to be dereferenceable.
7639 The number of bytes known to be dereferenceable is specified by the integer
7640 value in the metadata node. This is analogous to the ''dereferenceable''
7641 attribute on parameters and return values. This metadata can only be applied
7642 to loads of a pointer type.
7644 The optional ``!dereferenceable_or_null`` metadata must reference a single
7645 metadata name ``<deref_bytes_node>`` corresponding to a metadata node with one
7646 ``i64`` entry. The existence of the ``!dereferenceable_or_null`` metadata on the
7647 instruction tells the optimizer that the value loaded is known to be either
7648 dereferenceable or null.
7649 The number of bytes known to be dereferenceable is specified by the integer
7650 value in the metadata node. This is analogous to the ''dereferenceable_or_null''
7651 attribute on parameters and return values. This metadata can only be applied
7652 to loads of a pointer type.
7654 The optional ``!align`` metadata must reference a single metadata name
7655 ``<align_node>`` corresponding to a metadata node with one ``i64`` entry.
7656 The existence of the ``!align`` metadata on the instruction tells the
7657 optimizer that the value loaded is known to be aligned to a boundary specified
7658 by the integer value in the metadata node. The alignment must be a power of 2.
7659 This is analogous to the ''align'' attribute on parameters and return values.
7660 This metadata can only be applied to loads of a pointer type.
7665 The location of memory pointed to is loaded. If the value being loaded
7666 is of scalar type then the number of bytes read does not exceed the
7667 minimum number of bytes needed to hold all bits of the type. For
7668 example, loading an ``i24`` reads at most three bytes. When loading a
7669 value of a type like ``i20`` with a size that is not an integral number
7670 of bytes, the result is undefined if the value was not originally
7671 written using a store of the same type.
7676 .. code-block:: llvm
7678 %ptr = alloca i32 ; yields i32*:ptr
7679 store i32 3, i32* %ptr ; yields void
7680 %val = load i32, i32* %ptr ; yields i32:val = i32 3
7684 '``store``' Instruction
7685 ^^^^^^^^^^^^^^^^^^^^^^^
7692 store [volatile] <ty> <value>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>][, !invariant.group !<index>] ; yields void
7693 store atomic [volatile] <ty> <value>, <ty>* <pointer> [syncscope("<target-scope>")] <ordering>, align <alignment> [, !invariant.group !<index>] ; yields void
7698 The '``store``' instruction is used to write to memory.
7703 There are two arguments to the ``store`` instruction: a value to store and an
7704 address at which to store it. The type of the ``<pointer>`` operand must be a
7705 pointer to the :ref:`first class <t_firstclass>` type of the ``<value>``
7706 operand. If the ``store`` is marked as ``volatile``, then the optimizer is not
7707 allowed to modify the number or order of execution of this ``store`` with other
7708 :ref:`volatile operations <volatile>`. Only values of :ref:`first class
7709 <t_firstclass>` types of known size (i.e. not containing an :ref:`opaque
7710 structural type <t_opaque>`) can be stored.
7712 If the ``store`` is marked as ``atomic``, it takes an extra :ref:`ordering
7713 <ordering>` and optional ``syncscope("<target-scope>")`` argument. The
7714 ``acquire`` and ``acq_rel`` orderings aren't valid on ``store`` instructions.
7715 Atomic loads produce :ref:`defined <memmodel>` results when they may see
7716 multiple atomic stores. The type of the pointee must be an integer, pointer, or
7717 floating-point type whose bit width is a power of two greater than or equal to
7718 eight and less than or equal to a target-specific size limit. ``align`` must be
7719 explicitly specified on atomic stores, and the store has undefined behavior if
7720 the alignment is not set to a value which is at least the size in bytes of the
7721 pointee. ``!nontemporal`` does not have any defined semantics for atomic stores.
7723 The optional constant ``align`` argument specifies the alignment of the
7724 operation (that is, the alignment of the memory address). A value of 0
7725 or an omitted ``align`` argument means that the operation has the ABI
7726 alignment for the target. It is the responsibility of the code emitter
7727 to ensure that the alignment information is correct. Overestimating the
7728 alignment results in undefined behavior. Underestimating the
7729 alignment may produce less efficient code. An alignment of 1 is always
7730 safe. The maximum possible alignment is ``1 << 29``. An alignment
7731 value higher than the size of the stored type implies memory up to the
7732 alignment value bytes can be stored to without trapping in the default
7733 address space. Storing to the higher bytes however may result in data
7734 races if another thread can access the same address. Introducing a
7735 data race is not allowed. Storing to the extra bytes is not allowed
7736 even in situations where a data race is known to not exist if the
7737 function has the ``sanitize_address`` attribute.
7739 The optional ``!nontemporal`` metadata must reference a single metadata
7740 name ``<index>`` corresponding to a metadata node with one ``i32`` entry of
7741 value 1. The existence of the ``!nontemporal`` metadata on the instruction
7742 tells the optimizer and code generator that this load is not expected to
7743 be reused in the cache. The code generator may select special
7744 instructions to save cache bandwidth, such as the ``MOVNT`` instruction on
7747 The optional ``!invariant.group`` metadata must reference a
7748 single metadata name ``<index>``. See ``invariant.group`` metadata.
7753 The contents of memory are updated to contain ``<value>`` at the
7754 location specified by the ``<pointer>`` operand. If ``<value>`` is
7755 of scalar type then the number of bytes written does not exceed the
7756 minimum number of bytes needed to hold all bits of the type. For
7757 example, storing an ``i24`` writes at most three bytes. When writing a
7758 value of a type like ``i20`` with a size that is not an integral number
7759 of bytes, it is unspecified what happens to the extra bits that do not
7760 belong to the type, but they will typically be overwritten.
7765 .. code-block:: llvm
7767 %ptr = alloca i32 ; yields i32*:ptr
7768 store i32 3, i32* %ptr ; yields void
7769 %val = load i32, i32* %ptr ; yields i32:val = i32 3
7773 '``fence``' Instruction
7774 ^^^^^^^^^^^^^^^^^^^^^^^
7781 fence [syncscope("<target-scope>")] <ordering> ; yields void
7786 The '``fence``' instruction is used to introduce happens-before edges
7792 '``fence``' instructions take an :ref:`ordering <ordering>` argument which
7793 defines what *synchronizes-with* edges they add. They can only be given
7794 ``acquire``, ``release``, ``acq_rel``, and ``seq_cst`` orderings.
7799 A fence A which has (at least) ``release`` ordering semantics
7800 *synchronizes with* a fence B with (at least) ``acquire`` ordering
7801 semantics if and only if there exist atomic operations X and Y, both
7802 operating on some atomic object M, such that A is sequenced before X, X
7803 modifies M (either directly or through some side effect of a sequence
7804 headed by X), Y is sequenced before B, and Y observes M. This provides a
7805 *happens-before* dependency between A and B. Rather than an explicit
7806 ``fence``, one (but not both) of the atomic operations X or Y might
7807 provide a ``release`` or ``acquire`` (resp.) ordering constraint and
7808 still *synchronize-with* the explicit ``fence`` and establish the
7809 *happens-before* edge.
7811 A ``fence`` which has ``seq_cst`` ordering, in addition to having both
7812 ``acquire`` and ``release`` semantics specified above, participates in
7813 the global program order of other ``seq_cst`` operations and/or fences.
7815 A ``fence`` instruction can also take an optional
7816 ":ref:`syncscope <syncscope>`" argument.
7821 .. code-block:: text
7823 fence acquire ; yields void
7824 fence syncscope("singlethread") seq_cst ; yields void
7825 fence syncscope("agent") seq_cst ; yields void
7829 '``cmpxchg``' Instruction
7830 ^^^^^^^^^^^^^^^^^^^^^^^^^
7837 cmpxchg [weak] [volatile] <ty>* <pointer>, <ty> <cmp>, <ty> <new> [syncscope("<target-scope>")] <success ordering> <failure ordering> ; yields { ty, i1 }
7842 The '``cmpxchg``' instruction is used to atomically modify memory. It
7843 loads a value in memory and compares it to a given value. If they are
7844 equal, it tries to store a new value into the memory.
7849 There are three arguments to the '``cmpxchg``' instruction: an address
7850 to operate on, a value to compare to the value currently be at that
7851 address, and a new value to place at that address if the compared values
7852 are equal. The type of '<cmp>' must be an integer or pointer type whose
7853 bit width is a power of two greater than or equal to eight and less
7854 than or equal to a target-specific size limit. '<cmp>' and '<new>' must
7855 have the same type, and the type of '<pointer>' must be a pointer to
7856 that type. If the ``cmpxchg`` is marked as ``volatile``, then the
7857 optimizer is not allowed to modify the number or order of execution of
7858 this ``cmpxchg`` with other :ref:`volatile operations <volatile>`.
7860 The success and failure :ref:`ordering <ordering>` arguments specify how this
7861 ``cmpxchg`` synchronizes with other atomic operations. Both ordering parameters
7862 must be at least ``monotonic``, the ordering constraint on failure must be no
7863 stronger than that on success, and the failure ordering cannot be either
7864 ``release`` or ``acq_rel``.
7866 A ``cmpxchg`` instruction can also take an optional
7867 ":ref:`syncscope <syncscope>`" argument.
7869 The pointer passed into cmpxchg must have alignment greater than or
7870 equal to the size in memory of the operand.
7875 The contents of memory at the location specified by the '``<pointer>``' operand
7876 is read and compared to '``<cmp>``'; if the values are equal, '``<new>``' is
7877 written to the location. The original value at the location is returned,
7878 together with a flag indicating success (true) or failure (false).
7880 If the cmpxchg operation is marked as ``weak`` then a spurious failure is
7881 permitted: the operation may not write ``<new>`` even if the comparison
7884 If the cmpxchg operation is strong (the default), the i1 value is 1 if and only
7885 if the value loaded equals ``cmp``.
7887 A successful ``cmpxchg`` is a read-modify-write instruction for the purpose of
7888 identifying release sequences. A failed ``cmpxchg`` is equivalent to an atomic
7889 load with an ordering parameter determined the second ordering parameter.
7894 .. code-block:: llvm
7897 %orig = load atomic i32, i32* %ptr unordered, align 4 ; yields i32
7901 %cmp = phi i32 [ %orig, %entry ], [%value_loaded, %loop]
7902 %squared = mul i32 %cmp, %cmp
7903 %val_success = cmpxchg i32* %ptr, i32 %cmp, i32 %squared acq_rel monotonic ; yields { i32, i1 }
7904 %value_loaded = extractvalue { i32, i1 } %val_success, 0
7905 %success = extractvalue { i32, i1 } %val_success, 1
7906 br i1 %success, label %done, label %loop
7913 '``atomicrmw``' Instruction
7914 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7921 atomicrmw [volatile] <operation> <ty>* <pointer>, <ty> <value> [syncscope("<target-scope>")] <ordering> ; yields ty
7926 The '``atomicrmw``' instruction is used to atomically modify memory.
7931 There are three arguments to the '``atomicrmw``' instruction: an
7932 operation to apply, an address whose value to modify, an argument to the
7933 operation. The operation must be one of the following keywords:
7947 The type of '<value>' must be an integer type whose bit width is a power
7948 of two greater than or equal to eight and less than or equal to a
7949 target-specific size limit. The type of the '``<pointer>``' operand must
7950 be a pointer to that type. If the ``atomicrmw`` is marked as
7951 ``volatile``, then the optimizer is not allowed to modify the number or
7952 order of execution of this ``atomicrmw`` with other :ref:`volatile
7953 operations <volatile>`.
7955 A ``atomicrmw`` instruction can also take an optional
7956 ":ref:`syncscope <syncscope>`" argument.
7961 The contents of memory at the location specified by the '``<pointer>``'
7962 operand are atomically read, modified, and written back. The original
7963 value at the location is returned. The modification is specified by the
7966 - xchg: ``*ptr = val``
7967 - add: ``*ptr = *ptr + val``
7968 - sub: ``*ptr = *ptr - val``
7969 - and: ``*ptr = *ptr & val``
7970 - nand: ``*ptr = ~(*ptr & val)``
7971 - or: ``*ptr = *ptr | val``
7972 - xor: ``*ptr = *ptr ^ val``
7973 - max: ``*ptr = *ptr > val ? *ptr : val`` (using a signed comparison)
7974 - min: ``*ptr = *ptr < val ? *ptr : val`` (using a signed comparison)
7975 - umax: ``*ptr = *ptr > val ? *ptr : val`` (using an unsigned
7977 - umin: ``*ptr = *ptr < val ? *ptr : val`` (using an unsigned
7983 .. code-block:: llvm
7985 %old = atomicrmw add i32* %ptr, i32 1 acquire ; yields i32
7987 .. _i_getelementptr:
7989 '``getelementptr``' Instruction
7990 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7997 <result> = getelementptr <ty>, <ty>* <ptrval>{, [inrange] <ty> <idx>}*
7998 <result> = getelementptr inbounds <ty>, <ty>* <ptrval>{, [inrange] <ty> <idx>}*
7999 <result> = getelementptr <ty>, <ptr vector> <ptrval>, [inrange] <vector index type> <idx>
8004 The '``getelementptr``' instruction is used to get the address of a
8005 subelement of an :ref:`aggregate <t_aggregate>` data structure. It performs
8006 address calculation only and does not access memory. The instruction can also
8007 be used to calculate a vector of such addresses.
8012 The first argument is always a type used as the basis for the calculations.
8013 The second argument is always a pointer or a vector of pointers, and is the
8014 base address to start from. The remaining arguments are indices
8015 that indicate which of the elements of the aggregate object are indexed.
8016 The interpretation of each index is dependent on the type being indexed
8017 into. The first index always indexes the pointer value given as the
8018 second argument, the second index indexes a value of the type pointed to
8019 (not necessarily the value directly pointed to, since the first index
8020 can be non-zero), etc. The first type indexed into must be a pointer
8021 value, subsequent types can be arrays, vectors, and structs. Note that
8022 subsequent types being indexed into can never be pointers, since that
8023 would require loading the pointer before continuing calculation.
8025 The type of each index argument depends on the type it is indexing into.
8026 When indexing into a (optionally packed) structure, only ``i32`` integer
8027 **constants** are allowed (when using a vector of indices they must all
8028 be the **same** ``i32`` integer constant). When indexing into an array,
8029 pointer or vector, integers of any width are allowed, and they are not
8030 required to be constant. These integers are treated as signed values
8033 For example, let's consider a C code fragment and how it gets compiled
8049 int *foo(struct ST *s) {
8050 return &s[1].Z.B[5][13];
8053 The LLVM code generated by Clang is:
8055 .. code-block:: llvm
8057 %struct.RT = type { i8, [10 x [20 x i32]], i8 }
8058 %struct.ST = type { i32, double, %struct.RT }
8060 define i32* @foo(%struct.ST* %s) nounwind uwtable readnone optsize ssp {
8062 %arrayidx = getelementptr inbounds %struct.ST, %struct.ST* %s, i64 1, i32 2, i32 1, i64 5, i64 13
8069 In the example above, the first index is indexing into the
8070 '``%struct.ST*``' type, which is a pointer, yielding a '``%struct.ST``'
8071 = '``{ i32, double, %struct.RT }``' type, a structure. The second index
8072 indexes into the third element of the structure, yielding a
8073 '``%struct.RT``' = '``{ i8 , [10 x [20 x i32]], i8 }``' type, another
8074 structure. The third index indexes into the second element of the
8075 structure, yielding a '``[10 x [20 x i32]]``' type, an array. The two
8076 dimensions of the array are subscripted into, yielding an '``i32``'
8077 type. The '``getelementptr``' instruction returns a pointer to this
8078 element, thus computing a value of '``i32*``' type.
8080 Note that it is perfectly legal to index partially through a structure,
8081 returning a pointer to an inner element. Because of this, the LLVM code
8082 for the given testcase is equivalent to:
8084 .. code-block:: llvm
8086 define i32* @foo(%struct.ST* %s) {
8087 %t1 = getelementptr %struct.ST, %struct.ST* %s, i32 1 ; yields %struct.ST*:%t1
8088 %t2 = getelementptr %struct.ST, %struct.ST* %t1, i32 0, i32 2 ; yields %struct.RT*:%t2
8089 %t3 = getelementptr %struct.RT, %struct.RT* %t2, i32 0, i32 1 ; yields [10 x [20 x i32]]*:%t3
8090 %t4 = getelementptr [10 x [20 x i32]], [10 x [20 x i32]]* %t3, i32 0, i32 5 ; yields [20 x i32]*:%t4
8091 %t5 = getelementptr [20 x i32], [20 x i32]* %t4, i32 0, i32 13 ; yields i32*:%t5
8095 If the ``inbounds`` keyword is present, the result value of the
8096 ``getelementptr`` is a :ref:`poison value <poisonvalues>` if the base
8097 pointer is not an *in bounds* address of an allocated object, or if any
8098 of the addresses that would be formed by successive addition of the
8099 offsets implied by the indices to the base address with infinitely
8100 precise signed arithmetic are not an *in bounds* address of that
8101 allocated object. The *in bounds* addresses for an allocated object are
8102 all the addresses that point into the object, plus the address one byte
8103 past the end. The only *in bounds* address for a null pointer in the
8104 default address-space is the null pointer itself. In cases where the
8105 base is a vector of pointers the ``inbounds`` keyword applies to each
8106 of the computations element-wise.
8108 If the ``inbounds`` keyword is not present, the offsets are added to the
8109 base address with silently-wrapping two's complement arithmetic. If the
8110 offsets have a different width from the pointer, they are sign-extended
8111 or truncated to the width of the pointer. The result value of the
8112 ``getelementptr`` may be outside the object pointed to by the base
8113 pointer. The result value may not necessarily be used to access memory
8114 though, even if it happens to point into allocated storage. See the
8115 :ref:`Pointer Aliasing Rules <pointeraliasing>` section for more
8118 If the ``inrange`` keyword is present before any index, loading from or
8119 storing to any pointer derived from the ``getelementptr`` has undefined
8120 behavior if the load or store would access memory outside of the bounds of
8121 the element selected by the index marked as ``inrange``. The result of a
8122 pointer comparison or ``ptrtoint`` (including ``ptrtoint``-like operations
8123 involving memory) involving a pointer derived from a ``getelementptr`` with
8124 the ``inrange`` keyword is undefined, with the exception of comparisons
8125 in the case where both operands are in the range of the element selected
8126 by the ``inrange`` keyword, inclusive of the address one past the end of
8127 that element. Note that the ``inrange`` keyword is currently only allowed
8128 in constant ``getelementptr`` expressions.
8130 The getelementptr instruction is often confusing. For some more insight
8131 into how it works, see :doc:`the getelementptr FAQ <GetElementPtr>`.
8136 .. code-block:: llvm
8138 ; yields [12 x i8]*:aptr
8139 %aptr = getelementptr {i32, [12 x i8]}, {i32, [12 x i8]}* %saptr, i64 0, i32 1
8141 %vptr = getelementptr {i32, <2 x i8>}, {i32, <2 x i8>}* %svptr, i64 0, i32 1, i32 1
8143 %eptr = getelementptr [12 x i8], [12 x i8]* %aptr, i64 0, i32 1
8145 %iptr = getelementptr [10 x i32], [10 x i32]* @arr, i16 0, i16 0
8150 The ``getelementptr`` returns a vector of pointers, instead of a single address,
8151 when one or more of its arguments is a vector. In such cases, all vector
8152 arguments should have the same number of elements, and every scalar argument
8153 will be effectively broadcast into a vector during address calculation.
8155 .. code-block:: llvm
8157 ; All arguments are vectors:
8158 ; A[i] = ptrs[i] + offsets[i]*sizeof(i8)
8159 %A = getelementptr i8, <4 x i8*> %ptrs, <4 x i64> %offsets
8161 ; Add the same scalar offset to each pointer of a vector:
8162 ; A[i] = ptrs[i] + offset*sizeof(i8)
8163 %A = getelementptr i8, <4 x i8*> %ptrs, i64 %offset
8165 ; Add distinct offsets to the same pointer:
8166 ; A[i] = ptr + offsets[i]*sizeof(i8)
8167 %A = getelementptr i8, i8* %ptr, <4 x i64> %offsets
8169 ; In all cases described above the type of the result is <4 x i8*>
8171 The two following instructions are equivalent:
8173 .. code-block:: llvm
8175 getelementptr %struct.ST, <4 x %struct.ST*> %s, <4 x i64> %ind1,
8176 <4 x i32> <i32 2, i32 2, i32 2, i32 2>,
8177 <4 x i32> <i32 1, i32 1, i32 1, i32 1>,
8179 <4 x i64> <i64 13, i64 13, i64 13, i64 13>
8181 getelementptr %struct.ST, <4 x %struct.ST*> %s, <4 x i64> %ind1,
8182 i32 2, i32 1, <4 x i32> %ind4, i64 13
8184 Let's look at the C code, where the vector version of ``getelementptr``
8189 // Let's assume that we vectorize the following loop:
8190 double *A, *B; int *C;
8191 for (int i = 0; i < size; ++i) {
8195 .. code-block:: llvm
8197 ; get pointers for 8 elements from array B
8198 %ptrs = getelementptr double, double* %B, <8 x i32> %C
8199 ; load 8 elements from array B into A
8200 %A = call <8 x double> @llvm.masked.gather.v8f64.v8p0f64(<8 x double*> %ptrs,
8201 i32 8, <8 x i1> %mask, <8 x double> %passthru)
8203 Conversion Operations
8204 ---------------------
8206 The instructions in this category are the conversion instructions
8207 (casting) which all take a single operand and a type. They perform
8208 various bit conversions on the operand.
8212 '``trunc .. to``' Instruction
8213 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8220 <result> = trunc <ty> <value> to <ty2> ; yields ty2
8225 The '``trunc``' instruction truncates its operand to the type ``ty2``.
8230 The '``trunc``' instruction takes a value to trunc, and a type to trunc
8231 it to. Both types must be of :ref:`integer <t_integer>` types, or vectors
8232 of the same number of integers. The bit size of the ``value`` must be
8233 larger than the bit size of the destination type, ``ty2``. Equal sized
8234 types are not allowed.
8239 The '``trunc``' instruction truncates the high order bits in ``value``
8240 and converts the remaining bits to ``ty2``. Since the source size must
8241 be larger than the destination size, ``trunc`` cannot be a *no-op cast*.
8242 It will always truncate bits.
8247 .. code-block:: llvm
8249 %X = trunc i32 257 to i8 ; yields i8:1
8250 %Y = trunc i32 123 to i1 ; yields i1:true
8251 %Z = trunc i32 122 to i1 ; yields i1:false
8252 %W = trunc <2 x i16> <i16 8, i16 7> to <2 x i8> ; yields <i8 8, i8 7>
8256 '``zext .. to``' Instruction
8257 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8264 <result> = zext <ty> <value> to <ty2> ; yields ty2
8269 The '``zext``' instruction zero extends its operand to type ``ty2``.
8274 The '``zext``' instruction takes a value to cast, and a type to cast it
8275 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
8276 the same number of integers. The bit size of the ``value`` must be
8277 smaller than the bit size of the destination type, ``ty2``.
8282 The ``zext`` fills the high order bits of the ``value`` with zero bits
8283 until it reaches the size of the destination type, ``ty2``.
8285 When zero extending from i1, the result will always be either 0 or 1.
8290 .. code-block:: llvm
8292 %X = zext i32 257 to i64 ; yields i64:257
8293 %Y = zext i1 true to i32 ; yields i32:1
8294 %Z = zext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
8298 '``sext .. to``' Instruction
8299 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8306 <result> = sext <ty> <value> to <ty2> ; yields ty2
8311 The '``sext``' sign extends ``value`` to the type ``ty2``.
8316 The '``sext``' instruction takes a value to cast, and a type to cast it
8317 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
8318 the same number of integers. The bit size of the ``value`` must be
8319 smaller than the bit size of the destination type, ``ty2``.
8324 The '``sext``' instruction performs a sign extension by copying the sign
8325 bit (highest order bit) of the ``value`` until it reaches the bit size
8326 of the type ``ty2``.
8328 When sign extending from i1, the extension always results in -1 or 0.
8333 .. code-block:: llvm
8335 %X = sext i8 -1 to i16 ; yields i16 :65535
8336 %Y = sext i1 true to i32 ; yields i32:-1
8337 %Z = sext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
8339 '``fptrunc .. to``' Instruction
8340 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8347 <result> = fptrunc <ty> <value> to <ty2> ; yields ty2
8352 The '``fptrunc``' instruction truncates ``value`` to type ``ty2``.
8357 The '``fptrunc``' instruction takes a :ref:`floating point <t_floating>`
8358 value to cast and a :ref:`floating point <t_floating>` type to cast it to.
8359 The size of ``value`` must be larger than the size of ``ty2``. This
8360 implies that ``fptrunc`` cannot be used to make a *no-op cast*.
