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 attribute indicates that this function should be optimized
1466 for maximum fuzzing signal.
1468 This function attribute indicates that most optimization passes will skip
1469 this function, with the exception of interprocedural optimization passes.
1470 Code generation defaults to the "fast" instruction selector.
1471 This attribute cannot be used together with the ``alwaysinline``
1472 attribute; this attribute is also incompatible
1473 with the ``minsize`` attribute and the ``optsize`` attribute.
1475 This attribute requires the ``noinline`` attribute to be specified on
1476 the function as well, so the function is never inlined into any caller.
1477 Only functions with the ``alwaysinline`` attribute are valid
1478 candidates for inlining into the body of this function.
1480 This attribute suggests that optimization passes and code generator
1481 passes make choices that keep the code size of this function low,
1482 and otherwise do optimizations specifically to reduce code size as
1483 long as they do not significantly impact runtime performance.
1484 ``"patchable-function"``
1485 This attribute tells the code generator that the code
1486 generated for this function needs to follow certain conventions that
1487 make it possible for a runtime function to patch over it later.
1488 The exact effect of this attribute depends on its string value,
1489 for which there currently is one legal possibility:
1491 * ``"prologue-short-redirect"`` - This style of patchable
1492 function is intended to support patching a function prologue to
1493 redirect control away from the function in a thread safe
1494 manner. It guarantees that the first instruction of the
1495 function will be large enough to accommodate a short jump
1496 instruction, and will be sufficiently aligned to allow being
1497 fully changed via an atomic compare-and-swap instruction.
1498 While the first requirement can be satisfied by inserting large
1499 enough NOP, LLVM can and will try to re-purpose an existing
1500 instruction (i.e. one that would have to be emitted anyway) as
1501 the patchable instruction larger than a short jump.
1503 ``"prologue-short-redirect"`` is currently only supported on
1506 This attribute by itself does not imply restrictions on
1507 inter-procedural optimizations. All of the semantic effects the
1508 patching may have to be separately conveyed via the linkage type.
1510 This attribute indicates that the function will trigger a guard region
1511 in the end of the stack. It ensures that accesses to the stack must be
1512 no further apart than the size of the guard region to a previous
1513 access of the stack. It takes one required string value, the name of
1514 the stack probing function that will be called.
1516 If a function that has a ``"probe-stack"`` attribute is inlined into
1517 a function with another ``"probe-stack"`` attribute, the resulting
1518 function has the ``"probe-stack"`` attribute of the caller. If a
1519 function that has a ``"probe-stack"`` attribute is inlined into a
1520 function that has no ``"probe-stack"`` attribute at all, the resulting
1521 function has the ``"probe-stack"`` attribute of the callee.
1523 On a function, this attribute indicates that the function computes its
1524 result (or decides to unwind an exception) based strictly on its arguments,
1525 without dereferencing any pointer arguments or otherwise accessing
1526 any mutable state (e.g. memory, control registers, etc) visible to
1527 caller functions. It does not write through any pointer arguments
1528 (including ``byval`` arguments) and never changes any state visible
1529 to callers. This means while it cannot unwind exceptions by calling
1530 the ``C++`` exception throwing methods (since they write to memory), there may
1531 be non-``C++`` mechanisms that throw exceptions without writing to LLVM
1534 On an argument, this attribute indicates that the function does not
1535 dereference that pointer argument, even though it may read or write the
1536 memory that the pointer points to if accessed through other pointers.
1538 On a function, this attribute indicates that the function does not write
1539 through any pointer arguments (including ``byval`` arguments) or otherwise
1540 modify any state (e.g. memory, control registers, etc) visible to
1541 caller functions. It may dereference pointer arguments and read
1542 state that may be set in the caller. A readonly function always
1543 returns the same value (or unwinds an exception identically) when
1544 called with the same set of arguments and global state. This means while it
1545 cannot unwind exceptions by calling the ``C++`` exception throwing methods
1546 (since they write to memory), there may be non-``C++`` mechanisms that throw
1547 exceptions without writing to LLVM visible memory.
1549 On an argument, this attribute indicates that the function does not write
1550 through this pointer argument, even though it may write to the memory that
1551 the pointer points to.
1552 ``"stack-probe-size"``
1553 This attribute controls the behavior of stack probes: either
1554 the ``"probe-stack"`` attribute, or ABI-required stack probes, if any.
1555 It defines the size of the guard region. It ensures that if the function
1556 may use more stack space than the size of the guard region, stack probing
1557 sequence will be emitted. It takes one required integer value, which
1560 If a function that has a ``"stack-probe-size"`` attribute is inlined into
1561 a function with another ``"stack-probe-size"`` attribute, the resulting
1562 function has the ``"stack-probe-size"`` attribute that has the lower
1563 numeric value. If a function that has a ``"stack-probe-size"`` attribute is
1564 inlined into a function that has no ``"stack-probe-size"`` attribute
1565 at all, the resulting function has the ``"stack-probe-size"`` attribute
1567 ``"no-stack-arg-probe"``
1568 This attribute disables ABI-required stack probes, if any.
1570 On a function, this attribute indicates that the function may write to but
1571 does not read from memory.
1573 On an argument, this attribute indicates that the function may write to but
1574 does not read through this pointer argument (even though it may read from
1575 the memory that the pointer points to).
1577 This attribute indicates that the only memory accesses inside function are
1578 loads and stores from objects pointed to by its pointer-typed arguments,
1579 with arbitrary offsets. Or in other words, all memory operations in the
1580 function can refer to memory only using pointers based on its function
1582 Note that ``argmemonly`` can be used together with ``readonly`` attribute
1583 in order to specify that function reads only from its arguments.
1585 This attribute indicates that this function can return twice. The C
1586 ``setjmp`` is an example of such a function. The compiler disables
1587 some optimizations (like tail calls) in the caller of these
1590 This attribute indicates that
1591 `SafeStack <http://clang.llvm.org/docs/SafeStack.html>`_
1592 protection is enabled for this function.
1594 If a function that has a ``safestack`` attribute is inlined into a
1595 function that doesn't have a ``safestack`` attribute or which has an
1596 ``ssp``, ``sspstrong`` or ``sspreq`` attribute, then the resulting
1597 function will have a ``safestack`` attribute.
1598 ``sanitize_address``
1599 This attribute indicates that AddressSanitizer checks
1600 (dynamic address safety analysis) are enabled for this function.
1602 This attribute indicates that MemorySanitizer checks (dynamic detection
1603 of accesses to uninitialized memory) are enabled for this function.
1605 This attribute indicates that ThreadSanitizer checks
1606 (dynamic thread safety analysis) are enabled for this function.
1607 ``sanitize_hwaddress``
1608 This attribute indicates that HWAddressSanitizer checks
1609 (dynamic address safety analysis based on tagged pointers) are enabled for
1612 This function attribute indicates that the function does not have any
1613 effects besides calculating its result and does not have undefined behavior.
1614 Note that ``speculatable`` is not enough to conclude that along any
1615 particular execution path the number of calls to this function will not be
1616 externally observable. This attribute is only valid on functions
1617 and declarations, not on individual call sites. If a function is
1618 incorrectly marked as speculatable and really does exhibit
1619 undefined behavior, the undefined behavior may be observed even
1620 if the call site is dead code.
1623 This attribute indicates that the function should emit a stack
1624 smashing protector. It is in the form of a "canary" --- a random value
1625 placed on the stack before the local variables that's checked upon
1626 return from the function to see if it has been overwritten. A
1627 heuristic is used to determine if a function needs stack protectors
1628 or not. The heuristic used will enable protectors for functions with:
1630 - Character arrays larger than ``ssp-buffer-size`` (default 8).
1631 - Aggregates containing character arrays larger than ``ssp-buffer-size``.
1632 - Calls to alloca() with variable sizes or constant sizes greater than
1633 ``ssp-buffer-size``.
1635 Variables that are identified as requiring a protector will be arranged
1636 on the stack such that they are adjacent to the stack protector guard.
1638 If a function that has an ``ssp`` attribute is inlined into a
1639 function that doesn't have an ``ssp`` attribute, then the resulting
1640 function will have an ``ssp`` attribute.
1642 This attribute indicates that the function should *always* emit a
1643 stack smashing protector. This overrides the ``ssp`` function
1646 Variables that are identified as requiring a protector will be arranged
1647 on the stack such that they are adjacent to the stack protector guard.
1648 The specific layout rules are:
1650 #. Large arrays and structures containing large arrays
1651 (``>= ssp-buffer-size``) are closest to the stack protector.
1652 #. Small arrays and structures containing small arrays
1653 (``< ssp-buffer-size``) are 2nd closest to the protector.
1654 #. Variables that have had their address taken are 3rd closest to the
1657 If a function that has an ``sspreq`` attribute is inlined into a
1658 function that doesn't have an ``sspreq`` attribute or which has an
1659 ``ssp`` or ``sspstrong`` attribute, then the resulting function will have
1660 an ``sspreq`` attribute.
1662 This attribute indicates that the function should emit a stack smashing
1663 protector. This attribute causes a strong heuristic to be used when
1664 determining if a function needs stack protectors. The strong heuristic
1665 will enable protectors for functions with:
1667 - Arrays of any size and type
1668 - Aggregates containing an array of any size and type.
1669 - Calls to alloca().
1670 - Local variables that have had their address taken.
1672 Variables that are identified as requiring a protector will be arranged
1673 on the stack such that they are adjacent to the stack protector guard.
1674 The specific layout rules are:
1676 #. Large arrays and structures containing large arrays
1677 (``>= ssp-buffer-size``) are closest to the stack protector.
1678 #. Small arrays and structures containing small arrays
1679 (``< ssp-buffer-size``) are 2nd closest to the protector.
1680 #. Variables that have had their address taken are 3rd closest to the
1683 This overrides the ``ssp`` function attribute.
1685 If a function that has an ``sspstrong`` attribute is inlined into a
1686 function that doesn't have an ``sspstrong`` attribute, then the
1687 resulting function will have an ``sspstrong`` attribute.
1689 This attribute indicates that the function was called from a scope that
1690 requires strict floating-point semantics. LLVM will not attempt any
1691 optimizations that require assumptions about the floating-point rounding
1692 mode or that might alter the state of floating-point status flags that
1693 might otherwise be set or cleared by calling this function.
1695 This attribute indicates that the function will delegate to some other
1696 function with a tail call. The prototype of a thunk should not be used for
1697 optimization purposes. The caller is expected to cast the thunk prototype to
1698 match the thunk target prototype.
1700 This attribute indicates that the ABI being targeted requires that
1701 an unwind table entry be produced for this function even if we can
1702 show that no exceptions passes by it. This is normally the case for
1703 the ELF x86-64 abi, but it can be disabled for some compilation
1706 This attribute indicates that no control-flow check will be perfomed on
1707 the attributed entity. It disables -fcf-protection=<> for a specific
1708 entity to fine grain the HW control flow protection mechanism. The flag
1709 is target independant and currently appertains to a function or function
1717 Attributes may be set to communicate additional information about a global variable.
1718 Unlike :ref:`function attributes <fnattrs>`, attributes on a global variable
1719 are grouped into a single :ref:`attribute group <attrgrp>`.
1726 Operand bundles are tagged sets of SSA values that can be associated
1727 with certain LLVM instructions (currently only ``call`` s and
1728 ``invoke`` s). In a way they are like metadata, but dropping them is
1729 incorrect and will change program semantics.
1733 operand bundle set ::= '[' operand bundle (, operand bundle )* ']'
1734 operand bundle ::= tag '(' [ bundle operand ] (, bundle operand )* ')'
1735 bundle operand ::= SSA value
1736 tag ::= string constant
1738 Operand bundles are **not** part of a function's signature, and a
1739 given function may be called from multiple places with different kinds
1740 of operand bundles. This reflects the fact that the operand bundles
1741 are conceptually a part of the ``call`` (or ``invoke``), not the
1742 callee being dispatched to.
1744 Operand bundles are a generic mechanism intended to support
1745 runtime-introspection-like functionality for managed languages. While
1746 the exact semantics of an operand bundle depend on the bundle tag,
1747 there are certain limitations to how much the presence of an operand
1748 bundle can influence the semantics of a program. These restrictions
1749 are described as the semantics of an "unknown" operand bundle. As
1750 long as the behavior of an operand bundle is describable within these
1751 restrictions, LLVM does not need to have special knowledge of the
1752 operand bundle to not miscompile programs containing it.
1754 - The bundle operands for an unknown operand bundle escape in unknown
1755 ways before control is transferred to the callee or invokee.
1756 - Calls and invokes with operand bundles have unknown read / write
1757 effect on the heap on entry and exit (even if the call target is
1758 ``readnone`` or ``readonly``), unless they're overridden with
1759 callsite specific attributes.
1760 - An operand bundle at a call site cannot change the implementation
1761 of the called function. Inter-procedural optimizations work as
1762 usual as long as they take into account the first two properties.
1764 More specific types of operand bundles are described below.
1766 .. _deopt_opbundles:
1768 Deoptimization Operand Bundles
1769 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1771 Deoptimization operand bundles are characterized by the ``"deopt"``
1772 operand bundle tag. These operand bundles represent an alternate
1773 "safe" continuation for the call site they're attached to, and can be
1774 used by a suitable runtime to deoptimize the compiled frame at the
1775 specified call site. There can be at most one ``"deopt"`` operand
1776 bundle attached to a call site. Exact details of deoptimization is
1777 out of scope for the language reference, but it usually involves
1778 rewriting a compiled frame into a set of interpreted frames.
1780 From the compiler's perspective, deoptimization operand bundles make
1781 the call sites they're attached to at least ``readonly``. They read
1782 through all of their pointer typed operands (even if they're not
1783 otherwise escaped) and the entire visible heap. Deoptimization
1784 operand bundles do not capture their operands except during
1785 deoptimization, in which case control will not be returned to the
1788 The inliner knows how to inline through calls that have deoptimization
1789 operand bundles. Just like inlining through a normal call site
1790 involves composing the normal and exceptional continuations, inlining
1791 through a call site with a deoptimization operand bundle needs to
1792 appropriately compose the "safe" deoptimization continuation. The
1793 inliner does this by prepending the parent's deoptimization
1794 continuation to every deoptimization continuation in the inlined body.
1795 E.g. inlining ``@f`` into ``@g`` in the following example
1797 .. code-block:: llvm
1800 call void @x() ;; no deopt state
1801 call void @y() [ "deopt"(i32 10) ]
1802 call void @y() [ "deopt"(i32 10), "unknown"(i8* null) ]
1807 call void @f() [ "deopt"(i32 20) ]
1813 .. code-block:: llvm
1816 call void @x() ;; still no deopt state
1817 call void @y() [ "deopt"(i32 20, i32 10) ]
1818 call void @y() [ "deopt"(i32 20, i32 10), "unknown"(i8* null) ]
1822 It is the frontend's responsibility to structure or encode the
1823 deoptimization state in a way that syntactically prepending the
1824 caller's deoptimization state to the callee's deoptimization state is
1825 semantically equivalent to composing the caller's deoptimization
1826 continuation after the callee's deoptimization continuation.
1830 Funclet Operand Bundles
1831 ^^^^^^^^^^^^^^^^^^^^^^^
1833 Funclet operand bundles are characterized by the ``"funclet"``
1834 operand bundle tag. These operand bundles indicate that a call site
1835 is within a particular funclet. There can be at most one
1836 ``"funclet"`` operand bundle attached to a call site and it must have
1837 exactly one bundle operand.
1839 If any funclet EH pads have been "entered" but not "exited" (per the
1840 `description in the EH doc\ <ExceptionHandling.html#wineh-constraints>`_),
1841 it is undefined behavior to execute a ``call`` or ``invoke`` which:
1843 * does not have a ``"funclet"`` bundle and is not a ``call`` to a nounwind
1845 * has a ``"funclet"`` bundle whose operand is not the most-recently-entered
1846 not-yet-exited funclet EH pad.
1848 Similarly, if no funclet EH pads have been entered-but-not-yet-exited,
1849 executing a ``call`` or ``invoke`` with a ``"funclet"`` bundle is undefined behavior.
1851 GC Transition Operand Bundles
1852 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1854 GC transition operand bundles are characterized by the
1855 ``"gc-transition"`` operand bundle tag. These operand bundles mark a
1856 call as a transition between a function with one GC strategy to a
1857 function with a different GC strategy. If coordinating the transition
1858 between GC strategies requires additional code generation at the call
1859 site, these bundles may contain any values that are needed by the
1860 generated code. For more details, see :ref:`GC Transitions
1861 <gc_transition_args>`.
1865 Module-Level Inline Assembly
1866 ----------------------------
1868 Modules may contain "module-level inline asm" blocks, which corresponds
1869 to the GCC "file scope inline asm" blocks. These blocks are internally
1870 concatenated by LLVM and treated as a single unit, but may be separated
1871 in the ``.ll`` file if desired. The syntax is very simple:
1873 .. code-block:: llvm
1875 module asm "inline asm code goes here"
1876 module asm "more can go here"
1878 The strings can contain any character by escaping non-printable
1879 characters. The escape sequence used is simply "\\xx" where "xx" is the
1880 two digit hex code for the number.
1882 Note that the assembly string *must* be parseable by LLVM's integrated assembler
1883 (unless it is disabled), even when emitting a ``.s`` file.
1885 .. _langref_datalayout:
1890 A module may specify a target specific data layout string that specifies
1891 how data is to be laid out in memory. The syntax for the data layout is
1894 .. code-block:: llvm
1896 target datalayout = "layout specification"
1898 The *layout specification* consists of a list of specifications
1899 separated by the minus sign character ('-'). Each specification starts
1900 with a letter and may include other information after the letter to
1901 define some aspect of the data layout. The specifications accepted are
1905 Specifies that the target lays out data in big-endian form. That is,
1906 the bits with the most significance have the lowest address
1909 Specifies that the target lays out data in little-endian form. That
1910 is, the bits with the least significance have the lowest address
1913 Specifies the natural alignment of the stack in bits. Alignment
1914 promotion of stack variables is limited to the natural stack
1915 alignment to avoid dynamic stack realignment. The stack alignment
1916 must be a multiple of 8-bits. If omitted, the natural stack
1917 alignment defaults to "unspecified", which does not prevent any
1918 alignment promotions.
1919 ``P<address space>``
1920 Specifies the address space that corresponds to program memory.
1921 Harvard architectures can use this to specify what space LLVM
1922 should place things such as functions into. If omitted, the
1923 program memory space defaults to the default address space of 0,
1924 which corresponds to a Von Neumann architecture that has code
1925 and data in the same space.
1926 ``A<address space>``
1927 Specifies the address space of objects created by '``alloca``'.
1928 Defaults to the default address space of 0.
1929 ``p[n]:<size>:<abi>:<pref>:<idx>``
1930 This specifies the *size* of a pointer and its ``<abi>`` and
1931 ``<pref>``\erred alignments for address space ``n``. The fourth parameter
1932 ``<idx>`` is a size of index that used for address calculation. If not
1933 specified, the default index size is equal to the pointer size. All sizes
1934 are in bits. The address space, ``n``, is optional, and if not specified,
1935 denotes the default address space 0. The value of ``n`` must be
1936 in the range [1,2^23).
1937 ``i<size>:<abi>:<pref>``
1938 This specifies the alignment for an integer type of a given bit
1939 ``<size>``. The value of ``<size>`` must be in the range [1,2^23).
1940 ``v<size>:<abi>:<pref>``
1941 This specifies the alignment for a vector type of a given bit
1943 ``f<size>:<abi>:<pref>``
1944 This specifies the alignment for a floating-point type of a given bit
1945 ``<size>``. Only values of ``<size>`` that are supported by the target
1946 will work. 32 (float) and 64 (double) are supported on all targets; 80
1947 or 128 (different flavors of long double) are also supported on some
1950 This specifies the alignment for an object of aggregate type.
1952 If present, specifies that llvm names are mangled in the output. Symbols
1953 prefixed with the mangling escape character ``\01`` are passed through
1954 directly to the assembler without the escape character. The mangling style
1957 * ``e``: ELF mangling: Private symbols get a ``.L`` prefix.
1958 * ``m``: Mips mangling: Private symbols get a ``$`` prefix.
1959 * ``o``: Mach-O mangling: Private symbols get ``L`` prefix. Other
1960 symbols get a ``_`` prefix.
1961 * ``x``: Windows x86 COFF mangling: Private symbols get the usual prefix.
1962 Regular C symbols get a ``_`` prefix. Functions with ``__stdcall``,
1963 ``__fastcall``, and ``__vectorcall`` have custom mangling that appends
1964 ``@N`` where N is the number of bytes used to pass parameters. C++ symbols
1965 starting with ``?`` are not mangled in any way.
1966 * ``w``: Windows COFF mangling: Similar to ``x``, except that normal C
1967 symbols do not receive a ``_`` prefix.
1968 ``n<size1>:<size2>:<size3>...``
1969 This specifies a set of native integer widths for the target CPU in
1970 bits. For example, it might contain ``n32`` for 32-bit PowerPC,
1971 ``n32:64`` for PowerPC 64, or ``n8:16:32:64`` for X86-64. Elements of
1972 this set are considered to support most general arithmetic operations
1974 ``ni:<address space0>:<address space1>:<address space2>...``
1975 This specifies pointer types with the specified address spaces
1976 as :ref:`Non-Integral Pointer Type <nointptrtype>` s. The ``0``
1977 address space cannot be specified as non-integral.
1979 On every specification that takes a ``<abi>:<pref>``, specifying the
1980 ``<pref>`` alignment is optional. If omitted, the preceding ``:``
1981 should be omitted too and ``<pref>`` will be equal to ``<abi>``.
1983 When constructing the data layout for a given target, LLVM starts with a
1984 default set of specifications which are then (possibly) overridden by
1985 the specifications in the ``datalayout`` keyword. The default
1986 specifications are given in this list:
1988 - ``E`` - big endian
1989 - ``p:64:64:64`` - 64-bit pointers with 64-bit alignment.
1990 - ``p[n]:64:64:64`` - Other address spaces are assumed to be the
1991 same as the default address space.
1992 - ``S0`` - natural stack alignment is unspecified
1993 - ``i1:8:8`` - i1 is 8-bit (byte) aligned
1994 - ``i8:8:8`` - i8 is 8-bit (byte) aligned
1995 - ``i16:16:16`` - i16 is 16-bit aligned
1996 - ``i32:32:32`` - i32 is 32-bit aligned
1997 - ``i64:32:64`` - i64 has ABI alignment of 32-bits but preferred
1998 alignment of 64-bits
1999 - ``f16:16:16`` - half is 16-bit aligned
2000 - ``f32:32:32`` - float is 32-bit aligned
2001 - ``f64:64:64`` - double is 64-bit aligned
2002 - ``f128:128:128`` - quad is 128-bit aligned
2003 - ``v64:64:64`` - 64-bit vector is 64-bit aligned
2004 - ``v128:128:128`` - 128-bit vector is 128-bit aligned
2005 - ``a:0:64`` - aggregates are 64-bit aligned
2007 When LLVM is determining the alignment for a given type, it uses the
2010 #. If the type sought is an exact match for one of the specifications,
2011 that specification is used.
2012 #. If no match is found, and the type sought is an integer type, then
2013 the smallest integer type that is larger than the bitwidth of the
2014 sought type is used. If none of the specifications are larger than
2015 the bitwidth then the largest integer type is used. For example,
2016 given the default specifications above, the i7 type will use the
2017 alignment of i8 (next largest) while both i65 and i256 will use the
2018 alignment of i64 (largest specified).
2019 #. If no match is found, and the type sought is a vector type, then the
2020 largest vector type that is smaller than the sought vector type will
2021 be used as a fall back. This happens because <128 x double> can be
2022 implemented in terms of 64 <2 x double>, for example.
2024 The function of the data layout string may not be what you expect.
2025 Notably, this is not a specification from the frontend of what alignment
2026 the code generator should use.
2028 Instead, if specified, the target data layout is required to match what
2029 the ultimate *code generator* expects. This string is used by the
2030 mid-level optimizers to improve code, and this only works if it matches
2031 what the ultimate code generator uses. There is no way to generate IR
2032 that does not embed this target-specific detail into the IR. If you
2033 don't specify the string, the default specifications will be used to
2034 generate a Data Layout and the optimization phases will operate
2035 accordingly and introduce target specificity into the IR with respect to
2036 these default specifications.
2043 A module may specify a target triple string that describes the target
2044 host. The syntax for the target triple is simply:
2046 .. code-block:: llvm
2048 target triple = "x86_64-apple-macosx10.7.0"
2050 The *target triple* string consists of a series of identifiers delimited
2051 by the minus sign character ('-'). The canonical forms are:
2055 ARCHITECTURE-VENDOR-OPERATING_SYSTEM
2056 ARCHITECTURE-VENDOR-OPERATING_SYSTEM-ENVIRONMENT
2058 This information is passed along to the backend so that it generates
2059 code for the proper architecture. It's possible to override this on the
2060 command line with the ``-mtriple`` command line option.