8365 The '``fptrunc``' instruction casts a ``value`` from a larger
8366 :ref:`floating point <t_floating>` type to a smaller :ref:`floating
8367 point <t_floating>` type. If the value cannot fit (i.e. overflows) within the
8368 destination type, ``ty2``, then the results are undefined. If the cast produces
8369 an inexact result, how rounding is performed (e.g. truncation, also known as
8370 round to zero) is undefined.
8375 .. code-block:: llvm
8377 %X = fptrunc double 123.0 to float ; yields float:123.0
8378 %Y = fptrunc double 1.0E+300 to float ; yields undefined
8380 '``fpext .. to``' Instruction
8381 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8388 <result> = fpext <ty> <value> to <ty2> ; yields ty2
8393 The '``fpext``' extends a floating point ``value`` to a larger floating
8399 The '``fpext``' instruction takes a :ref:`floating point <t_floating>`
8400 ``value`` to cast, and a :ref:`floating point <t_floating>` type to cast it
8401 to. The source type must be smaller than the destination type.
8406 The '``fpext``' instruction extends the ``value`` from a smaller
8407 :ref:`floating point <t_floating>` type to a larger :ref:`floating
8408 point <t_floating>` type. The ``fpext`` cannot be used to make a
8409 *no-op cast* because it always changes bits. Use ``bitcast`` to make a
8410 *no-op cast* for a floating point cast.
8415 .. code-block:: llvm
8417 %X = fpext float 3.125 to double ; yields double:3.125000e+00
8418 %Y = fpext double %X to fp128 ; yields fp128:0xL00000000000000004000900000000000
8420 '``fptoui .. to``' Instruction
8421 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8428 <result> = fptoui <ty> <value> to <ty2> ; yields ty2
8433 The '``fptoui``' converts a floating point ``value`` to its unsigned
8434 integer equivalent of type ``ty2``.
8439 The '``fptoui``' instruction takes a value to cast, which must be a
8440 scalar or vector :ref:`floating point <t_floating>` value, and a type to
8441 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
8442 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
8443 type with the same number of elements as ``ty``
8448 The '``fptoui``' instruction converts its :ref:`floating
8449 point <t_floating>` operand into the nearest (rounding towards zero)
8450 unsigned integer value. If the value cannot fit in ``ty2``, the results
8456 .. code-block:: llvm
8458 %X = fptoui double 123.0 to i32 ; yields i32:123
8459 %Y = fptoui float 1.0E+300 to i1 ; yields undefined:1
8460 %Z = fptoui float 1.04E+17 to i8 ; yields undefined:1
8462 '``fptosi .. to``' Instruction
8463 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8470 <result> = fptosi <ty> <value> to <ty2> ; yields ty2
8475 The '``fptosi``' instruction converts :ref:`floating point <t_floating>`
8476 ``value`` to type ``ty2``.
8481 The '``fptosi``' instruction takes a value to cast, which must be a
8482 scalar or vector :ref:`floating point <t_floating>` value, and a type to
8483 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
8484 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
8485 type with the same number of elements as ``ty``
8490 The '``fptosi``' instruction converts its :ref:`floating
8491 point <t_floating>` operand into the nearest (rounding towards zero)
8492 signed integer value. If the value cannot fit in ``ty2``, the results
8498 .. code-block:: llvm
8500 %X = fptosi double -123.0 to i32 ; yields i32:-123
8501 %Y = fptosi float 1.0E-247 to i1 ; yields undefined:1
8502 %Z = fptosi float 1.04E+17 to i8 ; yields undefined:1
8504 '``uitofp .. to``' Instruction
8505 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8512 <result> = uitofp <ty> <value> to <ty2> ; yields ty2
8517 The '``uitofp``' instruction regards ``value`` as an unsigned integer
8518 and converts that value to the ``ty2`` type.
8523 The '``uitofp``' instruction takes a value to cast, which must be a
8524 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
8525 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
8526 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
8527 type with the same number of elements as ``ty``
8532 The '``uitofp``' instruction interprets its operand as an unsigned
8533 integer quantity and converts it to the corresponding floating point
8534 value. If the value cannot fit in the floating point value, the results
8540 .. code-block:: llvm
8542 %X = uitofp i32 257 to float ; yields float:257.0
8543 %Y = uitofp i8 -1 to double ; yields double:255.0
8545 '``sitofp .. to``' Instruction
8546 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8553 <result> = sitofp <ty> <value> to <ty2> ; yields ty2
8558 The '``sitofp``' instruction regards ``value`` as a signed integer and
8559 converts that value to the ``ty2`` type.
8564 The '``sitofp``' instruction takes a value to cast, which must be a
8565 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
8566 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
8567 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
8568 type with the same number of elements as ``ty``
8573 The '``sitofp``' instruction interprets its operand as a signed integer
8574 quantity and converts it to the corresponding floating point value. If
8575 the value cannot fit in the floating point value, the results are
8581 .. code-block:: llvm
8583 %X = sitofp i32 257 to float ; yields float:257.0
8584 %Y = sitofp i8 -1 to double ; yields double:-1.0
8588 '``ptrtoint .. to``' Instruction
8589 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8596 <result> = ptrtoint <ty> <value> to <ty2> ; yields ty2
8601 The '``ptrtoint``' instruction converts the pointer or a vector of
8602 pointers ``value`` to the integer (or vector of integers) type ``ty2``.
8607 The '``ptrtoint``' instruction takes a ``value`` to cast, which must be
8608 a value of type :ref:`pointer <t_pointer>` or a vector of pointers, and a
8609 type to cast it to ``ty2``, which must be an :ref:`integer <t_integer>` or
8610 a vector of integers type.
8615 The '``ptrtoint``' instruction converts ``value`` to integer type
8616 ``ty2`` by interpreting the pointer value as an integer and either
8617 truncating or zero extending that value to the size of the integer type.
8618 If ``value`` is smaller than ``ty2`` then a zero extension is done. If
8619 ``value`` is larger than ``ty2`` then a truncation is done. If they are
8620 the same size, then nothing is done (*no-op cast*) other than a type
8626 .. code-block:: llvm
8628 %X = ptrtoint i32* %P to i8 ; yields truncation on 32-bit architecture
8629 %Y = ptrtoint i32* %P to i64 ; yields zero extension on 32-bit architecture
8630 %Z = ptrtoint <4 x i32*> %P to <4 x i64>; yields vector zero extension for a vector of addresses on 32-bit architecture
8634 '``inttoptr .. to``' Instruction
8635 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8642 <result> = inttoptr <ty> <value> to <ty2> ; yields ty2
8647 The '``inttoptr``' instruction converts an integer ``value`` to a
8648 pointer type, ``ty2``.
8653 The '``inttoptr``' instruction takes an :ref:`integer <t_integer>` value to
8654 cast, and a type to cast it to, which must be a :ref:`pointer <t_pointer>`
8660 The '``inttoptr``' instruction converts ``value`` to type ``ty2`` by
8661 applying either a zero extension or a truncation depending on the size
8662 of the integer ``value``. If ``value`` is larger than the size of a
8663 pointer then a truncation is done. If ``value`` is smaller than the size
8664 of a pointer then a zero extension is done. If they are the same size,
8665 nothing is done (*no-op cast*).
8670 .. code-block:: llvm
8672 %X = inttoptr i32 255 to i32* ; yields zero extension on 64-bit architecture
8673 %Y = inttoptr i32 255 to i32* ; yields no-op on 32-bit architecture
8674 %Z = inttoptr i64 0 to i32* ; yields truncation on 32-bit architecture
8675 %Z = inttoptr <4 x i32> %G to <4 x i8*>; yields truncation of vector G to four pointers
8679 '``bitcast .. to``' Instruction
8680 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8687 <result> = bitcast <ty> <value> to <ty2> ; yields ty2
8692 The '``bitcast``' instruction converts ``value`` to type ``ty2`` without
8698 The '``bitcast``' instruction takes a value to cast, which must be a
8699 non-aggregate first class value, and a type to cast it to, which must
8700 also be a non-aggregate :ref:`first class <t_firstclass>` type. The
8701 bit sizes of ``value`` and the destination type, ``ty2``, must be
8702 identical. If the source type is a pointer, the destination type must
8703 also be a pointer of the same size. This instruction supports bitwise
8704 conversion of vectors to integers and to vectors of other types (as
8705 long as they have the same size).
8710 The '``bitcast``' instruction converts ``value`` to type ``ty2``. It
8711 is always a *no-op cast* because no bits change with this
8712 conversion. The conversion is done as if the ``value`` had been stored
8713 to memory and read back as type ``ty2``. Pointer (or vector of
8714 pointers) types may only be converted to other pointer (or vector of
8715 pointers) types with the same address space through this instruction.
8716 To convert pointers to other types, use the :ref:`inttoptr <i_inttoptr>`
8717 or :ref:`ptrtoint <i_ptrtoint>` instructions first.
8722 .. code-block:: text
8724 %X = bitcast i8 255 to i8 ; yields i8 :-1
8725 %Y = bitcast i32* %x to sint* ; yields sint*:%x
8726 %Z = bitcast <2 x int> %V to i64; ; yields i64: %V
8727 %Z = bitcast <2 x i32*> %V to <2 x i64*> ; yields <2 x i64*>
8729 .. _i_addrspacecast:
8731 '``addrspacecast .. to``' Instruction
8732 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8739 <result> = addrspacecast <pty> <ptrval> to <pty2> ; yields pty2
8744 The '``addrspacecast``' instruction converts ``ptrval`` from ``pty`` in
8745 address space ``n`` to type ``pty2`` in address space ``m``.
8750 The '``addrspacecast``' instruction takes a pointer or vector of pointer value
8751 to cast and a pointer type to cast it to, which must have a different
8757 The '``addrspacecast``' instruction converts the pointer value
8758 ``ptrval`` to type ``pty2``. It can be a *no-op cast* or a complex
8759 value modification, depending on the target and the address space
8760 pair. Pointer conversions within the same address space must be
8761 performed with the ``bitcast`` instruction. Note that if the address space
8762 conversion is legal then both result and operand refer to the same memory
8768 .. code-block:: llvm
8770 %X = addrspacecast i32* %x to i32 addrspace(1)* ; yields i32 addrspace(1)*:%x
8771 %Y = addrspacecast i32 addrspace(1)* %y to i64 addrspace(2)* ; yields i64 addrspace(2)*:%y
8772 %Z = addrspacecast <4 x i32*> %z to <4 x float addrspace(3)*> ; yields <4 x float addrspace(3)*>:%z
8779 The instructions in this category are the "miscellaneous" instructions,
8780 which defy better classification.
8784 '``icmp``' Instruction
8785 ^^^^^^^^^^^^^^^^^^^^^^
8792 <result> = icmp <cond> <ty> <op1>, <op2> ; yields i1 or <N x i1>:result
8797 The '``icmp``' instruction returns a boolean value or a vector of
8798 boolean values based on comparison of its two integer, integer vector,
8799 pointer, or pointer vector operands.
8804 The '``icmp``' instruction takes three operands. The first operand is
8805 the condition code indicating the kind of comparison to perform. It is
8806 not a value, just a keyword. The possible condition codes are:
8809 #. ``ne``: not equal
8810 #. ``ugt``: unsigned greater than
8811 #. ``uge``: unsigned greater or equal
8812 #. ``ult``: unsigned less than
8813 #. ``ule``: unsigned less or equal
8814 #. ``sgt``: signed greater than
8815 #. ``sge``: signed greater or equal
8816 #. ``slt``: signed less than
8817 #. ``sle``: signed less or equal
8819 The remaining two arguments must be :ref:`integer <t_integer>` or
8820 :ref:`pointer <t_pointer>` or integer :ref:`vector <t_vector>` typed. They
8821 must also be identical types.
8826 The '``icmp``' compares ``op1`` and ``op2`` according to the condition
8827 code given as ``cond``. The comparison performed always yields either an
8828 :ref:`i1 <t_integer>` or vector of ``i1`` result, as follows:
8830 #. ``eq``: yields ``true`` if the operands are equal, ``false``
8831 otherwise. No sign interpretation is necessary or performed.
8832 #. ``ne``: yields ``true`` if the operands are unequal, ``false``
8833 otherwise. No sign interpretation is necessary or performed.
8834 #. ``ugt``: interprets the operands as unsigned values and yields
8835 ``true`` if ``op1`` is greater than ``op2``.
8836 #. ``uge``: interprets the operands as unsigned values and yields
8837 ``true`` if ``op1`` is greater than or equal to ``op2``.
8838 #. ``ult``: interprets the operands as unsigned values and yields
8839 ``true`` if ``op1`` is less than ``op2``.
8840 #. ``ule``: interprets the operands as unsigned values and yields
8841 ``true`` if ``op1`` is less than or equal to ``op2``.
8842 #. ``sgt``: interprets the operands as signed values and yields ``true``
8843 if ``op1`` is greater than ``op2``.
8844 #. ``sge``: interprets the operands as signed values and yields ``true``
8845 if ``op1`` is greater than or equal to ``op2``.
8846 #. ``slt``: interprets the operands as signed values and yields ``true``
8847 if ``op1`` is less than ``op2``.
8848 #. ``sle``: interprets the operands as signed values and yields ``true``
8849 if ``op1`` is less than or equal to ``op2``.
8851 If the operands are :ref:`pointer <t_pointer>` typed, the pointer values
8852 are compared as if they were integers.
8854 If the operands are integer vectors, then they are compared element by
8855 element. The result is an ``i1`` vector with the same number of elements
8856 as the values being compared. Otherwise, the result is an ``i1``.
8861 .. code-block:: text
8863 <result> = icmp eq i32 4, 5 ; yields: result=false
8864 <result> = icmp ne float* %X, %X ; yields: result=false
8865 <result> = icmp ult i16 4, 5 ; yields: result=true
8866 <result> = icmp sgt i16 4, 5 ; yields: result=false
8867 <result> = icmp ule i16 -4, 5 ; yields: result=false
8868 <result> = icmp sge i16 4, 5 ; yields: result=false
8872 '``fcmp``' Instruction
8873 ^^^^^^^^^^^^^^^^^^^^^^
8880 <result> = fcmp [fast-math flags]* <cond> <ty> <op1>, <op2> ; yields i1 or <N x i1>:result
8885 The '``fcmp``' instruction returns a boolean value or vector of boolean
8886 values based on comparison of its operands.
8888 If the operands are floating point scalars, then the result type is a
8889 boolean (:ref:`i1 <t_integer>`).
8891 If the operands are floating point vectors, then the result type is a
8892 vector of boolean with the same number of elements as the operands being
8898 The '``fcmp``' instruction takes three operands. The first operand is
8899 the condition code indicating the kind of comparison to perform. It is
8900 not a value, just a keyword. The possible condition codes are:
8902 #. ``false``: no comparison, always returns false
8903 #. ``oeq``: ordered and equal
8904 #. ``ogt``: ordered and greater than
8905 #. ``oge``: ordered and greater than or equal
8906 #. ``olt``: ordered and less than
8907 #. ``ole``: ordered and less than or equal
8908 #. ``one``: ordered and not equal
8909 #. ``ord``: ordered (no nans)
8910 #. ``ueq``: unordered or equal
8911 #. ``ugt``: unordered or greater than
8912 #. ``uge``: unordered or greater than or equal
8913 #. ``ult``: unordered or less than
8914 #. ``ule``: unordered or less than or equal
8915 #. ``une``: unordered or not equal
8916 #. ``uno``: unordered (either nans)
8917 #. ``true``: no comparison, always returns true
8919 *Ordered* means that neither operand is a QNAN while *unordered* means
8920 that either operand may be a QNAN.
8922 Each of ``val1`` and ``val2`` arguments must be either a :ref:`floating
8923 point <t_floating>` type or a :ref:`vector <t_vector>` of floating point
8924 type. They must have identical types.
8929 The '``fcmp``' instruction compares ``op1`` and ``op2`` according to the
8930 condition code given as ``cond``. If the operands are vectors, then the
8931 vectors are compared element by element. Each comparison performed
8932 always yields an :ref:`i1 <t_integer>` result, as follows:
8934 #. ``false``: always yields ``false``, regardless of operands.
8935 #. ``oeq``: yields ``true`` if both operands are not a QNAN and ``op1``
8936 is equal to ``op2``.
8937 #. ``ogt``: yields ``true`` if both operands are not a QNAN and ``op1``
8938 is greater than ``op2``.
8939 #. ``oge``: yields ``true`` if both operands are not a QNAN and ``op1``
8940 is greater than or equal to ``op2``.
8941 #. ``olt``: yields ``true`` if both operands are not a QNAN and ``op1``
8942 is less than ``op2``.
8943 #. ``ole``: yields ``true`` if both operands are not a QNAN and ``op1``
8944 is less than or equal to ``op2``.
8945 #. ``one``: yields ``true`` if both operands are not a QNAN and ``op1``
8946 is not equal to ``op2``.
8947 #. ``ord``: yields ``true`` if both operands are not a QNAN.
8948 #. ``ueq``: yields ``true`` if either operand is a QNAN or ``op1`` is
8950 #. ``ugt``: yields ``true`` if either operand is a QNAN or ``op1`` is
8951 greater than ``op2``.
8952 #. ``uge``: yields ``true`` if either operand is a QNAN or ``op1`` is
8953 greater than or equal to ``op2``.
8954 #. ``ult``: yields ``true`` if either operand is a QNAN or ``op1`` is
8956 #. ``ule``: yields ``true`` if either operand is a QNAN or ``op1`` is
8957 less than or equal to ``op2``.
8958 #. ``une``: yields ``true`` if either operand is a QNAN or ``op1`` is
8959 not equal to ``op2``.
8960 #. ``uno``: yields ``true`` if either operand is a QNAN.
8961 #. ``true``: always yields ``true``, regardless of operands.
8963 The ``fcmp`` instruction can also optionally take any number of
8964 :ref:`fast-math flags <fastmath>`, which are optimization hints to enable
8965 otherwise unsafe floating point optimizations.
8967 Any set of fast-math flags are legal on an ``fcmp`` instruction, but the
8968 only flags that have any effect on its semantics are those that allow
8969 assumptions to be made about the values of input arguments; namely
8970 ``nnan``, ``ninf``, and ``nsz``. See :ref:`fastmath` for more information.
8975 .. code-block:: text
8977 <result> = fcmp oeq float 4.0, 5.0 ; yields: result=false
8978 <result> = fcmp one float 4.0, 5.0 ; yields: result=true
8979 <result> = fcmp olt float 4.0, 5.0 ; yields: result=true
8980 <result> = fcmp ueq double 1.0, 2.0 ; yields: result=false
8984 '``phi``' Instruction
8985 ^^^^^^^^^^^^^^^^^^^^^
8992 <result> = phi <ty> [ <val0>, <label0>], ...
8997 The '``phi``' instruction is used to implement the φ node in the SSA
8998 graph representing the function.
9003 The type of the incoming values is specified with the first type field.
9004 After this, the '``phi``' instruction takes a list of pairs as
9005 arguments, with one pair for each predecessor basic block of the current
9006 block. Only values of :ref:`first class <t_firstclass>` type may be used as
9007 the value arguments to the PHI node. Only labels may be used as the
9010 There must be no non-phi instructions between the start of a basic block
9011 and the PHI instructions: i.e. PHI instructions must be first in a basic
9014 For the purposes of the SSA form, the use of each incoming value is
9015 deemed to occur on the edge from the corresponding predecessor block to
9016 the current block (but after any definition of an '``invoke``'
9017 instruction's return value on the same edge).
9022 At runtime, the '``phi``' instruction logically takes on the value
9023 specified by the pair corresponding to the predecessor basic block that
9024 executed just prior to the current block.
9029 .. code-block:: llvm
9031 Loop: ; Infinite loop that counts from 0 on up...
9032 %indvar = phi i32 [ 0, %LoopHeader ], [ %nextindvar, %Loop ]
9033 %nextindvar = add i32 %indvar, 1
9038 '``select``' Instruction
9039 ^^^^^^^^^^^^^^^^^^^^^^^^
9046 <result> = select selty <cond>, <ty> <val1>, <ty> <val2> ; yields ty
9048 selty is either i1 or {<N x i1>}
9053 The '``select``' instruction is used to choose one value based on a
9054 condition, without IR-level branching.
9059 The '``select``' instruction requires an 'i1' value or a vector of 'i1'
9060 values indicating the condition, and two values of the same :ref:`first
9061 class <t_firstclass>` type.
9066 If the condition is an i1 and it evaluates to 1, the instruction returns
9067 the first value argument; otherwise, it returns the second value
9070 If the condition is a vector of i1, then the value arguments must be
9071 vectors of the same size, and the selection is done element by element.
9073 If the condition is an i1 and the value arguments are vectors of the
9074 same size, then an entire vector is selected.
9079 .. code-block:: llvm
9081 %X = select i1 true, i8 17, i8 42 ; yields i8:17
9085 '``call``' Instruction
9086 ^^^^^^^^^^^^^^^^^^^^^^
9093 <result> = [tail | musttail | notail ] call [fast-math flags] [cconv] [ret attrs] <ty>|<fnty> <fnptrval>(<function args>) [fn attrs]
9099 The '``call``' instruction represents a simple function call.
9104 This instruction requires several arguments:
9106 #. The optional ``tail`` and ``musttail`` markers indicate that the optimizers
9107 should perform tail call optimization. The ``tail`` marker is a hint that
9108 `can be ignored <CodeGenerator.html#sibcallopt>`_. The ``musttail`` marker
9109 means that the call must be tail call optimized in order for the program to
9110 be correct. The ``musttail`` marker provides these guarantees:
9112 #. The call will not cause unbounded stack growth if it is part of a
9113 recursive cycle in the call graph.
9114 #. Arguments with the :ref:`inalloca <attr_inalloca>` attribute are
9117 Both markers imply that the callee does not access allocas from the caller.
9118 The ``tail`` marker additionally implies that the callee does not access
9119 varargs from the caller, while ``musttail`` implies that varargs from the
9120 caller are passed to the callee. Calls marked ``musttail`` must obey the
9121 following additional rules:
9123 - The call must immediately precede a :ref:`ret <i_ret>` instruction,
9124 or a pointer bitcast followed by a ret instruction.
9125 - The ret instruction must return the (possibly bitcasted) value
9126 produced by the call or void.
9127 - The caller and callee prototypes must match. Pointer types of
9128 parameters or return types may differ in pointee type, but not
9130 - The calling conventions of the caller and callee must match.
9131 - All ABI-impacting function attributes, such as sret, byval, inreg,
9132 returned, and inalloca, must match.
9133 - The callee must be varargs iff the caller is varargs. Bitcasting a
9134 non-varargs function to the appropriate varargs type is legal so
9135 long as the non-varargs prefixes obey the other rules.
9137 Tail call optimization for calls marked ``tail`` is guaranteed to occur if
9138 the following conditions are met:
9140 - Caller and callee both have the calling convention ``fastcc``.
9141 - The call is in tail position (ret immediately follows call and ret
9142 uses value of call or is void).
9143 - Option ``-tailcallopt`` is enabled, or
9144 ``llvm::GuaranteedTailCallOpt`` is ``true``.
9145 - `Platform-specific constraints are
9146 met. <CodeGenerator.html#tailcallopt>`_
9148 #. The optional ``notail`` marker indicates that the optimizers should not add
9149 ``tail`` or ``musttail`` markers to the call. It is used to prevent tail
9150 call optimization from being performed on the call.
9152 #. The optional ``fast-math flags`` marker indicates that the call has one or more
9153 :ref:`fast-math flags <fastmath>`, which are optimization hints to enable
9154 otherwise unsafe floating-point optimizations. Fast-math flags are only valid
9155 for calls that return a floating-point scalar or vector type.
9157 #. The optional "cconv" marker indicates which :ref:`calling
9158 convention <callingconv>` the call should use. If none is
9159 specified, the call defaults to using C calling conventions. The
9160 calling convention of the call must match the calling convention of
9161 the target function, or else the behavior is undefined.
9162 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
9163 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
9165 #. '``ty``': the type of the call instruction itself which is also the
9166 type of the return value. Functions that return no value are marked
9168 #. '``fnty``': shall be the signature of the function being called. The
9169 argument types must match the types implied by this signature. This
9170 type can be omitted if the function is not varargs.
9171 #. '``fnptrval``': An LLVM value containing a pointer to a function to
9172 be called. In most cases, this is a direct function call, but
9173 indirect ``call``'s are just as possible, calling an arbitrary pointer
9175 #. '``function args``': argument list whose types match the function
9176 signature argument types and parameter attributes. All arguments must
9177 be of :ref:`first class <t_firstclass>` type. If the function signature
9178 indicates the function accepts a variable number of arguments, the
9179 extra arguments can be specified.
9180 #. The optional :ref:`function attributes <fnattrs>` list.
9181 #. The optional :ref:`operand bundles <opbundles>` list.
9186 The '``call``' instruction is used to cause control flow to transfer to
9187 a specified function, with its incoming arguments bound to the specified
9188 values. Upon a '``ret``' instruction in the called function, control
9189 flow continues with the instruction after the function call, and the
9190 return value of the function is bound to the result argument.
9195 .. code-block:: llvm
9197 %retval = call i32 @test(i32 %argc)
9198 call i32 (i8*, ...)* @printf(i8* %msg, i32 12, i8 42) ; yields i32
9199 %X = tail call i32 @foo() ; yields i32
9200 %Y = tail call fastcc i32 @foo() ; yields i32
9201 call void %foo(i8 97 signext)
9203 %struct.A = type { i32, i8 }
9204 %r = call %struct.A @foo() ; yields { i32, i8 }
9205 %gr = extractvalue %struct.A %r, 0 ; yields i32
9206 %gr1 = extractvalue %struct.A %r, 1 ; yields i8
9207 %Z = call void @foo() noreturn ; indicates that %foo never returns normally
9208 %ZZ = call zeroext i32 @bar() ; Return value is %zero extended
9210 llvm treats calls to some functions with names and arguments that match
9211 the standard C99 library as being the C99 library functions, and may
9212 perform optimizations or generate code for them under that assumption.
9213 This is something we'd like to change in the future to provide better
9214 support for freestanding environments and non-C-based languages.
9218 '``va_arg``' Instruction
9219 ^^^^^^^^^^^^^^^^^^^^^^^^
9226 <resultval> = va_arg <va_list*> <arglist>, <argty>
9231 The '``va_arg``' instruction is used to access arguments passed through
9232 the "variable argument" area of a function call. It is used to implement
9233 the ``va_arg`` macro in C.
9238 This instruction takes a ``va_list*`` value and the type of the
9239 argument. It returns a value of the specified argument type and
9240 increments the ``va_list`` to point to the next argument. The actual
9241 type of ``va_list`` is target specific.
9246 The '``va_arg``' instruction loads an argument of the specified type
9247 from the specified ``va_list`` and causes the ``va_list`` to point to
9248 the next argument. For more information, see the variable argument
9249 handling :ref:`Intrinsic Functions <int_varargs>`.
9251 It is legal for this instruction to be called in a function which does
9252 not take a variable number of arguments, for example, the ``vfprintf``
9255 ``va_arg`` is an LLVM instruction instead of an :ref:`intrinsic
9256 function <intrinsics>` because it takes a type as an argument.
9261 See the :ref:`variable argument processing <int_varargs>` section.
9263 Note that the code generator does not yet fully support va\_arg on many
9264 targets. Also, it does not currently support va\_arg with aggregate
9265 types on any target.
9269 '``landingpad``' Instruction
9270 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9277 <resultval> = landingpad <resultty> <clause>+
9278 <resultval> = landingpad <resultty> cleanup <clause>*
9280 <clause> := catch <type> <value>
9281 <clause> := filter <array constant type> <array constant>
9286 The '``landingpad``' instruction is used by `LLVM's exception handling
9287 system <ExceptionHandling.html#overview>`_ to specify that a basic block
9288 is a landing pad --- one where the exception lands, and corresponds to the
9289 code found in the ``catch`` portion of a ``try``/``catch`` sequence. It
9290 defines values supplied by the :ref:`personality function <personalityfn>` upon
9291 re-entry to the function. The ``resultval`` has the type ``resultty``.
9297 ``cleanup`` flag indicates that the landing pad block is a cleanup.
9299 A ``clause`` begins with the clause type --- ``catch`` or ``filter`` --- and
9300 contains the global variable representing the "type" that may be caught
9301 or filtered respectively. Unlike the ``catch`` clause, the ``filter``
9302 clause takes an array constant as its argument. Use
9303 "``[0 x i8**] undef``" for a filter which cannot throw. The
9304 '``landingpad``' instruction must contain *at least* one ``clause`` or
9305 the ``cleanup`` flag.
9310 The '``landingpad``' instruction defines the values which are set by the
9311 :ref:`personality function <personalityfn>` upon re-entry to the function, and
9312 therefore the "result type" of the ``landingpad`` instruction. As with
9313 calling conventions, how the personality function results are
9314 represented in LLVM IR is target specific.
9316 The clauses are applied in order from top to bottom. If two
9317 ``landingpad`` instructions are merged together through inlining, the
9318 clauses from the calling function are appended to the list of clauses.
9319 When the call stack is being unwound due to an exception being thrown,
9320 the exception is compared against each ``clause`` in turn. If it doesn't
9321 match any of the clauses, and the ``cleanup`` flag is not set, then
9322 unwinding continues further up the call stack.
9324 The ``landingpad`` instruction has several restrictions:
9326 - A landing pad block is a basic block which is the unwind destination
9327 of an '``invoke``' instruction.
9328 - A landing pad block must have a '``landingpad``' instruction as its
9329 first non-PHI instruction.
9330 - There can be only one '``landingpad``' instruction within the landing
9332 - A basic block that is not a landing pad block may not include a
9333 '``landingpad``' instruction.
9338 .. code-block:: llvm
9340 ;; A landing pad which can catch an integer.
9341 %res = landingpad { i8*, i32 }
9343 ;; A landing pad that is a cleanup.
9344 %res = landingpad { i8*, i32 }
9346 ;; A landing pad which can catch an integer and can only throw a double.
9347 %res = landingpad { i8*, i32 }
9349 filter [1 x i8**] [@_ZTId]
9353 '``catchpad``' Instruction
9354 ^^^^^^^^^^^^^^^^^^^^^^^^^^
9361 <resultval> = catchpad within <catchswitch> [<args>*]
9366 The '``catchpad``' instruction is used by `LLVM's exception handling
9367 system <ExceptionHandling.html#overview>`_ to specify that a basic block
9368 begins a catch handler --- one where a personality routine attempts to transfer
9369 control to catch an exception.
9374 The ``catchswitch`` operand must always be a token produced by a
9375 :ref:`catchswitch <i_catchswitch>` instruction in a predecessor block. This
9376 ensures that each ``catchpad`` has exactly one predecessor block, and it always
9377 terminates in a ``catchswitch``.
9379 The ``args`` correspond to whatever information the personality routine
9380 requires to know if this is an appropriate handler for the exception. Control
9381 will transfer to the ``catchpad`` if this is the first appropriate handler for
9384 The ``resultval`` has the type :ref:`token <t_token>` and is used to match the
9385 ``catchpad`` to corresponding :ref:`catchrets <i_catchret>` and other nested EH
9391 When the call stack is being unwound due to an exception being thrown, the
9392 exception is compared against the ``args``. If it doesn't match, control will
9393 not reach the ``catchpad`` instruction. The representation of ``args`` is
9394 entirely target and personality function-specific.
9396 Like the :ref:`landingpad <i_landingpad>` instruction, the ``catchpad``
9397 instruction must be the first non-phi of its parent basic block.
9399 The meaning of the tokens produced and consumed by ``catchpad`` and other "pad"
9400 instructions is described in the
9401 `Windows exception handling documentation\ <ExceptionHandling.html#wineh>`_.
9403 When a ``catchpad`` has been "entered" but not yet "exited" (as
9404 described in the `EH documentation\ <ExceptionHandling.html#wineh-constraints>`_),
9405 it is undefined behavior to execute a :ref:`call <i_call>` or :ref:`invoke <i_invoke>`
9406 that does not carry an appropriate :ref:`"funclet" bundle <ob_funclet>`.
9411 .. code-block:: text
9414 %cs = catchswitch within none [label %handler0] unwind to caller
9415 ;; A catch block which can catch an integer.
9417 %tok = catchpad within %cs [i8** @_ZTIi]
9421 '``cleanuppad``' Instruction
9422 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9429 <resultval> = cleanuppad within <parent> [<args>*]
9434 The '``cleanuppad``' instruction is used by `LLVM's exception handling
9435 system <ExceptionHandling.html#overview>`_ to specify that a basic block
9436 is a cleanup block --- one where a personality routine attempts to
9437 transfer control to run cleanup actions.
9438 The ``args`` correspond to whatever additional
9439 information the :ref:`personality function <personalityfn>` requires to
9440 execute the cleanup.
9441 The ``resultval`` has the type :ref:`token <t_token>` and is used to
9442 match the ``cleanuppad`` to corresponding :ref:`cleanuprets <i_cleanupret>`.
9443 The ``parent`` argument is the token of the funclet that contains the
9444 ``cleanuppad`` instruction. If the ``cleanuppad`` is not inside a funclet,
9445 this operand may be the token ``none``.
9450 The instruction takes a list of arbitrary values which are interpreted
9451 by the :ref:`personality function <personalityfn>`.
9456 When the call stack is being unwound due to an exception being thrown,
9457 the :ref:`personality function <personalityfn>` transfers control to the
9458 ``cleanuppad`` with the aid of the personality-specific arguments.
9459 As with calling conventions, how the personality function results are
9460 represented in LLVM IR is target specific.
9462 The ``cleanuppad`` instruction has several restrictions:
9464 - A cleanup block is a basic block which is the unwind destination of
9465 an exceptional instruction.
9466 - A cleanup block must have a '``cleanuppad``' instruction as its
9467 first non-PHI instruction.
9468 - There can be only one '``cleanuppad``' instruction within the
9470 - A basic block that is not a cleanup block may not include a
9471 '``cleanuppad``' instruction.
9473 When a ``cleanuppad`` has been "entered" but not yet "exited" (as
9474 described in the `EH documentation\ <ExceptionHandling.html#wineh-constraints>`_),
9475 it is undefined behavior to execute a :ref:`call <i_call>` or :ref:`invoke <i_invoke>`
9476 that does not carry an appropriate :ref:`"funclet" bundle <ob_funclet>`.
9481 .. code-block:: text
9483 %tok = cleanuppad within %cs []
9490 LLVM supports the notion of an "intrinsic function". These functions
9491 have well known names and semantics and are required to follow certain
9492 restrictions. Overall, these intrinsics represent an extension mechanism
9493 for the LLVM language that does not require changing all of the
9494 transformations in LLVM when adding to the language (or the bitcode
9495 reader/writer, the parser, etc...).
9497 Intrinsic function names must all start with an "``llvm.``" prefix. This
9498 prefix is reserved in LLVM for intrinsic names; thus, function names may
9499 not begin with this prefix. Intrinsic functions must always be external
9500 functions: you cannot define the body of intrinsic functions. Intrinsic
9501 functions may only be used in call or invoke instructions: it is illegal
9502 to take the address of an intrinsic function. Additionally, because
9503 intrinsic functions are part of the LLVM language, it is required if any
9504 are added that they be documented here.
9506 Some intrinsic functions can be overloaded, i.e., the intrinsic
9507 represents a family of functions that perform the same operation but on
9508 different data types. Because LLVM can represent over 8 million
9509 different integer types, overloading is used commonly to allow an
9510 intrinsic function to operate on any integer type. One or more of the
9511 argument types or the result type can be overloaded to accept any
9512 integer type. Argument types may also be defined as exactly matching a
9513 previous argument's type or the result type. This allows an intrinsic
9514 function which accepts multiple arguments, but needs all of them to be
9515 of the same type, to only be overloaded with respect to a single
9516 argument or the result.
9518 Overloaded intrinsics will have the names of its overloaded argument
9519 types encoded into its function name, each preceded by a period. Only
9520 those types which are overloaded result in a name suffix. Arguments
9521 whose type is matched against another type do not. For example, the
9522 ``llvm.ctpop`` function can take an integer of any width and returns an
9523 integer of exactly the same integer width. This leads to a family of
9524 functions such as ``i8 @llvm.ctpop.i8(i8 %val)`` and
9525 ``i29 @llvm.ctpop.i29(i29 %val)``. Only one type, the return type, is
9526 overloaded, and only one type suffix is required. Because the argument's
9527 type is matched against the return type, it does not require its own
9530 To learn how to add an intrinsic function, please see the `Extending
9531 LLVM Guide <ExtendingLLVM.html>`_.
9535 Variable Argument Handling Intrinsics
9536 -------------------------------------
9538 Variable argument support is defined in LLVM with the
9539 :ref:`va_arg <i_va_arg>` instruction and these three intrinsic
9540 functions. These functions are related to the similarly named macros
9541 defined in the ``<stdarg.h>`` header file.
9543 All of these functions operate on arguments that use a target-specific
9544 value type "``va_list``". The LLVM assembly language reference manual
9545 does not define what this type is, so all transformations should be
9546 prepared to handle these functions regardless of the type used.
9548 This example shows how the :ref:`va_arg <i_va_arg>` instruction and the
9549 variable argument handling intrinsic functions are used.
9551 .. code-block:: llvm
9553 ; This struct is different for every platform. For most platforms,
9554 ; it is merely an i8*.
9555 %struct.va_list = type { i8* }
9557 ; For Unix x86_64 platforms, va_list is the following struct:
9558 ; %struct.va_list = type { i32, i32, i8*, i8* }
9560 define i32 @test(i32 %X, ...) {
9561 ; Initialize variable argument processing
9562 %ap = alloca %struct.va_list
9563 %ap2 = bitcast %struct.va_list* %ap to i8*
9564 call void @llvm.va_start(i8* %ap2)
9566 ; Read a single integer argument
9567 %tmp = va_arg i8* %ap2, i32
9569 ; Demonstrate usage of llvm.va_copy and llvm.va_end
9571 %aq2 = bitcast i8** %aq to i8*
9572 call void @llvm.va_copy(i8* %aq2, i8* %ap2)
9573 call void @llvm.va_end(i8* %aq2)
9575 ; Stop processing of arguments.
9576 call void @llvm.va_end(i8* %ap2)
9580 declare void @llvm.va_start(i8*)
9581 declare void @llvm.va_copy(i8*, i8*)
9582 declare void @llvm.va_end(i8*)
9586 '``llvm.va_start``' Intrinsic
9587 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9594 declare void @llvm.va_start(i8* <arglist>)
9599 The '``llvm.va_start``' intrinsic initializes ``*<arglist>`` for
9600 subsequent use by ``va_arg``.
9605 The argument is a pointer to a ``va_list`` element to initialize.
9610 The '``llvm.va_start``' intrinsic works just like the ``va_start`` macro
9611 available in C. In a target-dependent way, it initializes the
9612 ``va_list`` element to which the argument points, so that the next call
9613 to ``va_arg`` will produce the first variable argument passed to the
9614 function. Unlike the C ``va_start`` macro, this intrinsic does not need
9615 to know the last argument of the function as the compiler can figure
9618 '``llvm.va_end``' Intrinsic
9619 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9626 declare void @llvm.va_end(i8* <arglist>)
9631 The '``llvm.va_end``' intrinsic destroys ``*<arglist>``, which has been
9632 initialized previously with ``llvm.va_start`` or ``llvm.va_copy``.
9637 The argument is a pointer to a ``va_list`` to destroy.
9642 The '``llvm.va_end``' intrinsic works just like the ``va_end`` macro
9643 available in C. In a target-dependent way, it destroys the ``va_list``
9644 element to which the argument points. Calls to
9645 :ref:`llvm.va_start <int_va_start>` and
9646 :ref:`llvm.va_copy <int_va_copy>` must be matched exactly with calls to
9651 '``llvm.va_copy``' Intrinsic
9652 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9659 declare void @llvm.va_copy(i8* <destarglist>, i8* <srcarglist>)
9664 The '``llvm.va_copy``' intrinsic copies the current argument position
9665 from the source argument list to the destination argument list.
9670 The first argument is a pointer to a ``va_list`` element to initialize.
9671 The second argument is a pointer to a ``va_list`` element to copy from.
9676 The '``llvm.va_copy``' intrinsic works just like the ``va_copy`` macro
9677 available in C. In a target-dependent way, it copies the source
9678 ``va_list`` element into the destination ``va_list`` element. This
9679 intrinsic is necessary because the `` llvm.va_start`` intrinsic may be
9680 arbitrarily complex and require, for example, memory allocation.
9682 Accurate Garbage Collection Intrinsics
9683 --------------------------------------
9685 LLVM's support for `Accurate Garbage Collection <GarbageCollection.html>`_
9686 (GC) requires the frontend to generate code containing appropriate intrinsic
9687 calls and select an appropriate GC strategy which knows how to lower these
9688 intrinsics in a manner which is appropriate for the target collector.
9690 These intrinsics allow identification of :ref:`GC roots on the
9691 stack <int_gcroot>`, as well as garbage collector implementations that
9692 require :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers.
9693 Frontends for type-safe garbage collected languages should generate
9694 these intrinsics to make use of the LLVM garbage collectors. For more
9695 details, see `Garbage Collection with LLVM <GarbageCollection.html>`_.
9697 Experimental Statepoint Intrinsics
9698 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9700 LLVM provides an second experimental set of intrinsics for describing garbage
9701 collection safepoints in compiled code. These intrinsics are an alternative
9702 to the ``llvm.gcroot`` intrinsics, but are compatible with the ones for
9703 :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers. The
9704 differences in approach are covered in the `Garbage Collection with LLVM
9705 <GarbageCollection.html>`_ documentation. The intrinsics themselves are
9706 described in :doc:`Statepoints`.
9710 '``llvm.gcroot``' Intrinsic
9711 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9718 declare void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
9723 The '``llvm.gcroot``' intrinsic declares the existence of a GC root to
9724 the code generator, and allows some metadata to be associated with it.
9729 The first argument specifies the address of a stack object that contains
9730 the root pointer. The second pointer (which must be either a constant or
9731 a global value address) contains the meta-data to be associated with the
9737 At runtime, a call to this intrinsic stores a null pointer into the
9738 "ptrloc" location. At compile-time, the code generator generates
9739 information to allow the runtime to find the pointer at GC safe points.
9740 The '``llvm.gcroot``' intrinsic may only be used in a function which
9741 :ref:`specifies a GC algorithm <gc>`.
9745 '``llvm.gcread``' Intrinsic
9746 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9753 declare i8* @llvm.gcread(i8* %ObjPtr, i8** %Ptr)
9758 The '``llvm.gcread``' intrinsic identifies reads of references from heap
9759 locations, allowing garbage collector implementations that require read
9765 The second argument is the address to read from, which should be an
9766 address allocated from the garbage collector. The first object is a
9767 pointer to the start of the referenced object, if needed by the language
9768 runtime (otherwise null).
9773 The '``llvm.gcread``' intrinsic has the same semantics as a load
9774 instruction, but may be replaced with substantially more complex code by
9775 the garbage collector runtime, as needed. The '``llvm.gcread``'
9776 intrinsic may only be used in a function which :ref:`specifies a GC
9781 '``llvm.gcwrite``' Intrinsic
9782 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9789 declare void @llvm.gcwrite(i8* %P1, i8* %Obj, i8** %P2)
9794 The '``llvm.gcwrite``' intrinsic identifies writes of references to heap
9795 locations, allowing garbage collector implementations that require write
9796 barriers (such as generational or reference counting collectors).
9801 The first argument is the reference to store, the second is the start of
9802 the object to store it to, and the third is the address of the field of
9803 Obj to store to. If the runtime does not require a pointer to the
9804 object, Obj may be null.
9809 The '``llvm.gcwrite``' intrinsic has the same semantics as a store
9810 instruction, but may be replaced with substantially more complex code by
9811 the garbage collector runtime, as needed. The '``llvm.gcwrite``'
9812 intrinsic may only be used in a function which :ref:`specifies a GC
9815 Code Generator Intrinsics
9816 -------------------------
9818 These intrinsics are provided by LLVM to expose special features that
9819 may only be implemented with code generator support.
9821 '``llvm.returnaddress``' Intrinsic
9822 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9829 declare i8* @llvm.returnaddress(i32 <level>)
9834 The '``llvm.returnaddress``' intrinsic attempts to compute a
9835 target-specific value indicating the return address of the current
9836 function or one of its callers.
9841 The argument to this intrinsic indicates which function to return the
9842 address for. Zero indicates the calling function, one indicates its
9843 caller, etc. The argument is **required** to be a constant integer
9849 The '``llvm.returnaddress``' intrinsic either returns a pointer
9850 indicating the return address of the specified call frame, or zero if it
9851 cannot be identified. The value returned by this intrinsic is likely to
9852 be incorrect or 0 for arguments other than zero, so it should only be
9853 used for debugging purposes.
9855 Note that calling this intrinsic does not prevent function inlining or
9856 other aggressive transformations, so the value returned may not be that
9857 of the obvious source-language caller.
9859 '``llvm.addressofreturnaddress``' Intrinsic
9860 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9867 declare i8* @llvm.addressofreturnaddress()
9872 The '``llvm.addressofreturnaddress``' intrinsic returns a target-specific
9873 pointer to the place in the stack frame where the return address of the
9874 current function is stored.
9879 Note that calling this intrinsic does not prevent function inlining or
9880 other aggressive transformations, so the value returned may not be that
9881 of the obvious source-language caller.
9883 This intrinsic is only implemented for x86.
9885 '``llvm.frameaddress``' Intrinsic
9886 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9893 declare i8* @llvm.frameaddress(i32 <level>)
9898 The '``llvm.frameaddress``' intrinsic attempts to return the
9899 target-specific frame pointer value for the specified stack frame.
9904 The argument to this intrinsic indicates which function to return the
9905 frame pointer for. Zero indicates the calling function, one indicates
9906 its caller, etc. The argument is **required** to be a constant integer
9912 The '``llvm.frameaddress``' intrinsic either returns a pointer
9913 indicating the frame address of the specified call frame, or zero if it
9914 cannot be identified. The value returned by this intrinsic is likely to
9915 be incorrect or 0 for arguments other than zero, so it should only be
9916 used for debugging purposes.
9918 Note that calling this intrinsic does not prevent function inlining or
9919 other aggressive transformations, so the value returned may not be that
9920 of the obvious source-language caller.
9922 '``llvm.localescape``' and '``llvm.localrecover``' Intrinsics
9923 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9930 declare void @llvm.localescape(...)
9931 declare i8* @llvm.localrecover(i8* %func, i8* %fp, i32 %idx)
9936 The '``llvm.localescape``' intrinsic escapes offsets of a collection of static
9937 allocas, and the '``llvm.localrecover``' intrinsic applies those offsets to a
9938 live frame pointer to recover the address of the allocation. The offset is
9939 computed during frame layout of the caller of ``llvm.localescape``.
9944 All arguments to '``llvm.localescape``' must be pointers to static allocas or
9945 casts of static allocas. Each function can only call '``llvm.localescape``'
9946 once, and it can only do so from the entry block.
9948 The ``func`` argument to '``llvm.localrecover``' must be a constant
9949 bitcasted pointer to a function defined in the current module. The code
9950 generator cannot determine the frame allocation offset of functions defined in
9953 The ``fp`` argument to '``llvm.localrecover``' must be a frame pointer of a
9954 call frame that is currently live. The return value of '``llvm.localaddress``'
9955 is one way to produce such a value, but various runtimes also expose a suitable
9956 pointer in platform-specific ways.
9958 The ``idx`` argument to '``llvm.localrecover``' indicates which alloca passed to
9959 '``llvm.localescape``' to recover. It is zero-indexed.
9964 These intrinsics allow a group of functions to share access to a set of local
9965 stack allocations of a one parent function. The parent function may call the
9966 '``llvm.localescape``' intrinsic once from the function entry block, and the
9967 child functions can use '``llvm.localrecover``' to access the escaped allocas.
9968 The '``llvm.localescape``' intrinsic blocks inlining, as inlining changes where
9969 the escaped allocas are allocated, which would break attempts to use
9970 '``llvm.localrecover``'.
9972 .. _int_read_register:
9973 .. _int_write_register:
9975 '``llvm.read_register``' and '``llvm.write_register``' Intrinsics
9976 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9983 declare i32 @llvm.read_register.i32(metadata)
9984 declare i64 @llvm.read_register.i64(metadata)
9985 declare void @llvm.write_register.i32(metadata, i32 @value)
9986 declare void @llvm.write_register.i64(metadata, i64 @value)
9992 The '``llvm.read_register``' and '``llvm.write_register``' intrinsics
9993 provides access to the named register. The register must be valid on
9994 the architecture being compiled to. The type needs to be compatible
9995 with the register being read.
10000 The '``llvm.read_register``' intrinsic returns the current value of the
10001 register, where possible. The '``llvm.write_register``' intrinsic sets
10002 the current value of the register, where possible.
10004 This is useful to implement named register global variables that need
10005 to always be mapped to a specific register, as is common practice on
10006 bare-metal programs including OS kernels.
10008 The compiler doesn't check for register availability or use of the used
10009 register in surrounding code, including inline assembly. Because of that,
10010 allocatable registers are not supported.
10012 Warning: So far it only works with the stack pointer on selected
10013 architectures (ARM, AArch64, PowerPC and x86_64). Significant amount of
10014 work is needed to support other registers and even more so, allocatable
10019 '``llvm.stacksave``' Intrinsic
10020 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10027 declare i8* @llvm.stacksave()
10032 The '``llvm.stacksave``' intrinsic is used to remember the current state
10033 of the function stack, for use with
10034 :ref:`llvm.stackrestore <int_stackrestore>`. This is useful for
10035 implementing language features like scoped automatic variable sized
10041 This intrinsic returns a opaque pointer value that can be passed to
10042 :ref:`llvm.stackrestore <int_stackrestore>`. When an
10043 ``llvm.stackrestore`` intrinsic is executed with a value saved from
10044 ``llvm.stacksave``, it effectively restores the state of the stack to
10045 the state it was in when the ``llvm.stacksave`` intrinsic executed. In
10046 practice, this pops any :ref:`alloca <i_alloca>` blocks from the stack that
10047 were allocated after the ``llvm.stacksave`` was executed.