2062 .. _pointeraliasing:
2064 Pointer Aliasing Rules
2065 ----------------------
2067 Any memory access must be done through a pointer value associated with
2068 an address range of the memory access, otherwise the behavior is
2069 undefined. Pointer values are associated with address ranges according
2070 to the following rules:
2072 - A pointer value is associated with the addresses associated with any
2073 value it is *based* on.
2074 - An address of a global variable is associated with the address range
2075 of the variable's storage.
2076 - The result value of an allocation instruction is associated with the
2077 address range of the allocated storage.
2078 - A null pointer in the default address-space is associated with no
2080 - An integer constant other than zero or a pointer value returned from
2081 a function not defined within LLVM may be associated with address
2082 ranges allocated through mechanisms other than those provided by
2083 LLVM. Such ranges shall not overlap with any ranges of addresses
2084 allocated by mechanisms provided by LLVM.
2086 A pointer value is *based* on another pointer value according to the
2089 - A pointer value formed from a scalar ``getelementptr`` operation is *based* on
2090 the pointer-typed operand of the ``getelementptr``.
2091 - The pointer in lane *l* of the result of a vector ``getelementptr`` operation
2092 is *based* on the pointer in lane *l* of the vector-of-pointers-typed operand
2093 of the ``getelementptr``.
2094 - The result value of a ``bitcast`` is *based* on the operand of the
2096 - A pointer value formed by an ``inttoptr`` is *based* on all pointer
2097 values that contribute (directly or indirectly) to the computation of
2098 the pointer's value.
2099 - The "*based* on" relationship is transitive.
2101 Note that this definition of *"based"* is intentionally similar to the
2102 definition of *"based"* in C99, though it is slightly weaker.
2104 LLVM IR does not associate types with memory. The result type of a
2105 ``load`` merely indicates the size and alignment of the memory from
2106 which to load, as well as the interpretation of the value. The first
2107 operand type of a ``store`` similarly only indicates the size and
2108 alignment of the store.
2110 Consequently, type-based alias analysis, aka TBAA, aka
2111 ``-fstrict-aliasing``, is not applicable to general unadorned LLVM IR.
2112 :ref:`Metadata <metadata>` may be used to encode additional information
2113 which specialized optimization passes may use to implement type-based
2118 Volatile Memory Accesses
2119 ------------------------
2121 Certain memory accesses, such as :ref:`load <i_load>`'s,
2122 :ref:`store <i_store>`'s, and :ref:`llvm.memcpy <int_memcpy>`'s may be
2123 marked ``volatile``. The optimizers must not change the number of
2124 volatile operations or change their order of execution relative to other
2125 volatile operations. The optimizers *may* change the order of volatile
2126 operations relative to non-volatile operations. This is not Java's
2127 "volatile" and has no cross-thread synchronization behavior.
2129 IR-level volatile loads and stores cannot safely be optimized into
2130 llvm.memcpy or llvm.memmove intrinsics even when those intrinsics are
2131 flagged volatile. Likewise, the backend should never split or merge
2132 target-legal volatile load/store instructions.
2134 .. admonition:: Rationale
2136 Platforms may rely on volatile loads and stores of natively supported
2137 data width to be executed as single instruction. For example, in C
2138 this holds for an l-value of volatile primitive type with native
2139 hardware support, but not necessarily for aggregate types. The
2140 frontend upholds these expectations, which are intentionally
2141 unspecified in the IR. The rules above ensure that IR transformations
2142 do not violate the frontend's contract with the language.
2146 Memory Model for Concurrent Operations
2147 --------------------------------------
2149 The LLVM IR does not define any way to start parallel threads of
2150 execution or to register signal handlers. Nonetheless, there are
2151 platform-specific ways to create them, and we define LLVM IR's behavior
2152 in their presence. This model is inspired by the C++0x memory model.
2154 For a more informal introduction to this model, see the :doc:`Atomics`.
2156 We define a *happens-before* partial order as the least partial order
2159 - Is a superset of single-thread program order, and
2160 - When a *synchronizes-with* ``b``, includes an edge from ``a`` to
2161 ``b``. *Synchronizes-with* pairs are introduced by platform-specific
2162 techniques, like pthread locks, thread creation, thread joining,
2163 etc., and by atomic instructions. (See also :ref:`Atomic Memory Ordering
2164 Constraints <ordering>`).
2166 Note that program order does not introduce *happens-before* edges
2167 between a thread and signals executing inside that thread.
2169 Every (defined) read operation (load instructions, memcpy, atomic
2170 loads/read-modify-writes, etc.) R reads a series of bytes written by
2171 (defined) write operations (store instructions, atomic
2172 stores/read-modify-writes, memcpy, etc.). For the purposes of this
2173 section, initialized globals are considered to have a write of the
2174 initializer which is atomic and happens before any other read or write
2175 of the memory in question. For each byte of a read R, R\ :sub:`byte`
2176 may see any write to the same byte, except:
2178 - If write\ :sub:`1` happens before write\ :sub:`2`, and
2179 write\ :sub:`2` happens before R\ :sub:`byte`, then
2180 R\ :sub:`byte` does not see write\ :sub:`1`.
2181 - If R\ :sub:`byte` happens before write\ :sub:`3`, then
2182 R\ :sub:`byte` does not see write\ :sub:`3`.
2184 Given that definition, R\ :sub:`byte` is defined as follows:
2186 - If R is volatile, the result is target-dependent. (Volatile is
2187 supposed to give guarantees which can support ``sig_atomic_t`` in
2188 C/C++, and may be used for accesses to addresses that do not behave
2189 like normal memory. It does not generally provide cross-thread
2191 - Otherwise, if there is no write to the same byte that happens before
2192 R\ :sub:`byte`, R\ :sub:`byte` returns ``undef`` for that byte.
2193 - Otherwise, if R\ :sub:`byte` may see exactly one write,
2194 R\ :sub:`byte` returns the value written by that write.
2195 - Otherwise, if R is atomic, and all the writes R\ :sub:`byte` may
2196 see are atomic, it chooses one of the values written. See the :ref:`Atomic
2197 Memory Ordering Constraints <ordering>` section for additional
2198 constraints on how the choice is made.
2199 - Otherwise R\ :sub:`byte` returns ``undef``.
2201 R returns the value composed of the series of bytes it read. This
2202 implies that some bytes within the value may be ``undef`` **without**
2203 the entire value being ``undef``. Note that this only defines the
2204 semantics of the operation; it doesn't mean that targets will emit more
2205 than one instruction to read the series of bytes.
2207 Note that in cases where none of the atomic intrinsics are used, this
2208 model places only one restriction on IR transformations on top of what
2209 is required for single-threaded execution: introducing a store to a byte
2210 which might not otherwise be stored is not allowed in general.
2211 (Specifically, in the case where another thread might write to and read
2212 from an address, introducing a store can change a load that may see
2213 exactly one write into a load that may see multiple writes.)
2217 Atomic Memory Ordering Constraints
2218 ----------------------------------
2220 Atomic instructions (:ref:`cmpxchg <i_cmpxchg>`,
2221 :ref:`atomicrmw <i_atomicrmw>`, :ref:`fence <i_fence>`,
2222 :ref:`atomic load <i_load>`, and :ref:`atomic store <i_store>`) take
2223 ordering parameters that determine which other atomic instructions on
2224 the same address they *synchronize with*. These semantics are borrowed
2225 from Java and C++0x, but are somewhat more colloquial. If these
2226 descriptions aren't precise enough, check those specs (see spec
2227 references in the :doc:`atomics guide <Atomics>`).
2228 :ref:`fence <i_fence>` instructions treat these orderings somewhat
2229 differently since they don't take an address. See that instruction's
2230 documentation for details.
2232 For a simpler introduction to the ordering constraints, see the
2236 The set of values that can be read is governed by the happens-before
2237 partial order. A value cannot be read unless some operation wrote
2238 it. This is intended to provide a guarantee strong enough to model
2239 Java's non-volatile shared variables. This ordering cannot be
2240 specified for read-modify-write operations; it is not strong enough
2241 to make them atomic in any interesting way.
2243 In addition to the guarantees of ``unordered``, there is a single
2244 total order for modifications by ``monotonic`` operations on each
2245 address. All modification orders must be compatible with the
2246 happens-before order. There is no guarantee that the modification
2247 orders can be combined to a global total order for the whole program
2248 (and this often will not be possible). The read in an atomic
2249 read-modify-write operation (:ref:`cmpxchg <i_cmpxchg>` and
2250 :ref:`atomicrmw <i_atomicrmw>`) reads the value in the modification
2251 order immediately before the value it writes. If one atomic read
2252 happens before another atomic read of the same address, the later
2253 read must see the same value or a later value in the address's
2254 modification order. This disallows reordering of ``monotonic`` (or
2255 stronger) operations on the same address. If an address is written
2256 ``monotonic``-ally by one thread, and other threads ``monotonic``-ally
2257 read that address repeatedly, the other threads must eventually see
2258 the write. This corresponds to the C++0x/C1x
2259 ``memory_order_relaxed``.
2261 In addition to the guarantees of ``monotonic``, a
2262 *synchronizes-with* edge may be formed with a ``release`` operation.
2263 This is intended to model C++'s ``memory_order_acquire``.
2265 In addition to the guarantees of ``monotonic``, if this operation
2266 writes a value which is subsequently read by an ``acquire``
2267 operation, it *synchronizes-with* that operation. (This isn't a
2268 complete description; see the C++0x definition of a release
2269 sequence.) This corresponds to the C++0x/C1x
2270 ``memory_order_release``.
2271 ``acq_rel`` (acquire+release)
2272 Acts as both an ``acquire`` and ``release`` operation on its
2273 address. This corresponds to the C++0x/C1x ``memory_order_acq_rel``.
2274 ``seq_cst`` (sequentially consistent)
2275 In addition to the guarantees of ``acq_rel`` (``acquire`` for an
2276 operation that only reads, ``release`` for an operation that only
2277 writes), there is a global total order on all
2278 sequentially-consistent operations on all addresses, which is
2279 consistent with the *happens-before* partial order and with the
2280 modification orders of all the affected addresses. Each
2281 sequentially-consistent read sees the last preceding write to the
2282 same address in this global order. This corresponds to the C++0x/C1x
2283 ``memory_order_seq_cst`` and Java volatile.
2287 If an atomic operation is marked ``syncscope("singlethread")``, it only
2288 *synchronizes with* and only participates in the seq\_cst total orderings of
2289 other operations running in the same thread (for example, in signal handlers).
2291 If an atomic operation is marked ``syncscope("<target-scope>")``, where
2292 ``<target-scope>`` is a target specific synchronization scope, then it is target
2293 dependent if it *synchronizes with* and participates in the seq\_cst total
2294 orderings of other operations.
2296 Otherwise, an atomic operation that is not marked ``syncscope("singlethread")``
2297 or ``syncscope("<target-scope>")`` *synchronizes with* and participates in the
2298 seq\_cst total orderings of other operations that are not marked
2299 ``syncscope("singlethread")`` or ``syncscope("<target-scope>")``.
2303 Floating-Point Environment
2304 --------------------------
2306 The default LLVM floating-point environment assumes that floating-point
2307 instructions do not have side effects. Results assume the round-to-nearest
2308 rounding mode. No floating-point exception state is maintained in this
2309 environment. Therefore, there is no attempt to create or preserve invalid
2310 operation (SNaN) or division-by-zero exceptions in these examples:
2312 .. code-block:: llvm
2314 %A = fdiv 0x7ff0000000000001, %X ; 64-bit SNaN hex value
2320 The benefit of this exception-free assumption is that floating-point
2321 operations may be speculated freely without any other fast-math relaxations
2322 to the floating-point model.
2324 Code that requires different behavior than this should use the
2325 :ref:`Constrained Floating-Point Intrinsics <constrainedfp>`.
2332 LLVM IR floating-point operations (:ref:`fadd <i_fadd>`,
2333 :ref:`fsub <i_fsub>`, :ref:`fmul <i_fmul>`, :ref:`fdiv <i_fdiv>`,
2334 :ref:`frem <i_frem>`, :ref:`fcmp <i_fcmp>`) and :ref:`call <i_call>`
2335 may use the following flags to enable otherwise unsafe
2336 floating-point transformations.
2339 No NaNs - Allow optimizations to assume the arguments and result are not
2340 NaN. Such optimizations are required to retain defined behavior over
2341 NaNs, but the value of the result is undefined.
2344 No Infs - Allow optimizations to assume the arguments and result are not
2345 +/-Inf. Such optimizations are required to retain defined behavior over
2346 +/-Inf, but the value of the result is undefined.
2349 No Signed Zeros - Allow optimizations to treat the sign of a zero
2350 argument or result as insignificant.
2353 Allow Reciprocal - Allow optimizations to use the reciprocal of an
2354 argument rather than perform division.
2357 Allow floating-point contraction (e.g. fusing a multiply followed by an
2358 addition into a fused multiply-and-add).
2361 Approximate functions - Allow substitution of approximate calculations for
2362 functions (sin, log, sqrt, etc). See floating-point intrinsic definitions
2363 for places where this can apply to LLVM's intrinsic math functions.
2366 Allow reassociation transformations for floating-point instructions.
2367 This may dramatically change results in floating-point.
2370 This flag implies all of the others.
2374 Use-list Order Directives
2375 -------------------------
2377 Use-list directives encode the in-memory order of each use-list, allowing the
2378 order to be recreated. ``<order-indexes>`` is a comma-separated list of
2379 indexes that are assigned to the referenced value's uses. The referenced
2380 value's use-list is immediately sorted by these indexes.
2382 Use-list directives may appear at function scope or global scope. They are not
2383 instructions, and have no effect on the semantics of the IR. When they're at
2384 function scope, they must appear after the terminator of the final basic block.
2386 If basic blocks have their address taken via ``blockaddress()`` expressions,
2387 ``uselistorder_bb`` can be used to reorder their use-lists from outside their
2394 uselistorder <ty> <value>, { <order-indexes> }
2395 uselistorder_bb @function, %block { <order-indexes> }
2401 define void @foo(i32 %arg1, i32 %arg2) {
2403 ; ... instructions ...
2405 ; ... instructions ...
2407 ; At function scope.
2408 uselistorder i32 %arg1, { 1, 0, 2 }
2409 uselistorder label %bb, { 1, 0 }
2413 uselistorder i32* @global, { 1, 2, 0 }
2414 uselistorder i32 7, { 1, 0 }
2415 uselistorder i32 (i32) @bar, { 1, 0 }
2416 uselistorder_bb @foo, %bb, { 5, 1, 3, 2, 0, 4 }
2418 .. _source_filename:
2423 The *source filename* string is set to the original module identifier,
2424 which will be the name of the compiled source file when compiling from
2425 source through the clang front end, for example. It is then preserved through
2428 This is currently necessary to generate a consistent unique global
2429 identifier for local functions used in profile data, which prepends the
2430 source file name to the local function name.
2432 The syntax for the source file name is simply:
2434 .. code-block:: text
2436 source_filename = "/path/to/source.c"
2443 The LLVM type system is one of the most important features of the
2444 intermediate representation. Being typed enables a number of
2445 optimizations to be performed on the intermediate representation
2446 directly, without having to do extra analyses on the side before the
2447 transformation. A strong type system makes it easier to read the
2448 generated code and enables novel analyses and transformations that are
2449 not feasible to perform on normal three address code representations.
2459 The void type does not represent any value and has no size.
2477 The function type can be thought of as a function signature. It consists of a
2478 return type and a list of formal parameter types. The return type of a function
2479 type is a void type or first class type --- except for :ref:`label <t_label>`
2480 and :ref:`metadata <t_metadata>` types.
2486 <returntype> (<parameter list>)
2488 ...where '``<parameter list>``' is a comma-separated list of type
2489 specifiers. Optionally, the parameter list may include a type ``...``, which
2490 indicates that the function takes a variable number of arguments. Variable
2491 argument functions can access their arguments with the :ref:`variable argument
2492 handling intrinsic <int_varargs>` functions. '``<returntype>``' is any type
2493 except :ref:`label <t_label>` and :ref:`metadata <t_metadata>`.
2497 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2498 | ``i32 (i32)`` | function taking an ``i32``, returning an ``i32`` |
2499 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2500 | ``float (i16, i32 *) *`` | :ref:`Pointer <t_pointer>` to a function that takes an ``i16`` and a :ref:`pointer <t_pointer>` to ``i32``, returning ``float``. |
2501 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2502 | ``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. |
2503 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2504 | ``{i32, i32} (i32)`` | A function taking an ``i32``, returning a :ref:`structure <t_struct>` containing two ``i32`` values |
2505 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2512 The :ref:`first class <t_firstclass>` types are perhaps the most important.
2513 Values of these types are the only ones which can be produced by
2521 These are the types that are valid in registers from CodeGen's perspective.
2530 The integer type is a very simple type that simply specifies an
2531 arbitrary bit width for the integer type desired. Any bit width from 1
2532 bit to 2\ :sup:`23`\ -1 (about 8 million) can be specified.
2540 The number of bits the integer will occupy is specified by the ``N``
2546 +----------------+------------------------------------------------+
2547 | ``i1`` | a single-bit integer. |
2548 +----------------+------------------------------------------------+
2549 | ``i32`` | a 32-bit integer. |
2550 +----------------+------------------------------------------------+
2551 | ``i1942652`` | a really big integer of over 1 million bits. |
2552 +----------------+------------------------------------------------+
2556 Floating-Point Types
2557 """"""""""""""""""""
2566 - 16-bit floating-point value
2569 - 32-bit floating-point value
2572 - 64-bit floating-point value
2575 - 128-bit floating-point value (112-bit mantissa)
2578 - 80-bit floating-point value (X87)
2581 - 128-bit floating-point value (two 64-bits)
2583 The binary format of half, float, double, and fp128 correspond to the
2584 IEEE-754-2008 specifications for binary16, binary32, binary64, and binary128
2592 The x86_mmx type represents a value held in an MMX register on an x86
2593 machine. The operations allowed on it are quite limited: parameters and
2594 return values, load and store, and bitcast. User-specified MMX
2595 instructions are represented as intrinsic or asm calls with arguments
2596 and/or results of this type. There are no arrays, vectors or constants
2613 The pointer type is used to specify memory locations. Pointers are
2614 commonly used to reference objects in memory.
2616 Pointer types may have an optional address space attribute defining the
2617 numbered address space where the pointed-to object resides. The default
2618 address space is number zero. The semantics of non-zero address spaces
2619 are target-specific.
2621 Note that LLVM does not permit pointers to void (``void*``) nor does it
2622 permit pointers to labels (``label*``). Use ``i8*`` instead.
2632 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2633 | ``[4 x i32]*`` | A :ref:`pointer <t_pointer>` to :ref:`array <t_array>` of four ``i32`` values. |
2634 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2635 | ``i32 (i32*) *`` | A :ref:`pointer <t_pointer>` to a :ref:`function <t_function>` that takes an ``i32*``, returning an ``i32``. |
2636 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2637 | ``i32 addrspace(5)*`` | A :ref:`pointer <t_pointer>` to an ``i32`` value that resides in address space #5. |
2638 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2647 A vector type is a simple derived type that represents a vector of
2648 elements. Vector types are used when multiple primitive data are
2649 operated in parallel using a single instruction (SIMD). A vector type
2650 requires a size (number of elements) and an underlying primitive data
2651 type. Vector types are considered :ref:`first class <t_firstclass>`.
2657 < <# elements> x <elementtype> >
2659 The number of elements is a constant integer value larger than 0;
2660 elementtype may be any integer, floating-point or pointer type. Vectors
2661 of size zero are not allowed.
2665 +-------------------+--------------------------------------------------+
2666 | ``<4 x i32>`` | Vector of 4 32-bit integer values. |
2667 +-------------------+--------------------------------------------------+
2668 | ``<8 x float>`` | Vector of 8 32-bit floating-point values. |
2669 +-------------------+--------------------------------------------------+
2670 | ``<2 x i64>`` | Vector of 2 64-bit integer values. |
2671 +-------------------+--------------------------------------------------+
2672 | ``<4 x i64*>`` | Vector of 4 pointers to 64-bit integer values. |
2673 +-------------------+--------------------------------------------------+
2682 The label type represents code labels.
2697 The token type is used when a value is associated with an instruction
2698 but all uses of the value must not attempt to introspect or obscure it.
2699 As such, it is not appropriate to have a :ref:`phi <i_phi>` or
2700 :ref:`select <i_select>` of type token.
2717 The metadata type represents embedded metadata. No derived types may be
2718 created from metadata except for :ref:`function <t_function>` arguments.
2731 Aggregate Types are a subset of derived types that can contain multiple
2732 member types. :ref:`Arrays <t_array>` and :ref:`structs <t_struct>` are
2733 aggregate types. :ref:`Vectors <t_vector>` are not considered to be
2743 The array type is a very simple derived type that arranges elements
2744 sequentially in memory. The array type requires a size (number of
2745 elements) and an underlying data type.
2751 [<# elements> x <elementtype>]
2753 The number of elements is a constant integer value; ``elementtype`` may
2754 be any type with a size.
2758 +------------------+--------------------------------------+
2759 | ``[40 x i32]`` | Array of 40 32-bit integer values. |
2760 +------------------+--------------------------------------+
2761 | ``[41 x i32]`` | Array of 41 32-bit integer values. |
2762 +------------------+--------------------------------------+
2763 | ``[4 x i8]`` | Array of 4 8-bit integer values. |
2764 +------------------+--------------------------------------+
2766 Here are some examples of multidimensional arrays:
2768 +-----------------------------+----------------------------------------------------------+
2769 | ``[3 x [4 x i32]]`` | 3x4 array of 32-bit integer values. |
2770 +-----------------------------+----------------------------------------------------------+
2771 | ``[12 x [10 x float]]`` | 12x10 array of single precision floating-point values. |
2772 +-----------------------------+----------------------------------------------------------+
2773 | ``[2 x [3 x [4 x i16]]]`` | 2x3x4 array of 16-bit integer values. |
2774 +-----------------------------+----------------------------------------------------------+
2776 There is no restriction on indexing beyond the end of the array implied
2777 by a static type (though there are restrictions on indexing beyond the
2778 bounds of an allocated object in some cases). This means that
2779 single-dimension 'variable sized array' addressing can be implemented in
2780 LLVM with a zero length array type. An implementation of 'pascal style
2781 arrays' in LLVM could use the type "``{ i32, [0 x float]}``", for
2791 The structure type is used to represent a collection of data members
2792 together in memory. The elements of a structure may be any type that has
2795 Structures in memory are accessed using '``load``' and '``store``' by
2796 getting a pointer to a field with the '``getelementptr``' instruction.
2797 Structures in registers are accessed using the '``extractvalue``' and
2798 '``insertvalue``' instructions.
2800 Structures may optionally be "packed" structures, which indicate that
2801 the alignment of the struct is one byte, and that there is no padding
2802 between the elements. In non-packed structs, padding between field types
2803 is inserted as defined by the DataLayout string in the module, which is
2804 required to match what the underlying code generator expects.
2806 Structures can either be "literal" or "identified". A literal structure
2807 is defined inline with other types (e.g. ``{i32, i32}*``) whereas
2808 identified types are always defined at the top level with a name.
2809 Literal types are uniqued by their contents and can never be recursive
2810 or opaque since there is no way to write one. Identified types can be
2811 recursive, can be opaqued, and are never uniqued.
2817 %T1 = type { <type list> } ; Identified normal struct type
2818 %T2 = type <{ <type list> }> ; Identified packed struct type
2822 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2823 | ``{ i32, i32, i32 }`` | A triple of three ``i32`` values |
2824 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2825 | ``{ 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``. |
2826 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2827 | ``<{ i8, i32 }>`` | A packed struct known to be 5 bytes in size. |
2828 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2832 Opaque Structure Types
2833 """"""""""""""""""""""
2837 Opaque structure types are used to represent named structure types that
2838 do not have a body specified. This corresponds (for example) to the C
2839 notion of a forward declared structure.
2850 +--------------+-------------------+
2851 | ``opaque`` | An opaque type. |
2852 +--------------+-------------------+
2859 LLVM has several different basic types of constants. This section
2860 describes them all and their syntax.
2865 **Boolean constants**
2866 The two strings '``true``' and '``false``' are both valid constants
2868 **Integer constants**
2869 Standard integers (such as '4') are constants of the
2870 :ref:`integer <t_integer>` type. Negative numbers may be used with
2872 **Floating-point constants**
2873 Floating-point constants use standard decimal notation (e.g.
2874 123.421), exponential notation (e.g. 1.23421e+2), or a more precise
2875 hexadecimal notation (see below). The assembler requires the exact
2876 decimal value of a floating-point constant. For example, the
2877 assembler accepts 1.25 but rejects 1.3 because 1.3 is a repeating
2878 decimal in binary. Floating-point constants must have a
2879 :ref:`floating-point <t_floating>` type.
2880 **Null pointer constants**
2881 The identifier '``null``' is recognized as a null pointer constant
2882 and must be of :ref:`pointer type <t_pointer>`.