10049 .. _int_stackrestore:
10051 '``llvm.stackrestore``' Intrinsic
10052 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10059 declare void @llvm.stackrestore(i8* %ptr)
10064 The '``llvm.stackrestore``' intrinsic is used to restore the state of
10065 the function stack to the state it was in when the corresponding
10066 :ref:`llvm.stacksave <int_stacksave>` intrinsic executed. This is
10067 useful for implementing language features like scoped automatic variable
10068 sized arrays in C99.
10073 See the description for :ref:`llvm.stacksave <int_stacksave>`.
10075 .. _int_get_dynamic_area_offset:
10077 '``llvm.get.dynamic.area.offset``' Intrinsic
10078 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10085 declare i32 @llvm.get.dynamic.area.offset.i32()
10086 declare i64 @llvm.get.dynamic.area.offset.i64()
10091 The '``llvm.get.dynamic.area.offset.*``' intrinsic family is used to
10092 get the offset from native stack pointer to the address of the most
10093 recent dynamic alloca on the caller's stack. These intrinsics are
10094 intendend for use in combination with
10095 :ref:`llvm.stacksave <int_stacksave>` to get a
10096 pointer to the most recent dynamic alloca. This is useful, for example,
10097 for AddressSanitizer's stack unpoisoning routines.
10102 These intrinsics return a non-negative integer value that can be used to
10103 get the address of the most recent dynamic alloca, allocated by :ref:`alloca <i_alloca>`
10104 on the caller's stack. In particular, for targets where stack grows downwards,
10105 adding this offset to the native stack pointer would get the address of the most
10106 recent dynamic alloca. For targets where stack grows upwards, the situation is a bit more
10107 complicated, because subtracting this value from stack pointer would get the address
10108 one past the end of the most recent dynamic alloca.
10110 Although for most targets `llvm.get.dynamic.area.offset <int_get_dynamic_area_offset>`
10111 returns just a zero, for others, such as PowerPC and PowerPC64, it returns a
10112 compile-time-known constant value.
10114 The return value type of :ref:`llvm.get.dynamic.area.offset <int_get_dynamic_area_offset>`
10115 must match the target's default address space's (address space 0) pointer type.
10117 '``llvm.prefetch``' Intrinsic
10118 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10125 declare void @llvm.prefetch(i8* <address>, i32 <rw>, i32 <locality>, i32 <cache type>)
10130 The '``llvm.prefetch``' intrinsic is a hint to the code generator to
10131 insert a prefetch instruction if supported; otherwise, it is a noop.
10132 Prefetches have no effect on the behavior of the program but can change
10133 its performance characteristics.
10138 ``address`` is the address to be prefetched, ``rw`` is the specifier
10139 determining if the fetch should be for a read (0) or write (1), and
10140 ``locality`` is a temporal locality specifier ranging from (0) - no
10141 locality, to (3) - extremely local keep in cache. The ``cache type``
10142 specifies whether the prefetch is performed on the data (1) or
10143 instruction (0) cache. The ``rw``, ``locality`` and ``cache type``
10144 arguments must be constant integers.
10149 This intrinsic does not modify the behavior of the program. In
10150 particular, prefetches cannot trap and do not produce a value. On
10151 targets that support this intrinsic, the prefetch can provide hints to
10152 the processor cache for better performance.
10154 '``llvm.pcmarker``' Intrinsic
10155 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10162 declare void @llvm.pcmarker(i32 <id>)
10167 The '``llvm.pcmarker``' intrinsic is a method to export a Program
10168 Counter (PC) in a region of code to simulators and other tools. The
10169 method is target specific, but it is expected that the marker will use
10170 exported symbols to transmit the PC of the marker. The marker makes no
10171 guarantees that it will remain with any specific instruction after
10172 optimizations. It is possible that the presence of a marker will inhibit
10173 optimizations. The intended use is to be inserted after optimizations to
10174 allow correlations of simulation runs.
10179 ``id`` is a numerical id identifying the marker.
10184 This intrinsic does not modify the behavior of the program. Backends
10185 that do not support this intrinsic may ignore it.
10187 '``llvm.readcyclecounter``' Intrinsic
10188 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10195 declare i64 @llvm.readcyclecounter()
10200 The '``llvm.readcyclecounter``' intrinsic provides access to the cycle
10201 counter register (or similar low latency, high accuracy clocks) on those
10202 targets that support it. On X86, it should map to RDTSC. On Alpha, it
10203 should map to RPCC. As the backing counters overflow quickly (on the
10204 order of 9 seconds on alpha), this should only be used for small
10210 When directly supported, reading the cycle counter should not modify any
10211 memory. Implementations are allowed to either return a application
10212 specific value or a system wide value. On backends without support, this
10213 is lowered to a constant 0.
10215 Note that runtime support may be conditional on the privilege-level code is
10216 running at and the host platform.
10218 '``llvm.clear_cache``' Intrinsic
10219 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10226 declare void @llvm.clear_cache(i8*, i8*)
10231 The '``llvm.clear_cache``' intrinsic ensures visibility of modifications
10232 in the specified range to the execution unit of the processor. On
10233 targets with non-unified instruction and data cache, the implementation
10234 flushes the instruction cache.
10239 On platforms with coherent instruction and data caches (e.g. x86), this
10240 intrinsic is a nop. On platforms with non-coherent instruction and data
10241 cache (e.g. ARM, MIPS), the intrinsic is lowered either to appropriate
10242 instructions or a system call, if cache flushing requires special
10245 The default behavior is to emit a call to ``__clear_cache`` from the run
10248 This instrinsic does *not* empty the instruction pipeline. Modifications
10249 of the current function are outside the scope of the intrinsic.
10251 '``llvm.instrprof.increment``' Intrinsic
10252 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10259 declare void @llvm.instrprof.increment(i8* <name>, i64 <hash>,
10260 i32 <num-counters>, i32 <index>)
10265 The '``llvm.instrprof.increment``' intrinsic can be emitted by a
10266 frontend for use with instrumentation based profiling. These will be
10267 lowered by the ``-instrprof`` pass to generate execution counts of a
10268 program at runtime.
10273 The first argument is a pointer to a global variable containing the
10274 name of the entity being instrumented. This should generally be the
10275 (mangled) function name for a set of counters.
10277 The second argument is a hash value that can be used by the consumer
10278 of the profile data to detect changes to the instrumented source, and
10279 the third is the number of counters associated with ``name``. It is an
10280 error if ``hash`` or ``num-counters`` differ between two instances of
10281 ``instrprof.increment`` that refer to the same name.
10283 The last argument refers to which of the counters for ``name`` should
10284 be incremented. It should be a value between 0 and ``num-counters``.
10289 This intrinsic represents an increment of a profiling counter. It will
10290 cause the ``-instrprof`` pass to generate the appropriate data
10291 structures and the code to increment the appropriate value, in a
10292 format that can be written out by a compiler runtime and consumed via
10293 the ``llvm-profdata`` tool.
10295 '``llvm.instrprof.increment.step``' Intrinsic
10296 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10303 declare void @llvm.instrprof.increment.step(i8* <name>, i64 <hash>,
10304 i32 <num-counters>,
10305 i32 <index>, i64 <step>)
10310 The '``llvm.instrprof.increment.step``' intrinsic is an extension to
10311 the '``llvm.instrprof.increment``' intrinsic with an additional fifth
10312 argument to specify the step of the increment.
10316 The first four arguments are the same as '``llvm.instrprof.increment``'
10319 The last argument specifies the value of the increment of the counter variable.
10323 See description of '``llvm.instrprof.increment``' instrinsic.
10326 '``llvm.instrprof.value.profile``' Intrinsic
10327 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10334 declare void @llvm.instrprof.value.profile(i8* <name>, i64 <hash>,
10335 i64 <value>, i32 <value_kind>,
10341 The '``llvm.instrprof.value.profile``' intrinsic can be emitted by a
10342 frontend for use with instrumentation based profiling. This will be
10343 lowered by the ``-instrprof`` pass to find out the target values,
10344 instrumented expressions take in a program at runtime.
10349 The first argument is a pointer to a global variable containing the
10350 name of the entity being instrumented. ``name`` should generally be the
10351 (mangled) function name for a set of counters.
10353 The second argument is a hash value that can be used by the consumer
10354 of the profile data to detect changes to the instrumented source. It
10355 is an error if ``hash`` differs between two instances of
10356 ``llvm.instrprof.*`` that refer to the same name.
10358 The third argument is the value of the expression being profiled. The profiled
10359 expression's value should be representable as an unsigned 64-bit value. The
10360 fourth argument represents the kind of value profiling that is being done. The
10361 supported value profiling kinds are enumerated through the
10362 ``InstrProfValueKind`` type declared in the
10363 ``<include/llvm/ProfileData/InstrProf.h>`` header file. The last argument is the
10364 index of the instrumented expression within ``name``. It should be >= 0.
10369 This intrinsic represents the point where a call to a runtime routine
10370 should be inserted for value profiling of target expressions. ``-instrprof``
10371 pass will generate the appropriate data structures and replace the
10372 ``llvm.instrprof.value.profile`` intrinsic with the call to the profile
10373 runtime library with proper arguments.
10375 '``llvm.thread.pointer``' Intrinsic
10376 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10383 declare i8* @llvm.thread.pointer()
10388 The '``llvm.thread.pointer``' intrinsic returns the value of the thread
10394 The '``llvm.thread.pointer``' intrinsic returns a pointer to the TLS area
10395 for the current thread. The exact semantics of this value are target
10396 specific: it may point to the start of TLS area, to the end, or somewhere
10397 in the middle. Depending on the target, this intrinsic may read a register,
10398 call a helper function, read from an alternate memory space, or perform
10399 other operations necessary to locate the TLS area. Not all targets support
10402 Standard C Library Intrinsics
10403 -----------------------------
10405 LLVM provides intrinsics for a few important standard C library
10406 functions. These intrinsics allow source-language front-ends to pass
10407 information about the alignment of the pointer arguments to the code
10408 generator, providing opportunity for more efficient code generation.
10412 '``llvm.memcpy``' Intrinsic
10413 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10418 This is an overloaded intrinsic. You can use ``llvm.memcpy`` on any
10419 integer bit width and for different address spaces. Not all targets
10420 support all bit widths however.
10424 declare void @llvm.memcpy.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
10425 i32 <len>, i1 <isvolatile>)
10426 declare void @llvm.memcpy.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
10427 i64 <len>, i1 <isvolatile>)
10432 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
10433 source location to the destination location.
10435 Note that, unlike the standard libc function, the ``llvm.memcpy.*``
10436 intrinsics do not return a value, takes extra isvolatile
10437 arguments and the pointers can be in specified address spaces.
10442 The first argument is a pointer to the destination, the second is a
10443 pointer to the source. The third argument is an integer argument
10444 specifying the number of bytes to copy, and the fourth is a
10445 boolean indicating a volatile access.
10447 The :ref:`align <attr_align>` parameter attribute can be provided
10448 for the first and second arguments.
10450 If the ``isvolatile`` parameter is ``true``, the ``llvm.memcpy`` call is
10451 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
10452 very cleanly specified and it is unwise to depend on it.
10457 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
10458 source location to the destination location, which are not allowed to
10459 overlap. It copies "len" bytes of memory over. If the argument is known
10460 to be aligned to some boundary, this can be specified as the fourth
10461 argument, otherwise it should be set to 0 or 1 (both meaning no alignment).
10465 '``llvm.memmove``' Intrinsic
10466 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10471 This is an overloaded intrinsic. You can use llvm.memmove on any integer
10472 bit width and for different address space. Not all targets support all
10473 bit widths however.
10477 declare void @llvm.memmove.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
10478 i32 <len>, i1 <isvolatile>)
10479 declare void @llvm.memmove.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
10480 i64 <len>, i1 <isvolatile>)
10485 The '``llvm.memmove.*``' intrinsics move a block of memory from the
10486 source location to the destination location. It is similar to the
10487 '``llvm.memcpy``' intrinsic but allows the two memory locations to
10490 Note that, unlike the standard libc function, the ``llvm.memmove.*``
10491 intrinsics do not return a value, takes an extra isvolatile
10492 argument and the pointers can be in specified address spaces.
10497 The first argument is a pointer to the destination, the second is a
10498 pointer to the source. The third argument is an integer argument
10499 specifying the number of bytes to copy, and the fourth is a
10500 boolean indicating a volatile access.
10502 The :ref:`align <attr_align>` parameter attribute can be provided
10503 for the first and second arguments.
10505 If the ``isvolatile`` parameter is ``true``, the ``llvm.memmove`` call
10506 is a :ref:`volatile operation <volatile>`. The detailed access behavior is
10507 not very cleanly specified and it is unwise to depend on it.
10512 The '``llvm.memmove.*``' intrinsics copy a block of memory from the
10513 source location to the destination location, which may overlap. It
10514 copies "len" bytes of memory over. If the argument is known to be
10515 aligned to some boundary, this can be specified as the fourth argument,
10516 otherwise it should be set to 0 or 1 (both meaning no alignment).
10520 '``llvm.memset.*``' Intrinsics
10521 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10526 This is an overloaded intrinsic. You can use llvm.memset on any integer
10527 bit width and for different address spaces. However, not all targets
10528 support all bit widths.
10532 declare void @llvm.memset.p0i8.i32(i8* <dest>, i8 <val>,
10533 i32 <len>, i1 <isvolatile>)
10534 declare void @llvm.memset.p0i8.i64(i8* <dest>, i8 <val>,
10535 i64 <len>, i1 <isvolatile>)
10540 The '``llvm.memset.*``' intrinsics fill a block of memory with a
10541 particular byte value.
10543 Note that, unlike the standard libc function, the ``llvm.memset``
10544 intrinsic does not return a value and takes an extra volatile
10545 argument. Also, the destination can be in an arbitrary address space.
10550 The first argument is a pointer to the destination to fill, the second
10551 is the byte value with which to fill it, the third argument is an
10552 integer argument specifying the number of bytes to fill, and the fourth
10553 is a boolean indicating a volatile access.
10555 The :ref:`align <attr_align>` parameter attribute can be provided
10556 for the first arguments.
10558 If the ``isvolatile`` parameter is ``true``, the ``llvm.memset`` call is
10559 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
10560 very cleanly specified and it is unwise to depend on it.
10565 The '``llvm.memset.*``' intrinsics fill "len" bytes of memory starting
10566 at the destination location.
10568 '``llvm.sqrt.*``' Intrinsic
10569 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10574 This is an overloaded intrinsic. You can use ``llvm.sqrt`` on any
10575 floating-point or vector of floating-point type. Not all targets support
10580 declare float @llvm.sqrt.f32(float %Val)
10581 declare double @llvm.sqrt.f64(double %Val)
10582 declare x86_fp80 @llvm.sqrt.f80(x86_fp80 %Val)
10583 declare fp128 @llvm.sqrt.f128(fp128 %Val)
10584 declare ppc_fp128 @llvm.sqrt.ppcf128(ppc_fp128 %Val)
10589 The '``llvm.sqrt``' intrinsics return the square root of the specified value.
10594 The argument and return value are floating-point numbers of the same type.
10599 Return the same value as a corresponding libm '``sqrt``' function but without
10600 trapping or setting ``errno``. For types specified by IEEE-754, the result
10601 matches a conforming libm implementation.
10603 When specified with the fast-math-flag 'afn', the result may be approximated
10604 using a less accurate calculation.
10606 '``llvm.powi.*``' Intrinsic
10607 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10612 This is an overloaded intrinsic. You can use ``llvm.powi`` on any
10613 floating point or vector of floating point type. Not all targets support
10618 declare float @llvm.powi.f32(float %Val, i32 %power)
10619 declare double @llvm.powi.f64(double %Val, i32 %power)
10620 declare x86_fp80 @llvm.powi.f80(x86_fp80 %Val, i32 %power)
10621 declare fp128 @llvm.powi.f128(fp128 %Val, i32 %power)
10622 declare ppc_fp128 @llvm.powi.ppcf128(ppc_fp128 %Val, i32 %power)
10627 The '``llvm.powi.*``' intrinsics return the first operand raised to the
10628 specified (positive or negative) power. The order of evaluation of
10629 multiplications is not defined. When a vector of floating point type is
10630 used, the second argument remains a scalar integer value.
10635 The second argument is an integer power, and the first is a value to
10636 raise to that power.
10641 This function returns the first value raised to the second power with an
10642 unspecified sequence of rounding operations.
10644 '``llvm.sin.*``' Intrinsic
10645 ^^^^^^^^^^^^^^^^^^^^^^^^^^
10650 This is an overloaded intrinsic. You can use ``llvm.sin`` on any
10651 floating-point or vector of floating-point type. Not all targets support
10656 declare float @llvm.sin.f32(float %Val)
10657 declare double @llvm.sin.f64(double %Val)
10658 declare x86_fp80 @llvm.sin.f80(x86_fp80 %Val)
10659 declare fp128 @llvm.sin.f128(fp128 %Val)
10660 declare ppc_fp128 @llvm.sin.ppcf128(ppc_fp128 %Val)
10665 The '``llvm.sin.*``' intrinsics return the sine of the operand.
10670 The argument and return value are floating-point numbers of the same type.
10675 Return the same value as a corresponding libm '``sin``' function but without
10676 trapping or setting ``errno``.
10678 When specified with the fast-math-flag 'afn', the result may be approximated
10679 using a less accurate calculation.
10681 '``llvm.cos.*``' Intrinsic
10682 ^^^^^^^^^^^^^^^^^^^^^^^^^^
10687 This is an overloaded intrinsic. You can use ``llvm.cos`` on any
10688 floating-point or vector of floating-point type. Not all targets support
10693 declare float @llvm.cos.f32(float %Val)
10694 declare double @llvm.cos.f64(double %Val)
10695 declare x86_fp80 @llvm.cos.f80(x86_fp80 %Val)
10696 declare fp128 @llvm.cos.f128(fp128 %Val)
10697 declare ppc_fp128 @llvm.cos.ppcf128(ppc_fp128 %Val)
10702 The '``llvm.cos.*``' intrinsics return the cosine of the operand.
10707 The argument and return value are floating-point numbers of the same type.
10712 Return the same value as a corresponding libm '``cos``' function but without
10713 trapping or setting ``errno``.
10715 When specified with the fast-math-flag 'afn', the result may be approximated
10716 using a less accurate calculation.
10718 '``llvm.pow.*``' Intrinsic
10719 ^^^^^^^^^^^^^^^^^^^^^^^^^^
10724 This is an overloaded intrinsic. You can use ``llvm.pow`` on any
10725 floating-point or vector of floating-point type. Not all targets support
10730 declare float @llvm.pow.f32(float %Val, float %Power)
10731 declare double @llvm.pow.f64(double %Val, double %Power)
10732 declare x86_fp80 @llvm.pow.f80(x86_fp80 %Val, x86_fp80 %Power)
10733 declare fp128 @llvm.pow.f128(fp128 %Val, fp128 %Power)
10734 declare ppc_fp128 @llvm.pow.ppcf128(ppc_fp128 %Val, ppc_fp128 Power)
10739 The '``llvm.pow.*``' intrinsics return the first operand raised to the
10740 specified (positive or negative) power.
10745 The arguments and return value are floating-point numbers of the same type.
10750 Return the same value as a corresponding libm '``pow``' function but without
10751 trapping or setting ``errno``.
10753 When specified with the fast-math-flag 'afn', the result may be approximated
10754 using a less accurate calculation.
10756 '``llvm.exp.*``' Intrinsic
10757 ^^^^^^^^^^^^^^^^^^^^^^^^^^
10762 This is an overloaded intrinsic. You can use ``llvm.exp`` on any
10763 floating-point or vector of floating-point type. Not all targets support
10768 declare float @llvm.exp.f32(float %Val)
10769 declare double @llvm.exp.f64(double %Val)
10770 declare x86_fp80 @llvm.exp.f80(x86_fp80 %Val)
10771 declare fp128 @llvm.exp.f128(fp128 %Val)
10772 declare ppc_fp128 @llvm.exp.ppcf128(ppc_fp128 %Val)
10777 The '``llvm.exp.*``' intrinsics compute the base-e exponential of the specified
10783 The argument and return value are floating-point numbers of the same type.
10788 Return the same value as a corresponding libm '``exp``' function but without
10789 trapping or setting ``errno``.
10791 When specified with the fast-math-flag 'afn', the result may be approximated
10792 using a less accurate calculation.
10794 '``llvm.exp2.*``' Intrinsic
10795 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10800 This is an overloaded intrinsic. You can use ``llvm.exp2`` on any
10801 floating-point or vector of floating-point type. Not all targets support
10806 declare float @llvm.exp2.f32(float %Val)
10807 declare double @llvm.exp2.f64(double %Val)
10808 declare x86_fp80 @llvm.exp2.f80(x86_fp80 %Val)
10809 declare fp128 @llvm.exp2.f128(fp128 %Val)
10810 declare ppc_fp128 @llvm.exp2.ppcf128(ppc_fp128 %Val)
10815 The '``llvm.exp2.*``' intrinsics compute the base-2 exponential of the
10821 The argument and return value are floating-point numbers of the same type.
10826 Return the same value as a corresponding libm '``exp2``' function but without
10827 trapping or setting ``errno``.
10829 When specified with the fast-math-flag 'afn', the result may be approximated
10830 using a less accurate calculation.
10832 '``llvm.log.*``' Intrinsic
10833 ^^^^^^^^^^^^^^^^^^^^^^^^^^
10838 This is an overloaded intrinsic. You can use ``llvm.log`` on any
10839 floating-point or vector of floating-point type. Not all targets support
10844 declare float @llvm.log.f32(float %Val)
10845 declare double @llvm.log.f64(double %Val)
10846 declare x86_fp80 @llvm.log.f80(x86_fp80 %Val)
10847 declare fp128 @llvm.log.f128(fp128 %Val)
10848 declare ppc_fp128 @llvm.log.ppcf128(ppc_fp128 %Val)
10853 The '``llvm.log.*``' intrinsics compute the base-e logarithm of the specified
10859 The argument and return value are floating-point numbers of the same type.
10864 Return the same value as a corresponding libm '``log``' function but without
10865 trapping or setting ``errno``.
10867 When specified with the fast-math-flag 'afn', the result may be approximated
10868 using a less accurate calculation.
10870 '``llvm.log10.*``' Intrinsic
10871 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10876 This is an overloaded intrinsic. You can use ``llvm.log10`` on any
10877 floating-point or vector of floating-point type. Not all targets support
10882 declare float @llvm.log10.f32(float %Val)
10883 declare double @llvm.log10.f64(double %Val)
10884 declare x86_fp80 @llvm.log10.f80(x86_fp80 %Val)
10885 declare fp128 @llvm.log10.f128(fp128 %Val)
10886 declare ppc_fp128 @llvm.log10.ppcf128(ppc_fp128 %Val)
10891 The '``llvm.log10.*``' intrinsics compute the base-10 logarithm of the
10897 The argument and return value are floating-point numbers of the same type.
10902 Return the same value as a corresponding libm '``log10``' function but without
10903 trapping or setting ``errno``.
10905 When specified with the fast-math-flag 'afn', the result may be approximated
10906 using a less accurate calculation.
10908 '``llvm.log2.*``' Intrinsic
10909 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10914 This is an overloaded intrinsic. You can use ``llvm.log2`` on any
10915 floating-point or vector of floating-point type. Not all targets support
10920 declare float @llvm.log2.f32(float %Val)
10921 declare double @llvm.log2.f64(double %Val)
10922 declare x86_fp80 @llvm.log2.f80(x86_fp80 %Val)
10923 declare fp128 @llvm.log2.f128(fp128 %Val)
10924 declare ppc_fp128 @llvm.log2.ppcf128(ppc_fp128 %Val)
10929 The '``llvm.log2.*``' intrinsics compute the base-2 logarithm of the specified
10935 The argument and return value are floating-point numbers of the same type.
10940 Return the same value as a corresponding libm '``log2``' function but without
10941 trapping or setting ``errno``.
10943 When specified with the fast-math-flag 'afn', the result may be approximated
10944 using a less accurate calculation.
10946 '``llvm.fma.*``' Intrinsic
10947 ^^^^^^^^^^^^^^^^^^^^^^^^^^
10952 This is an overloaded intrinsic. You can use ``llvm.fma`` on any
10953 floating-point or vector of floating-point type. Not all targets support
10958 declare float @llvm.fma.f32(float %a, float %b, float %c)
10959 declare double @llvm.fma.f64(double %a, double %b, double %c)
10960 declare x86_fp80 @llvm.fma.f80(x86_fp80 %a, x86_fp80 %b, x86_fp80 %c)
10961 declare fp128 @llvm.fma.f128(fp128 %a, fp128 %b, fp128 %c)
10962 declare ppc_fp128 @llvm.fma.ppcf128(ppc_fp128 %a, ppc_fp128 %b, ppc_fp128 %c)
10967 The '``llvm.fma.*``' intrinsics perform the fused multiply-add operation.
10972 The arguments and return value are floating-point numbers of the same type.
10977 Return the same value as a corresponding libm '``fma``' function but without
10978 trapping or setting ``errno``.
10980 When specified with the fast-math-flag 'afn', the result may be approximated
10981 using a less accurate calculation.
10983 '``llvm.fabs.*``' Intrinsic
10984 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10989 This is an overloaded intrinsic. You can use ``llvm.fabs`` on any
10990 floating point or vector of floating point type. Not all targets support
10995 declare float @llvm.fabs.f32(float %Val)
10996 declare double @llvm.fabs.f64(double %Val)
10997 declare x86_fp80 @llvm.fabs.f80(x86_fp80 %Val)
10998 declare fp128 @llvm.fabs.f128(fp128 %Val)
10999 declare ppc_fp128 @llvm.fabs.ppcf128(ppc_fp128 %Val)
11004 The '``llvm.fabs.*``' intrinsics return the absolute value of the
11010 The argument and return value are floating point numbers of the same
11016 This function returns the same values as the libm ``fabs`` functions
11017 would, and handles error conditions in the same way.
11019 '``llvm.minnum.*``' Intrinsic
11020 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11025 This is an overloaded intrinsic. You can use ``llvm.minnum`` on any
11026 floating point or vector of floating point type. Not all targets support
11031 declare float @llvm.minnum.f32(float %Val0, float %Val1)
11032 declare double @llvm.minnum.f64(double %Val0, double %Val1)
11033 declare x86_fp80 @llvm.minnum.f80(x86_fp80 %Val0, x86_fp80 %Val1)
11034 declare fp128 @llvm.minnum.f128(fp128 %Val0, fp128 %Val1)
11035 declare ppc_fp128 @llvm.minnum.ppcf128(ppc_fp128 %Val0, ppc_fp128 %Val1)
11040 The '``llvm.minnum.*``' intrinsics return the minimum of the two
11047 The arguments and return value are floating point numbers of the same
11053 Follows the IEEE-754 semantics for minNum, which also match for libm's
11056 If either operand is a NaN, returns the other non-NaN operand. Returns
11057 NaN only if both operands are NaN. If the operands compare equal,
11058 returns a value that compares equal to both operands. This means that
11059 fmin(+/-0.0, +/-0.0) could return either -0.0 or 0.0.
11061 '``llvm.maxnum.*``' Intrinsic
11062 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11067 This is an overloaded intrinsic. You can use ``llvm.maxnum`` on any
11068 floating point or vector of floating point type. Not all targets support
11073 declare float @llvm.maxnum.f32(float %Val0, float %Val1l)
11074 declare double @llvm.maxnum.f64(double %Val0, double %Val1)
11075 declare x86_fp80 @llvm.maxnum.f80(x86_fp80 %Val0, x86_fp80 %Val1)
11076 declare fp128 @llvm.maxnum.f128(fp128 %Val0, fp128 %Val1)
11077 declare ppc_fp128 @llvm.maxnum.ppcf128(ppc_fp128 %Val0, ppc_fp128 %Val1)
11082 The '``llvm.maxnum.*``' intrinsics return the maximum of the two
11089 The arguments and return value are floating point numbers of the same
11094 Follows the IEEE-754 semantics for maxNum, which also match for libm's
11097 If either operand is a NaN, returns the other non-NaN operand. Returns
11098 NaN only if both operands are NaN. If the operands compare equal,
11099 returns a value that compares equal to both operands. This means that
11100 fmax(+/-0.0, +/-0.0) could return either -0.0 or 0.0.
11102 '``llvm.copysign.*``' Intrinsic
11103 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11108 This is an overloaded intrinsic. You can use ``llvm.copysign`` on any
11109 floating point or vector of floating point type. Not all targets support
11114 declare float @llvm.copysign.f32(float %Mag, float %Sgn)
11115 declare double @llvm.copysign.f64(double %Mag, double %Sgn)
11116 declare x86_fp80 @llvm.copysign.f80(x86_fp80 %Mag, x86_fp80 %Sgn)
11117 declare fp128 @llvm.copysign.f128(fp128 %Mag, fp128 %Sgn)
11118 declare ppc_fp128 @llvm.copysign.ppcf128(ppc_fp128 %Mag, ppc_fp128 %Sgn)
11123 The '``llvm.copysign.*``' intrinsics return a value with the magnitude of the
11124 first operand and the sign of the second operand.
11129 The arguments and return value are floating point numbers of the same
11135 This function returns the same values as the libm ``copysign``
11136 functions would, and handles error conditions in the same way.
11138 '``llvm.floor.*``' Intrinsic
11139 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11144 This is an overloaded intrinsic. You can use ``llvm.floor`` on any
11145 floating point or vector of floating point type. Not all targets support
11150 declare float @llvm.floor.f32(float %Val)
11151 declare double @llvm.floor.f64(double %Val)
11152 declare x86_fp80 @llvm.floor.f80(x86_fp80 %Val)
11153 declare fp128 @llvm.floor.f128(fp128 %Val)
11154 declare ppc_fp128 @llvm.floor.ppcf128(ppc_fp128 %Val)
11159 The '``llvm.floor.*``' intrinsics return the floor of the operand.
11164 The argument and return value are floating point numbers of the same
11170 This function returns the same values as the libm ``floor`` functions
11171 would, and handles error conditions in the same way.
11173 '``llvm.ceil.*``' Intrinsic
11174 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
11179 This is an overloaded intrinsic. You can use ``llvm.ceil`` on any
11180 floating point or vector of floating point type. Not all targets support
11185 declare float @llvm.ceil.f32(float %Val)
11186 declare double @llvm.ceil.f64(double %Val)
11187 declare x86_fp80 @llvm.ceil.f80(x86_fp80 %Val)
11188 declare fp128 @llvm.ceil.f128(fp128 %Val)
11189 declare ppc_fp128 @llvm.ceil.ppcf128(ppc_fp128 %Val)
11194 The '``llvm.ceil.*``' intrinsics return the ceiling of the operand.
11199 The argument and return value are floating point numbers of the same
11205 This function returns the same values as the libm ``ceil`` functions
11206 would, and handles error conditions in the same way.
11208 '``llvm.trunc.*``' Intrinsic
11209 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11214 This is an overloaded intrinsic. You can use ``llvm.trunc`` on any
11215 floating point or vector of floating point type. Not all targets support
11220 declare float @llvm.trunc.f32(float %Val)
11221 declare double @llvm.trunc.f64(double %Val)
11222 declare x86_fp80 @llvm.trunc.f80(x86_fp80 %Val)
11223 declare fp128 @llvm.trunc.f128(fp128 %Val)
11224 declare ppc_fp128 @llvm.trunc.ppcf128(ppc_fp128 %Val)
11229 The '``llvm.trunc.*``' intrinsics returns the operand rounded to the
11230 nearest integer not larger in magnitude than the operand.
11235 The argument and return value are floating point numbers of the same
11241 This function returns the same values as the libm ``trunc`` functions
11242 would, and handles error conditions in the same way.
11244 '``llvm.rint.*``' Intrinsic
11245 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
11250 This is an overloaded intrinsic. You can use ``llvm.rint`` on any
11251 floating point or vector of floating point type. Not all targets support
11256 declare float @llvm.rint.f32(float %Val)
11257 declare double @llvm.rint.f64(double %Val)
11258 declare x86_fp80 @llvm.rint.f80(x86_fp80 %Val)
11259 declare fp128 @llvm.rint.f128(fp128 %Val)
11260 declare ppc_fp128 @llvm.rint.ppcf128(ppc_fp128 %Val)
11265 The '``llvm.rint.*``' intrinsics returns the operand rounded to the
11266 nearest integer. It may raise an inexact floating-point exception if the
11267 operand isn't an integer.
11272 The argument and return value are floating point numbers of the same
11278 This function returns the same values as the libm ``rint`` functions
11279 would, and handles error conditions in the same way.
11281 '``llvm.nearbyint.*``' Intrinsic
11282 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11287 This is an overloaded intrinsic. You can use ``llvm.nearbyint`` on any
11288 floating point or vector of floating point type. Not all targets support
11293 declare float @llvm.nearbyint.f32(float %Val)
11294 declare double @llvm.nearbyint.f64(double %Val)
11295 declare x86_fp80 @llvm.nearbyint.f80(x86_fp80 %Val)
11296 declare fp128 @llvm.nearbyint.f128(fp128 %Val)
11297 declare ppc_fp128 @llvm.nearbyint.ppcf128(ppc_fp128 %Val)
11302 The '``llvm.nearbyint.*``' intrinsics returns the operand rounded to the
11308 The argument and return value are floating point numbers of the same
11314 This function returns the same values as the libm ``nearbyint``
11315 functions would, and handles error conditions in the same way.
11317 '``llvm.round.*``' Intrinsic
11318 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11323 This is an overloaded intrinsic. You can use ``llvm.round`` on any
11324 floating point or vector of floating point type. Not all targets support
11329 declare float @llvm.round.f32(float %Val)
11330 declare double @llvm.round.f64(double %Val)
11331 declare x86_fp80 @llvm.round.f80(x86_fp80 %Val)
11332 declare fp128 @llvm.round.f128(fp128 %Val)
11333 declare ppc_fp128 @llvm.round.ppcf128(ppc_fp128 %Val)
11338 The '``llvm.round.*``' intrinsics returns the operand rounded to the
11344 The argument and return value are floating point numbers of the same
11350 This function returns the same values as the libm ``round``
11351 functions would, and handles error conditions in the same way.
11353 Bit Manipulation Intrinsics
11354 ---------------------------
11356 LLVM provides intrinsics for a few important bit manipulation
11357 operations. These allow efficient code generation for some algorithms.
11359 '``llvm.bitreverse.*``' Intrinsics
11360 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11365 This is an overloaded intrinsic function. You can use bitreverse on any
11370 declare i16 @llvm.bitreverse.i16(i16 <id>)
11371 declare i32 @llvm.bitreverse.i32(i32 <id>)
11372 declare i64 @llvm.bitreverse.i64(i64 <id>)
11377 The '``llvm.bitreverse``' family of intrinsics is used to reverse the
11378 bitpattern of an integer value; for example ``0b10110110`` becomes
11384 The ``llvm.bitreverse.iN`` intrinsic returns an iN value that has bit
11385 ``M`` in the input moved to bit ``N-M`` in the output.
11387 '``llvm.bswap.*``' Intrinsics
11388 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11393 This is an overloaded intrinsic function. You can use bswap on any
11394 integer type that is an even number of bytes (i.e. BitWidth % 16 == 0).
11398 declare i16 @llvm.bswap.i16(i16 <id>)
11399 declare i32 @llvm.bswap.i32(i32 <id>)
11400 declare i64 @llvm.bswap.i64(i64 <id>)
11405 The '``llvm.bswap``' family of intrinsics is used to byte swap integer
11406 values with an even number of bytes (positive multiple of 16 bits).
11407 These are useful for performing operations on data that is not in the
11408 target's native byte order.
11413 The ``llvm.bswap.i16`` intrinsic returns an i16 value that has the high
11414 and low byte of the input i16 swapped. Similarly, the ``llvm.bswap.i32``
11415 intrinsic returns an i32 value that has the four bytes of the input i32
11416 swapped, so that if the input bytes are numbered 0, 1, 2, 3 then the
11417 returned i32 will have its bytes in 3, 2, 1, 0 order. The
11418 ``llvm.bswap.i48``, ``llvm.bswap.i64`` and other intrinsics extend this
11419 concept to additional even-byte lengths (6 bytes, 8 bytes and more,
11422 '``llvm.ctpop.*``' Intrinsic
11423 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11428 This is an overloaded intrinsic. You can use llvm.ctpop on any integer
11429 bit width, or on any vector with integer elements. Not all targets
11430 support all bit widths or vector types, however.
11434 declare i8 @llvm.ctpop.i8(i8 <src>)
11435 declare i16 @llvm.ctpop.i16(i16 <src>)
11436 declare i32 @llvm.ctpop.i32(i32 <src>)
11437 declare i64 @llvm.ctpop.i64(i64 <src>)
11438 declare i256 @llvm.ctpop.i256(i256 <src>)
11439 declare <2 x i32> @llvm.ctpop.v2i32(<2 x i32> <src>)
11444 The '``llvm.ctpop``' family of intrinsics counts the number of bits set
11450 The only argument is the value to be counted. The argument may be of any
11451 integer type, or a vector with integer elements. The return type must
11452 match the argument type.
11457 The '``llvm.ctpop``' intrinsic counts the 1's in a variable, or within
11458 each element of a vector.
11460 '``llvm.ctlz.*``' Intrinsic
11461 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
11466 This is an overloaded intrinsic. You can use ``llvm.ctlz`` on any
11467 integer bit width, or any vector whose elements are integers. Not all
11468 targets support all bit widths or vector types, however.
11472 declare i8 @llvm.ctlz.i8 (i8 <src>, i1 <is_zero_undef>)
11473 declare i16 @llvm.ctlz.i16 (i16 <src>, i1 <is_zero_undef>)
11474 declare i32 @llvm.ctlz.i32 (i32 <src>, i1 <is_zero_undef>)
11475 declare i64 @llvm.ctlz.i64 (i64 <src>, i1 <is_zero_undef>)
11476 declare i256 @llvm.ctlz.i256(i256 <src>, i1 <is_zero_undef>)
11477 declare <2 x i32> @llvm.ctlz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
11482 The '``llvm.ctlz``' family of intrinsic functions counts the number of
11483 leading zeros in a variable.
11488 The first argument is the value to be counted. This argument may be of
11489 any integer type, or a vector with integer element type. The return
11490 type must match the first argument type.
11492 The second argument must be a constant and is a flag to indicate whether
11493 the intrinsic should ensure that a zero as the first argument produces a
11494 defined result. Historically some architectures did not provide a
11495 defined result for zero values as efficiently, and many algorithms are
11496 now predicated on avoiding zero-value inputs.
11501 The '``llvm.ctlz``' intrinsic counts the leading (most significant)
11502 zeros in a variable, or within each element of the vector. If
11503 ``src == 0`` then the result is the size in bits of the type of ``src``
11504 if ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
11505 ``llvm.ctlz(i32 2) = 30``.
11507 '``llvm.cttz.*``' Intrinsic
11508 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
11513 This is an overloaded intrinsic. You can use ``llvm.cttz`` on any
11514 integer bit width, or any vector of integer elements. Not all targets
11515 support all bit widths or vector types, however.
11519 declare i8 @llvm.cttz.i8 (i8 <src>, i1 <is_zero_undef>)
11520 declare i16 @llvm.cttz.i16 (i16 <src>, i1 <is_zero_undef>)
11521 declare i32 @llvm.cttz.i32 (i32 <src>, i1 <is_zero_undef>)
11522 declare i64 @llvm.cttz.i64 (i64 <src>, i1 <is_zero_undef>)
11523 declare i256 @llvm.cttz.i256(i256 <src>, i1 <is_zero_undef>)
11524 declare <2 x i32> @llvm.cttz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
11529 The '``llvm.cttz``' family of intrinsic functions counts the number of
11535 The first argument is the value to be counted. This argument may be of
11536 any integer type, or a vector with integer element type. The return
11537 type must match the first argument type.
11539 The second argument must be a constant and is a flag to indicate whether
11540 the intrinsic should ensure that a zero as the first argument produces a
11541 defined result. Historically some architectures did not provide a
11542 defined result for zero values as efficiently, and many algorithms are
11543 now predicated on avoiding zero-value inputs.
11548 The '``llvm.cttz``' intrinsic counts the trailing (least significant)
11549 zeros in a variable, or within each element of a vector. If ``src == 0``
11550 then the result is the size in bits of the type of ``src`` if
11551 ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
11552 ``llvm.cttz(2) = 1``.
11556 Arithmetic with Overflow Intrinsics
11557 -----------------------------------
11559 LLVM provides intrinsics for fast arithmetic overflow checking.
11561 Each of these intrinsics returns a two-element struct. The first
11562 element of this struct contains the result of the corresponding
11563 arithmetic operation modulo 2\ :sup:`n`\ , where n is the bit width of
11564 the result. Therefore, for example, the first element of the struct
11565 returned by ``llvm.sadd.with.overflow.i32`` is always the same as the
11566 result of a 32-bit ``add`` instruction with the same operands, where
11567 the ``add`` is *not* modified by an ``nsw`` or ``nuw`` flag.
11569 The second element of the result is an ``i1`` that is 1 if the
11570 arithmetic operation overflowed and 0 otherwise. An operation
11571 overflows if, for any values of its operands ``A`` and ``B`` and for
11572 any ``N`` larger than the operands' width, ``ext(A op B) to iN`` is
11573 not equal to ``(ext(A) to iN) op (ext(B) to iN)`` where ``ext`` is
11574 ``sext`` for signed overflow and ``zext`` for unsigned overflow, and
11575 ``op`` is the underlying arithmetic operation.
11577 The behavior of these intrinsics is well-defined for all argument
11580 '``llvm.sadd.with.overflow.*``' Intrinsics
11581 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11586 This is an overloaded intrinsic. You can use ``llvm.sadd.with.overflow``
11587 on any integer bit width.
11591 declare {i16, i1} @llvm.sadd.with.overflow.i16(i16 %a, i16 %b)
11592 declare {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
11593 declare {i64, i1} @llvm.sadd.with.overflow.i64(i64 %a, i64 %b)
11598 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
11599 a signed addition of the two arguments, and indicate whether an overflow
11600 occurred during the signed summation.
11605 The arguments (%a and %b) and the first element of the result structure
11606 may be of integer types of any bit width, but they must have the same
11607 bit width. The second element of the result structure must be of type
11608 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
11614 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
11615 a signed addition of the two variables. They return a structure --- the
11616 first element of which is the signed summation, and the second element
11617 of which is a bit specifying if the signed summation resulted in an
11623 .. code-block:: llvm
11625 %res = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
11626 %sum = extractvalue {i32, i1} %res, 0
11627 %obit = extractvalue {i32, i1} %res, 1
11628 br i1 %obit, label %overflow, label %normal
11630 '``llvm.uadd.with.overflow.*``' Intrinsics
11631 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11636 This is an overloaded intrinsic. You can use ``llvm.uadd.with.overflow``
11637 on any integer bit width.
11641 declare {i16, i1} @llvm.uadd.with.overflow.i16(i16 %a, i16 %b)
11642 declare {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
11643 declare {i64, i1} @llvm.uadd.with.overflow.i64(i64 %a, i64 %b)
11648 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
11649 an unsigned addition of the two arguments, and indicate whether a carry
11650 occurred during the unsigned summation.
11655 The arguments (%a and %b) and the first element of the result structure
11656 may be of integer types of any bit width, but they must have the same
11657 bit width. The second element of the result structure must be of type
11658 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
11664 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
11665 an unsigned addition of the two arguments. They return a structure --- the
11666 first element of which is the sum, and the second element of which is a
11667 bit specifying if the unsigned summation resulted in a carry.
11672 .. code-block:: llvm
11674 %res = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
11675 %sum = extractvalue {i32, i1} %res, 0
11676 %obit = extractvalue {i32, i1} %res, 1
11677 br i1 %obit, label %carry, label %normal
11679 '``llvm.ssub.with.overflow.*``' Intrinsics
11680 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11685 This is an overloaded intrinsic. You can use ``llvm.ssub.with.overflow``
11686 on any integer bit width.
11690 declare {i16, i1} @llvm.ssub.with.overflow.i16(i16 %a, i16 %b)
11691 declare {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
11692 declare {i64, i1} @llvm.ssub.with.overflow.i64(i64 %a, i64 %b)
11697 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
11698 a signed subtraction of the two arguments, and indicate whether an
11699 overflow occurred during the signed subtraction.
11704 The arguments (%a and %b) and the first element of the result structure
11705 may be of integer types of any bit width, but they must have the same
11706 bit width. The second element of the result structure must be of type
11707 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
11713 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
11714 a signed subtraction of the two arguments. They return a structure --- the
11715 first element of which is the subtraction, and the second element of
11716 which is a bit specifying if the signed subtraction resulted in an
11722 .. code-block:: llvm
11724 %res = call {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
11725 %sum = extractvalue {i32, i1} %res, 0
11726 %obit = extractvalue {i32, i1} %res, 1
11727 br i1 %obit, label %overflow, label %normal
11729 '``llvm.usub.with.overflow.*``' Intrinsics
11730 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11735 This is an overloaded intrinsic. You can use ``llvm.usub.with.overflow``
11736 on any integer bit width.
11740 declare {i16, i1} @llvm.usub.with.overflow.i16(i16 %a, i16 %b)
11741 declare {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
11742 declare {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
11747 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
11748 an unsigned subtraction of the two arguments, and indicate whether an
11749 overflow occurred during the unsigned subtraction.
11754 The arguments (%a and %b) and the first element of the result structure
11755 may be of integer types of any bit width, but they must have the same
11756 bit width. The second element of the result structure must be of type
11757 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
11763 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
11764 an unsigned subtraction of the two arguments. They return a structure ---
11765 the first element of which is the subtraction, and the second element of
11766 which is a bit specifying if the unsigned subtraction resulted in an
11772 .. code-block:: llvm
11774 %res = call {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
11775 %sum = extractvalue {i32, i1} %res, 0
11776 %obit = extractvalue {i32, i1} %res, 1
11777 br i1 %obit, label %overflow, label %normal
11779 '``llvm.smul.with.overflow.*``' Intrinsics
11780 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11785 This is an overloaded intrinsic. You can use ``llvm.smul.with.overflow``
11786 on any integer bit width.
11790 declare {i16, i1} @llvm.smul.with.overflow.i16(i16 %a, i16 %b)
11791 declare {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
11792 declare {i64, i1} @llvm.smul.with.overflow.i64(i64 %a, i64 %b)
11797 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
11798 a signed multiplication of the two arguments, and indicate whether an
11799 overflow occurred during the signed multiplication.
11804 The arguments (%a and %b) and the first element of the result structure
11805 may be of integer types of any bit width, but they must have the same
11806 bit width. The second element of the result structure must be of type
11807 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
11813 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
11814 a signed multiplication of the two arguments. They return a structure ---
11815 the first element of which is the multiplication, and the second element
11816 of which is a bit specifying if the signed multiplication resulted in an
11822 .. code-block:: llvm
11824 %res = call {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
11825 %sum = extractvalue {i32, i1} %res, 0
11826 %obit = extractvalue {i32, i1} %res, 1
11827 br i1 %obit, label %overflow, label %normal
11829 '``llvm.umul.with.overflow.*``' Intrinsics
11830 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11835 This is an overloaded intrinsic. You can use ``llvm.umul.with.overflow``
11836 on any integer bit width.
11840 declare {i16, i1} @llvm.umul.with.overflow.i16(i16 %a, i16 %b)
11841 declare {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
11842 declare {i64, i1} @llvm.umul.with.overflow.i64(i64 %a, i64 %b)
11847 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
11848 a unsigned multiplication of the two arguments, and indicate whether an
11849 overflow occurred during the unsigned multiplication.
11854 The arguments (%a and %b) and the first element of the result structure
11855 may be of integer types of any bit width, but they must have the same
11856 bit width. The second element of the result structure must be of type
11857 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
11863 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
11864 an unsigned multiplication of the two arguments. They return a structure ---
11865 the first element of which is the multiplication, and the second
11866 element of which is a bit specifying if the unsigned multiplication
11867 resulted in an overflow.
11872 .. code-block:: llvm
11874 %res = call {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
11875 %sum = extractvalue {i32, i1} %res, 0
11876 %obit = extractvalue {i32, i1} %res, 1
11877 br i1 %obit, label %overflow, label %normal
11879 Specialised Arithmetic Intrinsics
11880 ---------------------------------
11882 '``llvm.canonicalize.*``' Intrinsic
11883 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11890 declare float @llvm.canonicalize.f32(float %a)
11891 declare double @llvm.canonicalize.f64(double %b)
11896 The '``llvm.canonicalize.*``' intrinsic returns the platform specific canonical
11897 encoding of a floating point number. This canonicalization is useful for
11898 implementing certain numeric primitives such as frexp. The canonical encoding is
11899 defined by IEEE-754-2008 to be:
11903 2.1.8 canonical encoding: The preferred encoding of a floating-point
11904 representation in a format. Applied to declets, significands of finite
11905 numbers, infinities, and NaNs, especially in decimal formats.
11907 This operation can also be considered equivalent to the IEEE-754-2008
11908 conversion of a floating-point value to the same format. NaNs are handled
11909 according to section 6.2.
11911 Examples of non-canonical encodings:
11913 - x87 pseudo denormals, pseudo NaNs, pseudo Infinity, Unnormals. These are
11914 converted to a canonical representation per hardware-specific protocol.
11915 - Many normal decimal floating point numbers have non-canonical alternative
11917 - Some machines, like GPUs or ARMv7 NEON, do not support subnormal values.