2884 The identifier '``none``' is recognized as an empty token constant
2885 and must be of :ref:`token type <t_token>`.
2887 The one non-intuitive notation for constants is the hexadecimal form of
2888 floating-point constants. For example, the form
2889 '``double 0x432ff973cafa8000``' is equivalent to (but harder to read
2890 than) '``double 4.5e+15``'. The only time hexadecimal floating-point
2891 constants are required (and the only time that they are generated by the
2892 disassembler) is when a floating-point constant must be emitted but it
2893 cannot be represented as a decimal floating-point number in a reasonable
2894 number of digits. For example, NaN's, infinities, and other special
2895 values are represented in their IEEE hexadecimal format so that assembly
2896 and disassembly do not cause any bits to change in the constants.
2898 When using the hexadecimal form, constants of types half, float, and
2899 double are represented using the 16-digit form shown above (which
2900 matches the IEEE754 representation for double); half and float values
2901 must, however, be exactly representable as IEEE 754 half and single
2902 precision, respectively. Hexadecimal format is always used for long
2903 double, and there are three forms of long double. The 80-bit format used
2904 by x86 is represented as ``0xK`` followed by 20 hexadecimal digits. The
2905 128-bit format used by PowerPC (two adjacent doubles) is represented by
2906 ``0xM`` followed by 32 hexadecimal digits. The IEEE 128-bit format is
2907 represented by ``0xL`` followed by 32 hexadecimal digits. Long doubles
2908 will only work if they match the long double format on your target.
2909 The IEEE 16-bit format (half precision) is represented by ``0xH``
2910 followed by 4 hexadecimal digits. All hexadecimal formats are big-endian
2911 (sign bit at the left).
2913 There are no constants of type x86_mmx.
2915 .. _complexconstants:
2920 Complex constants are a (potentially recursive) combination of simple
2921 constants and smaller complex constants.
2923 **Structure constants**
2924 Structure constants are represented with notation similar to
2925 structure type definitions (a comma separated list of elements,
2926 surrounded by braces (``{}``)). For example:
2927 "``{ i32 4, float 17.0, i32* @G }``", where "``@G``" is declared as
2928 "``@G = external global i32``". Structure constants must have
2929 :ref:`structure type <t_struct>`, and the number and types of elements
2930 must match those specified by the type.
2932 Array constants are represented with notation similar to array type
2933 definitions (a comma separated list of elements, surrounded by
2934 square brackets (``[]``)). For example:
2935 "``[ i32 42, i32 11, i32 74 ]``". Array constants must have
2936 :ref:`array type <t_array>`, and the number and types of elements must
2937 match those specified by the type. As a special case, character array
2938 constants may also be represented as a double-quoted string using the ``c``
2939 prefix. For example: "``c"Hello World\0A\00"``".
2940 **Vector constants**
2941 Vector constants are represented with notation similar to vector
2942 type definitions (a comma separated list of elements, surrounded by
2943 less-than/greater-than's (``<>``)). For example:
2944 "``< i32 42, i32 11, i32 74, i32 100 >``". Vector constants
2945 must have :ref:`vector type <t_vector>`, and the number and types of
2946 elements must match those specified by the type.
2947 **Zero initialization**
2948 The string '``zeroinitializer``' can be used to zero initialize a
2949 value to zero of *any* type, including scalar and
2950 :ref:`aggregate <t_aggregate>` types. This is often used to avoid
2951 having to print large zero initializers (e.g. for large arrays) and
2952 is always exactly equivalent to using explicit zero initializers.
2954 A metadata node is a constant tuple without types. For example:
2955 "``!{!0, !{!2, !0}, !"test"}``". Metadata can reference constant values,
2956 for example: "``!{!0, i32 0, i8* @global, i64 (i64)* @function, !"str"}``".
2957 Unlike other typed constants that are meant to be interpreted as part of
2958 the instruction stream, metadata is a place to attach additional
2959 information such as debug info.
2961 Global Variable and Function Addresses
2962 --------------------------------------
2964 The addresses of :ref:`global variables <globalvars>` and
2965 :ref:`functions <functionstructure>` are always implicitly valid
2966 (link-time) constants. These constants are explicitly referenced when
2967 the :ref:`identifier for the global <identifiers>` is used and always have
2968 :ref:`pointer <t_pointer>` type. For example, the following is a legal LLVM
2971 .. code-block:: llvm
2975 @Z = global [2 x i32*] [ i32* @X, i32* @Y ]
2982 The string '``undef``' can be used anywhere a constant is expected, and
2983 indicates that the user of the value may receive an unspecified
2984 bit-pattern. Undefined values may be of any type (other than '``label``'
2985 or '``void``') and be used anywhere a constant is permitted.
2987 Undefined values are useful because they indicate to the compiler that
2988 the program is well defined no matter what value is used. This gives the
2989 compiler more freedom to optimize. Here are some examples of
2990 (potentially surprising) transformations that are valid (in pseudo IR):
2992 .. code-block:: llvm
3002 This is safe because all of the output bits are affected by the undef
3003 bits. Any output bit can have a zero or one depending on the input bits.
3005 .. code-block:: llvm
3013 %A = %X ;; By choosing undef as 0
3014 %B = %X ;; By choosing undef as -1
3019 These logical operations have bits that are not always affected by the
3020 input. For example, if ``%X`` has a zero bit, then the output of the
3021 '``and``' operation will always be a zero for that bit, no matter what
3022 the corresponding bit from the '``undef``' is. As such, it is unsafe to
3023 optimize or assume that the result of the '``and``' is '``undef``'.
3024 However, it is safe to assume that all bits of the '``undef``' could be
3025 0, and optimize the '``and``' to 0. Likewise, it is safe to assume that
3026 all the bits of the '``undef``' operand to the '``or``' could be set,
3027 allowing the '``or``' to be folded to -1.
3029 .. code-block:: llvm
3031 %A = select undef, %X, %Y
3032 %B = select undef, 42, %Y
3033 %C = select %X, %Y, undef
3043 This set of examples shows that undefined '``select``' (and conditional
3044 branch) conditions can go *either way*, but they have to come from one
3045 of the two operands. In the ``%A`` example, if ``%X`` and ``%Y`` were
3046 both known to have a clear low bit, then ``%A`` would have to have a
3047 cleared low bit. However, in the ``%C`` example, the optimizer is
3048 allowed to assume that the '``undef``' operand could be the same as
3049 ``%Y``, allowing the whole '``select``' to be eliminated.
3051 .. code-block:: text
3053 %A = xor undef, undef
3070 This example points out that two '``undef``' operands are not
3071 necessarily the same. This can be surprising to people (and also matches
3072 C semantics) where they assume that "``X^X``" is always zero, even if
3073 ``X`` is undefined. This isn't true for a number of reasons, but the
3074 short answer is that an '``undef``' "variable" can arbitrarily change
3075 its value over its "live range". This is true because the variable
3076 doesn't actually *have a live range*. Instead, the value is logically
3077 read from arbitrary registers that happen to be around when needed, so
3078 the value is not necessarily consistent over time. In fact, ``%A`` and
3079 ``%C`` need to have the same semantics or the core LLVM "replace all
3080 uses with" concept would not hold.
3082 .. code-block:: llvm
3090 These examples show the crucial difference between an *undefined value*
3091 and *undefined behavior*. An undefined value (like '``undef``') is
3092 allowed to have an arbitrary bit-pattern. This means that the ``%A``
3093 operation can be constant folded to '``0``', because the '``undef``'
3094 could be zero, and zero divided by any value is zero.
3095 However, in the second example, we can make a more aggressive
3096 assumption: because the ``undef`` is allowed to be an arbitrary value,
3097 we are allowed to assume that it could be zero. Since a divide by zero
3098 has *undefined behavior*, we are allowed to assume that the operation
3099 does not execute at all. This allows us to delete the divide and all
3100 code after it. Because the undefined operation "can't happen", the
3101 optimizer can assume that it occurs in dead code.
3103 .. code-block:: text
3105 a: store undef -> %X
3106 b: store %X -> undef
3111 A store *of* an undefined value can be assumed to not have any effect;
3112 we can assume that the value is overwritten with bits that happen to
3113 match what was already there. However, a store *to* an undefined
3114 location could clobber arbitrary memory, therefore, it has undefined
3122 Poison values are similar to :ref:`undef values <undefvalues>`, however
3123 they also represent the fact that an instruction or constant expression
3124 that cannot evoke side effects has nevertheless detected a condition
3125 that results in undefined behavior.
3127 There is currently no way of representing a poison value in the IR; they
3128 only exist when produced by operations such as :ref:`add <i_add>` with
3131 Poison value behavior is defined in terms of value *dependence*:
3133 - Values other than :ref:`phi <i_phi>` nodes depend on their operands.
3134 - :ref:`Phi <i_phi>` nodes depend on the operand corresponding to
3135 their dynamic predecessor basic block.
3136 - Function arguments depend on the corresponding actual argument values
3137 in the dynamic callers of their functions.
3138 - :ref:`Call <i_call>` instructions depend on the :ref:`ret <i_ret>`
3139 instructions that dynamically transfer control back to them.
3140 - :ref:`Invoke <i_invoke>` instructions depend on the
3141 :ref:`ret <i_ret>`, :ref:`resume <i_resume>`, or exception-throwing
3142 call instructions that dynamically transfer control back to them.
3143 - Non-volatile loads and stores depend on the most recent stores to all
3144 of the referenced memory addresses, following the order in the IR
3145 (including loads and stores implied by intrinsics such as
3146 :ref:`@llvm.memcpy <int_memcpy>`.)
3147 - An instruction with externally visible side effects depends on the
3148 most recent preceding instruction with externally visible side
3149 effects, following the order in the IR. (This includes :ref:`volatile
3150 operations <volatile>`.)
3151 - An instruction *control-depends* on a :ref:`terminator
3152 instruction <terminators>` if the terminator instruction has
3153 multiple successors and the instruction is always executed when
3154 control transfers to one of the successors, and may not be executed
3155 when control is transferred to another.
3156 - Additionally, an instruction also *control-depends* on a terminator
3157 instruction if the set of instructions it otherwise depends on would
3158 be different if the terminator had transferred control to a different
3160 - Dependence is transitive.
3162 Poison values have the same behavior as :ref:`undef values <undefvalues>`,
3163 with the additional effect that any instruction that has a *dependence*
3164 on a poison value has undefined behavior.
3166 Here are some examples:
3168 .. code-block:: llvm
3171 %poison = sub nuw i32 0, 1 ; Results in a poison value.
3172 %still_poison = and i32 %poison, 0 ; 0, but also poison.
3173 %poison_yet_again = getelementptr i32, i32* @h, i32 %still_poison
3174 store i32 0, i32* %poison_yet_again ; memory at @h[0] is poisoned
3176 store i32 %poison, i32* @g ; Poison value stored to memory.
3177 %poison2 = load i32, i32* @g ; Poison value loaded back from memory.
3179 store volatile i32 %poison, i32* @g ; External observation; undefined behavior.
3181 %narrowaddr = bitcast i32* @g to i16*
3182 %wideaddr = bitcast i32* @g to i64*
3183 %poison3 = load i16, i16* %narrowaddr ; Returns a poison value.
3184 %poison4 = load i64, i64* %wideaddr ; Returns a poison value.
3186 %cmp = icmp slt i32 %poison, 0 ; Returns a poison value.
3187 br i1 %cmp, label %true, label %end ; Branch to either destination.
3190 store volatile i32 0, i32* @g ; This is control-dependent on %cmp, so
3191 ; it has undefined behavior.
3195 %p = phi i32 [ 0, %entry ], [ 1, %true ]
3196 ; Both edges into this PHI are
3197 ; control-dependent on %cmp, so this
3198 ; always results in a poison value.
3200 store volatile i32 0, i32* @g ; This would depend on the store in %true
3201 ; if %cmp is true, or the store in %entry
3202 ; otherwise, so this is undefined behavior.
3204 br i1 %cmp, label %second_true, label %second_end
3205 ; The same branch again, but this time the
3206 ; true block doesn't have side effects.
3213 store volatile i32 0, i32* @g ; This time, the instruction always depends
3214 ; on the store in %end. Also, it is
3215 ; control-equivalent to %end, so this is
3216 ; well-defined (ignoring earlier undefined
3217 ; behavior in this example).
3221 Addresses of Basic Blocks
3222 -------------------------
3224 ``blockaddress(@function, %block)``
3226 The '``blockaddress``' constant computes the address of the specified
3227 basic block in the specified function, and always has an ``i8*`` type.
3228 Taking the address of the entry block is illegal.
3230 This value only has defined behavior when used as an operand to the
3231 ':ref:`indirectbr <i_indirectbr>`' instruction, or for comparisons
3232 against null. Pointer equality tests between labels addresses results in
3233 undefined behavior --- though, again, comparison against null is ok, and
3234 no label is equal to the null pointer. This may be passed around as an
3235 opaque pointer sized value as long as the bits are not inspected. This
3236 allows ``ptrtoint`` and arithmetic to be performed on these values so
3237 long as the original value is reconstituted before the ``indirectbr``
3240 Finally, some targets may provide defined semantics when using the value
3241 as the operand to an inline assembly, but that is target specific.
3245 Constant Expressions
3246 --------------------
3248 Constant expressions are used to allow expressions involving other
3249 constants to be used as constants. Constant expressions may be of any
3250 :ref:`first class <t_firstclass>` type and may involve any LLVM operation
3251 that does not have side effects (e.g. load and call are not supported).
3252 The following is the syntax for constant expressions:
3254 ``trunc (CST to TYPE)``
3255 Perform the :ref:`trunc operation <i_trunc>` on constants.
3256 ``zext (CST to TYPE)``
3257 Perform the :ref:`zext operation <i_zext>` on constants.
3258 ``sext (CST to TYPE)``
3259 Perform the :ref:`sext operation <i_sext>` on constants.
3260 ``fptrunc (CST to TYPE)``
3261 Truncate a floating-point constant to another floating-point type.
3262 The size of CST must be larger than the size of TYPE. Both types
3263 must be floating-point.
3264 ``fpext (CST to TYPE)``
3265 Floating-point extend a constant to another type. The size of CST
3266 must be smaller or equal to the size of TYPE. Both types must be
3268 ``fptoui (CST to TYPE)``
3269 Convert a floating-point constant to the corresponding unsigned
3270 integer constant. TYPE must be a scalar or vector integer type. CST
3271 must be of scalar or vector floating-point type. Both CST and TYPE
3272 must be scalars, or vectors of the same number of elements. If the
3273 value won't fit in the integer type, the results are undefined.
3274 ``fptosi (CST to TYPE)``
3275 Convert a floating-point constant to the corresponding signed
3276 integer constant. TYPE must be a scalar or vector integer type. CST
3277 must be of scalar or vector floating-point type. Both CST and TYPE
3278 must be scalars, or vectors of the same number of elements. If the
3279 value won't fit in the integer type, the results are undefined.
3280 ``uitofp (CST to TYPE)``
3281 Convert an unsigned integer constant to the corresponding
3282 floating-point constant. TYPE must be a scalar or vector floating-point
3283 type. CST must be of scalar or vector integer type. Both CST and TYPE must
3284 be scalars, or vectors of the same number of elements. If the value
3285 won't fit in the floating-point type, the results are undefined.
3286 ``sitofp (CST to TYPE)``
3287 Convert a signed integer constant to the corresponding floating-point
3288 constant. TYPE must be a scalar or vector floating-point type.
3289 CST must be of scalar or vector integer type. Both CST and TYPE must
3290 be scalars, or vectors of the same number of elements. If the value
3291 won't fit in the floating-point type, the results are undefined.
3292 ``ptrtoint (CST to TYPE)``
3293 Perform the :ref:`ptrtoint operation <i_ptrtoint>` on constants.
3294 ``inttoptr (CST to TYPE)``
3295 Perform the :ref:`inttoptr operation <i_inttoptr>` on constants.
3296 This one is *really* dangerous!
3297 ``bitcast (CST to TYPE)``
3298 Convert a constant, CST, to another TYPE.
3299 The constraints of the operands are the same as those for the
3300 :ref:`bitcast instruction <i_bitcast>`.
3301 ``addrspacecast (CST to TYPE)``
3302 Convert a constant pointer or constant vector of pointer, CST, to another
3303 TYPE in a different address space. The constraints of the operands are the
3304 same as those for the :ref:`addrspacecast instruction <i_addrspacecast>`.
3305 ``getelementptr (TY, CSTPTR, IDX0, IDX1, ...)``, ``getelementptr inbounds (TY, CSTPTR, IDX0, IDX1, ...)``
3306 Perform the :ref:`getelementptr operation <i_getelementptr>` on
3307 constants. As with the :ref:`getelementptr <i_getelementptr>`
3308 instruction, the index list may have one or more indexes, which are
3309 required to make sense for the type of "pointer to TY".
3310 ``select (COND, VAL1, VAL2)``
3311 Perform the :ref:`select operation <i_select>` on constants.
3312 ``icmp COND (VAL1, VAL2)``
3313 Perform the :ref:`icmp operation <i_icmp>` on constants.
3314 ``fcmp COND (VAL1, VAL2)``
3315 Perform the :ref:`fcmp operation <i_fcmp>` on constants.
3316 ``extractelement (VAL, IDX)``
3317 Perform the :ref:`extractelement operation <i_extractelement>` on
3319 ``insertelement (VAL, ELT, IDX)``
3320 Perform the :ref:`insertelement operation <i_insertelement>` on
3322 ``shufflevector (VEC1, VEC2, IDXMASK)``
3323 Perform the :ref:`shufflevector operation <i_shufflevector>` on
3325 ``extractvalue (VAL, IDX0, IDX1, ...)``
3326 Perform the :ref:`extractvalue operation <i_extractvalue>` on
3327 constants. The index list is interpreted in a similar manner as
3328 indices in a ':ref:`getelementptr <i_getelementptr>`' operation. At
3329 least one index value must be specified.
3330 ``insertvalue (VAL, ELT, IDX0, IDX1, ...)``
3331 Perform the :ref:`insertvalue operation <i_insertvalue>` on constants.
3332 The index list is interpreted in a similar manner as indices in a
3333 ':ref:`getelementptr <i_getelementptr>`' operation. At least one index
3334 value must be specified.
3335 ``OPCODE (LHS, RHS)``
3336 Perform the specified operation of the LHS and RHS constants. OPCODE
3337 may be any of the :ref:`binary <binaryops>` or :ref:`bitwise
3338 binary <bitwiseops>` operations. The constraints on operands are
3339 the same as those for the corresponding instruction (e.g. no bitwise
3340 operations on floating-point values are allowed).
3347 Inline Assembler Expressions
3348 ----------------------------
3350 LLVM supports inline assembler expressions (as opposed to :ref:`Module-Level
3351 Inline Assembly <moduleasm>`) through the use of a special value. This value
3352 represents the inline assembler as a template string (containing the
3353 instructions to emit), a list of operand constraints (stored as a string), a
3354 flag that indicates whether or not the inline asm expression has side effects,
3355 and a flag indicating whether the function containing the asm needs to align its
3356 stack conservatively.
3358 The template string supports argument substitution of the operands using "``$``"
3359 followed by a number, to indicate substitution of the given register/memory
3360 location, as specified by the constraint string. "``${NUM:MODIFIER}``" may also
3361 be used, where ``MODIFIER`` is a target-specific annotation for how to print the
3362 operand (See :ref:`inline-asm-modifiers`).
3364 A literal "``$``" may be included by using "``$$``" in the template. To include
3365 other special characters into the output, the usual "``\XX``" escapes may be
3366 used, just as in other strings. Note that after template substitution, the
3367 resulting assembly string is parsed by LLVM's integrated assembler unless it is
3368 disabled -- even when emitting a ``.s`` file -- and thus must contain assembly
3369 syntax known to LLVM.
3371 LLVM also supports a few more substitions useful for writing inline assembly:
3373 - ``${:uid}``: Expands to a decimal integer unique to this inline assembly blob.
3374 This substitution is useful when declaring a local label. Many standard
3375 compiler optimizations, such as inlining, may duplicate an inline asm blob.
3376 Adding a blob-unique identifier ensures that the two labels will not conflict
3377 during assembly. This is used to implement `GCC's %= special format
3378 string <https://gcc.gnu.org/onlinedocs/gcc/Extended-Asm.html>`_.
3379 - ``${:comment}``: Expands to the comment character of the current target's
3380 assembly dialect. This is usually ``#``, but many targets use other strings,
3381 such as ``;``, ``//``, or ``!``.
3382 - ``${:private}``: Expands to the assembler private label prefix. Labels with
3383 this prefix will not appear in the symbol table of the assembled object.
3384 Typically the prefix is ``L``, but targets may use other strings. ``.L`` is
3387 LLVM's support for inline asm is modeled closely on the requirements of Clang's
3388 GCC-compatible inline-asm support. Thus, the feature-set and the constraint and
3389 modifier codes listed here are similar or identical to those in GCC's inline asm
3390 support. However, to be clear, the syntax of the template and constraint strings
3391 described here is *not* the same as the syntax accepted by GCC and Clang, and,
3392 while most constraint letters are passed through as-is by Clang, some get
3393 translated to other codes when converting from the C source to the LLVM
3396 An example inline assembler expression is:
3398 .. code-block:: llvm
3400 i32 (i32) asm "bswap $0", "=r,r"
3402 Inline assembler expressions may **only** be used as the callee operand
3403 of a :ref:`call <i_call>` or an :ref:`invoke <i_invoke>` instruction.
3404 Thus, typically we have:
3406 .. code-block:: llvm
3408 %X = call i32 asm "bswap $0", "=r,r"(i32 %Y)
3410 Inline asms with side effects not visible in the constraint list must be
3411 marked as having side effects. This is done through the use of the
3412 '``sideeffect``' keyword, like so:
3414 .. code-block:: llvm
3416 call void asm sideeffect "eieio", ""()
3418 In some cases inline asms will contain code that will not work unless
3419 the stack is aligned in some way, such as calls or SSE instructions on
3420 x86, yet will not contain code that does that alignment within the asm.
3421 The compiler should make conservative assumptions about what the asm
3422 might contain and should generate its usual stack alignment code in the
3423 prologue if the '``alignstack``' keyword is present:
3425 .. code-block:: llvm
3427 call void asm alignstack "eieio", ""()
3429 Inline asms also support using non-standard assembly dialects. The
3430 assumed dialect is ATT. When the '``inteldialect``' keyword is present,
3431 the inline asm is using the Intel dialect. Currently, ATT and Intel are
3432 the only supported dialects. An example is:
3434 .. code-block:: llvm
3436 call void asm inteldialect "eieio", ""()
3438 If multiple keywords appear the '``sideeffect``' keyword must come
3439 first, the '``alignstack``' keyword second and the '``inteldialect``'
3442 Inline Asm Constraint String
3443 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3445 The constraint list is a comma-separated string, each element containing one or
3446 more constraint codes.
3448 For each element in the constraint list an appropriate register or memory
3449 operand will be chosen, and it will be made available to assembly template
3450 string expansion as ``$0`` for the first constraint in the list, ``$1`` for the
3453 There are three different types of constraints, which are distinguished by a
3454 prefix symbol in front of the constraint code: Output, Input, and Clobber. The
3455 constraints must always be given in that order: outputs first, then inputs, then
3456 clobbers. They cannot be intermingled.
3458 There are also three different categories of constraint codes:
3460 - Register constraint. This is either a register class, or a fixed physical
3461 register. This kind of constraint will allocate a register, and if necessary,
3462 bitcast the argument or result to the appropriate type.
3463 - Memory constraint. This kind of constraint is for use with an instruction
3464 taking a memory operand. Different constraints allow for different addressing
3465 modes used by the target.
3466 - Immediate value constraint. This kind of constraint is for an integer or other
3467 immediate value which can be rendered directly into an instruction. The
3468 various target-specific constraints allow the selection of a value in the
3469 proper range for the instruction you wish to use it with.
3474 Output constraints are specified by an "``=``" prefix (e.g. "``=r``"). This
3475 indicates that the assembly will write to this operand, and the operand will
3476 then be made available as a return value of the ``asm`` expression. Output
3477 constraints do not consume an argument from the call instruction. (Except, see
3478 below about indirect outputs).
3480 Normally, it is expected that no output locations are written to by the assembly
3481 expression until *all* of the inputs have been read. As such, LLVM may assign
3482 the same register to an output and an input. If this is not safe (e.g. if the
3483 assembly contains two instructions, where the first writes to one output, and
3484 the second reads an input and writes to a second output), then the "``&``"
3485 modifier must be used (e.g. "``=&r``") to specify that the output is an
3486 "early-clobber" output. Marking an output as "early-clobber" ensures that LLVM
3487 will not use the same register for any inputs (other than an input tied to this
3493 Input constraints do not have a prefix -- just the constraint codes. Each input
3494 constraint will consume one argument from the call instruction. It is not
3495 permitted for the asm to write to any input register or memory location (unless
3496 that input is tied to an output). Note also that multiple inputs may all be
3497 assigned to the same register, if LLVM can determine that they necessarily all
3498 contain the same value.