11918 These are treated as non-canonical encodings of zero and will be flushed to
11919 a zero of the same sign by this operation.
11921 Note that per IEEE-754-2008 6.2, systems that support signaling NaNs with
11922 default exception handling must signal an invalid exception, and produce a
11925 This function should always be implementable as multiplication by 1.0, provided
11926 that the compiler does not constant fold the operation. Likewise, division by
11927 1.0 and ``llvm.minnum(x, x)`` are possible implementations. Addition with
11928 -0.0 is also sufficient provided that the rounding mode is not -Infinity.
11930 ``@llvm.canonicalize`` must preserve the equality relation. That is:
11932 - ``(@llvm.canonicalize(x) == x)`` is equivalent to ``(x == x)``
11933 - ``(@llvm.canonicalize(x) == @llvm.canonicalize(y))`` is equivalent to
11936 Additionally, the sign of zero must be conserved:
11937 ``@llvm.canonicalize(-0.0) = -0.0`` and ``@llvm.canonicalize(+0.0) = +0.0``
11939 The payload bits of a NaN must be conserved, with two exceptions.
11940 First, environments which use only a single canonical representation of NaN
11941 must perform said canonicalization. Second, SNaNs must be quieted per the
11944 The canonicalization operation may be optimized away if:
11946 - The input is known to be canonical. For example, it was produced by a
11947 floating-point operation that is required by the standard to be canonical.
11948 - The result is consumed only by (or fused with) other floating-point
11949 operations. That is, the bits of the floating point value are not examined.
11951 '``llvm.fmuladd.*``' Intrinsic
11952 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11959 declare float @llvm.fmuladd.f32(float %a, float %b, float %c)
11960 declare double @llvm.fmuladd.f64(double %a, double %b, double %c)
11965 The '``llvm.fmuladd.*``' intrinsic functions represent multiply-add
11966 expressions that can be fused if the code generator determines that (a) the
11967 target instruction set has support for a fused operation, and (b) that the
11968 fused operation is more efficient than the equivalent, separate pair of mul
11969 and add instructions.
11974 The '``llvm.fmuladd.*``' intrinsics each take three arguments: two
11975 multiplicands, a and b, and an addend c.
11984 %0 = call float @llvm.fmuladd.f32(%a, %b, %c)
11986 is equivalent to the expression a \* b + c, except that rounding will
11987 not be performed between the multiplication and addition steps if the
11988 code generator fuses the operations. Fusion is not guaranteed, even if
11989 the target platform supports it. If a fused multiply-add is required the
11990 corresponding llvm.fma.\* intrinsic function should be used
11991 instead. This never sets errno, just as '``llvm.fma.*``'.
11996 .. code-block:: llvm
11998 %r2 = call float @llvm.fmuladd.f32(float %a, float %b, float %c) ; yields float:r2 = (a * b) + c
12001 Experimental Vector Reduction Intrinsics
12002 ----------------------------------------
12004 Horizontal reductions of vectors can be expressed using the following
12005 intrinsics. Each one takes a vector operand as an input and applies its
12006 respective operation across all elements of the vector, returning a single
12007 scalar result of the same element type.
12010 '``llvm.experimental.vector.reduce.add.*``' Intrinsic
12011 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12018 declare i32 @llvm.experimental.vector.reduce.add.i32.v4i32(<4 x i32> %a)
12019 declare i64 @llvm.experimental.vector.reduce.add.i64.v2i64(<2 x i64> %a)
12024 The '``llvm.experimental.vector.reduce.add.*``' intrinsics do an integer ``ADD``
12025 reduction of a vector, returning the result as a scalar. The return type matches
12026 the element-type of the vector input.
12030 The argument to this intrinsic must be a vector of integer values.
12032 '``llvm.experimental.vector.reduce.fadd.*``' Intrinsic
12033 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12040 declare float @llvm.experimental.vector.reduce.fadd.f32.v4f32(float %acc, <4 x float> %a)
12041 declare double @llvm.experimental.vector.reduce.fadd.f64.v2f64(double %acc, <2 x double> %a)
12046 The '``llvm.experimental.vector.reduce.fadd.*``' intrinsics do a floating point
12047 ``ADD`` reduction of a vector, returning the result as a scalar. The return type
12048 matches the element-type of the vector input.
12050 If the intrinsic call has fast-math flags, then the reduction will not preserve
12051 the associativity of an equivalent scalarized counterpart. If it does not have
12052 fast-math flags, then the reduction will be *ordered*, implying that the
12053 operation respects the associativity of a scalarized reduction.
12058 The first argument to this intrinsic is a scalar accumulator value, which is
12059 only used when there are no fast-math flags attached. This argument may be undef
12060 when fast-math flags are used.
12062 The second argument must be a vector of floating point values.
12067 .. code-block:: llvm
12069 %fast = call fast float @llvm.experimental.vector.reduce.fadd.f32.v4f32(float undef, <4 x float> %input) ; fast reduction
12070 %ord = call float @llvm.experimental.vector.reduce.fadd.f32.v4f32(float %acc, <4 x float> %input) ; ordered reduction
12073 '``llvm.experimental.vector.reduce.mul.*``' Intrinsic
12074 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12081 declare i32 @llvm.experimental.vector.reduce.mul.i32.v4i32(<4 x i32> %a)
12082 declare i64 @llvm.experimental.vector.reduce.mul.i64.v2i64(<2 x i64> %a)
12087 The '``llvm.experimental.vector.reduce.mul.*``' intrinsics do an integer ``MUL``
12088 reduction of a vector, returning the result as a scalar. The return type matches
12089 the element-type of the vector input.
12093 The argument to this intrinsic must be a vector of integer values.
12095 '``llvm.experimental.vector.reduce.fmul.*``' Intrinsic
12096 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12103 declare float @llvm.experimental.vector.reduce.fmul.f32.v4f32(float %acc, <4 x float> %a)
12104 declare double @llvm.experimental.vector.reduce.fmul.f64.v2f64(double %acc, <2 x double> %a)
12109 The '``llvm.experimental.vector.reduce.fmul.*``' intrinsics do a floating point
12110 ``MUL`` reduction of a vector, returning the result as a scalar. The return type
12111 matches the element-type of the vector input.
12113 If the intrinsic call has fast-math flags, then the reduction will not preserve
12114 the associativity of an equivalent scalarized counterpart. If it does not have
12115 fast-math flags, then the reduction will be *ordered*, implying that the
12116 operation respects the associativity of a scalarized reduction.
12121 The first argument to this intrinsic is a scalar accumulator value, which is
12122 only used when there are no fast-math flags attached. This argument may be undef
12123 when fast-math flags are used.
12125 The second argument must be a vector of floating point values.
12130 .. code-block:: llvm
12132 %fast = call fast float @llvm.experimental.vector.reduce.fmul.f32.v4f32(float undef, <4 x float> %input) ; fast reduction
12133 %ord = call float @llvm.experimental.vector.reduce.fmul.f32.v4f32(float %acc, <4 x float> %input) ; ordered reduction
12135 '``llvm.experimental.vector.reduce.and.*``' Intrinsic
12136 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12143 declare i32 @llvm.experimental.vector.reduce.and.i32.v4i32(<4 x i32> %a)
12148 The '``llvm.experimental.vector.reduce.and.*``' intrinsics do a bitwise ``AND``
12149 reduction of a vector, returning the result as a scalar. The return type matches
12150 the element-type of the vector input.
12154 The argument to this intrinsic must be a vector of integer values.
12156 '``llvm.experimental.vector.reduce.or.*``' Intrinsic
12157 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12164 declare i32 @llvm.experimental.vector.reduce.or.i32.v4i32(<4 x i32> %a)
12169 The '``llvm.experimental.vector.reduce.or.*``' intrinsics do a bitwise ``OR`` reduction
12170 of a vector, returning the result as a scalar. The return type matches the
12171 element-type of the vector input.
12175 The argument to this intrinsic must be a vector of integer values.
12177 '``llvm.experimental.vector.reduce.xor.*``' Intrinsic
12178 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12185 declare i32 @llvm.experimental.vector.reduce.xor.i32.v4i32(<4 x i32> %a)
12190 The '``llvm.experimental.vector.reduce.xor.*``' intrinsics do a bitwise ``XOR``
12191 reduction of a vector, returning the result as a scalar. The return type matches
12192 the element-type of the vector input.
12196 The argument to this intrinsic must be a vector of integer values.
12198 '``llvm.experimental.vector.reduce.smax.*``' Intrinsic
12199 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12206 declare i32 @llvm.experimental.vector.reduce.smax.i32.v4i32(<4 x i32> %a)
12211 The '``llvm.experimental.vector.reduce.smax.*``' intrinsics do a signed integer
12212 ``MAX`` reduction of a vector, returning the result as a scalar. The return type
12213 matches the element-type of the vector input.
12217 The argument to this intrinsic must be a vector of integer values.
12219 '``llvm.experimental.vector.reduce.smin.*``' Intrinsic
12220 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12227 declare i32 @llvm.experimental.vector.reduce.smin.i32.v4i32(<4 x i32> %a)
12232 The '``llvm.experimental.vector.reduce.smin.*``' intrinsics do a signed integer
12233 ``MIN`` reduction of a vector, returning the result as a scalar. The return type
12234 matches the element-type of the vector input.
12238 The argument to this intrinsic must be a vector of integer values.
12240 '``llvm.experimental.vector.reduce.umax.*``' Intrinsic
12241 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12248 declare i32 @llvm.experimental.vector.reduce.umax.i32.v4i32(<4 x i32> %a)
12253 The '``llvm.experimental.vector.reduce.umax.*``' intrinsics do an unsigned
12254 integer ``MAX`` reduction of a vector, returning the result as a scalar. The
12255 return type matches the element-type of the vector input.
12259 The argument to this intrinsic must be a vector of integer values.
12261 '``llvm.experimental.vector.reduce.umin.*``' Intrinsic
12262 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12269 declare i32 @llvm.experimental.vector.reduce.umin.i32.v4i32(<4 x i32> %a)
12274 The '``llvm.experimental.vector.reduce.umin.*``' intrinsics do an unsigned
12275 integer ``MIN`` reduction of a vector, returning the result as a scalar. The
12276 return type matches the element-type of the vector input.
12280 The argument to this intrinsic must be a vector of integer values.
12282 '``llvm.experimental.vector.reduce.fmax.*``' Intrinsic
12283 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12290 declare float @llvm.experimental.vector.reduce.fmax.f32.v4f32(<4 x float> %a)
12291 declare double @llvm.experimental.vector.reduce.fmax.f64.v2f64(<2 x double> %a)
12296 The '``llvm.experimental.vector.reduce.fmax.*``' intrinsics do a floating point
12297 ``MAX`` reduction of a vector, returning the result as a scalar. The return type
12298 matches the element-type of the vector input.
12300 If the intrinsic call has the ``nnan`` fast-math flag then the operation can
12301 assume that NaNs are not present in the input vector.
12305 The argument to this intrinsic must be a vector of floating point values.
12307 '``llvm.experimental.vector.reduce.fmin.*``' Intrinsic
12308 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12315 declare float @llvm.experimental.vector.reduce.fmin.f32.v4f32(<4 x float> %a)
12316 declare double @llvm.experimental.vector.reduce.fmin.f64.v2f64(<2 x double> %a)
12321 The '``llvm.experimental.vector.reduce.fmin.*``' intrinsics do a floating point
12322 ``MIN`` reduction of a vector, returning the result as a scalar. The return type
12323 matches the element-type of the vector input.
12325 If the intrinsic call has the ``nnan`` fast-math flag then the operation can
12326 assume that NaNs are not present in the input vector.
12330 The argument to this intrinsic must be a vector of floating point values.
12332 Half Precision Floating Point Intrinsics
12333 ----------------------------------------
12335 For most target platforms, half precision floating point is a
12336 storage-only format. This means that it is a dense encoding (in memory)
12337 but does not support computation in the format.
12339 This means that code must first load the half-precision floating point
12340 value as an i16, then convert it to float with
12341 :ref:`llvm.convert.from.fp16 <int_convert_from_fp16>`. Computation can
12342 then be performed on the float value (including extending to double
12343 etc). To store the value back to memory, it is first converted to float
12344 if needed, then converted to i16 with
12345 :ref:`llvm.convert.to.fp16 <int_convert_to_fp16>`, then storing as an
12348 .. _int_convert_to_fp16:
12350 '``llvm.convert.to.fp16``' Intrinsic
12351 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12358 declare i16 @llvm.convert.to.fp16.f32(float %a)
12359 declare i16 @llvm.convert.to.fp16.f64(double %a)
12364 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion from a
12365 conventional floating point type to half precision floating point format.
12370 The intrinsic function contains single argument - the value to be
12376 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion from a
12377 conventional floating point format to half precision floating point format. The
12378 return value is an ``i16`` which contains the converted number.
12383 .. code-block:: llvm
12385 %res = call i16 @llvm.convert.to.fp16.f32(float %a)
12386 store i16 %res, i16* @x, align 2
12388 .. _int_convert_from_fp16:
12390 '``llvm.convert.from.fp16``' Intrinsic
12391 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12398 declare float @llvm.convert.from.fp16.f32(i16 %a)
12399 declare double @llvm.convert.from.fp16.f64(i16 %a)
12404 The '``llvm.convert.from.fp16``' intrinsic function performs a
12405 conversion from half precision floating point format to single precision
12406 floating point format.
12411 The intrinsic function contains single argument - the value to be
12417 The '``llvm.convert.from.fp16``' intrinsic function performs a
12418 conversion from half single precision floating point format to single
12419 precision floating point format. The input half-float value is
12420 represented by an ``i16`` value.
12425 .. code-block:: llvm
12427 %a = load i16, i16* @x, align 2
12428 %res = call float @llvm.convert.from.fp16(i16 %a)
12430 .. _dbg_intrinsics:
12432 Debugger Intrinsics
12433 -------------------
12435 The LLVM debugger intrinsics (which all start with ``llvm.dbg.``
12436 prefix), are described in the `LLVM Source Level
12437 Debugging <SourceLevelDebugging.html#format-common-intrinsics>`_
12440 Exception Handling Intrinsics
12441 -----------------------------
12443 The LLVM exception handling intrinsics (which all start with
12444 ``llvm.eh.`` prefix), are described in the `LLVM Exception
12445 Handling <ExceptionHandling.html#format-common-intrinsics>`_ document.
12447 .. _int_trampoline:
12449 Trampoline Intrinsics
12450 ---------------------
12452 These intrinsics make it possible to excise one parameter, marked with
12453 the :ref:`nest <nest>` attribute, from a function. The result is a
12454 callable function pointer lacking the nest parameter - the caller does
12455 not need to provide a value for it. Instead, the value to use is stored
12456 in advance in a "trampoline", a block of memory usually allocated on the
12457 stack, which also contains code to splice the nest value into the
12458 argument list. This is used to implement the GCC nested function address
12461 For example, if the function is ``i32 f(i8* nest %c, i32 %x, i32 %y)``
12462 then the resulting function pointer has signature ``i32 (i32, i32)*``.
12463 It can be created as follows:
12465 .. code-block:: llvm
12467 %tramp = alloca [10 x i8], align 4 ; size and alignment only correct for X86
12468 %tramp1 = getelementptr [10 x i8], [10 x i8]* %tramp, i32 0, i32 0
12469 call i8* @llvm.init.trampoline(i8* %tramp1, i8* bitcast (i32 (i8*, i32, i32)* @f to i8*), i8* %nval)
12470 %p = call i8* @llvm.adjust.trampoline(i8* %tramp1)
12471 %fp = bitcast i8* %p to i32 (i32, i32)*
12473 The call ``%val = call i32 %fp(i32 %x, i32 %y)`` is then equivalent to
12474 ``%val = call i32 %f(i8* %nval, i32 %x, i32 %y)``.
12478 '``llvm.init.trampoline``' Intrinsic
12479 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12486 declare void @llvm.init.trampoline(i8* <tramp>, i8* <func>, i8* <nval>)
12491 This fills the memory pointed to by ``tramp`` with executable code,
12492 turning it into a trampoline.
12497 The ``llvm.init.trampoline`` intrinsic takes three arguments, all
12498 pointers. The ``tramp`` argument must point to a sufficiently large and
12499 sufficiently aligned block of memory; this memory is written to by the
12500 intrinsic. Note that the size and the alignment are target-specific -
12501 LLVM currently provides no portable way of determining them, so a
12502 front-end that generates this intrinsic needs to have some
12503 target-specific knowledge. The ``func`` argument must hold a function
12504 bitcast to an ``i8*``.
12509 The block of memory pointed to by ``tramp`` is filled with target
12510 dependent code, turning it into a function. Then ``tramp`` needs to be
12511 passed to :ref:`llvm.adjust.trampoline <int_at>` to get a pointer which can
12512 be :ref:`bitcast (to a new function) and called <int_trampoline>`. The new
12513 function's signature is the same as that of ``func`` with any arguments
12514 marked with the ``nest`` attribute removed. At most one such ``nest``
12515 argument is allowed, and it must be of pointer type. Calling the new
12516 function is equivalent to calling ``func`` with the same argument list,
12517 but with ``nval`` used for the missing ``nest`` argument. If, after
12518 calling ``llvm.init.trampoline``, the memory pointed to by ``tramp`` is
12519 modified, then the effect of any later call to the returned function
12520 pointer is undefined.
12524 '``llvm.adjust.trampoline``' Intrinsic
12525 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12532 declare i8* @llvm.adjust.trampoline(i8* <tramp>)
12537 This performs any required machine-specific adjustment to the address of
12538 a trampoline (passed as ``tramp``).
12543 ``tramp`` must point to a block of memory which already has trampoline
12544 code filled in by a previous call to
12545 :ref:`llvm.init.trampoline <int_it>`.
12550 On some architectures the address of the code to be executed needs to be
12551 different than the address where the trampoline is actually stored. This
12552 intrinsic returns the executable address corresponding to ``tramp``
12553 after performing the required machine specific adjustments. The pointer
12554 returned can then be :ref:`bitcast and executed <int_trampoline>`.
12556 .. _int_mload_mstore:
12558 Masked Vector Load and Store Intrinsics
12559 ---------------------------------------
12561 LLVM provides intrinsics for predicated vector load and store operations. The predicate is specified by a mask operand, which holds one bit per vector element, switching the associated vector lane on or off. The memory addresses corresponding to the "off" lanes are not accessed. When all bits of the mask are on, the intrinsic is identical to a regular vector load or store. When all bits are off, no memory is accessed.
12565 '``llvm.masked.load.*``' Intrinsics
12566 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12570 This is an overloaded intrinsic. The loaded data is a vector of any integer, floating point or pointer data type.
12574 declare <16 x float> @llvm.masked.load.v16f32.p0v16f32 (<16 x float>* <ptr>, i32 <alignment>, <16 x i1> <mask>, <16 x float> <passthru>)
12575 declare <2 x double> @llvm.masked.load.v2f64.p0v2f64 (<2 x double>* <ptr>, i32 <alignment>, <2 x i1> <mask>, <2 x double> <passthru>)
12576 ;; The data is a vector of pointers to double
12577 declare <8 x double*> @llvm.masked.load.v8p0f64.p0v8p0f64 (<8 x double*>* <ptr>, i32 <alignment>, <8 x i1> <mask>, <8 x double*> <passthru>)
12578 ;; The data is a vector of function pointers
12579 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>)
12584 Reads a vector from memory according to the provided mask. The mask holds a bit for each vector lane, and is used to prevent memory accesses to the masked-off lanes. The masked-off lanes in the result vector are taken from the corresponding lanes of the '``passthru``' operand.
12590 The first operand is the base pointer for the load. The second operand is the alignment of the source location. It must be a constant integer value. The third operand, mask, is a vector of boolean values with the same number of elements as the return type. The fourth is a pass-through value that is used to fill the masked-off lanes of the result. The return type, underlying type of the base pointer and the type of the '``passthru``' operand are the same vector types.
12596 The '``llvm.masked.load``' intrinsic is designed for conditional reading of selected vector elements in a single IR operation. It is useful for targets that support vector masked loads and allows vectorizing predicated basic blocks on these targets. Other targets may support this intrinsic differently, for example by lowering it into a sequence of branches that guard scalar load operations.
12597 The result of this operation is equivalent to a regular vector load instruction followed by a 'select' between the loaded and the passthru values, predicated on the same mask. However, using this intrinsic prevents exceptions on memory access to masked-off lanes.
12602 %res = call <16 x float> @llvm.masked.load.v16f32.p0v16f32 (<16 x float>* %ptr, i32 4, <16 x i1>%mask, <16 x float> %passthru)
12604 ;; The result of the two following instructions is identical aside from potential memory access exception
12605 %loadlal = load <16 x float>, <16 x float>* %ptr, align 4
12606 %res = select <16 x i1> %mask, <16 x float> %loadlal, <16 x float> %passthru
12610 '``llvm.masked.store.*``' Intrinsics
12611 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12615 This is an overloaded intrinsic. The data stored in memory is a vector of any integer, floating point or pointer data type.
12619 declare void @llvm.masked.store.v8i32.p0v8i32 (<8 x i32> <value>, <8 x i32>* <ptr>, i32 <alignment>, <8 x i1> <mask>)
12620 declare void @llvm.masked.store.v16f32.p0v16f32 (<16 x float> <value>, <16 x float>* <ptr>, i32 <alignment>, <16 x i1> <mask>)
12621 ;; The data is a vector of pointers to double
12622 declare void @llvm.masked.store.v8p0f64.p0v8p0f64 (<8 x double*> <value>, <8 x double*>* <ptr>, i32 <alignment>, <8 x i1> <mask>)
12623 ;; The data is a vector of function pointers
12624 declare void @llvm.masked.store.v4p0f_i32f.p0v4p0f_i32f (<4 x i32 ()*> <value>, <4 x i32 ()*>* <ptr>, i32 <alignment>, <4 x i1> <mask>)
12629 Writes a vector to memory according to the provided mask. The mask holds a bit for each vector lane, and is used to prevent memory accesses to the masked-off lanes.
12634 The first operand is the vector value to be written to memory. The second operand is the base pointer for the store, it has the same underlying type as the value operand. The third operand is the alignment of the destination location. The fourth operand, mask, is a vector of boolean values. The types of the mask and the value operand must have the same number of vector elements.
12640 The '``llvm.masked.store``' intrinsics is designed for conditional writing of selected vector elements in a single IR operation. It is useful for targets that support vector masked store and allows vectorizing predicated basic blocks on these targets. Other targets may support this intrinsic differently, for example by lowering it into a sequence of branches that guard scalar store operations.
12641 The result of this operation is equivalent to a load-modify-store sequence. However, using this intrinsic prevents exceptions and data races on memory access to masked-off lanes.
12645 call void @llvm.masked.store.v16f32.p0v16f32(<16 x float> %value, <16 x float>* %ptr, i32 4, <16 x i1> %mask)
12647 ;; The result of the following instructions is identical aside from potential data races and memory access exceptions
12648 %oldval = load <16 x float>, <16 x float>* %ptr, align 4
12649 %res = select <16 x i1> %mask, <16 x float> %value, <16 x float> %oldval
12650 store <16 x float> %res, <16 x float>* %ptr, align 4
12653 Masked Vector Gather and Scatter Intrinsics
12654 -------------------------------------------
12656 LLVM provides intrinsics for vector gather and scatter operations. They are similar to :ref:`Masked Vector Load and Store <int_mload_mstore>`, except they are designed for arbitrary memory accesses, rather than sequential memory accesses. Gather and scatter also employ a mask operand, which holds one bit per vector element, switching the associated vector lane on or off. The memory addresses corresponding to the "off" lanes are not accessed. When all bits are off, no memory is accessed.
12660 '``llvm.masked.gather.*``' Intrinsics
12661 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12665 This is an overloaded intrinsic. The loaded data are multiple scalar values of any integer, floating point or pointer data type gathered together into one vector.
12669 declare <16 x float> @llvm.masked.gather.v16f32.v16p0f32 (<16 x float*> <ptrs>, i32 <alignment>, <16 x i1> <mask>, <16 x float> <passthru>)
12670 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>)
12671 declare <8 x float*> @llvm.masked.gather.v8p0f32.v8p0p0f32 (<8 x float**> <ptrs>, i32 <alignment>, <8 x i1> <mask>, <8 x float*> <passthru>)
12676 Reads scalar values from arbitrary memory locations and gathers them into one vector. The memory locations are provided in the vector of pointers '``ptrs``'. The memory is accessed according to the provided mask. The mask holds a bit for each vector lane, and is used to prevent memory accesses to the masked-off lanes. The masked-off lanes in the result vector are taken from the corresponding lanes of the '``passthru``' operand.
12682 The first operand is a vector of pointers which holds all memory addresses to read. The second operand is an alignment of the source addresses. It must be a constant integer value. The third operand, mask, is a vector of boolean values with the same number of elements as the return type. The fourth is a pass-through value that is used to fill the masked-off lanes of the result. The return type, underlying type of the vector of pointers and the type of the '``passthru``' operand are the same vector types.
12688 The '``llvm.masked.gather``' intrinsic is designed for conditional reading of multiple scalar values from arbitrary memory locations in a single IR operation. It is useful for targets that support vector masked gathers and allows vectorizing basic blocks with data and control divergence. Other targets may support this intrinsic differently, for example by lowering it into a sequence of scalar load operations.
12689 The semantics of this operation are equivalent to a sequence of conditional scalar loads with subsequent gathering all loaded values into a single vector. The mask restricts memory access to certain lanes and facilitates vectorization of predicated basic blocks.
12694 %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)
12696 ;; The gather with all-true mask is equivalent to the following instruction sequence
12697 %ptr0 = extractelement <4 x double*> %ptrs, i32 0
12698 %ptr1 = extractelement <4 x double*> %ptrs, i32 1
12699 %ptr2 = extractelement <4 x double*> %ptrs, i32 2
12700 %ptr3 = extractelement <4 x double*> %ptrs, i32 3
12702 %val0 = load double, double* %ptr0, align 8
12703 %val1 = load double, double* %ptr1, align 8
12704 %val2 = load double, double* %ptr2, align 8
12705 %val3 = load double, double* %ptr3, align 8
12707 %vec0 = insertelement <4 x double>undef, %val0, 0
12708 %vec01 = insertelement <4 x double>%vec0, %val1, 1
12709 %vec012 = insertelement <4 x double>%vec01, %val2, 2
12710 %vec0123 = insertelement <4 x double>%vec012, %val3, 3
12714 '``llvm.masked.scatter.*``' Intrinsics
12715 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12719 This is an overloaded intrinsic. The data stored in memory is a vector of any integer, floating point or pointer data type. Each vector element is stored in an arbitrary memory address. Scatter with overlapping addresses is guaranteed to be ordered from least-significant to most-significant element.
12723 declare void @llvm.masked.scatter.v8i32.v8p0i32 (<8 x i32> <value>, <8 x i32*> <ptrs>, i32 <alignment>, <8 x i1> <mask>)
12724 declare void @llvm.masked.scatter.v16f32.v16p1f32 (<16 x float> <value>, <16 x float addrspace(1)*> <ptrs>, i32 <alignment>, <16 x i1> <mask>)
12725 declare void @llvm.masked.scatter.v4p0f64.v4p0p0f64 (<4 x double*> <value>, <4 x double**> <ptrs>, i32 <alignment>, <4 x i1> <mask>)
12730 Writes each element from the value vector to the corresponding memory address. The memory addresses are represented as a vector of pointers. Writing is done according to the provided mask. The mask holds a bit for each vector lane, and is used to prevent memory accesses to the masked-off lanes.
12735 The first operand is a vector value to be written to memory. The second operand is a vector of pointers, pointing to where the value elements should be stored. It has the same underlying type as the value operand. The third operand is an alignment of the destination addresses. The fourth operand, mask, is a vector of boolean values. The types of the mask and the value operand must have the same number of vector elements.
12741 The '``llvm.masked.scatter``' intrinsics is designed for writing selected vector elements to arbitrary memory addresses in a single IR operation. The operation may be conditional, when not all bits in the mask are switched on. It is useful for targets that support vector masked scatter and allows vectorizing basic blocks with data and control divergence. Other targets may support this intrinsic differently, for example by lowering it into a sequence of branches that guard scalar store operations.
12745 ;; This instruction unconditionally stores data vector in multiple addresses
12746 call @llvm.masked.scatter.v8i32.v8p0i32 (<8 x i32> %value, <8 x i32*> %ptrs, i32 4, <8 x i1> <true, true, .. true>)
12748 ;; It is equivalent to a list of scalar stores
12749 %val0 = extractelement <8 x i32> %value, i32 0
12750 %val1 = extractelement <8 x i32> %value, i32 1
12752 %val7 = extractelement <8 x i32> %value, i32 7
12753 %ptr0 = extractelement <8 x i32*> %ptrs, i32 0
12754 %ptr1 = extractelement <8 x i32*> %ptrs, i32 1
12756 %ptr7 = extractelement <8 x i32*> %ptrs, i32 7
12757 ;; Note: the order of the following stores is important when they overlap:
12758 store i32 %val0, i32* %ptr0, align 4
12759 store i32 %val1, i32* %ptr1, align 4
12761 store i32 %val7, i32* %ptr7, align 4
12767 This class of intrinsics provides information about the lifetime of
12768 memory objects and ranges where variables are immutable.
12772 '``llvm.lifetime.start``' Intrinsic
12773 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12780 declare void @llvm.lifetime.start(i64 <size>, i8* nocapture <ptr>)
12785 The '``llvm.lifetime.start``' intrinsic specifies the start of a memory
12791 The first argument is a constant integer representing the size of the
12792 object, or -1 if it is variable sized. The second argument is a pointer
12798 This intrinsic indicates that before this point in the code, the value
12799 of the memory pointed to by ``ptr`` is dead. This means that it is known
12800 to never be used and has an undefined value. A load from the pointer
12801 that precedes this intrinsic can be replaced with ``'undef'``.
12805 '``llvm.lifetime.end``' Intrinsic
12806 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12813 declare void @llvm.lifetime.end(i64 <size>, i8* nocapture <ptr>)
12818 The '``llvm.lifetime.end``' intrinsic specifies the end of a memory
12824 The first argument is a constant integer representing the size of the
12825 object, or -1 if it is variable sized. The second argument is a pointer
12831 This intrinsic indicates that after this point in the code, the value of
12832 the memory pointed to by ``ptr`` is dead. This means that it is known to
12833 never be used and has an undefined value. Any stores into the memory
12834 object following this intrinsic may be removed as dead.
12836 '``llvm.invariant.start``' Intrinsic
12837 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12841 This is an overloaded intrinsic. The memory object can belong to any address space.
12845 declare {}* @llvm.invariant.start.p0i8(i64 <size>, i8* nocapture <ptr>)
12850 The '``llvm.invariant.start``' intrinsic specifies that the contents of
12851 a memory object will not change.
12856 The first argument is a constant integer representing the size of the
12857 object, or -1 if it is variable sized. The second argument is a pointer
12863 This intrinsic indicates that until an ``llvm.invariant.end`` that uses
12864 the return value, the referenced memory location is constant and
12867 '``llvm.invariant.end``' Intrinsic
12868 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12872 This is an overloaded intrinsic. The memory object can belong to any address space.
12876 declare void @llvm.invariant.end.p0i8({}* <start>, i64 <size>, i8* nocapture <ptr>)
12881 The '``llvm.invariant.end``' intrinsic specifies that the contents of a
12882 memory object are mutable.
12887 The first argument is the matching ``llvm.invariant.start`` intrinsic.
12888 The second argument is a constant integer representing the size of the
12889 object, or -1 if it is variable sized and the third argument is a
12890 pointer to the object.
12895 This intrinsic indicates that the memory is mutable again.
12897 '``llvm.invariant.group.barrier``' Intrinsic
12898 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12902 This is an overloaded intrinsic. The memory object can belong to any address
12903 space. The returned pointer must belong to the same address space as the
12908 declare i8* @llvm.invariant.group.barrier.p0i8(i8* <ptr>)
12913 The '``llvm.invariant.group.barrier``' intrinsic can be used when an invariant
12914 established by invariant.group metadata no longer holds, to obtain a new pointer
12915 value that does not carry the invariant information.
12921 The ``llvm.invariant.group.barrier`` takes only one argument, which is
12922 the pointer to the memory for which the ``invariant.group`` no longer holds.
12927 Returns another pointer that aliases its argument but which is considered different
12928 for the purposes of ``load``/``store`` ``invariant.group`` metadata.
12932 Constrained Floating Point Intrinsics
12933 -------------------------------------
12935 These intrinsics are used to provide special handling of floating point
12936 operations when specific rounding mode or floating point exception behavior is
12937 required. By default, LLVM optimization passes assume that the rounding mode is
12938 round-to-nearest and that floating point exceptions will not be monitored.
12939 Constrained FP intrinsics are used to support non-default rounding modes and
12940 accurately preserve exception behavior without compromising LLVM's ability to
12941 optimize FP code when the default behavior is used.
12943 Each of these intrinsics corresponds to a normal floating point operation. The
12944 first two arguments and the return value are the same as the corresponding FP
12947 The third argument is a metadata argument specifying the rounding mode to be
12948 assumed. This argument must be one of the following strings:
12958 If this argument is "round.dynamic" optimization passes must assume that the
12959 rounding mode is unknown and may change at runtime. No transformations that
12960 depend on rounding mode may be performed in this case.
12962 The other possible values for the rounding mode argument correspond to the
12963 similarly named IEEE rounding modes. If the argument is any of these values
12964 optimization passes may perform transformations as long as they are consistent
12965 with the specified rounding mode.
12967 For example, 'x-0'->'x' is not a valid transformation if the rounding mode is
12968 "round.downward" or "round.dynamic" because if the value of 'x' is +0 then
12969 'x-0' should evaluate to '-0' when rounding downward. However, this
12970 transformation is legal for all other rounding modes.
12972 For values other than "round.dynamic" optimization passes may assume that the
12973 actual runtime rounding mode (as defined in a target-specific manner) matches
12974 the specified rounding mode, but this is not guaranteed. Using a specific
12975 non-dynamic rounding mode which does not match the actual rounding mode at
12976 runtime results in undefined behavior.
12978 The fourth argument to the constrained floating point intrinsics specifies the
12979 required exception behavior. This argument must be one of the following
12988 If this argument is "fpexcept.ignore" optimization passes may assume that the
12989 exception status flags will not be read and that floating point exceptions will
12990 be masked. This allows transformations to be performed that may change the
12991 exception semantics of the original code. For example, FP operations may be
12992 speculatively executed in this case whereas they must not be for either of the
12993 other possible values of this argument.
12995 If the exception behavior argument is "fpexcept.maytrap" optimization passes
12996 must avoid transformations that may raise exceptions that would not have been
12997 raised by the original code (such as speculatively executing FP operations), but
12998 passes are not required to preserve all exceptions that are implied by the
12999 original code. For example, exceptions may be potentially hidden by constant
13002 If the exception behavior argument is "fpexcept.strict" all transformations must
13003 strictly preserve the floating point exception semantics of the original code.
13004 Any FP exception that would have been raised by the original code must be raised
13005 by the transformed code, and the transformed code must not raise any FP
13006 exceptions that would not have been raised by the original code. This is the
13007 exception behavior argument that will be used if the code being compiled reads
13008 the FP exception status flags, but this mode can also be used with code that
13009 unmasks FP exceptions.
13011 The number and order of floating point exceptions is NOT guaranteed. For
13012 example, a series of FP operations that each may raise exceptions may be
13013 vectorized into a single instruction that raises each unique exception a single
13017 '``llvm.experimental.constrained.fadd``' Intrinsic
13018 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13026 @llvm.experimental.constrained.fadd(<type> <op1>, <type> <op2>,
13027 metadata <rounding mode>,
13028 metadata <exception behavior>)
13033 The '``llvm.experimental.constrained.fadd``' intrinsic returns the sum of its
13040 The first two arguments to the '``llvm.experimental.constrained.fadd``'
13041 intrinsic must be :ref:`floating point <t_floating>` or :ref:`vector <t_vector>`
13042 of floating point values. Both arguments must have identical types.
13044 The third and fourth arguments specify the rounding mode and exception
13045 behavior as described above.
13050 The value produced is the floating point sum of the two value operands and has
13051 the same type as the operands.
13054 '``llvm.experimental.constrained.fsub``' Intrinsic
13055 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13063 @llvm.experimental.constrained.fsub(<type> <op1>, <type> <op2>,
13064 metadata <rounding mode>,
13065 metadata <exception behavior>)
13070 The '``llvm.experimental.constrained.fsub``' intrinsic returns the difference
13071 of its two operands.
13077 The first two arguments to the '``llvm.experimental.constrained.fsub``'
13078 intrinsic must be :ref:`floating point <t_floating>` or :ref:`vector <t_vector>`
13079 of floating point values. Both arguments must have identical types.
13081 The third and fourth arguments specify the rounding mode and exception
13082 behavior as described above.
13087 The value produced is the floating point difference of the two value operands
13088 and has the same type as the operands.
13091 '``llvm.experimental.constrained.fmul``' Intrinsic
13092 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13100 @llvm.experimental.constrained.fmul(<type> <op1>, <type> <op2>,
13101 metadata <rounding mode>,
13102 metadata <exception behavior>)
13107 The '``llvm.experimental.constrained.fmul``' intrinsic returns the product of
13114 The first two arguments to the '``llvm.experimental.constrained.fmul``'
13115 intrinsic must be :ref:`floating point <t_floating>` or :ref:`vector <t_vector>`
13116 of floating point values. Both arguments must have identical types.
13118 The third and fourth arguments specify the rounding mode and exception
13119 behavior as described above.
13124 The value produced is the floating point product of the two value operands and
13125 has the same type as the operands.
13128 '``llvm.experimental.constrained.fdiv``' Intrinsic
13129 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13137 @llvm.experimental.constrained.fdiv(<type> <op1>, <type> <op2>,
13138 metadata <rounding mode>,
13139 metadata <exception behavior>)
13144 The '``llvm.experimental.constrained.fdiv``' intrinsic returns the quotient of
13151 The first two arguments to the '``llvm.experimental.constrained.fdiv``'
13152 intrinsic must be :ref:`floating point <t_floating>` or :ref:`vector <t_vector>`
13153 of floating point values. Both arguments must have identical types.
13155 The third and fourth arguments specify the rounding mode and exception
13156 behavior as described above.
13161 The value produced is the floating point quotient of the two value operands and
13162 has the same type as the operands.
13165 '``llvm.experimental.constrained.frem``' Intrinsic
13166 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13174 @llvm.experimental.constrained.frem(<type> <op1>, <type> <op2>,
13175 metadata <rounding mode>,
13176 metadata <exception behavior>)
13181 The '``llvm.experimental.constrained.frem``' intrinsic returns the remainder
13182 from the division of its two operands.
13188 The first two arguments to the '``llvm.experimental.constrained.frem``'
13189 intrinsic must be :ref:`floating point <t_floating>` or :ref:`vector <t_vector>`
13190 of floating point values. Both arguments must have identical types.
13192 The third and fourth arguments specify the rounding mode and exception
13193 behavior as described above. The rounding mode argument has no effect, since
13194 the result of frem is never rounded, but the argument is included for
13195 consistency with the other constrained floating point intrinsics.
13200 The value produced is the floating point remainder from the division of the two
13201 value operands and has the same type as the operands. The remainder has the
13202 same sign as the dividend.
13204 '``llvm.experimental.constrained.fma``' Intrinsic
13205 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13213 @llvm.experimental.constrained.fma(<type> <op1>, <type> <op2>, <type> <op3>,
13214 metadata <rounding mode>,
13215 metadata <exception behavior>)
13220 The '``llvm.experimental.constrained.fma``' intrinsic returns the result of a
13221 fused-multiply-add operation on its operands.
13226 The first three arguments to the '``llvm.experimental.constrained.fma``'
13227 intrinsic must be :ref:`floating point <t_floating>` or :ref:`vector
13228 <t_vector>` of floating point values. All arguments must have identical types.
13230 The fourth and fifth arguments specify the rounding mode and exception behavior
13231 as described above.
13236 The result produced is the product of the first two operands added to the third
13237 operand computed with infinite precision, and then rounded to the target
13240 Constrained libm-equivalent Intrinsics
13241 --------------------------------------
13243 In addition to the basic floating point operations for which constrained
13244 intrinsics are described above, there are constrained versions of various
13245 operations which provide equivalent behavior to a corresponding libm function.
13246 These intrinsics allow the precise behavior of these operations with respect to
13247 rounding mode and exception behavior to be controlled.
13249 As with the basic constrained floating point intrinsics, the rounding mode
13250 and exception behavior arguments only control the behavior of the optimizer.
13251 They do not change the runtime floating point environment.
13254 '``llvm.experimental.constrained.sqrt``' Intrinsic
13255 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13263 @llvm.experimental.constrained.sqrt(<type> <op1>,
13264 metadata <rounding mode>,
13265 metadata <exception behavior>)
13270 The '``llvm.experimental.constrained.sqrt``' intrinsic returns the square root
13271 of the specified value, returning the same value as the libm '``sqrt``'
13272 functions would, but without setting ``errno``.
13277 The first argument and the return type are floating point numbers of the same
13280 The second and third arguments specify the rounding mode and exception
13281 behavior as described above.
13286 This function returns the nonnegative square root of the specified value.
13287 If the value is less than negative zero, a floating point exception occurs
13288 and the return value is architecture specific.
13291 '``llvm.experimental.constrained.pow``' Intrinsic
13292 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13300 @llvm.experimental.constrained.pow(<type> <op1>, <type> <op2>,
13301 metadata <rounding mode>,
13302 metadata <exception behavior>)
13307 The '``llvm.experimental.constrained.pow``' intrinsic returns the first operand
13308 raised to the (positive or negative) power specified by the second operand.
13313 The first two arguments and the return value are floating point numbers of the
13314 same type. The second argument specifies the power to which the first argument
13317 The third and fourth arguments specify the rounding mode and exception
13318 behavior as described above.
13323 This function returns the first value raised to the second power,
13324 returning the same values as the libm ``pow`` functions would, and
13325 handles error conditions in the same way.
13328 '``llvm.experimental.constrained.powi``' Intrinsic
13329 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13337 @llvm.experimental.constrained.powi(<type> <op1>, i32 <op2>,
13338 metadata <rounding mode>,
13339 metadata <exception behavior>)
13344 The '``llvm.experimental.constrained.powi``' intrinsic returns the first operand
13345 raised to the (positive or negative) power specified by the second operand. The
13346 order of evaluation of multiplications is not defined. When a vector of floating
13347 point type is used, the second argument remains a scalar integer value.
13353 The first argument and the return value are floating point numbers of the same
13354 type. The second argument is a 32-bit signed integer specifying the power to
13355 which the first argument should be raised.
13357 The third and fourth arguments specify the rounding mode and exception
13358 behavior as described above.
13363 This function returns the first value raised to the second power with an
13364 unspecified sequence of rounding operations.
13367 '``llvm.experimental.constrained.sin``' Intrinsic
13368 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13376 @llvm.experimental.constrained.sin(<type> <op1>,
13377 metadata <rounding mode>,
13378 metadata <exception behavior>)
13383 The '``llvm.experimental.constrained.sin``' intrinsic returns the sine of the
13389 The first argument and the return type are floating point numbers of the same
13392 The second and third arguments specify the rounding mode and exception
13393 behavior as described above.
13398 This function returns the sine of the specified operand, returning the
13399 same values as the libm ``sin`` functions would, and handles error
13400 conditions in the same way.
13403 '``llvm.experimental.constrained.cos``' Intrinsic
13404 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13412 @llvm.experimental.constrained.cos(<type> <op1>,
13413 metadata <rounding mode>,
13414 metadata <exception behavior>)
13419 The '``llvm.experimental.constrained.cos``' intrinsic returns the cosine of the
13425 The first argument and the return type are floating point numbers of the same
13428 The second and third arguments specify the rounding mode and exception
13429 behavior as described above.
13434 This function returns the cosine of the specified operand, returning the
13435 same values as the libm ``cos`` functions would, and handles error
13436 conditions in the same way.
13439 '``llvm.experimental.constrained.exp``' Intrinsic
13440 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13448 @llvm.experimental.constrained.exp(<type> <op1>,
13449 metadata <rounding mode>,
13450 metadata <exception behavior>)
13455 The '``llvm.experimental.constrained.exp``' intrinsic computes the base-e
13456 exponential of the specified value.
13461 The first argument and the return value are floating point numbers of the same
13464 The second and third arguments specify the rounding mode and exception
13465 behavior as described above.
13470 This function returns the same values as the libm ``exp`` functions
13471 would, and handles error conditions in the same way.
13474 '``llvm.experimental.constrained.exp2``' Intrinsic
13475 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13483 @llvm.experimental.constrained.exp2(<type> <op1>,
13484 metadata <rounding mode>,
13485 metadata <exception behavior>)
13490 The '``llvm.experimental.constrained.exp2``' intrinsic computes the base-2
13491 exponential of the specified value.
13497 The first argument and the return value are floating point numbers of the same
13500 The second and third arguments specify the rounding mode and exception
13501 behavior as described above.
13506 This function returns the same values as the libm ``exp2`` functions
13507 would, and handles error conditions in the same way.
13510 '``llvm.experimental.constrained.log``' Intrinsic
13511 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13519 @llvm.experimental.constrained.log(<type> <op1>,
13520 metadata <rounding mode>,
13521 metadata <exception behavior>)
13526 The '``llvm.experimental.constrained.log``' intrinsic computes the base-e
13527 logarithm of the specified value.
13532 The first argument and the return value are floating point numbers of the same
13535 The second and third arguments specify the rounding mode and exception
13536 behavior as described above.
13542 This function returns the same values as the libm ``log`` functions
13543 would, and handles error conditions in the same way.
13546 '``llvm.experimental.constrained.log10``' Intrinsic
13547 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13555 @llvm.experimental.constrained.log10(<type> <op1>,
13556 metadata <rounding mode>,
13557 metadata <exception behavior>)
13562 The '``llvm.experimental.constrained.log10``' intrinsic computes the base-10
13563 logarithm of the specified value.
13568 The first argument and the return value are floating point numbers of the same
13571 The second and third arguments specify the rounding mode and exception
13572 behavior as described above.
13577 This function returns the same values as the libm ``log10`` functions
13578 would, and handles error conditions in the same way.
13581 '``llvm.experimental.constrained.log2``' Intrinsic
13582 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13590 @llvm.experimental.constrained.log2(<type> <op1>,
13591 metadata <rounding mode>,
13592 metadata <exception behavior>)
13597 The '``llvm.experimental.constrained.log2``' intrinsic computes the base-2
13598 logarithm of the specified value.
13603 The first argument and the return value are floating point numbers of the same
13606 The second and third arguments specify the rounding mode and exception
13607 behavior as described above.
13612 This function returns the same values as the libm ``log2`` functions
13613 would, and handles error conditions in the same way.
13616 '``llvm.experimental.constrained.rint``' Intrinsic
13617 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13625 @llvm.experimental.constrained.rint(<type> <op1>,
13626 metadata <rounding mode>,
13627 metadata <exception behavior>)
13632 The '``llvm.experimental.constrained.rint``' intrinsic returns the first
13633 operand rounded to the nearest integer. It may raise an inexact floating point
13634 exception if the operand is not an integer.
13639 The first argument and the return value are floating point numbers of the same
13642 The second and third arguments specify the rounding mode and exception
13643 behavior as described above.
13648 This function returns the same values as the libm ``rint`` functions
13649 would, and handles error conditions in the same way. The rounding mode is
13650 described, not determined, by the rounding mode argument. The actual rounding
13651 mode is determined by the runtime floating point environment. The rounding
13652 mode argument is only intended as information to the compiler.
13655 '``llvm.experimental.constrained.nearbyint``' Intrinsic
13656 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13664 @llvm.experimental.constrained.nearbyint(<type> <op1>,
13665 metadata <rounding mode>,
13666 metadata <exception behavior>)
13671 The '``llvm.experimental.constrained.nearbyint``' intrinsic returns the first
13672 operand rounded to the nearest integer. It will not raise an inexact floating
13673 point exception if the operand is not an integer.
13679 The first argument and the return value are floating point numbers of the same
13682 The second and third arguments specify the rounding mode and exception
13683 behavior as described above.
13688 This function returns the same values as the libm ``nearbyint`` functions
13689 would, and handles error conditions in the same way. The rounding mode is
13690 described, not determined, by the rounding mode argument. The actual rounding
13691 mode is determined by the runtime floating point environment. The rounding
13692 mode argument is only intended as information to the compiler.
13698 This class of intrinsics is designed to be generic and has no specific
13701 '``llvm.var.annotation``' Intrinsic
13702 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13709 declare void @llvm.var.annotation(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
13714 The '``llvm.var.annotation``' intrinsic.
13719 The first argument is a pointer to a value, the second is a pointer to a
13720 global string, the third is a pointer to a global string which is the
13721 source file name, and the last argument is the line number.
13726 This intrinsic allows annotation of local variables with arbitrary
13727 strings. This can be useful for special purpose optimizations that want
13728 to look for these annotations. These have no other defined use; they are
13729 ignored by code generation and optimization.
13731 '``llvm.ptr.annotation.*``' Intrinsic
13732 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13737 This is an overloaded intrinsic. You can use '``llvm.ptr.annotation``' on a
13738 pointer to an integer of any width. *NOTE* you must specify an address space for
13739 the pointer. The identifier for the default address space is the integer
13744 declare i8* @llvm.ptr.annotation.p<address space>i8(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
13745 declare i16* @llvm.ptr.annotation.p<address space>i16(i16* <val>, i8* <str>, i8* <str>, i32 <int>)
13746 declare i32* @llvm.ptr.annotation.p<address space>i32(i32* <val>, i8* <str>, i8* <str>, i32 <int>)
13747 declare i64* @llvm.ptr.annotation.p<address space>i64(i64* <val>, i8* <str>, i8* <str>, i32 <int>)
13748 declare i256* @llvm.ptr.annotation.p<address space>i256(i256* <val>, i8* <str>, i8* <str>, i32 <int>)
13753 The '``llvm.ptr.annotation``' intrinsic.
13758 The first argument is a pointer to an integer value of arbitrary bitwidth
13759 (result of some expression), the second is a pointer to a global string, the
13760 third is a pointer to a global string which is the source file name, and the
13761 last argument is the line number. It returns the value of the first argument.