3500 Instead of providing a Constraint Code, input constraints may also "tie"
3501 themselves to an output constraint, by providing an integer as the constraint
3502 string. Tied inputs still consume an argument from the call instruction, and
3503 take up a position in the asm template numbering as is usual -- they will simply
3504 be constrained to always use the same register as the output they've been tied
3505 to. For example, a constraint string of "``=r,0``" says to assign a register for
3506 output, and use that register as an input as well (it being the 0'th
3509 It is permitted to tie an input to an "early-clobber" output. In that case, no
3510 *other* input may share the same register as the input tied to the early-clobber
3511 (even when the other input has the same value).
3513 You may only tie an input to an output which has a register constraint, not a
3514 memory constraint. Only a single input may be tied to an output.
3516 There is also an "interesting" feature which deserves a bit of explanation: if a
3517 register class constraint allocates a register which is too small for the value
3518 type operand provided as input, the input value will be split into multiple
3519 registers, and all of them passed to the inline asm.
3521 However, this feature is often not as useful as you might think.
3523 Firstly, the registers are *not* guaranteed to be consecutive. So, on those
3524 architectures that have instructions which operate on multiple consecutive
3525 instructions, this is not an appropriate way to support them. (e.g. the 32-bit
3526 SparcV8 has a 64-bit load, which instruction takes a single 32-bit register. The
3527 hardware then loads into both the named register, and the next register. This
3528 feature of inline asm would not be useful to support that.)
3530 A few of the targets provide a template string modifier allowing explicit access
3531 to the second register of a two-register operand (e.g. MIPS ``L``, ``M``, and
3532 ``D``). On such an architecture, you can actually access the second allocated
3533 register (yet, still, not any subsequent ones). But, in that case, you're still
3534 probably better off simply splitting the value into two separate operands, for
3535 clarity. (e.g. see the description of the ``A`` constraint on X86, which,
3536 despite existing only for use with this feature, is not really a good idea to
3539 Indirect inputs and outputs
3540 """""""""""""""""""""""""""
3542 Indirect output or input constraints can be specified by the "``*``" modifier
3543 (which goes after the "``=``" in case of an output). This indicates that the asm
3544 will write to or read from the contents of an *address* provided as an input
3545 argument. (Note that in this way, indirect outputs act more like an *input* than
3546 an output: just like an input, they consume an argument of the call expression,
3547 rather than producing a return value. An indirect output constraint is an
3548 "output" only in that the asm is expected to write to the contents of the input
3549 memory location, instead of just read from it).
3551 This is most typically used for memory constraint, e.g. "``=*m``", to pass the
3552 address of a variable as a value.
3554 It is also possible to use an indirect *register* constraint, but only on output
3555 (e.g. "``=*r``"). This will cause LLVM to allocate a register for an output
3556 value normally, and then, separately emit a store to the address provided as
3557 input, after the provided inline asm. (It's not clear what value this
3558 functionality provides, compared to writing the store explicitly after the asm
3559 statement, and it can only produce worse code, since it bypasses many
3560 optimization passes. I would recommend not using it.)
3566 A clobber constraint is indicated by a "``~``" prefix. A clobber does not
3567 consume an input operand, nor generate an output. Clobbers cannot use any of the
3568 general constraint code letters -- they may use only explicit register
3569 constraints, e.g. "``~{eax}``". The one exception is that a clobber string of
3570 "``~{memory}``" indicates that the assembly writes to arbitrary undeclared
3571 memory locations -- not only the memory pointed to by a declared indirect
3574 Note that clobbering named registers that are also present in output
3575 constraints is not legal.
3580 After a potential prefix comes constraint code, or codes.
3582 A Constraint Code is either a single letter (e.g. "``r``"), a "``^``" character
3583 followed by two letters (e.g. "``^wc``"), or "``{``" register-name "``}``"
3586 The one and two letter constraint codes are typically chosen to be the same as
3587 GCC's constraint codes.
3589 A single constraint may include one or more than constraint code in it, leaving
3590 it up to LLVM to choose which one to use. This is included mainly for
3591 compatibility with the translation of GCC inline asm coming from clang.
3593 There are two ways to specify alternatives, and either or both may be used in an
3594 inline asm constraint list:
3596 1) Append the codes to each other, making a constraint code set. E.g. "``im``"
3597 or "``{eax}m``". This means "choose any of the options in the set". The
3598 choice of constraint is made independently for each constraint in the
3601 2) Use "``|``" between constraint code sets, creating alternatives. Every
3602 constraint in the constraint list must have the same number of alternative
3603 sets. With this syntax, the same alternative in *all* of the items in the
3604 constraint list will be chosen together.
3606 Putting those together, you might have a two operand constraint string like
3607 ``"rm|r,ri|rm"``. This indicates that if operand 0 is ``r`` or ``m``, then
3608 operand 1 may be one of ``r`` or ``i``. If operand 0 is ``r``, then operand 1
3609 may be one of ``r`` or ``m``. But, operand 0 and 1 cannot both be of type m.
3611 However, the use of either of the alternatives features is *NOT* recommended, as
3612 LLVM is not able to make an intelligent choice about which one to use. (At the
3613 point it currently needs to choose, not enough information is available to do so
3614 in a smart way.) Thus, it simply tries to make a choice that's most likely to
3615 compile, not one that will be optimal performance. (e.g., given "``rm``", it'll
3616 always choose to use memory, not registers). And, if given multiple registers,
3617 or multiple register classes, it will simply choose the first one. (In fact, it
3618 doesn't currently even ensure explicitly specified physical registers are
3619 unique, so specifying multiple physical registers as alternatives, like
3620 ``{r11}{r12},{r11}{r12}``, will assign r11 to both operands, not at all what was
3623 Supported Constraint Code List
3624 """"""""""""""""""""""""""""""
3626 The constraint codes are, in general, expected to behave the same way they do in
3627 GCC. LLVM's support is often implemented on an 'as-needed' basis, to support C
3628 inline asm code which was supported by GCC. A mismatch in behavior between LLVM
3629 and GCC likely indicates a bug in LLVM.
3631 Some constraint codes are typically supported by all targets:
3633 - ``r``: A register in the target's general purpose register class.
3634 - ``m``: A memory address operand. It is target-specific what addressing modes
3635 are supported, typical examples are register, or register + register offset,
3636 or register + immediate offset (of some target-specific size).
3637 - ``i``: An integer constant (of target-specific width). Allows either a simple
3638 immediate, or a relocatable value.
3639 - ``n``: An integer constant -- *not* including relocatable values.
3640 - ``s``: An integer constant, but allowing *only* relocatable values.
3641 - ``X``: Allows an operand of any kind, no constraint whatsoever. Typically
3642 useful to pass a label for an asm branch or call.
3644 .. FIXME: but that surely isn't actually okay to jump out of an asm
3645 block without telling llvm about the control transfer???)
3647 - ``{register-name}``: Requires exactly the named physical register.
3649 Other constraints are target-specific:
3653 - ``z``: An immediate integer 0. Outputs ``WZR`` or ``XZR``, as appropriate.
3654 - ``I``: An immediate integer valid for an ``ADD`` or ``SUB`` instruction,
3655 i.e. 0 to 4095 with optional shift by 12.
3656 - ``J``: An immediate integer that, when negated, is valid for an ``ADD`` or
3657 ``SUB`` instruction, i.e. -1 to -4095 with optional left shift by 12.
3658 - ``K``: An immediate integer that is valid for the 'bitmask immediate 32' of a
3659 logical instruction like ``AND``, ``EOR``, or ``ORR`` with a 32-bit register.
3660 - ``L``: An immediate integer that is valid for the 'bitmask immediate 64' of a
3661 logical instruction like ``AND``, ``EOR``, or ``ORR`` with a 64-bit register.
3662 - ``M``: An immediate integer for use with the ``MOV`` assembly alias on a
3663 32-bit register. This is a superset of ``K``: in addition to the bitmask
3664 immediate, also allows immediate integers which can be loaded with a single
3665 ``MOVZ`` or ``MOVL`` instruction.
3666 - ``N``: An immediate integer for use with the ``MOV`` assembly alias on a
3667 64-bit register. This is a superset of ``L``.
3668 - ``Q``: Memory address operand must be in a single register (no
3669 offsets). (However, LLVM currently does this for the ``m`` constraint as
3671 - ``r``: A 32 or 64-bit integer register (W* or X*).
3672 - ``w``: A 32, 64, or 128-bit floating-point/SIMD register.
3673 - ``x``: A lower 128-bit floating-point/SIMD register (``V0`` to ``V15``).
3677 - ``r``: A 32 or 64-bit integer register.
3678 - ``[0-9]v``: The 32-bit VGPR register, number 0-9.
3679 - ``[0-9]s``: The 32-bit SGPR register, number 0-9.
3684 - ``Q``, ``Um``, ``Un``, ``Uq``, ``Us``, ``Ut``, ``Uv``, ``Uy``: Memory address
3685 operand. Treated the same as operand ``m``, at the moment.
3687 ARM and ARM's Thumb2 mode:
3689 - ``j``: An immediate integer between 0 and 65535 (valid for ``MOVW``)
3690 - ``I``: An immediate integer valid for a data-processing instruction.
3691 - ``J``: An immediate integer between -4095 and 4095.
3692 - ``K``: An immediate integer whose bitwise inverse is valid for a
3693 data-processing instruction. (Can be used with template modifier "``B``" to
3694 print the inverted value).
3695 - ``L``: An immediate integer whose negation is valid for a data-processing
3696 instruction. (Can be used with template modifier "``n``" to print the negated
3698 - ``M``: A power of two or a integer between 0 and 32.
3699 - ``N``: Invalid immediate constraint.
3700 - ``O``: Invalid immediate constraint.
3701 - ``r``: A general-purpose 32-bit integer register (``r0-r15``).
3702 - ``l``: In Thumb2 mode, low 32-bit GPR registers (``r0-r7``). In ARM mode, same
3704 - ``h``: In Thumb2 mode, a high 32-bit GPR register (``r8-r15``). In ARM mode,
3706 - ``w``: A 32, 64, or 128-bit floating-point/SIMD register: ``s0-s31``,
3707 ``d0-d31``, or ``q0-q15``.
3708 - ``x``: A 32, 64, or 128-bit floating-point/SIMD register: ``s0-s15``,
3709 ``d0-d7``, or ``q0-q3``.
3710 - ``t``: A low floating-point/SIMD register: ``s0-s31``, ``d0-d16``, or
3715 - ``I``: An immediate integer between 0 and 255.
3716 - ``J``: An immediate integer between -255 and -1.
3717 - ``K``: An immediate integer between 0 and 255, with optional left-shift by
3719 - ``L``: An immediate integer between -7 and 7.
3720 - ``M``: An immediate integer which is a multiple of 4 between 0 and 1020.
3721 - ``N``: An immediate integer between 0 and 31.
3722 - ``O``: An immediate integer which is a multiple of 4 between -508 and 508.
3723 - ``r``: A low 32-bit GPR register (``r0-r7``).
3724 - ``l``: A low 32-bit GPR register (``r0-r7``).
3725 - ``h``: A high GPR register (``r0-r7``).
3726 - ``w``: A 32, 64, or 128-bit floating-point/SIMD register: ``s0-s31``,
3727 ``d0-d31``, or ``q0-q15``.
3728 - ``x``: A 32, 64, or 128-bit floating-point/SIMD register: ``s0-s15``,
3729 ``d0-d7``, or ``q0-q3``.
3730 - ``t``: A low floating-point/SIMD register: ``s0-s31``, ``d0-d16``, or
3736 - ``o``, ``v``: A memory address operand, treated the same as constraint ``m``,
3738 - ``r``: A 32 or 64-bit register.
3742 - ``r``: An 8 or 16-bit register.
3746 - ``I``: An immediate signed 16-bit integer.
3747 - ``J``: An immediate integer zero.
3748 - ``K``: An immediate unsigned 16-bit integer.
3749 - ``L``: An immediate 32-bit integer, where the lower 16 bits are 0.
3750 - ``N``: An immediate integer between -65535 and -1.
3751 - ``O``: An immediate signed 15-bit integer.
3752 - ``P``: An immediate integer between 1 and 65535.
3753 - ``m``: A memory address operand. In MIPS-SE mode, allows a base address
3754 register plus 16-bit immediate offset. In MIPS mode, just a base register.
3755 - ``R``: A memory address operand. In MIPS-SE mode, allows a base address
3756 register plus a 9-bit signed offset. In MIPS mode, the same as constraint
3758 - ``ZC``: A memory address operand, suitable for use in a ``pref``, ``ll``, or
3759 ``sc`` instruction on the given subtarget (details vary).
3760 - ``r``, ``d``, ``y``: A 32 or 64-bit GPR register.
3761 - ``f``: A 32 or 64-bit FPU register (``F0-F31``), or a 128-bit MSA register
3762 (``W0-W31``). In the case of MSA registers, it is recommended to use the ``w``
3763 argument modifier for compatibility with GCC.
3764 - ``c``: A 32-bit or 64-bit GPR register suitable for indirect jump (always
3766 - ``l``: The ``lo`` register, 32 or 64-bit.
3771 - ``b``: A 1-bit integer register.
3772 - ``c`` or ``h``: A 16-bit integer register.
3773 - ``r``: A 32-bit integer register.
3774 - ``l`` or ``N``: A 64-bit integer register.
3775 - ``f``: A 32-bit float register.
3776 - ``d``: A 64-bit float register.
3781 - ``I``: An immediate signed 16-bit integer.
3782 - ``J``: An immediate unsigned 16-bit integer, shifted left 16 bits.
3783 - ``K``: An immediate unsigned 16-bit integer.
3784 - ``L``: An immediate signed 16-bit integer, shifted left 16 bits.
3785 - ``M``: An immediate integer greater than 31.
3786 - ``N``: An immediate integer that is an exact power of 2.
3787 - ``O``: The immediate integer constant 0.
3788 - ``P``: An immediate integer constant whose negation is a signed 16-bit
3790 - ``es``, ``o``, ``Q``, ``Z``, ``Zy``: A memory address operand, currently
3791 treated the same as ``m``.
3792 - ``r``: A 32 or 64-bit integer register.
3793 - ``b``: A 32 or 64-bit integer register, excluding ``R0`` (that is:
3795 - ``f``: A 32 or 64-bit float register (``F0-F31``), or when QPX is enabled, a
3796 128 or 256-bit QPX register (``Q0-Q31``; aliases the ``F`` registers).
3797 - ``v``: For ``4 x f32`` or ``4 x f64`` types, when QPX is enabled, a
3798 128 or 256-bit QPX register (``Q0-Q31``), otherwise a 128-bit
3799 altivec vector register (``V0-V31``).
3801 .. FIXME: is this a bug that v accepts QPX registers? I think this
3802 is supposed to only use the altivec vector registers?
3804 - ``y``: Condition register (``CR0-CR7``).
3805 - ``wc``: An individual CR bit in a CR register.
3806 - ``wa``, ``wd``, ``wf``: Any 128-bit VSX vector register, from the full VSX
3807 register set (overlapping both the floating-point and vector register files).
3808 - ``ws``: A 32 or 64-bit floating-point register, from the full VSX register
3813 - ``I``: An immediate 13-bit signed integer.
3814 - ``r``: A 32-bit integer register.
3815 - ``f``: Any floating-point register on SparcV8, or a floating-point
3816 register in the "low" half of the registers on SparcV9.
3817 - ``e``: Any floating-point register. (Same as ``f`` on SparcV8.)
3821 - ``I``: An immediate unsigned 8-bit integer.
3822 - ``J``: An immediate unsigned 12-bit integer.
3823 - ``K``: An immediate signed 16-bit integer.
3824 - ``L``: An immediate signed 20-bit integer.
3825 - ``M``: An immediate integer 0x7fffffff.
3826 - ``Q``: A memory address operand with a base address and a 12-bit immediate
3827 unsigned displacement.
3828 - ``R``: A memory address operand with a base address, a 12-bit immediate
3829 unsigned displacement, and an index register.
3830 - ``S``: A memory address operand with a base address and a 20-bit immediate
3831 signed displacement.
3832 - ``T``: A memory address operand with a base address, a 20-bit immediate
3833 signed displacement, and an index register.
3834 - ``r`` or ``d``: A 32, 64, or 128-bit integer register.
3835 - ``a``: A 32, 64, or 128-bit integer address register (excludes R0, which in an
3836 address context evaluates as zero).
3837 - ``h``: A 32-bit value in the high part of a 64bit data register
3839 - ``f``: A 32, 64, or 128-bit floating-point register.
3843 - ``I``: An immediate integer between 0 and 31.
3844 - ``J``: An immediate integer between 0 and 64.
3845 - ``K``: An immediate signed 8-bit integer.
3846 - ``L``: An immediate integer, 0xff or 0xffff or (in 64-bit mode only)
3848 - ``M``: An immediate integer between 0 and 3.
3849 - ``N``: An immediate unsigned 8-bit integer.
3850 - ``O``: An immediate integer between 0 and 127.
3851 - ``e``: An immediate 32-bit signed integer.
3852 - ``Z``: An immediate 32-bit unsigned integer.
3853 - ``o``, ``v``: Treated the same as ``m``, at the moment.
3854 - ``q``: An 8, 16, 32, or 64-bit register which can be accessed as an 8-bit
3855 ``l`` integer register. On X86-32, this is the ``a``, ``b``, ``c``, and ``d``
3856 registers, and on X86-64, it is all of the integer registers.
3857 - ``Q``: An 8, 16, 32, or 64-bit register which can be accessed as an 8-bit
3858 ``h`` integer register. This is the ``a``, ``b``, ``c``, and ``d`` registers.
3859 - ``r`` or ``l``: An 8, 16, 32, or 64-bit integer register.
3860 - ``R``: An 8, 16, 32, or 64-bit "legacy" integer register -- one which has
3861 existed since i386, and can be accessed without the REX prefix.
3862 - ``f``: A 32, 64, or 80-bit '387 FPU stack pseudo-register.
3863 - ``y``: A 64-bit MMX register, if MMX is enabled.
3864 - ``x``: If SSE is enabled: a 32 or 64-bit scalar operand, or 128-bit vector
3865 operand in a SSE register. If AVX is also enabled, can also be a 256-bit
3866 vector operand in an AVX register. If AVX-512 is also enabled, can also be a
3867 512-bit vector operand in an AVX512 register, Otherwise, an error.
3868 - ``Y``: The same as ``x``, if *SSE2* is enabled, otherwise an error.
3869 - ``A``: Special case: allocates EAX first, then EDX, for a single operand (in
3870 32-bit mode, a 64-bit integer operand will get split into two registers). It
3871 is not recommended to use this constraint, as in 64-bit mode, the 64-bit
3872 operand will get allocated only to RAX -- if two 32-bit operands are needed,
3873 you're better off splitting it yourself, before passing it to the asm
3878 - ``r``: A 32-bit integer register.
3881 .. _inline-asm-modifiers:
3883 Asm template argument modifiers
3884 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3886 In the asm template string, modifiers can be used on the operand reference, like
3889 The modifiers are, in general, expected to behave the same way they do in
3890 GCC. LLVM's support is often implemented on an 'as-needed' basis, to support C
3891 inline asm code which was supported by GCC. A mismatch in behavior between LLVM
3892 and GCC likely indicates a bug in LLVM.
3896 - ``c``: Print an immediate integer constant unadorned, without
3897 the target-specific immediate punctuation (e.g. no ``$`` prefix).
3898 - ``n``: Negate and print immediate integer constant unadorned, without the
3899 target-specific immediate punctuation (e.g. no ``$`` prefix).
3900 - ``l``: Print as an unadorned label, without the target-specific label
3901 punctuation (e.g. no ``$`` prefix).
3905 - ``w``: Print a GPR register with a ``w*`` name instead of ``x*`` name. E.g.,
3906 instead of ``x30``, print ``w30``.
3907 - ``x``: Print a GPR register with a ``x*`` name. (this is the default, anyhow).
3908 - ``b``, ``h``, ``s``, ``d``, ``q``: Print a floating-point/SIMD register with a
3909 ``b*``, ``h*``, ``s*``, ``d*``, or ``q*`` name, rather than the default of
3918 - ``a``: Print an operand as an address (with ``[`` and ``]`` surrounding a
3922 - ``y``: Print a VFP single-precision register as an indexed double (e.g. print
3923 as ``d4[1]`` instead of ``s9``)
3924 - ``B``: Bitwise invert and print an immediate integer constant without ``#``
3926 - ``L``: Print the low 16-bits of an immediate integer constant.
3927 - ``M``: Print as a register set suitable for ldm/stm. Also prints *all*
3928 register operands subsequent to the specified one (!), so use carefully.
3929 - ``Q``: Print the low-order register of a register-pair, or the low-order
3930 register of a two-register operand.
3931 - ``R``: Print the high-order register of a register-pair, or the high-order
3932 register of a two-register operand.
3933 - ``H``: Print the second register of a register-pair. (On a big-endian system,
3934 ``H`` is equivalent to ``Q``, and on little-endian system, ``H`` is equivalent
3937 .. FIXME: H doesn't currently support printing the second register
3938 of a two-register operand.
3940 - ``e``: Print the low doubleword register of a NEON quad register.
3941 - ``f``: Print the high doubleword register of a NEON quad register.
3942 - ``m``: Print the base register of a memory operand without the ``[`` and ``]``
3947 - ``L``: Print the second register of a two-register operand. Requires that it
3948 has been allocated consecutively to the first.
3950 .. FIXME: why is it restricted to consecutive ones? And there's
3951 nothing that ensures that happens, is there?
3953 - ``I``: Print the letter 'i' if the operand is an integer constant, otherwise
3954 nothing. Used to print 'addi' vs 'add' instructions.
3958 No additional modifiers.
3962 - ``X``: Print an immediate integer as hexadecimal
3963 - ``x``: Print the low 16 bits of an immediate integer as hexadecimal.
3964 - ``d``: Print an immediate integer as decimal.
3965 - ``m``: Subtract one and print an immediate integer as decimal.
3966 - ``z``: Print $0 if an immediate zero, otherwise print normally.
3967 - ``L``: Print the low-order register of a two-register operand, or prints the
3968 address of the low-order word of a double-word memory operand.
3970 .. FIXME: L seems to be missing memory operand support.
3972 - ``M``: Print the high-order register of a two-register operand, or prints the
3973 address of the high-order word of a double-word memory operand.
3975 .. FIXME: M seems to be missing memory operand support.
3977 - ``D``: Print the second register of a two-register operand, or prints the
3978 second word of a double-word memory operand. (On a big-endian system, ``D`` is
3979 equivalent to ``L``, and on little-endian system, ``D`` is equivalent to
3981 - ``w``: No effect. Provided for compatibility with GCC which requires this
3982 modifier in order to print MSA registers (``W0-W31``) with the ``f``
3991 - ``L``: Print the second register of a two-register operand. Requires that it
3992 has been allocated consecutively to the first.
3994 .. FIXME: why is it restricted to consecutive ones? And there's
3995 nothing that ensures that happens, is there?
3997 - ``I``: Print the letter 'i' if the operand is an integer constant, otherwise
3998 nothing. Used to print 'addi' vs 'add' instructions.
3999 - ``y``: For a memory operand, prints formatter for a two-register X-form
4000 instruction. (Currently always prints ``r0,OPERAND``).
4001 - ``U``: Prints 'u' if the memory operand is an update form, and nothing
4002 otherwise. (NOTE: LLVM does not support update form, so this will currently
4003 always print nothing)
4004 - ``X``: Prints 'x' if the memory operand is an indexed form. (NOTE: LLVM does
4005 not support indexed form, so this will currently always print nothing)
4013 SystemZ implements only ``n``, and does *not* support any of the other
4014 target-independent modifiers.
4018 - ``c``: Print an unadorned integer or symbol name. (The latter is
4019 target-specific behavior for this typically target-independent modifier).
4020 - ``A``: Print a register name with a '``*``' before it.
4021 - ``b``: Print an 8-bit register name (e.g. ``al``); do nothing on a memory
4023 - ``h``: Print the upper 8-bit register name (e.g. ``ah``); do nothing on a
4025 - ``w``: Print the 16-bit register name (e.g. ``ax``); do nothing on a memory
4027 - ``k``: Print the 32-bit register name (e.g. ``eax``); do nothing on a memory
4029 - ``q``: Print the 64-bit register name (e.g. ``rax``), if 64-bit registers are
4030 available, otherwise the 32-bit register name; do nothing on a memory operand.
4031 - ``n``: Negate and print an unadorned integer, or, for operands other than an
4032 immediate integer (e.g. a relocatable symbol expression), print a '-' before
4033 the operand. (The behavior for relocatable symbol expressions is a
4034 target-specific behavior for this typically target-independent modifier)
4035 - ``H``: Print a memory reference with additional offset +8.
4036 - ``P``: Print a memory reference or operand for use as the argument of a call
4037 instruction. (E.g. omit ``(rip)``, even though it's PC-relative.)
4041 No additional modifiers.
4047 The call instructions that wrap inline asm nodes may have a
4048 "``!srcloc``" MDNode attached to it that contains a list of constant
4049 integers. If present, the code generator will use the integer as the
4050 location cookie value when report errors through the ``LLVMContext``
4051 error reporting mechanisms. This allows a front-end to correlate backend
4052 errors that occur with inline asm back to the source code that produced
4055 .. code-block:: llvm
4057 call void asm sideeffect "something bad", ""(), !srcloc !42
4059 !42 = !{ i32 1234567 }
4061 It is up to the front-end to make sense of the magic numbers it places
4062 in the IR. If the MDNode contains multiple constants, the code generator
4063 will use the one that corresponds to the line of the asm that the error
4071 LLVM IR allows metadata to be attached to instructions in the program
4072 that can convey extra information about the code to the optimizers and
4073 code generator. One example application of metadata is source-level
4074 debug information. There are two metadata primitives: strings and nodes.
4076 Metadata does not have a type, and is not a value. If referenced from a
4077 ``call`` instruction, it uses the ``metadata`` type.
4079 All metadata are identified in syntax by a exclamation point ('``!``').
4081 .. _metadata-string:
4083 Metadata Nodes and Metadata Strings
4084 -----------------------------------
4086 A metadata string is a string surrounded by double quotes. It can
4087 contain any character by escaping non-printable characters with
4088 "``\xx``" where "``xx``" is the two digit hex code. For example:
4091 Metadata nodes are represented with notation similar to structure
4092 constants (a comma separated list of elements, surrounded by braces and
4093 preceded by an exclamation point). Metadata nodes can have any values as
4094 their operand. For example:
4096 .. code-block:: llvm
4098 !{ !"test\00", i32 10}
4100 Metadata nodes that aren't uniqued use the ``distinct`` keyword. For example:
4102 .. code-block:: text
4104 !0 = distinct !{!"test\00", i32 10}
4106 ``distinct`` nodes are useful when nodes shouldn't be merged based on their
4107 content. They can also occur when transformations cause uniquing collisions
4108 when metadata operands change.
4110 A :ref:`named metadata <namedmetadatastructure>` is a collection of
4111 metadata nodes, which can be looked up in the module symbol table. For
4114 .. code-block:: llvm
4118 Metadata can be used as function arguments. Here the ``llvm.dbg.value``
4119 intrinsic is using three metadata arguments:
4121 .. code-block:: llvm
4123 call void @llvm.dbg.value(metadata !24, metadata !25, metadata !26)
4125 Metadata can be attached to an instruction. Here metadata ``!21`` is attached
4126 to the ``add`` instruction using the ``!dbg`` identifier:
4128 .. code-block:: llvm
4130 %indvar.next = add i64 %indvar, 1, !dbg !21
4132 Metadata can also be attached to a function or a global variable. Here metadata
4133 ``!22`` is attached to the ``f1`` and ``f2 functions, and the globals ``g1``
4134 and ``g2`` using the ``!dbg`` identifier:
4136 .. code-block:: llvm
4138 declare !dbg !22 void @f1()
4139 define void @f2() !dbg !22 {
4143 @g1 = global i32 0, !dbg !22
4144 @g2 = external global i32, !dbg !22
4146 A transformation is required to drop any metadata attachment that it does not
4147 know or know it can't preserve. Currently there is an exception for metadata
4148 attachment to globals for ``!type`` and ``!absolute_symbol`` which can't be
4149 unconditionally dropped unless the global is itself deleted.
4151 Metadata attached to a module using named metadata may not be dropped, with
4152 the exception of debug metadata (named metadata with the name ``!llvm.dbg.*``).
4154 More information about specific metadata nodes recognized by the
4155 optimizers and code generator is found below.
4157 .. _specialized-metadata:
4159 Specialized Metadata Nodes
4160 ^^^^^^^^^^^^^^^^^^^^^^^^^^
4162 Specialized metadata nodes are custom data structures in metadata (as opposed
4163 to generic tuples). Their fields are labelled, and can be specified in any
4166 These aren't inherently debug info centric, but currently all the specialized
4167 metadata nodes are related to debug info.
4174 ``DICompileUnit`` nodes represent a compile unit. The ``enums:``,
4175 ``retainedTypes:``, ``globals:``, ``imports:`` and ``macros:`` fields are tuples
4176 containing the debug info to be emitted along with the compile unit, regardless
4177 of code optimizations (some nodes are only emitted if there are references to
4178 them from instructions). The ``debugInfoForProfiling:`` field is a boolean
4179 indicating whether or not line-table discriminators are updated to provide
4180 more-accurate debug info for profiling results.
4182 .. code-block:: text
4184 !0 = !DICompileUnit(language: DW_LANG_C99, file: !1, producer: "clang",
4185 isOptimized: true, flags: "-O2", runtimeVersion: 2,
4186 splitDebugFilename: "abc.debug", emissionKind: FullDebug,
4187 enums: !2, retainedTypes: !3, globals: !4, imports: !5,
4188 macros: !6, dwoId: 0x0abcd)
4190 Compile unit descriptors provide the root scope for objects declared in a
4191 specific compilation unit. File descriptors are defined using this scope. These
4192 descriptors are collected by a named metadata node ``!llvm.dbg.cu``. They keep
4193 track of global variables, type information, and imported entities (declarations
4201 ``DIFile`` nodes represent files. The ``filename:`` can include slashes.
4203 .. code-block:: none
4205 !0 = !DIFile(filename: "path/to/file", directory: "/path/to/dir",
4206 checksumkind: CSK_MD5,
4207 checksum: "000102030405060708090a0b0c0d0e0f")
4209 Files are sometimes used in ``scope:`` fields, and are the only valid target
4210 for ``file:`` fields.
4211 Valid values for ``checksumkind:`` field are: {CSK_None, CSK_MD5, CSK_SHA1}
4218 ``DIBasicType`` nodes represent primitive types, such as ``int``, ``bool`` and
4219 ``float``. ``tag:`` defaults to ``DW_TAG_base_type``.
4221 .. code-block:: text
4223 !0 = !DIBasicType(name: "unsigned char", size: 8, align: 8,
4224 encoding: DW_ATE_unsigned_char)
4225 !1 = !DIBasicType(tag: DW_TAG_unspecified_type, name: "decltype(nullptr)")
4227 The ``encoding:`` describes the details of the type. Usually it's one of the
4230 .. code-block:: text
4236 DW_ATE_signed_char = 6
4238 DW_ATE_unsigned_char = 8
4240 .. _DISubroutineType:
4245 ``DISubroutineType`` nodes represent subroutine types. Their ``types:`` field
4246 refers to a tuple; the first operand is the return type, while the rest are the
4247 types of the formal arguments in order. If the first operand is ``null``, that
4248 represents a function with no return value (such as ``void foo() {}`` in C++).
4250 .. code-block:: text
4252 !0 = !BasicType(name: "int", size: 32, align: 32, DW_ATE_signed)
4253 !1 = !BasicType(name: "char", size: 8, align: 8, DW_ATE_signed_char)
4254 !2 = !DISubroutineType(types: !{null, !0, !1}) ; void (int, char)
4261 ``DIDerivedType`` nodes represent types derived from other types, such as
4264 .. code-block:: text
4266 !0 = !DIBasicType(name: "unsigned char", size: 8, align: 8,
4267 encoding: DW_ATE_unsigned_char)
4268 !1 = !DIDerivedType(tag: DW_TAG_pointer_type, baseType: !0, size: 32,
4271 The following ``tag:`` values are valid:
4273 .. code-block:: text
4276 DW_TAG_pointer_type = 15
4277 DW_TAG_reference_type = 16
4279 DW_TAG_inheritance = 28
4280 DW_TAG_ptr_to_member_type = 31
4281 DW_TAG_const_type = 38
4283 DW_TAG_volatile_type = 53
4284 DW_TAG_restrict_type = 55
4285 DW_TAG_atomic_type = 71
4287 .. _DIDerivedTypeMember:
4289 ``DW_TAG_member`` is used to define a member of a :ref:`composite type
4290 <DICompositeType>`. The type of the member is the ``baseType:``. The
4291 ``offset:`` is the member's bit offset. If the composite type has an ODR
4292 ``identifier:`` and does not set ``flags: DIFwdDecl``, then the member is
4293 uniqued based only on its ``name:`` and ``scope:``.
4295 ``DW_TAG_inheritance`` and ``DW_TAG_friend`` are used in the ``elements:``
4296 field of :ref:`composite types <DICompositeType>` to describe parents and
4299 ``DW_TAG_typedef`` is used to provide a name for the ``baseType:``.
4301 ``DW_TAG_pointer_type``, ``DW_TAG_reference_type``, ``DW_TAG_const_type``,
4302 ``DW_TAG_volatile_type``, ``DW_TAG_restrict_type`` and ``DW_TAG_atomic_type``
4303 are used to qualify the ``baseType:``.
4305 Note that the ``void *`` type is expressed as a type derived from NULL.
4307 .. _DICompositeType:
4312 ``DICompositeType`` nodes represent types composed of other types, like
4313 structures and unions. ``elements:`` points to a tuple of the composed types.
4315 If the source language supports ODR, the ``identifier:`` field gives the unique
4316 identifier used for type merging between modules. When specified,
4317 :ref:`subprogram declarations <DISubprogramDeclaration>` and :ref:`member
4318 derived types <DIDerivedTypeMember>` that reference the ODR-type in their
4319 ``scope:`` change uniquing rules.
4321 For a given ``identifier:``, there should only be a single composite type that
4322 does not have ``flags: DIFlagFwdDecl`` set. LLVM tools that link modules
4323 together will unique such definitions at parse time via the ``identifier:``
4324 field, even if the nodes are ``distinct``.
4326 .. code-block:: text
4328 !0 = !DIEnumerator(name: "SixKind", value: 7)
4329 !1 = !DIEnumerator(name: "SevenKind", value: 7)
4330 !2 = !DIEnumerator(name: "NegEightKind", value: -8)
4331 !3 = !DICompositeType(tag: DW_TAG_enumeration_type, name: "Enum", file: !12,
4332 line: 2, size: 32, align: 32, identifier: "_M4Enum",
4333 elements: !{!0, !1, !2})
4335 The following ``tag:`` values are valid:
4337 .. code-block:: text
4339 DW_TAG_array_type = 1
4340 DW_TAG_class_type = 2
4341 DW_TAG_enumeration_type = 4
4342 DW_TAG_structure_type = 19
4343 DW_TAG_union_type = 23
4345 For ``DW_TAG_array_type``, the ``elements:`` should be :ref:`subrange
4346 descriptors <DISubrange>`, each representing the range of subscripts at that
4347 level of indexing. The ``DIFlagVector`` flag to ``flags:`` indicates that an
4348 array type is a native packed vector.
4350 For ``DW_TAG_enumeration_type``, the ``elements:`` should be :ref:`enumerator
4351 descriptors <DIEnumerator>`, each representing the definition of an enumeration
4352 value for the set. All enumeration type descriptors are collected in the
4353 ``enums:`` field of the :ref:`compile unit <DICompileUnit>`.
4355 For ``DW_TAG_structure_type``, ``DW_TAG_class_type``, and
4356 ``DW_TAG_union_type``, the ``elements:`` should be :ref:`derived types
4357 <DIDerivedType>` with ``tag: DW_TAG_member``, ``tag: DW_TAG_inheritance``, or
4358 ``tag: DW_TAG_friend``; or :ref:`subprograms <DISubprogram>` with
4359 ``isDefinition: false``.
4366 ``DISubrange`` nodes are the elements for ``DW_TAG_array_type`` variants of
4367 :ref:`DICompositeType`.
4369 - ``count: -1`` indicates an empty array.
4370 - ``count: !9`` describes the count with a :ref:`DILocalVariable`.
4371 - ``count: !11`` describes the count with a :ref:`DIGlobalVariable`.
4373 .. code-block:: llvm
4375 !0 = !DISubrange(count: 5, lowerBound: 0) ; array counting from 0
4376 !1 = !DISubrange(count: 5, lowerBound: 1) ; array counting from 1
4377 !2 = !DISubrange(count: -1) ; empty array.
4379 ; Scopes used in rest of example
4380 !6 = !DIFile(filename: "vla.c", directory: "/path/to/file")
4381 !7 = distinct !DICompileUnit(language: DW_LANG_C99, ...
4382 !8 = distinct !DISubprogram(name: "foo", scope: !7, file: !6, line: 5, ...
4384 ; Use of local variable as count value
4385 !9 = !DIBasicType(name: "int", size: 32, encoding: DW_ATE_signed)
4386 !10 = !DILocalVariable(name: "count", scope: !8, file: !6, line: 42, type: !9)
4387 !11 = !DISubrange(count !10, lowerBound: 0)
4389 ; Use of global variable as count value
4390 !12 = !DIGlobalVariable(name: "count", scope: !8, file: !6, line: 22, type: !9)
4391 !13 = !DISubrange(count !12, lowerBound: 0)
4398 ``DIEnumerator`` nodes are the elements for ``DW_TAG_enumeration_type``
4399 variants of :ref:`DICompositeType`.
4401 .. code-block:: llvm
4403 !0 = !DIEnumerator(name: "SixKind", value: 7)
4404 !1 = !DIEnumerator(name: "SevenKind", value: 7)
4405 !2 = !DIEnumerator(name: "NegEightKind", value: -8)
4407 DITemplateTypeParameter
4408 """""""""""""""""""""""
4410 ``DITemplateTypeParameter`` nodes represent type parameters to generic source
4411 language constructs. They are used (optionally) in :ref:`DICompositeType` and
4412 :ref:`DISubprogram` ``templateParams:`` fields.
4414 .. code-block:: llvm
4416 !0 = !DITemplateTypeParameter(name: "Ty", type: !1)
4418 DITemplateValueParameter
4419 """"""""""""""""""""""""
4421 ``DITemplateValueParameter`` nodes represent value parameters to generic source
4422 language constructs. ``tag:`` defaults to ``DW_TAG_template_value_parameter``,
4423 but if specified can also be set to ``DW_TAG_GNU_template_template_param`` or
4424 ``DW_TAG_GNU_template_param_pack``. They are used (optionally) in
4425 :ref:`DICompositeType` and :ref:`DISubprogram` ``templateParams:`` fields.
4427 .. code-block:: llvm
4429 !0 = !DITemplateValueParameter(name: "Ty", type: !1, value: i32 7)
4434 ``DINamespace`` nodes represent namespaces in the source language.
4436 .. code-block:: llvm
4438 !0 = !DINamespace(name: "myawesomeproject", scope: !1, file: !2, line: 7)
4440 .. _DIGlobalVariable:
4445 ``DIGlobalVariable`` nodes represent global variables in the source language.
4447 .. code-block:: llvm
4449 !0 = !DIGlobalVariable(name: "foo", linkageName: "foo", scope: !1,
4450 file: !2, line: 7, type: !3, isLocal: true,
4451 isDefinition: false, variable: i32* @foo,
4454 All global variables should be referenced by the `globals:` field of a
4455 :ref:`compile unit <DICompileUnit>`.
4462 ``DISubprogram`` nodes represent functions from the source language. A
4463 ``DISubprogram`` may be attached to a function definition using ``!dbg``
4464 metadata. The ``variables:`` field points at :ref:`variables <DILocalVariable>`
4465 that must be retained, even if their IR counterparts are optimized out of
4466 the IR. The ``type:`` field must point at an :ref:`DISubroutineType`.
4468 .. _DISubprogramDeclaration:
4470 When ``isDefinition: false``, subprograms describe a declaration in the type
4471 tree as opposed to a definition of a function. If the scope is a composite
4472 type with an ODR ``identifier:`` and that does not set ``flags: DIFwdDecl``,
4473 then the subprogram declaration is uniqued based only on its ``linkageName:``
4476 .. code-block:: text
4478 define void @_Z3foov() !dbg !0 {
4482 !0 = distinct !DISubprogram(name: "foo", linkageName: "_Zfoov", scope: !1,
4483 file: !2, line: 7, type: !3, isLocal: true,
4484 isDefinition: true, scopeLine: 8,
4486 virtuality: DW_VIRTUALITY_pure_virtual,
4487 virtualIndex: 10, flags: DIFlagPrototyped,
4488 isOptimized: true, unit: !5, templateParams: !6,
4489 declaration: !7, variables: !8, thrownTypes: !9)
4496 ``DILexicalBlock`` nodes describe nested blocks within a :ref:`subprogram
4497 <DISubprogram>`. The line number and column numbers are used to distinguish
4498 two lexical blocks at same depth. They are valid targets for ``scope:``
4501 .. code-block:: text
4503 !0 = distinct !DILexicalBlock(scope: !1, file: !2, line: 7, column: 35)
4505 Usually lexical blocks are ``distinct`` to prevent node merging based on
4508 .. _DILexicalBlockFile:
4513 ``DILexicalBlockFile`` nodes are used to discriminate between sections of a
4514 :ref:`lexical block <DILexicalBlock>`. The ``file:`` field can be changed to
4515 indicate textual inclusion, or the ``discriminator:`` field can be used to
4516 discriminate between control flow within a single block in the source language.
4518 .. code-block:: llvm
4520 !0 = !DILexicalBlock(scope: !3, file: !4, line: 7, column: 35)
4521 !1 = !DILexicalBlockFile(scope: !0, file: !4, discriminator: 0)
4522 !2 = !DILexicalBlockFile(scope: !0, file: !4, discriminator: 1)
4529 ``DILocation`` nodes represent source debug locations. The ``scope:`` field is
4530 mandatory, and points at an :ref:`DILexicalBlockFile`, an
4531 :ref:`DILexicalBlock`, or an :ref:`DISubprogram`.
4533 .. code-block:: llvm
4535 !0 = !DILocation(line: 2900, column: 42, scope: !1, inlinedAt: !2)
4537 .. _DILocalVariable:
4542 ``DILocalVariable`` nodes represent local variables in the source language. If
4543 the ``arg:`` field is set to non-zero, then this variable is a subprogram
4544 parameter, and it will be included in the ``variables:`` field of its
4545 :ref:`DISubprogram`.
4547 .. code-block:: text
4549 !0 = !DILocalVariable(name: "this", arg: 1, scope: !3, file: !2, line: 7,
4550 type: !3, flags: DIFlagArtificial)
4551 !1 = !DILocalVariable(name: "x", arg: 2, scope: !4, file: !2, line: 7,
4553 !2 = !DILocalVariable(name: "y", scope: !5, file: !2, line: 7, type: !3)
4558 ``DIExpression`` nodes represent expressions that are inspired by the DWARF
4559 expression language. They are used in :ref:`debug intrinsics<dbg_intrinsics>`
4560 (such as ``llvm.dbg.declare`` and ``llvm.dbg.value``) to describe how the
4561 referenced LLVM variable relates to the source language variable.
4563 The current supported vocabulary is limited:
4565 - ``DW_OP_deref`` dereferences the top of the expression stack.
4566 - ``DW_OP_plus`` pops the last two entries from the expression stack, adds
4567 them together and appends the result to the expression stack.
4568 - ``DW_OP_minus`` pops the last two entries from the expression stack, subtracts
4569 the last entry from the second last entry and appends the result to the
4571 - ``DW_OP_plus_uconst, 93`` adds ``93`` to the working expression.
4572 - ``DW_OP_LLVM_fragment, 16, 8`` specifies the offset and size (``16`` and ``8``
4573 here, respectively) of the variable fragment from the working expression. Note
4574 that contrary to DW_OP_bit_piece, the offset is describing the location
4575 within the described source variable.
4576 - ``DW_OP_swap`` swaps top two stack entries.
4577 - ``DW_OP_xderef`` provides extended dereference mechanism. The entry at the top
4578 of the stack is treated as an address. The second stack entry is treated as an
4579 address space identifier.
4580 - ``DW_OP_stack_value`` marks a constant value.
4582 DWARF specifies three kinds of simple location descriptions: Register, memory,
4583 and implicit location descriptions. Register and memory location descriptions
4584 describe the *location* of a source variable (in the sense that a debugger might
4585 modify its value), whereas implicit locations describe merely the *value* of a
4586 source variable. DIExpressions also follow this model: A DIExpression that
4587 doesn't have a trailing ``DW_OP_stack_value`` will describe an *address* when
4588 combined with a concrete location.
4590 .. code-block:: text
4592 !0 = !DIExpression(DW_OP_deref)
4593 !1 = !DIExpression(DW_OP_plus_uconst, 3)
4594 !1 = !DIExpression(DW_OP_constu, 3, DW_OP_plus)
4595 !2 = !DIExpression(DW_OP_bit_piece, 3, 7)
4596 !3 = !DIExpression(DW_OP_deref, DW_OP_constu, 3, DW_OP_plus, DW_OP_LLVM_fragment, 3, 7)
4597 !4 = !DIExpression(DW_OP_constu, 2, DW_OP_swap, DW_OP_xderef)
4598 !5 = !DIExpression(DW_OP_constu, 42, DW_OP_stack_value)
4603 ``DIObjCProperty`` nodes represent Objective-C property nodes.
4605 .. code-block:: llvm
4607 !3 = !DIObjCProperty(name: "foo", file: !1, line: 7, setter: "setFoo",
4608 getter: "getFoo", attributes: 7, type: !2)
4613 ``DIImportedEntity`` nodes represent entities (such as modules) imported into a
4616 .. code-block:: text
4618 !2 = !DIImportedEntity(tag: DW_TAG_imported_module, name: "foo", scope: !0,
4619 entity: !1, line: 7)
4624 ``DIMacro`` nodes represent definition or undefinition of a macro identifiers.
4625 The ``name:`` field is the macro identifier, followed by macro parameters when
4626 defining a function-like macro, and the ``value`` field is the token-string
4627 used to expand the macro identifier.
4629 .. code-block:: text
4631 !2 = !DIMacro(macinfo: DW_MACINFO_define, line: 7, name: "foo(x)",
4633 !3 = !DIMacro(macinfo: DW_MACINFO_undef, line: 30, name: "foo")
4638 ``DIMacroFile`` nodes represent inclusion of source files.
4639 The ``nodes:`` field is a list of ``DIMacro`` and ``DIMacroFile`` nodes that
4640 appear in the included source file.
4642 .. code-block:: text
4644 !2 = !DIMacroFile(macinfo: DW_MACINFO_start_file, line: 7, file: !2,
4650 In LLVM IR, memory does not have types, so LLVM's own type system is not
4651 suitable for doing type based alias analysis (TBAA). Instead, metadata is
4652 added to the IR to describe a type system of a higher level language. This
4653 can be used to implement C/C++ strict type aliasing rules, but it can also
4654 be used to implement custom alias analysis behavior for other languages.
4656 This description of LLVM's TBAA system is broken into two parts:
4657 :ref:`Semantics<tbaa_node_semantics>` talks about high level issues, and
4658 :ref:`Representation<tbaa_node_representation>` talks about the metadata
4659 encoding of various entities.
4661 It is always possible to trace any TBAA node to a "root" TBAA node (details
4662 in the :ref:`Representation<tbaa_node_representation>` section). TBAA
4663 nodes with different roots have an unknown aliasing relationship, and LLVM
4664 conservatively infers ``MayAlias`` between them. The rules mentioned in
4665 this section only pertain to TBAA nodes living under the same root.
4667 .. _tbaa_node_semantics:
4672 The TBAA metadata system, referred to as "struct path TBAA" (not to be
4673 confused with ``tbaa.struct``), consists of the following high level
4674 concepts: *Type Descriptors*, further subdivided into scalar type
4675 descriptors and struct type descriptors; and *Access Tags*.
4677 **Type descriptors** describe the type system of the higher level language
4678 being compiled. **Scalar type descriptors** describe types that do not
4679 contain other types. Each scalar type has a parent type, which must also
4680 be a scalar type or the TBAA root. Via this parent relation, scalar types
4681 within a TBAA root form a tree. **Struct type descriptors** denote types
4682 that contain a sequence of other type descriptors, at known offsets. These
4683 contained type descriptors can either be struct type descriptors themselves
4684 or scalar type descriptors.
4686 **Access tags** are metadata nodes attached to load and store instructions.
4687 Access tags use type descriptors to describe the *location* being accessed
4688 in terms of the type system of the higher level language. Access tags are
4689 tuples consisting of a base type, an access type and an offset. The base
4690 type is a scalar type descriptor or a struct type descriptor, the access
4691 type is a scalar type descriptor, and the offset is a constant integer.
4693 The access tag ``(BaseTy, AccessTy, Offset)`` can describe one of two
4696 * If ``BaseTy`` is a struct type, the tag describes a memory access (load
4697 or store) of a value of type ``AccessTy`` contained in the struct type
4698 ``BaseTy`` at offset ``Offset``.
4700 * If ``BaseTy`` is a scalar type, ``Offset`` must be 0 and ``BaseTy`` and
4701 ``AccessTy`` must be the same; and the access tag describes a scalar
4702 access with scalar type ``AccessTy``.