13766 This intrinsic allows annotation of a pointer to an integer with arbitrary
13767 strings. This can be useful for special purpose optimizations that want to look
13768 for these annotations. These have no other defined use; they are ignored by code
13769 generation and optimization.
13771 '``llvm.annotation.*``' Intrinsic
13772 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13777 This is an overloaded intrinsic. You can use '``llvm.annotation``' on
13778 any integer bit width.
13782 declare i8 @llvm.annotation.i8(i8 <val>, i8* <str>, i8* <str>, i32 <int>)
13783 declare i16 @llvm.annotation.i16(i16 <val>, i8* <str>, i8* <str>, i32 <int>)
13784 declare i32 @llvm.annotation.i32(i32 <val>, i8* <str>, i8* <str>, i32 <int>)
13785 declare i64 @llvm.annotation.i64(i64 <val>, i8* <str>, i8* <str>, i32 <int>)
13786 declare i256 @llvm.annotation.i256(i256 <val>, i8* <str>, i8* <str>, i32 <int>)
13791 The '``llvm.annotation``' intrinsic.
13796 The first argument is an integer value (result of some expression), the
13797 second is a pointer to a global string, the third is a pointer to a
13798 global string which is the source file name, and the last argument is
13799 the line number. It returns the value of the first argument.
13804 This intrinsic allows annotations to be put on arbitrary expressions
13805 with arbitrary strings. This can be useful for special purpose
13806 optimizations that want to look for these annotations. These have no
13807 other defined use; they are ignored by code generation and optimization.
13809 '``llvm.codeview.annotation``' Intrinsic
13810 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13815 This annotation emits a label at its program point and an associated
13816 ``S_ANNOTATION`` codeview record with some additional string metadata. This is
13817 used to implement MSVC's ``__annotation`` intrinsic. It is marked
13818 ``noduplicate``, so calls to this intrinsic prevent inlining and should be
13819 considered expensive.
13823 declare void @llvm.codeview.annotation(metadata)
13828 The argument should be an MDTuple containing any number of MDStrings.
13830 '``llvm.trap``' Intrinsic
13831 ^^^^^^^^^^^^^^^^^^^^^^^^^
13838 declare void @llvm.trap() noreturn nounwind
13843 The '``llvm.trap``' intrinsic.
13853 This intrinsic is lowered to the target dependent trap instruction. If
13854 the target does not have a trap instruction, this intrinsic will be
13855 lowered to a call of the ``abort()`` function.
13857 '``llvm.debugtrap``' Intrinsic
13858 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13865 declare void @llvm.debugtrap() nounwind
13870 The '``llvm.debugtrap``' intrinsic.
13880 This intrinsic is lowered to code which is intended to cause an
13881 execution trap with the intention of requesting the attention of a
13884 '``llvm.stackprotector``' Intrinsic
13885 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13892 declare void @llvm.stackprotector(i8* <guard>, i8** <slot>)
13897 The ``llvm.stackprotector`` intrinsic takes the ``guard`` and stores it
13898 onto the stack at ``slot``. The stack slot is adjusted to ensure that it
13899 is placed on the stack before local variables.
13904 The ``llvm.stackprotector`` intrinsic requires two pointer arguments.
13905 The first argument is the value loaded from the stack guard
13906 ``@__stack_chk_guard``. The second variable is an ``alloca`` that has
13907 enough space to hold the value of the guard.
13912 This intrinsic causes the prologue/epilogue inserter to force the position of
13913 the ``AllocaInst`` stack slot to be before local variables on the stack. This is
13914 to ensure that if a local variable on the stack is overwritten, it will destroy
13915 the value of the guard. When the function exits, the guard on the stack is
13916 checked against the original guard by ``llvm.stackprotectorcheck``. If they are
13917 different, then ``llvm.stackprotectorcheck`` causes the program to abort by
13918 calling the ``__stack_chk_fail()`` function.
13920 '``llvm.stackguard``' Intrinsic
13921 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13928 declare i8* @llvm.stackguard()
13933 The ``llvm.stackguard`` intrinsic returns the system stack guard value.
13935 It should not be generated by frontends, since it is only for internal usage.
13936 The reason why we create this intrinsic is that we still support IR form Stack
13937 Protector in FastISel.
13947 On some platforms, the value returned by this intrinsic remains unchanged
13948 between loads in the same thread. On other platforms, it returns the same
13949 global variable value, if any, e.g. ``@__stack_chk_guard``.
13951 Currently some platforms have IR-level customized stack guard loading (e.g.
13952 X86 Linux) that is not handled by ``llvm.stackguard()``, while they should be
13955 '``llvm.objectsize``' Intrinsic
13956 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13963 declare i32 @llvm.objectsize.i32(i8* <object>, i1 <min>, i1 <nullunknown>)
13964 declare i64 @llvm.objectsize.i64(i8* <object>, i1 <min>, i1 <nullunknown>)
13969 The ``llvm.objectsize`` intrinsic is designed to provide information to
13970 the optimizers to determine at compile time whether a) an operation
13971 (like memcpy) will overflow a buffer that corresponds to an object, or
13972 b) that a runtime check for overflow isn't necessary. An object in this
13973 context means an allocation of a specific class, structure, array, or
13979 The ``llvm.objectsize`` intrinsic takes three arguments. The first argument is
13980 a pointer to or into the ``object``. The second argument determines whether
13981 ``llvm.objectsize`` returns 0 (if true) or -1 (if false) when the object size
13982 is unknown. The third argument controls how ``llvm.objectsize`` acts when
13983 ``null`` is used as its pointer argument. If it's true and the pointer is in
13984 address space 0, ``null`` is treated as an opaque value with an unknown number
13985 of bytes. Otherwise, ``llvm.objectsize`` reports 0 bytes available when given
13988 The second and third arguments only accept constants.
13993 The ``llvm.objectsize`` intrinsic is lowered to a constant representing
13994 the size of the object concerned. If the size cannot be determined at
13995 compile time, ``llvm.objectsize`` returns ``i32/i64 -1 or 0`` (depending
13996 on the ``min`` argument).
13998 '``llvm.expect``' Intrinsic
13999 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
14004 This is an overloaded intrinsic. You can use ``llvm.expect`` on any
14009 declare i1 @llvm.expect.i1(i1 <val>, i1 <expected_val>)
14010 declare i32 @llvm.expect.i32(i32 <val>, i32 <expected_val>)
14011 declare i64 @llvm.expect.i64(i64 <val>, i64 <expected_val>)
14016 The ``llvm.expect`` intrinsic provides information about expected (the
14017 most probable) value of ``val``, which can be used by optimizers.
14022 The ``llvm.expect`` intrinsic takes two arguments. The first argument is
14023 a value. The second argument is an expected value, this needs to be a
14024 constant value, variables are not allowed.
14029 This intrinsic is lowered to the ``val``.
14033 '``llvm.assume``' Intrinsic
14034 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
14041 declare void @llvm.assume(i1 %cond)
14046 The ``llvm.assume`` allows the optimizer to assume that the provided
14047 condition is true. This information can then be used in simplifying other parts
14053 The condition which the optimizer may assume is always true.
14058 The intrinsic allows the optimizer to assume that the provided condition is
14059 always true whenever the control flow reaches the intrinsic call. No code is
14060 generated for this intrinsic, and instructions that contribute only to the
14061 provided condition are not used for code generation. If the condition is
14062 violated during execution, the behavior is undefined.
14064 Note that the optimizer might limit the transformations performed on values
14065 used by the ``llvm.assume`` intrinsic in order to preserve the instructions
14066 only used to form the intrinsic's input argument. This might prove undesirable
14067 if the extra information provided by the ``llvm.assume`` intrinsic does not cause
14068 sufficient overall improvement in code quality. For this reason,
14069 ``llvm.assume`` should not be used to document basic mathematical invariants
14070 that the optimizer can otherwise deduce or facts that are of little use to the
14075 '``llvm.ssa_copy``' Intrinsic
14076 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
14083 declare type @llvm.ssa_copy(type %operand) returned(1) readnone
14088 The first argument is an operand which is used as the returned value.
14093 The ``llvm.ssa_copy`` intrinsic can be used to attach information to
14094 operations by copying them and giving them new names. For example,
14095 the PredicateInfo utility uses it to build Extended SSA form, and
14096 attach various forms of information to operands that dominate specific
14097 uses. It is not meant for general use, only for building temporary
14098 renaming forms that require value splits at certain points.
14102 '``llvm.type.test``' Intrinsic
14103 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
14110 declare i1 @llvm.type.test(i8* %ptr, metadata %type) nounwind readnone
14116 The first argument is a pointer to be tested. The second argument is a
14117 metadata object representing a :doc:`type identifier <TypeMetadata>`.
14122 The ``llvm.type.test`` intrinsic tests whether the given pointer is associated
14123 with the given type identifier.
14125 '``llvm.type.checked.load``' Intrinsic
14126 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
14133 declare {i8*, i1} @llvm.type.checked.load(i8* %ptr, i32 %offset, metadata %type) argmemonly nounwind readonly
14139 The first argument is a pointer from which to load a function pointer. The
14140 second argument is the byte offset from which to load the function pointer. The
14141 third argument is a metadata object representing a :doc:`type identifier
14147 The ``llvm.type.checked.load`` intrinsic safely loads a function pointer from a
14148 virtual table pointer using type metadata. This intrinsic is used to implement
14149 control flow integrity in conjunction with virtual call optimization. The
14150 virtual call optimization pass will optimize away ``llvm.type.checked.load``
14151 intrinsics associated with devirtualized calls, thereby removing the type
14152 check in cases where it is not needed to enforce the control flow integrity
14155 If the given pointer is associated with a type metadata identifier, this
14156 function returns true as the second element of its return value. (Note that
14157 the function may also return true if the given pointer is not associated
14158 with a type metadata identifier.) If the function's return value's second
14159 element is true, the following rules apply to the first element:
14161 - If the given pointer is associated with the given type metadata identifier,
14162 it is the function pointer loaded from the given byte offset from the given
14165 - If the given pointer is not associated with the given type metadata
14166 identifier, it is one of the following (the choice of which is unspecified):
14168 1. The function pointer that would have been loaded from an arbitrarily chosen
14169 (through an unspecified mechanism) pointer associated with the type
14172 2. If the function has a non-void return type, a pointer to a function that
14173 returns an unspecified value without causing side effects.
14175 If the function's return value's second element is false, the value of the
14176 first element is undefined.
14179 '``llvm.donothing``' Intrinsic
14180 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
14187 declare void @llvm.donothing() nounwind readnone
14192 The ``llvm.donothing`` intrinsic doesn't perform any operation. It's one of only
14193 three intrinsics (besides ``llvm.experimental.patchpoint`` and
14194 ``llvm.experimental.gc.statepoint``) that can be called with an invoke
14205 This intrinsic does nothing, and it's removed by optimizers and ignored
14208 '``llvm.experimental.deoptimize``' Intrinsic
14209 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
14216 declare type @llvm.experimental.deoptimize(...) [ "deopt"(...) ]
14221 This intrinsic, together with :ref:`deoptimization operand bundles
14222 <deopt_opbundles>`, allow frontends to express transfer of control and
14223 frame-local state from the currently executing (typically more specialized,
14224 hence faster) version of a function into another (typically more generic, hence
14227 In languages with a fully integrated managed runtime like Java and JavaScript
14228 this intrinsic can be used to implement "uncommon trap" or "side exit" like
14229 functionality. In unmanaged languages like C and C++, this intrinsic can be
14230 used to represent the slow paths of specialized functions.
14236 The intrinsic takes an arbitrary number of arguments, whose meaning is
14237 decided by the :ref:`lowering strategy<deoptimize_lowering>`.
14242 The ``@llvm.experimental.deoptimize`` intrinsic executes an attached
14243 deoptimization continuation (denoted using a :ref:`deoptimization
14244 operand bundle <deopt_opbundles>`) and returns the value returned by
14245 the deoptimization continuation. Defining the semantic properties of
14246 the continuation itself is out of scope of the language reference --
14247 as far as LLVM is concerned, the deoptimization continuation can
14248 invoke arbitrary side effects, including reading from and writing to
14251 Deoptimization continuations expressed using ``"deopt"`` operand bundles always
14252 continue execution to the end of the physical frame containing them, so all
14253 calls to ``@llvm.experimental.deoptimize`` must be in "tail position":
14255 - ``@llvm.experimental.deoptimize`` cannot be invoked.
14256 - The call must immediately precede a :ref:`ret <i_ret>` instruction.
14257 - The ``ret`` instruction must return the value produced by the
14258 ``@llvm.experimental.deoptimize`` call if there is one, or void.
14260 Note that the above restrictions imply that the return type for a call to
14261 ``@llvm.experimental.deoptimize`` will match the return type of its immediate
14264 The inliner composes the ``"deopt"`` continuations of the caller into the
14265 ``"deopt"`` continuations present in the inlinee, and also updates calls to this
14266 intrinsic to return directly from the frame of the function it inlined into.
14268 All declarations of ``@llvm.experimental.deoptimize`` must share the
14269 same calling convention.
14271 .. _deoptimize_lowering:
14276 Calls to ``@llvm.experimental.deoptimize`` are lowered to calls to the
14277 symbol ``__llvm_deoptimize`` (it is the frontend's responsibility to
14278 ensure that this symbol is defined). The call arguments to
14279 ``@llvm.experimental.deoptimize`` are lowered as if they were formal
14280 arguments of the specified types, and not as varargs.
14283 '``llvm.experimental.guard``' Intrinsic
14284 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
14291 declare void @llvm.experimental.guard(i1, ...) [ "deopt"(...) ]
14296 This intrinsic, together with :ref:`deoptimization operand bundles
14297 <deopt_opbundles>`, allows frontends to express guards or checks on
14298 optimistic assumptions made during compilation. The semantics of
14299 ``@llvm.experimental.guard`` is defined in terms of
14300 ``@llvm.experimental.deoptimize`` -- its body is defined to be
14303 .. code-block:: text
14305 define void @llvm.experimental.guard(i1 %pred, <args...>) {
14306 %realPred = and i1 %pred, undef
14307 br i1 %realPred, label %continue, label %leave [, !make.implicit !{}]
14310 call void @llvm.experimental.deoptimize(<args...>) [ "deopt"() ]
14318 with the optional ``[, !make.implicit !{}]`` present if and only if it
14319 is present on the call site. For more details on ``!make.implicit``,
14320 see :doc:`FaultMaps`.
14322 In words, ``@llvm.experimental.guard`` executes the attached
14323 ``"deopt"`` continuation if (but **not** only if) its first argument
14324 is ``false``. Since the optimizer is allowed to replace the ``undef``
14325 with an arbitrary value, it can optimize guard to fail "spuriously",
14326 i.e. without the original condition being false (hence the "not only
14327 if"); and this allows for "check widening" type optimizations.
14329 ``@llvm.experimental.guard`` cannot be invoked.
14332 '``llvm.load.relative``' Intrinsic
14333 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
14340 declare i8* @llvm.load.relative.iN(i8* %ptr, iN %offset) argmemonly nounwind readonly
14345 This intrinsic loads a 32-bit value from the address ``%ptr + %offset``,
14346 adds ``%ptr`` to that value and returns it. The constant folder specifically
14347 recognizes the form of this intrinsic and the constant initializers it may
14348 load from; if a loaded constant initializer is known to have the form
14349 ``i32 trunc(x - %ptr)``, the intrinsic call is folded to ``x``.
14351 LLVM provides that the calculation of such a constant initializer will
14352 not overflow at link time under the medium code model if ``x`` is an
14353 ``unnamed_addr`` function. However, it does not provide this guarantee for
14354 a constant initializer folded into a function body. This intrinsic can be
14355 used to avoid the possibility of overflows when loading from such a constant.
14357 '``llvm.sideeffect``' Intrinsic
14358 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
14365 declare void @llvm.sideeffect() inaccessiblememonly nounwind
14370 The ``llvm.sideeffect`` intrinsic doesn't perform any operation. Optimizers
14371 treat it as having side effects, so it can be inserted into a loop to
14372 indicate that the loop shouldn't be assumed to terminate (which could
14373 potentially lead to the loop being optimized away entirely), even if it's
14374 an infinite loop with no other side effects.
14384 This intrinsic actually does nothing, but optimizers must assume that it
14385 has externally observable side effects.
14387 Stack Map Intrinsics
14388 --------------------
14390 LLVM provides experimental intrinsics to support runtime patching
14391 mechanisms commonly desired in dynamic language JITs. These intrinsics
14392 are described in :doc:`StackMaps`.
14394 Element Wise Atomic Memory Intrinsics
14395 -------------------------------------
14397 These intrinsics are similar to the standard library memory intrinsics except
14398 that they perform memory transfer as a sequence of atomic memory accesses.
14400 .. _int_memcpy_element_unordered_atomic:
14402 '``llvm.memcpy.element.unordered.atomic``' Intrinsic
14403 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
14408 This is an overloaded intrinsic. You can use ``llvm.memcpy.element.unordered.atomic`` on
14409 any integer bit width and for different address spaces. Not all targets
14410 support all bit widths however.
14414 declare void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i32(i8* <dest>,
14417 i32 <element_size>)
14418 declare void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* <dest>,
14421 i32 <element_size>)
14426 The '``llvm.memcpy.element.unordered.atomic.*``' intrinsic is a specialization of the
14427 '``llvm.memcpy.*``' intrinsic. It differs in that the ``dest`` and ``src`` are treated
14428 as arrays with elements that are exactly ``element_size`` bytes, and the copy between
14429 buffers uses a sequence of :ref:`unordered atomic <ordering>` load/store operations
14430 that are a positive integer multiple of the ``element_size`` in size.
14435 The first three arguments are the same as they are in the :ref:`@llvm.memcpy <int_memcpy>`
14436 intrinsic, with the added constraint that ``len`` is required to be a positive integer
14437 multiple of the ``element_size``. If ``len`` is not a positive integer multiple of
14438 ``element_size``, then the behaviour of the intrinsic is undefined.
14440 ``element_size`` must be a compile-time constant positive power of two no greater than
14441 target-specific atomic access size limit.
14443 For each of the input pointers ``align`` parameter attribute must be specified. It
14444 must be a power of two no less than the ``element_size``. Caller guarantees that
14445 both the source and destination pointers are aligned to that boundary.
14450 The '``llvm.memcpy.element.unordered.atomic.*``' intrinsic copies ``len`` bytes of
14451 memory from the source location to the destination location. These locations are not
14452 allowed to overlap. The memory copy is performed as a sequence of load/store operations
14453 where each access is guaranteed to be a multiple of ``element_size`` bytes wide and
14454 aligned at an ``element_size`` boundary.
14456 The order of the copy is unspecified. The same value may be read from the source
14457 buffer many times, but only one write is issued to the destination buffer per
14458 element. It is well defined to have concurrent reads and writes to both source and
14459 destination provided those reads and writes are unordered atomic when specified.
14461 This intrinsic does not provide any additional ordering guarantees over those
14462 provided by a set of unordered loads from the source location and stores to the
14468 In the most general case call to the '``llvm.memcpy.element.unordered.atomic.*``' is
14469 lowered to a call to the symbol ``__llvm_memcpy_element_unordered_atomic_*``. Where '*'
14470 is replaced with an actual element size.
14472 Optimizer is allowed to inline memory copy when it's profitable to do so.
14474 '``llvm.memmove.element.unordered.atomic``' Intrinsic
14475 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
14480 This is an overloaded intrinsic. You can use
14481 ``llvm.memmove.element.unordered.atomic`` on any integer bit width and for
14482 different address spaces. Not all targets support all bit widths however.
14486 declare void @llvm.memmove.element.unordered.atomic.p0i8.p0i8.i32(i8* <dest>,
14489 i32 <element_size>)
14490 declare void @llvm.memmove.element.unordered.atomic.p0i8.p0i8.i64(i8* <dest>,
14493 i32 <element_size>)
14498 The '``llvm.memmove.element.unordered.atomic.*``' intrinsic is a specialization
14499 of the '``llvm.memmove.*``' intrinsic. It differs in that the ``dest`` and
14500 ``src`` are treated as arrays with elements that are exactly ``element_size``
14501 bytes, and the copy between buffers uses a sequence of
14502 :ref:`unordered atomic <ordering>` load/store operations that are a positive
14503 integer multiple of the ``element_size`` in size.
14508 The first three arguments are the same as they are in the
14509 :ref:`@llvm.memmove <int_memmove>` intrinsic, with the added constraint that
14510 ``len`` is required to be a positive integer multiple of the ``element_size``.
14511 If ``len`` is not a positive integer multiple of ``element_size``, then the
14512 behaviour of the intrinsic is undefined.
14514 ``element_size`` must be a compile-time constant positive power of two no
14515 greater than a target-specific atomic access size limit.
14517 For each of the input pointers the ``align`` parameter attribute must be
14518 specified. It must be a power of two no less than the ``element_size``. Caller
14519 guarantees that both the source and destination pointers are aligned to that
14525 The '``llvm.memmove.element.unordered.atomic.*``' intrinsic copies ``len`` bytes
14526 of memory from the source location to the destination location. These locations
14527 are allowed to overlap. The memory copy is performed as a sequence of load/store
14528 operations where each access is guaranteed to be a multiple of ``element_size``
14529 bytes wide and aligned at an ``element_size`` boundary.
14531 The order of the copy is unspecified. The same value may be read from the source
14532 buffer many times, but only one write is issued to the destination buffer per
14533 element. It is well defined to have concurrent reads and writes to both source
14534 and destination provided those reads and writes are unordered atomic when
14537 This intrinsic does not provide any additional ordering guarantees over those
14538 provided by a set of unordered loads from the source location and stores to the
14544 In the most general case call to the
14545 '``llvm.memmove.element.unordered.atomic.*``' is lowered to a call to the symbol
14546 ``__llvm_memmove_element_unordered_atomic_*``. Where '*' is replaced with an
14547 actual element size.
14549 The optimizer is allowed to inline the memory copy when it's profitable to do so.
14551 .. _int_memset_element_unordered_atomic:
14553 '``llvm.memset.element.unordered.atomic``' Intrinsic
14554 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
14559 This is an overloaded intrinsic. You can use ``llvm.memset.element.unordered.atomic`` on
14560 any integer bit width and for different address spaces. Not all targets
14561 support all bit widths however.
14565 declare void @llvm.memset.element.unordered.atomic.p0i8.i32(i8* <dest>,
14568 i32 <element_size>)
14569 declare void @llvm.memset.element.unordered.atomic.p0i8.i64(i8* <dest>,
14572 i32 <element_size>)
14577 The '``llvm.memset.element.unordered.atomic.*``' intrinsic is a specialization of the
14578 '``llvm.memset.*``' intrinsic. It differs in that the ``dest`` is treated as an array
14579 with elements that are exactly ``element_size`` bytes, and the assignment to that array
14580 uses uses a sequence of :ref:`unordered atomic <ordering>` store operations
14581 that are a positive integer multiple of the ``element_size`` in size.
14586 The first three arguments are the same as they are in the :ref:`@llvm.memset <int_memset>`
14587 intrinsic, with the added constraint that ``len`` is required to be a positive integer
14588 multiple of the ``element_size``. If ``len`` is not a positive integer multiple of
14589 ``element_size``, then the behaviour of the intrinsic is undefined.
14591 ``element_size`` must be a compile-time constant positive power of two no greater than
14592 target-specific atomic access size limit.
14594 The ``dest`` input pointer must have the ``align`` parameter attribute specified. It
14595 must be a power of two no less than the ``element_size``. Caller guarantees that
14596 the destination pointer is aligned to that boundary.
14601 The '``llvm.memset.element.unordered.atomic.*``' intrinsic sets the ``len`` bytes of
14602 memory starting at the destination location to the given ``value``. The memory is
14603 set with a sequence of store operations where each access is guaranteed to be a
14604 multiple of ``element_size`` bytes wide and aligned at an ``element_size`` boundary.
14606 The order of the assignment is unspecified. Only one write is issued to the
14607 destination buffer per element. It is well defined to have concurrent reads and
14608 writes to the destination provided those reads and writes are unordered atomic
14611 This intrinsic does not provide any additional ordering guarantees over those
14612 provided by a set of unordered stores to the destination.
14617 In the most general case call to the '``llvm.memset.element.unordered.atomic.*``' is
14618 lowered to a call to the symbol ``__llvm_memset_element_unordered_atomic_*``. Where '*'
14619 is replaced with an actual element size.
14621 The optimizer is allowed to inline the memory assignment when it's profitable to do so.