4704 We first define an ``ImmediateParent`` relation on ``(BaseTy, Offset)``
4707 * If ``BaseTy`` is a scalar type then ``ImmediateParent(BaseTy, 0)`` is
4708 ``(ParentTy, 0)`` where ``ParentTy`` is the parent of the scalar type as
4709 described in the TBAA metadata. ``ImmediateParent(BaseTy, Offset)`` is
4710 undefined if ``Offset`` is non-zero.
4712 * If ``BaseTy`` is a struct type then ``ImmediateParent(BaseTy, Offset)``
4713 is ``(NewTy, NewOffset)`` where ``NewTy`` is the type contained in
4714 ``BaseTy`` at offset ``Offset`` and ``NewOffset`` is ``Offset`` adjusted
4715 to be relative within that inner type.
4717 A memory access with an access tag ``(BaseTy1, AccessTy1, Offset1)``
4718 aliases a memory access with an access tag ``(BaseTy2, AccessTy2,
4719 Offset2)`` if either ``(BaseTy1, Offset1)`` is reachable from ``(Base2,
4720 Offset2)`` via the ``Parent`` relation or vice versa.
4722 As a concrete example, the type descriptor graph for the following program
4728 float f; // offset 4
4732 float f; // offset 0
4733 double d; // offset 4
4734 struct Inner inner_a; // offset 12
4737 void f(struct Outer* outer, struct Inner* inner, float* f, int* i, char* c) {
4738 outer->f = 0; // tag0: (OuterStructTy, FloatScalarTy, 0)
4739 outer->inner_a.i = 0; // tag1: (OuterStructTy, IntScalarTy, 12)
4740 outer->inner_a.f = 0.0; // tag2: (OuterStructTy, IntScalarTy, 16)
4741 *f = 0.0; // tag3: (FloatScalarTy, FloatScalarTy, 0)
4744 is (note that in C and C++, ``char`` can be used to access any arbitrary
4747 .. code-block:: text
4750 CharScalarTy = ("char", Root, 0)
4751 FloatScalarTy = ("float", CharScalarTy, 0)
4752 DoubleScalarTy = ("double", CharScalarTy, 0)
4753 IntScalarTy = ("int", CharScalarTy, 0)
4754 InnerStructTy = {"Inner" (IntScalarTy, 0), (FloatScalarTy, 4)}
4755 OuterStructTy = {"Outer", (FloatScalarTy, 0), (DoubleScalarTy, 4),
4756 (InnerStructTy, 12)}
4759 with (e.g.) ``ImmediateParent(OuterStructTy, 12)`` = ``(InnerStructTy,
4760 0)``, ``ImmediateParent(InnerStructTy, 0)`` = ``(IntScalarTy, 0)``, and
4761 ``ImmediateParent(IntScalarTy, 0)`` = ``(CharScalarTy, 0)``.
4763 .. _tbaa_node_representation:
4768 The root node of a TBAA type hierarchy is an ``MDNode`` with 0 operands or
4769 with exactly one ``MDString`` operand.
4771 Scalar type descriptors are represented as an ``MDNode`` s with two
4772 operands. The first operand is an ``MDString`` denoting the name of the
4773 struct type. LLVM does not assign meaning to the value of this operand, it
4774 only cares about it being an ``MDString``. The second operand is an
4775 ``MDNode`` which points to the parent for said scalar type descriptor,
4776 which is either another scalar type descriptor or the TBAA root. Scalar
4777 type descriptors can have an optional third argument, but that must be the
4778 constant integer zero.
4780 Struct type descriptors are represented as ``MDNode`` s with an odd number
4781 of operands greater than 1. The first operand is an ``MDString`` denoting
4782 the name of the struct type. Like in scalar type descriptors the actual
4783 value of this name operand is irrelevant to LLVM. After the name operand,
4784 the struct type descriptors have a sequence of alternating ``MDNode`` and
4785 ``ConstantInt`` operands. With N starting from 1, the 2N - 1 th operand,
4786 an ``MDNode``, denotes a contained field, and the 2N th operand, a
4787 ``ConstantInt``, is the offset of the said contained field. The offsets
4788 must be in non-decreasing order.
4790 Access tags are represented as ``MDNode`` s with either 3 or 4 operands.
4791 The first operand is an ``MDNode`` pointing to the node representing the
4792 base type. The second operand is an ``MDNode`` pointing to the node
4793 representing the access type. The third operand is a ``ConstantInt`` that
4794 states the offset of the access. If a fourth field is present, it must be
4795 a ``ConstantInt`` valued at 0 or 1. If it is 1 then the access tag states
4796 that the location being accessed is "constant" (meaning
4797 ``pointsToConstantMemory`` should return true; see `other useful
4798 AliasAnalysis methods <AliasAnalysis.html#OtherItfs>`_). The TBAA root of
4799 the access type and the base type of an access tag must be the same, and
4800 that is the TBAA root of the access tag.
4802 '``tbaa.struct``' Metadata
4803 ^^^^^^^^^^^^^^^^^^^^^^^^^^
4805 The :ref:`llvm.memcpy <int_memcpy>` is often used to implement
4806 aggregate assignment operations in C and similar languages, however it
4807 is defined to copy a contiguous region of memory, which is more than
4808 strictly necessary for aggregate types which contain holes due to
4809 padding. Also, it doesn't contain any TBAA information about the fields
4812 ``!tbaa.struct`` metadata can describe which memory subregions in a
4813 memcpy are padding and what the TBAA tags of the struct are.
4815 The current metadata format is very simple. ``!tbaa.struct`` metadata
4816 nodes are a list of operands which are in conceptual groups of three.
4817 For each group of three, the first operand gives the byte offset of a
4818 field in bytes, the second gives its size in bytes, and the third gives
4821 .. code-block:: llvm
4823 !4 = !{ i64 0, i64 4, !1, i64 8, i64 4, !2 }
4825 This describes a struct with two fields. The first is at offset 0 bytes
4826 with size 4 bytes, and has tbaa tag !1. The second is at offset 8 bytes
4827 and has size 4 bytes and has tbaa tag !2.
4829 Note that the fields need not be contiguous. In this example, there is a
4830 4 byte gap between the two fields. This gap represents padding which
4831 does not carry useful data and need not be preserved.
4833 '``noalias``' and '``alias.scope``' Metadata
4834 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4836 ``noalias`` and ``alias.scope`` metadata provide the ability to specify generic
4837 noalias memory-access sets. This means that some collection of memory access
4838 instructions (loads, stores, memory-accessing calls, etc.) that carry
4839 ``noalias`` metadata can specifically be specified not to alias with some other
4840 collection of memory access instructions that carry ``alias.scope`` metadata.
4841 Each type of metadata specifies a list of scopes where each scope has an id and
4844 When evaluating an aliasing query, if for some domain, the set
4845 of scopes with that domain in one instruction's ``alias.scope`` list is a
4846 subset of (or equal to) the set of scopes for that domain in another
4847 instruction's ``noalias`` list, then the two memory accesses are assumed not to
4850 Because scopes in one domain don't affect scopes in other domains, separate
4851 domains can be used to compose multiple independent noalias sets. This is
4852 used for example during inlining. As the noalias function parameters are
4853 turned into noalias scope metadata, a new domain is used every time the
4854 function is inlined.
4856 The metadata identifying each domain is itself a list containing one or two
4857 entries. The first entry is the name of the domain. Note that if the name is a
4858 string then it can be combined across functions and translation units. A
4859 self-reference can be used to create globally unique domain names. A
4860 descriptive string may optionally be provided as a second list entry.
4862 The metadata identifying each scope is also itself a list containing two or
4863 three entries. The first entry is the name of the scope. Note that if the name
4864 is a string then it can be combined across functions and translation units. A
4865 self-reference can be used to create globally unique scope names. A metadata
4866 reference to the scope's domain is the second entry. A descriptive string may
4867 optionally be provided as a third list entry.
4871 .. code-block:: llvm
4873 ; Two scope domains:
4877 ; Some scopes in these domains:
4883 !5 = !{!4} ; A list containing only scope !4
4887 ; These two instructions don't alias:
4888 %0 = load float, float* %c, align 4, !alias.scope !5
4889 store float %0, float* %arrayidx.i, align 4, !noalias !5
4891 ; These two instructions also don't alias (for domain !1, the set of scopes
4892 ; in the !alias.scope equals that in the !noalias list):
4893 %2 = load float, float* %c, align 4, !alias.scope !5
4894 store float %2, float* %arrayidx.i2, align 4, !noalias !6
4896 ; These two instructions may alias (for domain !0, the set of scopes in
4897 ; the !noalias list is not a superset of, or equal to, the scopes in the
4898 ; !alias.scope list):
4899 %2 = load float, float* %c, align 4, !alias.scope !6
4900 store float %0, float* %arrayidx.i, align 4, !noalias !7
4902 '``fpmath``' Metadata
4903 ^^^^^^^^^^^^^^^^^^^^^
4905 ``fpmath`` metadata may be attached to any instruction of floating-point
4906 type. It can be used to express the maximum acceptable error in the
4907 result of that instruction, in ULPs, thus potentially allowing the
4908 compiler to use a more efficient but less accurate method of computing
4909 it. ULP is defined as follows:
4911 If ``x`` is a real number that lies between two finite consecutive
4912 floating-point numbers ``a`` and ``b``, without being equal to one
4913 of them, then ``ulp(x) = |b - a|``, otherwise ``ulp(x)`` is the
4914 distance between the two non-equal finite floating-point numbers
4915 nearest ``x``. Moreover, ``ulp(NaN)`` is ``NaN``.
4917 The metadata node shall consist of a single positive float type number
4918 representing the maximum relative error, for example:
4920 .. code-block:: llvm
4922 !0 = !{ float 2.5 } ; maximum acceptable inaccuracy is 2.5 ULPs
4926 '``range``' Metadata
4927 ^^^^^^^^^^^^^^^^^^^^
4929 ``range`` metadata may be attached only to ``load``, ``call`` and ``invoke`` of
4930 integer types. It expresses the possible ranges the loaded value or the value
4931 returned by the called function at this call site is in. The ranges are
4932 represented with a flattened list of integers. The loaded value or the value
4933 returned is known to be in the union of the ranges defined by each consecutive
4934 pair. Each pair has the following properties:
4936 - The type must match the type loaded by the instruction.
4937 - The pair ``a,b`` represents the range ``[a,b)``.
4938 - Both ``a`` and ``b`` are constants.
4939 - The range is allowed to wrap.
4940 - The range should not represent the full or empty set. That is,
4943 In addition, the pairs must be in signed order of the lower bound and
4944 they must be non-contiguous.
4948 .. code-block:: llvm
4950 %a = load i8, i8* %x, align 1, !range !0 ; Can only be 0 or 1
4951 %b = load i8, i8* %y, align 1, !range !1 ; Can only be 255 (-1), 0 or 1
4952 %c = call i8 @foo(), !range !2 ; Can only be 0, 1, 3, 4 or 5
4953 %d = invoke i8 @bar() to label %cont
4954 unwind label %lpad, !range !3 ; Can only be -2, -1, 3, 4 or 5
4956 !0 = !{ i8 0, i8 2 }
4957 !1 = !{ i8 255, i8 2 }
4958 !2 = !{ i8 0, i8 2, i8 3, i8 6 }
4959 !3 = !{ i8 -2, i8 0, i8 3, i8 6 }
4961 '``absolute_symbol``' Metadata
4962 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4964 ``absolute_symbol`` metadata may be attached to a global variable
4965 declaration. It marks the declaration as a reference to an absolute symbol,
4966 which causes the backend to use absolute relocations for the symbol even
4967 in position independent code, and expresses the possible ranges that the
4968 global variable's *address* (not its value) is in, in the same format as
4969 ``range`` metadata, with the extension that the pair ``all-ones,all-ones``
4970 may be used to represent the full set.
4972 Example (assuming 64-bit pointers):
4974 .. code-block:: llvm
4976 @a = external global i8, !absolute_symbol !0 ; Absolute symbol in range [0,256)
4977 @b = external global i8, !absolute_symbol !1 ; Absolute symbol in range [0,2^64)
4980 !0 = !{ i64 0, i64 256 }
4981 !1 = !{ i64 -1, i64 -1 }
4983 '``callees``' Metadata
4984 ^^^^^^^^^^^^^^^^^^^^^^
4986 ``callees`` metadata may be attached to indirect call sites. If ``callees``
4987 metadata is attached to a call site, and any callee is not among the set of
4988 functions provided by the metadata, the behavior is undefined. The intent of
4989 this metadata is to facilitate optimizations such as indirect-call promotion.
4990 For example, in the code below, the call instruction may only target the
4991 ``add`` or ``sub`` functions:
4993 .. code-block:: llvm
4995 %result = call i64 %binop(i64 %x, i64 %y), !callees !0
4998 !0 = !{i64 (i64, i64)* @add, i64 (i64, i64)* @sub}
5000 '``unpredictable``' Metadata
5001 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5003 ``unpredictable`` metadata may be attached to any branch or switch
5004 instruction. It can be used to express the unpredictability of control
5005 flow. Similar to the llvm.expect intrinsic, it may be used to alter
5006 optimizations related to compare and branch instructions. The metadata
5007 is treated as a boolean value; if it exists, it signals that the branch
5008 or switch that it is attached to is completely unpredictable.
5013 It is sometimes useful to attach information to loop constructs. Currently,
5014 loop metadata is implemented as metadata attached to the branch instruction
5015 in the loop latch block. This type of metadata refer to a metadata node that is
5016 guaranteed to be separate for each loop. The loop identifier metadata is
5017 specified with the name ``llvm.loop``.
5019 The loop identifier metadata is implemented using a metadata that refers to
5020 itself to avoid merging it with any other identifier metadata, e.g.,
5021 during module linkage or function inlining. That is, each loop should refer
5022 to their own identification metadata even if they reside in separate functions.
5023 The following example contains loop identifier metadata for two separate loop
5026 .. code-block:: llvm
5031 The loop identifier metadata can be used to specify additional
5032 per-loop metadata. Any operands after the first operand can be treated
5033 as user-defined metadata. For example the ``llvm.loop.unroll.count``
5034 suggests an unroll factor to the loop unroller:
5036 .. code-block:: llvm
5038 br i1 %exitcond, label %._crit_edge, label %.lr.ph, !llvm.loop !0
5041 !1 = !{!"llvm.loop.unroll.count", i32 4}
5043 '``llvm.loop.vectorize``' and '``llvm.loop.interleave``'
5044 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5046 Metadata prefixed with ``llvm.loop.vectorize`` or ``llvm.loop.interleave`` are
5047 used to control per-loop vectorization and interleaving parameters such as
5048 vectorization width and interleave count. These metadata should be used in
5049 conjunction with ``llvm.loop`` loop identification metadata. The
5050 ``llvm.loop.vectorize`` and ``llvm.loop.interleave`` metadata are only
5051 optimization hints and the optimizer will only interleave and vectorize loops if
5052 it believes it is safe to do so. The ``llvm.mem.parallel_loop_access`` metadata
5053 which contains information about loop-carried memory dependencies can be helpful
5054 in determining the safety of these transformations.
5056 '``llvm.loop.interleave.count``' Metadata
5057 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5059 This metadata suggests an interleave count to the loop interleaver.
5060 The first operand is the string ``llvm.loop.interleave.count`` and the
5061 second operand is an integer specifying the interleave count. For
5064 .. code-block:: llvm
5066 !0 = !{!"llvm.loop.interleave.count", i32 4}
5068 Note that setting ``llvm.loop.interleave.count`` to 1 disables interleaving
5069 multiple iterations of the loop. If ``llvm.loop.interleave.count`` is set to 0
5070 then the interleave count will be determined automatically.
5072 '``llvm.loop.vectorize.enable``' Metadata
5073 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5075 This metadata selectively enables or disables vectorization for the loop. The
5076 first operand is the string ``llvm.loop.vectorize.enable`` and the second operand
5077 is a bit. If the bit operand value is 1 vectorization is enabled. A value of
5078 0 disables vectorization:
5080 .. code-block:: llvm
5082 !0 = !{!"llvm.loop.vectorize.enable", i1 0}
5083 !1 = !{!"llvm.loop.vectorize.enable", i1 1}
5085 '``llvm.loop.vectorize.width``' Metadata
5086 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5088 This metadata sets the target width of the vectorizer. The first
5089 operand is the string ``llvm.loop.vectorize.width`` and the second
5090 operand is an integer specifying the width. For example:
5092 .. code-block:: llvm
5094 !0 = !{!"llvm.loop.vectorize.width", i32 4}
5096 Note that setting ``llvm.loop.vectorize.width`` to 1 disables
5097 vectorization of the loop. If ``llvm.loop.vectorize.width`` is set to
5098 0 or if the loop does not have this metadata the width will be
5099 determined automatically.
5101 '``llvm.loop.unroll``'
5102 ^^^^^^^^^^^^^^^^^^^^^^
5104 Metadata prefixed with ``llvm.loop.unroll`` are loop unrolling
5105 optimization hints such as the unroll factor. ``llvm.loop.unroll``
5106 metadata should be used in conjunction with ``llvm.loop`` loop
5107 identification metadata. The ``llvm.loop.unroll`` metadata are only
5108 optimization hints and the unrolling will only be performed if the
5109 optimizer believes it is safe to do so.
5111 '``llvm.loop.unroll.count``' Metadata
5112 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5114 This metadata suggests an unroll factor to the loop unroller. The
5115 first operand is the string ``llvm.loop.unroll.count`` and the second
5116 operand is a positive integer specifying the unroll factor. For
5119 .. code-block:: llvm
5121 !0 = !{!"llvm.loop.unroll.count", i32 4}
5123 If the trip count of the loop is less than the unroll count the loop
5124 will be partially unrolled.
5126 '``llvm.loop.unroll.disable``' Metadata
5127 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5129 This metadata disables loop unrolling. The metadata has a single operand
5130 which is the string ``llvm.loop.unroll.disable``. For example:
5132 .. code-block:: llvm
5134 !0 = !{!"llvm.loop.unroll.disable"}
5136 '``llvm.loop.unroll.runtime.disable``' Metadata
5137 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5139 This metadata disables runtime loop unrolling. The metadata has a single
5140 operand which is the string ``llvm.loop.unroll.runtime.disable``. For example:
5142 .. code-block:: llvm
5144 !0 = !{!"llvm.loop.unroll.runtime.disable"}
5146 '``llvm.loop.unroll.enable``' Metadata
5147 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5149 This metadata suggests that the loop should be fully unrolled if the trip count
5150 is known at compile time and partially unrolled if the trip count is not known
5151 at compile time. The metadata has a single operand which is the string
5152 ``llvm.loop.unroll.enable``. For example:
5154 .. code-block:: llvm
5156 !0 = !{!"llvm.loop.unroll.enable"}
5158 '``llvm.loop.unroll.full``' Metadata
5159 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5161 This metadata suggests that the loop should be unrolled fully. The
5162 metadata has a single operand which is the string ``llvm.loop.unroll.full``.
5165 .. code-block:: llvm
5167 !0 = !{!"llvm.loop.unroll.full"}
5169 '``llvm.loop.licm_versioning.disable``' Metadata
5170 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5172 This metadata indicates that the loop should not be versioned for the purpose
5173 of enabling loop-invariant code motion (LICM). The metadata has a single operand
5174 which is the string ``llvm.loop.licm_versioning.disable``. For example:
5176 .. code-block:: llvm
5178 !0 = !{!"llvm.loop.licm_versioning.disable"}
5180 '``llvm.loop.distribute.enable``' Metadata
5181 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5183 Loop distribution allows splitting a loop into multiple loops. Currently,
5184 this is only performed if the entire loop cannot be vectorized due to unsafe
5185 memory dependencies. The transformation will attempt to isolate the unsafe
5186 dependencies into their own loop.
5188 This metadata can be used to selectively enable or disable distribution of the
5189 loop. The first operand is the string ``llvm.loop.distribute.enable`` and the
5190 second operand is a bit. If the bit operand value is 1 distribution is
5191 enabled. A value of 0 disables distribution:
5193 .. code-block:: llvm
5195 !0 = !{!"llvm.loop.distribute.enable", i1 0}
5196 !1 = !{!"llvm.loop.distribute.enable", i1 1}
5198 This metadata should be used in conjunction with ``llvm.loop`` loop
5199 identification metadata.
5204 Metadata types used to annotate memory accesses with information helpful
5205 for optimizations are prefixed with ``llvm.mem``.
5207 '``llvm.mem.parallel_loop_access``' Metadata
5208 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5210 The ``llvm.mem.parallel_loop_access`` metadata refers to a loop identifier,
5211 or metadata containing a list of loop identifiers for nested loops.
5212 The metadata is attached to memory accessing instructions and denotes that
5213 no loop carried memory dependence exist between it and other instructions denoted
5214 with the same loop identifier. The metadata on memory reads also implies that
5215 if conversion (i.e. speculative execution within a loop iteration) is safe.
5217 Precisely, given two instructions ``m1`` and ``m2`` that both have the
5218 ``llvm.mem.parallel_loop_access`` metadata, with ``L1`` and ``L2`` being the
5219 set of loops associated with that metadata, respectively, then there is no loop
5220 carried dependence between ``m1`` and ``m2`` for loops in both ``L1`` and
5223 As a special case, if all memory accessing instructions in a loop have
5224 ``llvm.mem.parallel_loop_access`` metadata that refers to that loop, then the
5225 loop has no loop carried memory dependences and is considered to be a parallel
5228 Note that if not all memory access instructions have such metadata referring to
5229 the loop, then the loop is considered not being trivially parallel. Additional
5230 memory dependence analysis is required to make that determination. As a fail
5231 safe mechanism, this causes loops that were originally parallel to be considered
5232 sequential (if optimization passes that are unaware of the parallel semantics
5233 insert new memory instructions into the loop body).
5235 Example of a loop that is considered parallel due to its correct use of
5236 both ``llvm.loop`` and ``llvm.mem.parallel_loop_access``
5237 metadata types that refer to the same loop identifier metadata.
5239 .. code-block:: llvm
5243 %val0 = load i32, i32* %arrayidx, !llvm.mem.parallel_loop_access !0
5245 store i32 %val0, i32* %arrayidx1, !llvm.mem.parallel_loop_access !0
5247 br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !0
5253 It is also possible to have nested parallel loops. In that case the
5254 memory accesses refer to a list of loop identifier metadata nodes instead of
5255 the loop identifier metadata node directly:
5257 .. code-block:: llvm
5261 %val1 = load i32, i32* %arrayidx3, !llvm.mem.parallel_loop_access !2
5263 br label %inner.for.body
5267 %val0 = load i32, i32* %arrayidx1, !llvm.mem.parallel_loop_access !0
5269 store i32 %val0, i32* %arrayidx2, !llvm.mem.parallel_loop_access !0
5271 br i1 %exitcond, label %inner.for.end, label %inner.for.body, !llvm.loop !1
5275 store i32 %val1, i32* %arrayidx4, !llvm.mem.parallel_loop_access !2
5277 br i1 %exitcond, label %outer.for.end, label %outer.for.body, !llvm.loop !2
5279 outer.for.end: ; preds = %for.body
5281 !0 = !{!1, !2} ; a list of loop identifiers
5282 !1 = !{!1} ; an identifier for the inner loop
5283 !2 = !{!2} ; an identifier for the outer loop
5285 '``irr_loop``' Metadata
5286 ^^^^^^^^^^^^^^^^^^^^^^^
5288 ``irr_loop`` metadata may be attached to the terminator instruction of a basic
5289 block that's an irreducible loop header (note that an irreducible loop has more
5290 than once header basic blocks.) If ``irr_loop`` metadata is attached to the
5291 terminator instruction of a basic block that is not really an irreducible loop
5292 header, the behavior is undefined. The intent of this metadata is to improve the
5293 accuracy of the block frequency propagation. For example, in the code below, the
5294 block ``header0`` may have a loop header weight (relative to the other headers of
5295 the irreducible loop) of 100:
5297 .. code-block:: llvm
5301 br i1 %cmp, label %t1, label %t2, !irr_loop !0
5304 !0 = !{"loop_header_weight", i64 100}
5306 Irreducible loop header weights are typically based on profile data.
5308 '``invariant.group``' Metadata
5309 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5311 The ``invariant.group`` metadata may be attached to ``load``/``store`` instructions.
5312 The existence of the ``invariant.group`` metadata on the instruction tells
5313 the optimizer that every ``load`` and ``store`` to the same pointer operand
5314 within the same invariant group can be assumed to load or store the same
5315 value (but see the ``llvm.invariant.group.barrier`` intrinsic which affects
5316 when two pointers are considered the same). Pointers returned by bitcast or
5317 getelementptr with only zero indices are considered the same.
5321 .. code-block:: llvm
5323 @unknownPtr = external global i8
5326 store i8 42, i8* %ptr, !invariant.group !0
5327 call void @foo(i8* %ptr)
5329 %a = load i8, i8* %ptr, !invariant.group !0 ; Can assume that value under %ptr didn't change
5330 call void @foo(i8* %ptr)
5331 %b = load i8, i8* %ptr, !invariant.group !1 ; Can't assume anything, because group changed
5333 %newPtr = call i8* @getPointer(i8* %ptr)
5334 %c = load i8, i8* %newPtr, !invariant.group !0 ; Can't assume anything, because we only have information about %ptr
5336 %unknownValue = load i8, i8* @unknownPtr
5337 store i8 %unknownValue, i8* %ptr, !invariant.group !0 ; Can assume that %unknownValue == 42
5339 call void @foo(i8* %ptr)
5340 %newPtr2 = call i8* @llvm.invariant.group.barrier(i8* %ptr)
5341 %d = load i8, i8* %newPtr2, !invariant.group !0 ; Can't step through invariant.group.barrier to get value of %ptr
5344 declare void @foo(i8*)
5345 declare i8* @getPointer(i8*)
5346 declare i8* @llvm.invariant.group.barrier(i8*)
5348 !0 = !{!"magic ptr"}
5349 !1 = !{!"other ptr"}
5351 The invariant.group metadata must be dropped when replacing one pointer by
5352 another based on aliasing information. This is because invariant.group is tied
5353 to the SSA value of the pointer operand.
5355 .. code-block:: llvm
5357 %v = load i8, i8* %x, !invariant.group !0
5358 ; if %x mustalias %y then we can replace the above instruction with
5359 %v = load i8, i8* %y
5365 See :doc:`TypeMetadata`.
5367 '``associated``' Metadata
5368 ^^^^^^^^^^^^^^^^^^^^^^^^^
5370 The ``associated`` metadata may be attached to a global object
5371 declaration with a single argument that references another global object.
5373 This metadata prevents discarding of the global object in linker GC
5374 unless the referenced object is also discarded. The linker support for
5375 this feature is spotty. For best compatibility, globals carrying this
5378 - Be in a comdat with the referenced global.
5379 - Be in @llvm.compiler.used.
5380 - Have an explicit section with a name which is a valid C identifier.
5382 It does not have any effect on non-ELF targets.
5386 .. code-block:: text
5389 @a = global i32 1, comdat $a
5390 @b = internal global i32 2, comdat $a, section "abc", !associated !0
5397 The ``prof`` metadata is used to record profile data in the IR.
5398 The first operand of the metadata node indicates the profile metadata
5399 type. There are currently 3 types:
5400 :ref:`branch_weights<prof_node_branch_weights>`,
5401 :ref:`function_entry_count<prof_node_function_entry_count>`, and
5402 :ref:`VP<prof_node_VP>`.
5404 .. _prof_node_branch_weights:
5409 Branch weight metadata attached to a branch, select, switch or call instruction
5410 represents the likeliness of the associated branch being taken.
5411 For more information, see :doc:`BranchWeightMetadata`.
5413 .. _prof_node_function_entry_count:
5415 function_entry_count
5416 """"""""""""""""""""
5418 Function entry count metadata can be attached to function definitions
5419 to record the number of times the function is called. Used with BFI
5420 information, it is also used to derive the basic block profile count.
5421 For more information, see :doc:`BranchWeightMetadata`.
5428 VP (value profile) metadata can be attached to instructions that have
5429 value profile information. Currently this is indirect calls (where it
5430 records the hottest callees) and calls to memory intrinsics such as memcpy,
5431 memmove, and memset (where it records the hottest byte lengths).
5433 Each VP metadata node contains "VP" string, then a uint32_t value for the value
5434 profiling kind, a uint64_t value for the total number of times the instruction
5435 is executed, followed by uint64_t value and execution count pairs.
5436 The value profiling kind is 0 for indirect call targets and 1 for memory
5437 operations. For indirect call targets, each profile value is a hash
5438 of the callee function name, and for memory operations each value is the
5441 Note that the value counts do not need to add up to the total count
5442 listed in the third operand (in practice only the top hottest values
5443 are tracked and reported).
5445 Indirect call example:
5447 .. code-block:: llvm
5449 call void %f(), !prof !1
5450 !1 = !{!"VP", i32 0, i64 1600, i64 7651369219802541373, i64 1030, i64 -4377547752858689819, i64 410}
5452 Note that the VP type is 0 (the second operand), which indicates this is
5453 an indirect call value profile data. The third operand indicates that the
5454 indirect call executed 1600 times. The 4th and 6th operands give the
5455 hashes of the 2 hottest target functions' names (this is the same hash used
5456 to represent function names in the profile database), and the 5th and 7th
5457 operands give the execution count that each of the respective prior target
5458 functions was called.
5460 Module Flags Metadata
5461 =====================
5463 Information about the module as a whole is difficult to convey to LLVM's
5464 subsystems. The LLVM IR isn't sufficient to transmit this information.
5465 The ``llvm.module.flags`` named metadata exists in order to facilitate
5466 this. These flags are in the form of key / value pairs --- much like a
5467 dictionary --- making it easy for any subsystem who cares about a flag to
5470 The ``llvm.module.flags`` metadata contains a list of metadata triplets.
5471 Each triplet has the following form:
5473 - The first element is a *behavior* flag, which specifies the behavior
5474 when two (or more) modules are merged together, and it encounters two
5475 (or more) metadata with the same ID. The supported behaviors are
5477 - The second element is a metadata string that is a unique ID for the
5478 metadata. Each module may only have one flag entry for each unique ID (not
5479 including entries with the **Require** behavior).
5480 - The third element is the value of the flag.
5482 When two (or more) modules are merged together, the resulting
5483 ``llvm.module.flags`` metadata is the union of the modules' flags. That is, for
5484 each unique metadata ID string, there will be exactly one entry in the merged
5485 modules ``llvm.module.flags`` metadata table, and the value for that entry will
5486 be determined by the merge behavior flag, as described below. The only exception
5487 is that entries with the *Require* behavior are always preserved.
5489 The following behaviors are supported:
5500 Emits an error if two values disagree, otherwise the resulting value
5501 is that of the operands.
5505 Emits a warning if two values disagree. The result value will be the
5506 operand for the flag from the first module being linked.
5510 Adds a requirement that another module flag be present and have a
5511 specified value after linking is performed. The value must be a
5512 metadata pair, where the first element of the pair is the ID of the
5513 module flag to be restricted, and the second element of the pair is
5514 the value the module flag should be restricted to. This behavior can
5515 be used to restrict the allowable results (via triggering of an
5516 error) of linking IDs with the **Override** behavior.
5520 Uses the specified value, regardless of the behavior or value of the
5521 other module. If both modules specify **Override**, but the values
5522 differ, an error will be emitted.
5526 Appends the two values, which are required to be metadata nodes.
5530 Appends the two values, which are required to be metadata
5531 nodes. However, duplicate entries in the second list are dropped
5532 during the append operation.
5536 Takes the max of the two values, which are required to be integers.
5538 It is an error for a particular unique flag ID to have multiple behaviors,
5539 except in the case of **Require** (which adds restrictions on another metadata
5540 value) or **Override**.
5542 An example of module flags:
5544 .. code-block:: llvm
5546 !0 = !{ i32 1, !"foo", i32 1 }
5547 !1 = !{ i32 4, !"bar", i32 37 }
5548 !2 = !{ i32 2, !"qux", i32 42 }
5549 !3 = !{ i32 3, !"qux",
5554 !llvm.module.flags = !{ !0, !1, !2, !3 }
5556 - Metadata ``!0`` has the ID ``!"foo"`` and the value '1'. The behavior
5557 if two or more ``!"foo"`` flags are seen is to emit an error if their
5558 values are not equal.
5560 - Metadata ``!1`` has the ID ``!"bar"`` and the value '37'. The
5561 behavior if two or more ``!"bar"`` flags are seen is to use the value
5564 - Metadata ``!2`` has the ID ``!"qux"`` and the value '42'. The
5565 behavior if two or more ``!"qux"`` flags are seen is to emit a
5566 warning if their values are not equal.
5568 - Metadata ``!3`` has the ID ``!"qux"`` and the value:
5574 The behavior is to emit an error if the ``llvm.module.flags`` does not
5575 contain a flag with the ID ``!"foo"`` that has the value '1' after linking is
5578 Objective-C Garbage Collection Module Flags Metadata
5579 ----------------------------------------------------
5581 On the Mach-O platform, Objective-C stores metadata about garbage
5582 collection in a special section called "image info". The metadata
5583 consists of a version number and a bitmask specifying what types of
5584 garbage collection are supported (if any) by the file. If two or more
5585 modules are linked together their garbage collection metadata needs to
5586 be merged rather than appended together.
5588 The Objective-C garbage collection module flags metadata consists of the
5589 following key-value pairs:
5598 * - ``Objective-C Version``
5599 - **[Required]** --- The Objective-C ABI version. Valid values are 1 and 2.
5601 * - ``Objective-C Image Info Version``
5602 - **[Required]** --- The version of the image info section. Currently
5605 * - ``Objective-C Image Info Section``
5606 - **[Required]** --- The section to place the metadata. Valid values are
5607 ``"__OBJC, __image_info, regular"`` for Objective-C ABI version 1, and
5608 ``"__DATA,__objc_imageinfo, regular, no_dead_strip"`` for
5609 Objective-C ABI version 2.
5611 * - ``Objective-C Garbage Collection``
5612 - **[Required]** --- Specifies whether garbage collection is supported or
5613 not. Valid values are 0, for no garbage collection, and 2, for garbage
5614 collection supported.
5616 * - ``Objective-C GC Only``
5617 - **[Optional]** --- Specifies that only garbage collection is supported.
5618 If present, its value must be 6. This flag requires that the
5619 ``Objective-C Garbage Collection`` flag have the value 2.
5621 Some important flag interactions:
5623 - If a module with ``Objective-C Garbage Collection`` set to 0 is
5624 merged with a module with ``Objective-C Garbage Collection`` set to
5625 2, then the resulting module has the
5626 ``Objective-C Garbage Collection`` flag set to 0.
5627 - A module with ``Objective-C Garbage Collection`` set to 0 cannot be
5628 merged with a module with ``Objective-C GC Only`` set to 6.
5630 C type width Module Flags Metadata
5631 ----------------------------------
5633 The ARM backend emits a section into each generated object file describing the
5634 options that it was compiled with (in a compiler-independent way) to prevent
5635 linking incompatible objects, and to allow automatic library selection. Some
5636 of these options are not visible at the IR level, namely wchar_t width and enum
5639 To pass this information to the backend, these options are encoded in module
5640 flags metadata, using the following key-value pairs:
5650 - * 0 --- sizeof(wchar_t) == 4
5651 * 1 --- sizeof(wchar_t) == 2
5654 - * 0 --- Enums are at least as large as an ``int``.
5655 * 1 --- Enums are stored in the smallest integer type which can
5656 represent all of its values.
5658 For example, the following metadata section specifies that the module was
5659 compiled with a ``wchar_t`` width of 4 bytes, and the underlying type of an
5660 enum is the smallest type which can represent all of its values::
5662 !llvm.module.flags = !{!0, !1}
5663 !0 = !{i32 1, !"short_wchar", i32 1}
5664 !1 = !{i32 1, !"short_enum", i32 0}
5666 Automatic Linker Flags Named Metadata
5667 =====================================
5669 Some targets support embedding flags to the linker inside individual object
5670 files. Typically this is used in conjunction with language extensions which
5671 allow source files to explicitly declare the libraries they depend on, and have
5672 these automatically be transmitted to the linker via object files.
5674 These flags are encoded in the IR using named metadata with the name
5675 ``!llvm.linker.options``. Each operand is expected to be a metadata node
5676 which should be a list of other metadata nodes, each of which should be a
5677 list of metadata strings defining linker options.
5679 For example, the following metadata section specifies two separate sets of
5680 linker options, presumably to link against ``libz`` and the ``Cocoa``
5684 !1 = !{ !"-framework", !"Cocoa" } } }
5685 !llvm.linker.options = !{ !0, !1 }
5687 The metadata encoding as lists of lists of options, as opposed to a collapsed
5688 list of options, is chosen so that the IR encoding can use multiple option
5689 strings to specify e.g., a single library, while still having that specifier be
5690 preserved as an atomic element that can be recognized by a target specific
5691 assembly writer or object file emitter.
5693 Each individual option is required to be either a valid option for the target's
5694 linker, or an option that is reserved by the target specific assembly writer or
5695 object file emitter. No other aspect of these options is defined by the IR.
5697 .. _intrinsicglobalvariables:
5699 Intrinsic Global Variables
5700 ==========================
5702 LLVM has a number of "magic" global variables that contain data that
5703 affect code generation or other IR semantics. These are documented here.
5704 All globals of this sort should have a section specified as
5705 "``llvm.metadata``". This section and all globals that start with
5706 "``llvm.``" are reserved for use by LLVM.
5710 The '``llvm.used``' Global Variable
5711 -----------------------------------
5713 The ``@llvm.used`` global is an array which has
5714 :ref:`appending linkage <linkage_appending>`. This array contains a list of
5715 pointers to named global variables, functions and aliases which may optionally
5716 have a pointer cast formed of bitcast or getelementptr. For example, a legal
5719 .. code-block:: llvm
5724 @llvm.used = appending global [2 x i8*] [
5726 i8* bitcast (i32* @Y to i8*)
5727 ], section "llvm.metadata"
5729 If a symbol appears in the ``@llvm.used`` list, then the compiler, assembler,
5730 and linker are required to treat the symbol as if there is a reference to the
5731 symbol that it cannot see (which is why they have to be named). For example, if
5732 a variable has internal linkage and no references other than that from the
5733 ``@llvm.used`` list, it cannot be deleted. This is commonly used to represent
5734 references from inline asms and other things the compiler cannot "see", and
5735 corresponds to "``attribute((used))``" in GNU C.
5737 On some targets, the code generator must emit a directive to the
5738 assembler or object file to prevent the assembler and linker from
5739 molesting the symbol.
5741 .. _gv_llvmcompilerused:
5743 The '``llvm.compiler.used``' Global Variable
5744 --------------------------------------------
5746 The ``@llvm.compiler.used`` directive is the same as the ``@llvm.used``
5747 directive, except that it only prevents the compiler from touching the
5748 symbol. On targets that support it, this allows an intelligent linker to
5749 optimize references to the symbol without being impeded as it would be
5752 This is a rare construct that should only be used in rare circumstances,
5753 and should not be exposed to source languages.
5755 .. _gv_llvmglobalctors:
5757 The '``llvm.global_ctors``' Global Variable
5758 -------------------------------------------
5760 .. code-block:: llvm
5762 %0 = type { i32, void ()*, i8* }
5763 @llvm.global_ctors = appending global [1 x %0] [%0 { i32 65535, void ()* @ctor, i8* @data }]
5765 The ``@llvm.global_ctors`` array contains a list of constructor
5766 functions, priorities, and an optional associated global or function.
5767 The functions referenced by this array will be called in ascending order
5768 of priority (i.e. lowest first) when the module is loaded. The order of
5769 functions with the same priority is not defined.
5771 If the third field is present, non-null, and points to a global variable
5772 or function, the initializer function will only run if the associated
5773 data from the current module is not discarded.
5775 .. _llvmglobaldtors:
5777 The '``llvm.global_dtors``' Global Variable
5778 -------------------------------------------
5780 .. code-block:: llvm
5782 %0 = type { i32, void ()*, i8* }
5783 @llvm.global_dtors = appending global [1 x %0] [%0 { i32 65535, void ()* @dtor, i8* @data }]
5785 The ``@llvm.global_dtors`` array contains a list of destructor
5786 functions, priorities, and an optional associated global or function.
5787 The functions referenced by this array will be called in descending
5788 order of priority (i.e. highest first) when the module is unloaded. The
5789 order of functions with the same priority is not defined.
5791 If the third field is present, non-null, and points to a global variable
5792 or function, the destructor function will only run if the associated
5793 data from the current module is not discarded.
5795 Instruction Reference
5796 =====================
5798 The LLVM instruction set consists of several different classifications
5799 of instructions: :ref:`terminator instructions <terminators>`, :ref:`binary
5800 instructions <binaryops>`, :ref:`bitwise binary
5801 instructions <bitwiseops>`, :ref:`memory instructions <memoryops>`, and
5802 :ref:`other instructions <otherops>`.
5806 Terminator Instructions
5807 -----------------------
5809 As mentioned :ref:`previously <functionstructure>`, every basic block in a
5810 program ends with a "Terminator" instruction, which indicates which
5811 block should be executed after the current block is finished. These
5812 terminator instructions typically yield a '``void``' value: they produce
5813 control flow, not values (the one exception being the
5814 ':ref:`invoke <i_invoke>`' instruction).
5816 The terminator instructions are: ':ref:`ret <i_ret>`',
5817 ':ref:`br <i_br>`', ':ref:`switch <i_switch>`',
5818 ':ref:`indirectbr <i_indirectbr>`', ':ref:`invoke <i_invoke>`',
5819 ':ref:`resume <i_resume>`', ':ref:`catchswitch <i_catchswitch>`',
5820 ':ref:`catchret <i_catchret>`',
5821 ':ref:`cleanupret <i_cleanupret>`',
5822 and ':ref:`unreachable <i_unreachable>`'.
5826 '``ret``' Instruction
5827 ^^^^^^^^^^^^^^^^^^^^^
5834 ret <type> <value> ; Return a value from a non-void function
5835 ret void ; Return from void function
5840 The '``ret``' instruction is used to return control flow (and optionally
5841 a value) from a function back to the caller.
5843 There are two forms of the '``ret``' instruction: one that returns a
5844 value and then causes control flow, and one that just causes control
5850 The '``ret``' instruction optionally accepts a single argument, the
5851 return value. The type of the return value must be a ':ref:`first
5852 class <t_firstclass>`' type.
5854 A function is not :ref:`well formed <wellformed>` if it it has a non-void
5855 return type and contains a '``ret``' instruction with no return value or
5856 a return value with a type that does not match its type, or if it has a
5857 void return type and contains a '``ret``' instruction with a return
5863 When the '``ret``' instruction is executed, control flow returns back to
5864 the calling function's context. If the caller is a
5865 ":ref:`call <i_call>`" instruction, execution continues at the
5866 instruction after the call. If the caller was an
5867 ":ref:`invoke <i_invoke>`" instruction, execution continues at the
5868 beginning of the "normal" destination block. If the instruction returns
5869 a value, that value shall set the call or invoke instruction's return
5875 .. code-block:: llvm
5877 ret i32 5 ; Return an integer value of 5
5878 ret void ; Return from a void function
5879 ret { i32, i8 } { i32 4, i8 2 } ; Return a struct of values 4 and 2
5883 '``br``' Instruction
5884 ^^^^^^^^^^^^^^^^^^^^
5891 br i1 <cond>, label <iftrue>, label <iffalse>
5892 br label <dest> ; Unconditional branch
5897 The '``br``' instruction is used to cause control flow to transfer to a
5898 different basic block in the current function. There are two forms of
5899 this instruction, corresponding to a conditional branch and an
5900 unconditional branch.
5905 The conditional branch form of the '``br``' instruction takes a single
5906 '``i1``' value and two '``label``' values. The unconditional form of the
5907 '``br``' instruction takes a single '``label``' value as a target.
5912 Upon execution of a conditional '``br``' instruction, the '``i1``'
5913 argument is evaluated. If the value is ``true``, control flows to the
5914 '``iftrue``' ``label`` argument. If "cond" is ``false``, control flows
5915 to the '``iffalse``' ``label`` argument.
5920 .. code-block:: llvm
5923 %cond = icmp eq i32 %a, %b
5924 br i1 %cond, label %IfEqual, label %IfUnequal
5932 '``switch``' Instruction
5933 ^^^^^^^^^^^^^^^^^^^^^^^^
5940 switch <intty> <value>, label <defaultdest> [ <intty> <val>, label <dest> ... ]
5945 The '``switch``' instruction is used to transfer control flow to one of
5946 several different places. It is a generalization of the '``br``'
5947 instruction, allowing a branch to occur to one of many possible
5953 The '``switch``' instruction uses three parameters: an integer
5954 comparison value '``value``', a default '``label``' destination, and an
5955 array of pairs of comparison value constants and '``label``'s. The table
5956 is not allowed to contain duplicate constant entries.
5961 The ``switch`` instruction specifies a table of values and destinations.
5962 When the '``switch``' instruction is executed, this table is searched
5963 for the given value. If the value is found, control flow is transferred
5964 to the corresponding destination; otherwise, control flow is transferred
5965 to the default destination.
5970 Depending on properties of the target machine and the particular
5971 ``switch`` instruction, this instruction may be code generated in
5972 different ways. For example, it could be generated as a series of
5973 chained conditional branches or with a lookup table.
5978 .. code-block:: llvm
5980 ; Emulate a conditional br instruction
5981 %Val = zext i1 %value to i32
5982 switch i32 %Val, label %truedest [ i32 0, label %falsedest ]
5984 ; Emulate an unconditional br instruction
5985 switch i32 0, label %dest [ ]
5987 ; Implement a jump table:
5988 switch i32 %val, label %otherwise [ i32 0, label %onzero
5990 i32 2, label %ontwo ]
5994 '``indirectbr``' Instruction
5995 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6002 indirectbr <somety>* <address>, [ label <dest1>, label <dest2>, ... ]
6007 The '``indirectbr``' instruction implements an indirect branch to a
6008 label within the current function, whose address is specified by
6009 "``address``". Address must be derived from a
6010 :ref:`blockaddress <blockaddress>` constant.
6015 The '``address``' argument is the address of the label to jump to. The
6016 rest of the arguments indicate the full set of possible destinations
6017 that the address may point to. Blocks are allowed to occur multiple
6018 times in the destination list, though this isn't particularly useful.
6020 This destination list is required so that dataflow analysis has an
6021 accurate understanding of the CFG.
6026 Control transfers to the block specified in the address argument. All
6027 possible destination blocks must be listed in the label list, otherwise
6028 this instruction has undefined behavior. This implies that jumps to
6029 labels defined in other functions have undefined behavior as well.
6034 This is typically implemented with a jump through a register.
6039 .. code-block:: llvm
6041 indirectbr i8* %Addr, [ label %bb1, label %bb2, label %bb3 ]
6045 '``invoke``' Instruction
6046 ^^^^^^^^^^^^^^^^^^^^^^^^
6053 <result> = invoke [cconv] [ret attrs] <ty>|<fnty> <fnptrval>(<function args>) [fn attrs]
6054 [operand bundles] to label <normal label> unwind label <exception label>
6059 The '``invoke``' instruction causes control to transfer to a specified
6060 function, with the possibility of control flow transfer to either the
6061 '``normal``' label or the '``exception``' label. If the callee function
6062 returns with the "``ret``" instruction, control flow will return to the
6063 "normal" label. If the callee (or any indirect callees) returns via the
6064 ":ref:`resume <i_resume>`" instruction or other exception handling
6065 mechanism, control is interrupted and continued at the dynamically
6066 nearest "exception" label.
6068 The '``exception``' label is a `landing
6069 pad <ExceptionHandling.html#overview>`_ for the exception. As such,
6070 '``exception``' label is required to have the
6071 ":ref:`landingpad <i_landingpad>`" instruction, which contains the
6072 information about the behavior of the program after unwinding happens,
6073 as its first non-PHI instruction. The restrictions on the
6074 "``landingpad``" instruction's tightly couples it to the "``invoke``"
6075 instruction, so that the important information contained within the
6076 "``landingpad``" instruction can't be lost through normal code motion.
6081 This instruction requires several arguments:
6083 #. The optional "cconv" marker indicates which :ref:`calling
6084 convention <callingconv>` the call should use. If none is
6085 specified, the call defaults to using C calling conventions.
6086 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
6087 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
6089 #. '``ty``': the type of the call instruction itself which is also the
6090 type of the return value. Functions that return no value are marked
6092 #. '``fnty``': shall be the signature of the function being invoked. The
6093 argument types must match the types implied by this signature. This
6094 type can be omitted if the function is not varargs.
6095 #. '``fnptrval``': An LLVM value containing a pointer to a function to
6096 be invoked. In most cases, this is a direct function invocation, but
6097 indirect ``invoke``'s are just as possible, calling an arbitrary pointer
6099 #. '``function args``': argument list whose types match the function
6100 signature argument types and parameter attributes. All arguments must
6101 be of :ref:`first class <t_firstclass>` type. If the function signature
6102 indicates the function accepts a variable number of arguments, the
6103 extra arguments can be specified.
6104 #. '``normal label``': the label reached when the called function
6105 executes a '``ret``' instruction.
6106 #. '``exception label``': the label reached when a callee returns via
6107 the :ref:`resume <i_resume>` instruction or other exception handling
6109 #. The optional :ref:`function attributes <fnattrs>` list.
6110 #. The optional :ref:`operand bundles <opbundles>` list.
6115 This instruction is designed to operate as a standard '``call``'
6116 instruction in most regards. The primary difference is that it
6117 establishes an association with a label, which is used by the runtime
6118 library to unwind the stack.
6120 This instruction is used in languages with destructors to ensure that
6121 proper cleanup is performed in the case of either a ``longjmp`` or a
6122 thrown exception. Additionally, this is important for implementation of
6123 '``catch``' clauses in high-level languages that support them.
6125 For the purposes of the SSA form, the definition of the value returned
6126 by the '``invoke``' instruction is deemed to occur on the edge from the
6127 current block to the "normal" label. If the callee unwinds then no
6128 return value is available.
6133 .. code-block:: llvm
6135 %retval = invoke i32 @Test(i32 15) to label %Continue
6136 unwind label %TestCleanup ; i32:retval set
6137 %retval = invoke coldcc i32 %Testfnptr(i32 15) to label %Continue
6138 unwind label %TestCleanup ; i32:retval set
6142 '``resume``' Instruction
6143 ^^^^^^^^^^^^^^^^^^^^^^^^
6150 resume <type> <value>
6155 The '``resume``' instruction is a terminator instruction that has no
6161 The '``resume``' instruction requires one argument, which must have the
6162 same type as the result of any '``landingpad``' instruction in the same
6168 The '``resume``' instruction resumes propagation of an existing
6169 (in-flight) exception whose unwinding was interrupted with a
6170 :ref:`landingpad <i_landingpad>` instruction.
6175 .. code-block:: llvm
6177 resume { i8*, i32 } %exn
6181 '``catchswitch``' Instruction
6182 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6189 <resultval> = catchswitch within <parent> [ label <handler1>, label <handler2>, ... ] unwind to caller
6190 <resultval> = catchswitch within <parent> [ label <handler1>, label <handler2>, ... ] unwind label <default>
6195 The '``catchswitch``' instruction is used by `LLVM's exception handling system
6196 <ExceptionHandling.html#overview>`_ to describe the set of possible catch handlers
6197 that may be executed by the :ref:`EH personality routine <personalityfn>`.
6202 The ``parent`` argument is the token of the funclet that contains the
6203 ``catchswitch`` instruction. If the ``catchswitch`` is not inside a funclet,
6204 this operand may be the token ``none``.
6206 The ``default`` argument is the label of another basic block beginning with
6207 either a ``cleanuppad`` or ``catchswitch`` instruction. This unwind destination
6208 must be a legal target with respect to the ``parent`` links, as described in
6209 the `exception handling documentation\ <ExceptionHandling.html#wineh-constraints>`_.
6211 The ``handlers`` are a nonempty list of successor blocks that each begin with a
6212 :ref:`catchpad <i_catchpad>` instruction.
6217 Executing this instruction transfers control to one of the successors in
6218 ``handlers``, if appropriate, or continues to unwind via the unwind label if
6221 The ``catchswitch`` is both a terminator and a "pad" instruction, meaning that
6222 it must be both the first non-phi instruction and last instruction in the basic
6223 block. Therefore, it must be the only non-phi instruction in the block.
6228 .. code-block:: text
6231 %cs1 = catchswitch within none [label %handler0, label %handler1] unwind to caller
6233 %cs2 = catchswitch within %parenthandler [label %handler0] unwind label %cleanup
6237 '``catchret``' Instruction
6238 ^^^^^^^^^^^^^^^^^^^^^^^^^^
6245 catchret from <token> to label <normal>
6250 The '``catchret``' instruction is a terminator instruction that has a
6257 The first argument to a '``catchret``' indicates which ``catchpad`` it
6258 exits. It must be a :ref:`catchpad <i_catchpad>`.
6259 The second argument to a '``catchret``' specifies where control will
6265 The '``catchret``' instruction ends an existing (in-flight) exception whose
6266 unwinding was interrupted with a :ref:`catchpad <i_catchpad>` instruction. The
6267 :ref:`personality function <personalityfn>` gets a chance to execute arbitrary
6268 code to, for example, destroy the active exception. Control then transfers to
6271 The ``token`` argument must be a token produced by a ``catchpad`` instruction.
6272 If the specified ``catchpad`` is not the most-recently-entered not-yet-exited
6273 funclet pad (as described in the `EH documentation\ <ExceptionHandling.html#wineh-constraints>`_),
6274 the ``catchret``'s behavior is undefined.
6279 .. code-block:: text
6281 catchret from %catch label %continue
6285 '``cleanupret``' Instruction
6286 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6293 cleanupret from <value> unwind label <continue>
6294 cleanupret from <value> unwind to caller
6299 The '``cleanupret``' instruction is a terminator instruction that has
6300 an optional successor.
6306 The '``cleanupret``' instruction requires one argument, which indicates
6307 which ``cleanuppad`` it exits, and must be a :ref:`cleanuppad <i_cleanuppad>`.
6308 If the specified ``cleanuppad`` is not the most-recently-entered not-yet-exited
6309 funclet pad (as described in the `EH documentation\ <ExceptionHandling.html#wineh-constraints>`_),
6310 the ``cleanupret``'s behavior is undefined.
6312 The '``cleanupret``' instruction also has an optional successor, ``continue``,
6313 which must be the label of another basic block beginning with either a
6314 ``cleanuppad`` or ``catchswitch`` instruction. This unwind destination must
6315 be a legal target with respect to the ``parent`` links, as described in the
6316 `exception handling documentation\ <ExceptionHandling.html#wineh-constraints>`_.
6321 The '``cleanupret``' instruction indicates to the
6322 :ref:`personality function <personalityfn>` that one
6323 :ref:`cleanuppad <i_cleanuppad>` it transferred control to has ended.
6324 It transfers control to ``continue`` or unwinds out of the function.
6329 .. code-block:: text
6331 cleanupret from %cleanup unwind to caller
6332 cleanupret from %cleanup unwind label %continue
6336 '``unreachable``' Instruction
6337 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6349 The '``unreachable``' instruction has no defined semantics. This
6350 instruction is used to inform the optimizer that a particular portion of
6351 the code is not reachable. This can be used to indicate that the code
6352 after a no-return function cannot be reached, and other facts.
6357 The '``unreachable``' instruction has no defined semantics.
6364 Binary operators are used to do most of the computation in a program.
6365 They require two operands of the same type, execute an operation on
6366 them, and produce a single value. The operands might represent multiple
6367 data, as is the case with the :ref:`vector <t_vector>` data type. The
6368 result value has the same type as its operands.
6370 There are several different binary operators:
6374 '``add``' Instruction
6375 ^^^^^^^^^^^^^^^^^^^^^
6382 <result> = add <ty> <op1>, <op2> ; yields ty:result
6383 <result> = add nuw <ty> <op1>, <op2> ; yields ty:result
6384 <result> = add nsw <ty> <op1>, <op2> ; yields ty:result
6385 <result> = add nuw nsw <ty> <op1>, <op2> ; yields ty:result
6390 The '``add``' instruction returns the sum of its two operands.
6395 The two arguments to the '``add``' instruction must be
6396 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6397 arguments must have identical types.
6402 The value produced is the integer sum of the two operands.
6404 If the sum has unsigned overflow, the result returned is the
6405 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
6408 Because LLVM integers use a two's complement representation, this
6409 instruction is appropriate for both signed and unsigned integers.
6411 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
6412 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
6413 result value of the ``add`` is a :ref:`poison value <poisonvalues>` if
6414 unsigned and/or signed overflow, respectively, occurs.
6419 .. code-block:: text
6421 <result> = add i32 4, %var ; yields i32:result = 4 + %var
6425 '``fadd``' Instruction
6426 ^^^^^^^^^^^^^^^^^^^^^^
6433 <result> = fadd [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
6438 The '``fadd``' instruction returns the sum of its two operands.
6443 The two arguments to the '``fadd``' instruction must be
6444 :ref:`floating-point <t_floating>` or :ref:`vector <t_vector>` of
6445 floating-point values. Both arguments must have identical types.
6450 The value produced is the floating-point sum of the two operands.
6451 This instruction is assumed to execute in the default :ref:`floating-point
6452 environment <floatenv>`.
6453 This instruction can also take any number of :ref:`fast-math
6454 flags <fastmath>`, which are optimization hints to enable otherwise
6455 unsafe floating-point optimizations:
6460 .. code-block:: text
6462 <result> = fadd float 4.0, %var ; yields float:result = 4.0 + %var
6464 '``sub``' Instruction
6465 ^^^^^^^^^^^^^^^^^^^^^
6472 <result> = sub <ty> <op1>, <op2> ; yields ty:result
6473 <result> = sub nuw <ty> <op1>, <op2> ; yields ty:result
6474 <result> = sub nsw <ty> <op1>, <op2> ; yields ty:result
6475 <result> = sub nuw nsw <ty> <op1>, <op2> ; yields ty:result
6480 The '``sub``' instruction returns the difference of its two operands.
6482 Note that the '``sub``' instruction is used to represent the '``neg``'
6483 instruction present in most other intermediate representations.
6488 The two arguments to the '``sub``' instruction must be
6489 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6490 arguments must have identical types.
6495 The value produced is the integer difference of the two operands.
6497 If the difference has unsigned overflow, the result returned is the
6498 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
6501 Because LLVM integers use a two's complement representation, this
6502 instruction is appropriate for both signed and unsigned integers.
6504 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
6505 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
6506 result value of the ``sub`` is a :ref:`poison value <poisonvalues>` if
6507 unsigned and/or signed overflow, respectively, occurs.
6512 .. code-block:: text
6514 <result> = sub i32 4, %var ; yields i32:result = 4 - %var
6515 <result> = sub i32 0, %val ; yields i32:result = -%var
6519 '``fsub``' Instruction
6520 ^^^^^^^^^^^^^^^^^^^^^^
6527 <result> = fsub [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
6532 The '``fsub``' instruction returns the difference of its two operands.
6534 Note that the '``fsub``' instruction is used to represent the '``fneg``'
6535 instruction present in most other intermediate representations.
6540 The two arguments to the '``fsub``' instruction must be
6541 :ref:`floating-point <t_floating>` or :ref:`vector <t_vector>` of
6542 floating-point values. Both arguments must have identical types.
6547 The value produced is the floating-point difference of the two operands.
6548 This instruction is assumed to execute in the default :ref:`floating-point
6549 environment <floatenv>`.
6550 This instruction can also take any number of :ref:`fast-math
6551 flags <fastmath>`, which are optimization hints to enable otherwise
6552 unsafe floating-point optimizations:
6557 .. code-block:: text
6559 <result> = fsub float 4.0, %var ; yields float:result = 4.0 - %var
6560 <result> = fsub float -0.0, %val ; yields float:result = -%var
6562 '``mul``' Instruction
6563 ^^^^^^^^^^^^^^^^^^^^^
6570 <result> = mul <ty> <op1>, <op2> ; yields ty:result
6571 <result> = mul nuw <ty> <op1>, <op2> ; yields ty:result
6572 <result> = mul nsw <ty> <op1>, <op2> ; yields ty:result
6573 <result> = mul nuw nsw <ty> <op1>, <op2> ; yields ty:result
6578 The '``mul``' instruction returns the product of its two operands.
6583 The two arguments to the '``mul``' instruction must be
6584 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6585 arguments must have identical types.
6590 The value produced is the integer product of the two operands.
6592 If the result of the multiplication has unsigned overflow, the result
6593 returned is the mathematical result modulo 2\ :sup:`n`\ , where n is the
6594 bit width of the result.
6596 Because LLVM integers use a two's complement representation, and the
6597 result is the same width as the operands, this instruction returns the
6598 correct result for both signed and unsigned integers. If a full product
6599 (e.g. ``i32`` * ``i32`` -> ``i64``) is needed, the operands should be
6600 sign-extended or zero-extended as appropriate to the width of the full
6603 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
6604 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
6605 result value of the ``mul`` is a :ref:`poison value <poisonvalues>` if
6606 unsigned and/or signed overflow, respectively, occurs.
6611 .. code-block:: text
6613 <result> = mul i32 4, %var ; yields i32:result = 4 * %var
6617 '``fmul``' Instruction
6618 ^^^^^^^^^^^^^^^^^^^^^^
6625 <result> = fmul [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
6630 The '``fmul``' instruction returns the product of its two operands.
6635 The two arguments to the '``fmul``' instruction must be
6636 :ref:`floating-point <t_floating>` or :ref:`vector <t_vector>` of
6637 floating-point values. Both arguments must have identical types.
6642 The value produced is the floating-point product of the two operands.
6643 This instruction is assumed to execute in the default :ref:`floating-point
6644 environment <floatenv>`.
6645 This instruction can also take any number of :ref:`fast-math
6646 flags <fastmath>`, which are optimization hints to enable otherwise
6647 unsafe floating-point optimizations:
6652 .. code-block:: text
6654 <result> = fmul float 4.0, %var ; yields float:result = 4.0 * %var
6656 '``udiv``' Instruction
6657 ^^^^^^^^^^^^^^^^^^^^^^
6664 <result> = udiv <ty> <op1>, <op2> ; yields ty:result
6665 <result> = udiv exact <ty> <op1>, <op2> ; yields ty:result
6670 The '``udiv``' instruction returns the quotient of its two operands.
6675 The two arguments to the '``udiv``' instruction must be
6676 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6677 arguments must have identical types.
6682 The value produced is the unsigned integer quotient of the two operands.
6684 Note that unsigned integer division and signed integer division are
6685 distinct operations; for signed integer division, use '``sdiv``'.
6687 Division by zero is undefined behavior. For vectors, if any element
6688 of the divisor is zero, the operation has undefined behavior.
6691 If the ``exact`` keyword is present, the result value of the ``udiv`` is
6692 a :ref:`poison value <poisonvalues>` if %op1 is not a multiple of %op2 (as
6693 such, "((a udiv exact b) mul b) == a").
6698 .. code-block:: text
6700 <result> = udiv i32 4, %var ; yields i32:result = 4 / %var
6702 '``sdiv``' Instruction
6703 ^^^^^^^^^^^^^^^^^^^^^^
6710 <result> = sdiv <ty> <op1>, <op2> ; yields ty:result
6711 <result> = sdiv exact <ty> <op1>, <op2> ; yields ty:result
6716 The '``sdiv``' instruction returns the quotient of its two operands.
6721 The two arguments to the '``sdiv``' instruction must be
6722 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6723 arguments must have identical types.
6728 The value produced is the signed integer quotient of the two operands
6729 rounded towards zero.
6731 Note that signed integer division and unsigned integer division are
6732 distinct operations; for unsigned integer division, use '``udiv``'.
6734 Division by zero is undefined behavior. For vectors, if any element
6735 of the divisor is zero, the operation has undefined behavior.
6736 Overflow also leads to undefined behavior; this is a rare case, but can
6737 occur, for example, by doing a 32-bit division of -2147483648 by -1.
6739 If the ``exact`` keyword is present, the result value of the ``sdiv`` is
6740 a :ref:`poison value <poisonvalues>` if the result would be rounded.
6745 .. code-block:: text
6747 <result> = sdiv i32 4, %var ; yields i32:result = 4 / %var
6751 '``fdiv``' Instruction
6752 ^^^^^^^^^^^^^^^^^^^^^^
6759 <result> = fdiv [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
6764 The '``fdiv``' instruction returns the quotient of its two operands.
6769 The two arguments to the '``fdiv``' instruction must be
6770 :ref:`floating-point <t_floating>` or :ref:`vector <t_vector>` of
6771 floating-point values. Both arguments must have identical types.
6776 The value produced is the floating-point quotient of the two operands.
6777 This instruction is assumed to execute in the default :ref:`floating-point
6778 environment <floatenv>`.
6779 This instruction can also take any number of :ref:`fast-math
6780 flags <fastmath>`, which are optimization hints to enable otherwise
6781 unsafe floating-point optimizations:
6786 .. code-block:: text
6788 <result> = fdiv float 4.0, %var ; yields float:result = 4.0 / %var
6790 '``urem``' Instruction
6791 ^^^^^^^^^^^^^^^^^^^^^^
6798 <result> = urem <ty> <op1>, <op2> ; yields ty:result
6803 The '``urem``' instruction returns the remainder from the unsigned
6804 division of its two arguments.
6809 The two arguments to the '``urem``' instruction must be
6810 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6811 arguments must have identical types.
6816 This instruction returns the unsigned integer *remainder* of a division.
6817 This instruction always performs an unsigned division to get the
6820 Note that unsigned integer remainder and signed integer remainder are
6821 distinct operations; for signed integer remainder, use '``srem``'.
6823 Taking the remainder of a division by zero is undefined behavior.
6824 For vectors, if any element of the divisor is zero, the operation has
6830 .. code-block:: text
6832 <result> = urem i32 4, %var ; yields i32:result = 4 % %var
6834 '``srem``' Instruction
6835 ^^^^^^^^^^^^^^^^^^^^^^
6842 <result> = srem <ty> <op1>, <op2> ; yields ty:result
6847 The '``srem``' instruction returns the remainder from the signed
6848 division of its two operands. This instruction can also take
6849 :ref:`vector <t_vector>` versions of the values in which case the elements
6855 The two arguments to the '``srem``' instruction must be
6856 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6857 arguments must have identical types.
6862 This instruction returns the *remainder* of a division (where the result
6863 is either zero or has the same sign as the dividend, ``op1``), not the
6864 *modulo* operator (where the result is either zero or has the same sign
6865 as the divisor, ``op2``) of a value. For more information about the
6866 difference, see `The Math
6867 Forum <http://mathforum.org/dr.math/problems/anne.4.28.99.html>`_. For a
6868 table of how this is implemented in various languages, please see
6870 operation <http://en.wikipedia.org/wiki/Modulo_operation>`_.
6872 Note that signed integer remainder and unsigned integer remainder are
6873 distinct operations; for unsigned integer remainder, use '``urem``'.
6875 Taking the remainder of a division by zero is undefined behavior.
6876 For vectors, if any element of the divisor is zero, the operation has
6878 Overflow also leads to undefined behavior; this is a rare case, but can
6879 occur, for example, by taking the remainder of a 32-bit division of
6880 -2147483648 by -1. (The remainder doesn't actually overflow, but this
6881 rule lets srem be implemented using instructions that return both the
6882 result of the division and the remainder.)
6887 .. code-block:: text
6889 <result> = srem i32 4, %var ; yields i32:result = 4 % %var
6893 '``frem``' Instruction
6894 ^^^^^^^^^^^^^^^^^^^^^^
6901 <result> = frem [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
6906 The '``frem``' instruction returns the remainder from the division of
6912 The two arguments to the '``frem``' instruction must be
6913 :ref:`floating-point <t_floating>` or :ref:`vector <t_vector>` of
6914 floating-point values. Both arguments must have identical types.
6919 The value produced is the floating-point remainder of the two operands.
6920 This is the same output as a libm '``fmod``' function, but without any
6921 possibility of setting ``errno``. The remainder has the same sign as the
6923 This instruction is assumed to execute in the default :ref:`floating-point
6924 environment <floatenv>`.
6925 This instruction can also take any number of :ref:`fast-math
6926 flags <fastmath>`, which are optimization hints to enable otherwise
6927 unsafe floating-point optimizations:
6932 .. code-block:: text
6934 <result> = frem float 4.0, %var ; yields float:result = 4.0 % %var
6938 Bitwise Binary Operations
6939 -------------------------
6941 Bitwise binary operators are used to do various forms of bit-twiddling
6942 in a program. They are generally very efficient instructions and can
6943 commonly be strength reduced from other instructions. They require two
6944 operands of the same type, execute an operation on them, and produce a
6945 single value. The resulting value is the same type as its operands.
6947 '``shl``' Instruction
6948 ^^^^^^^^^^^^^^^^^^^^^
6955 <result> = shl <ty> <op1>, <op2> ; yields ty:result
6956 <result> = shl nuw <ty> <op1>, <op2> ; yields ty:result
6957 <result> = shl nsw <ty> <op1>, <op2> ; yields ty:result
6958 <result> = shl nuw nsw <ty> <op1>, <op2> ; yields ty:result
6963 The '``shl``' instruction returns the first operand shifted to the left
6964 a specified number of bits.
6969 Both arguments to the '``shl``' instruction must be the same
6970 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
6971 '``op2``' is treated as an unsigned value.
6976 The value produced is ``op1`` \* 2\ :sup:`op2` mod 2\ :sup:`n`,
6977 where ``n`` is the width of the result. If ``op2`` is (statically or
6978 dynamically) equal to or larger than the number of bits in
6979 ``op1``, this instruction returns a :ref:`poison value <poisonvalues>`.
6980 If the arguments are vectors, each vector element of ``op1`` is shifted
6981 by the corresponding shift amount in ``op2``.
6983 If the ``nuw`` keyword is present, then the shift produces a poison
6984 value if it shifts out any non-zero bits.
6985 If the ``nsw`` keyword is present, then the shift produces a poison
6986 value it shifts out any bits that disagree with the resultant sign bit.
6991 .. code-block:: text
6993 <result> = shl i32 4, %var ; yields i32: 4 << %var
6994 <result> = shl i32 4, 2 ; yields i32: 16
6995 <result> = shl i32 1, 10 ; yields i32: 1024
6996 <result> = shl i32 1, 32 ; undefined
6997 <result> = shl <2 x i32> < i32 1, i32 1>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 2, i32 4>
6999 '``lshr``' Instruction
7000 ^^^^^^^^^^^^^^^^^^^^^^
7007 <result> = lshr <ty> <op1>, <op2> ; yields ty:result
7008 <result> = lshr exact <ty> <op1>, <op2> ; yields ty:result
7013 The '``lshr``' instruction (logical shift right) returns the first
7014 operand shifted to the right a specified number of bits with zero fill.
7019 Both arguments to the '``lshr``' instruction must be the same
7020 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
7021 '``op2``' is treated as an unsigned value.
7026 This instruction always performs a logical shift right operation. The
7027 most significant bits of the result will be filled with zero bits after
7028 the shift. If ``op2`` is (statically or dynamically) equal to or larger
7029 than the number of bits in ``op1``, this instruction returns a :ref:`poison
7030 value <poisonvalues>`. If the arguments are vectors, each vector element
7031 of ``op1`` is shifted by the corresponding shift amount in ``op2``.
7033 If the ``exact`` keyword is present, the result value of the ``lshr`` is
7034 a poison value if any of the bits shifted out are non-zero.
7039 .. code-block:: text
7041 <result> = lshr i32 4, 1 ; yields i32:result = 2
7042 <result> = lshr i32 4, 2 ; yields i32:result = 1
7043 <result> = lshr i8 4, 3 ; yields i8:result = 0
7044 <result> = lshr i8 -2, 1 ; yields i8:result = 0x7F
7045 <result> = lshr i32 1, 32 ; undefined
7046 <result> = lshr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 0x7FFFFFFF, i32 1>
7048 '``ashr``' Instruction
7049 ^^^^^^^^^^^^^^^^^^^^^^
7056 <result> = ashr <ty> <op1>, <op2> ; yields ty:result
7057 <result> = ashr exact <ty> <op1>, <op2> ; yields ty:result
7062 The '``ashr``' instruction (arithmetic shift right) returns the first
7063 operand shifted to the right a specified number of bits with sign
7069 Both arguments to the '``ashr``' instruction must be the same
7070 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
7071 '``op2``' is treated as an unsigned value.
7076 This instruction always performs an arithmetic shift right operation,
7077 The most significant bits of the result will be filled with the sign bit
7078 of ``op1``. If ``op2`` is (statically or dynamically) equal to or larger
7079 than the number of bits in ``op1``, this instruction returns a :ref:`poison
7080 value <poisonvalues>`. If the arguments are vectors, each vector element
7081 of ``op1`` is shifted by the corresponding shift amount in ``op2``.
7083 If the ``exact`` keyword is present, the result value of the ``ashr`` is
7084 a poison value if any of the bits shifted out are non-zero.
7089 .. code-block:: text
7091 <result> = ashr i32 4, 1 ; yields i32:result = 2
7092 <result> = ashr i32 4, 2 ; yields i32:result = 1
7093 <result> = ashr i8 4, 3 ; yields i8:result = 0
7094 <result> = ashr i8 -2, 1 ; yields i8:result = -1
7095 <result> = ashr i32 1, 32 ; undefined
7096 <result> = ashr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 3> ; yields: result=<2 x i32> < i32 -1, i32 0>
7098 '``and``' Instruction
7099 ^^^^^^^^^^^^^^^^^^^^^
7106 <result> = and <ty> <op1>, <op2> ; yields ty:result
7111 The '``and``' instruction returns the bitwise logical and of its two
7117 The two arguments to the '``and``' instruction must be
7118 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
7119 arguments must have identical types.
7124 The truth table used for the '``and``' instruction is:
7141 .. code-block:: text
7143 <result> = and i32 4, %var ; yields i32:result = 4 & %var
7144 <result> = and i32 15, 40 ; yields i32:result = 8
7145 <result> = and i32 4, 8 ; yields i32:result = 0
7147 '``or``' Instruction
7148 ^^^^^^^^^^^^^^^^^^^^
7155 <result> = or <ty> <op1>, <op2> ; yields ty:result
7160 The '``or``' instruction returns the bitwise logical inclusive or of its
7166 The two arguments to the '``or``' instruction must be
7167 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
7168 arguments must have identical types.
7173 The truth table used for the '``or``' instruction is:
7192 <result> = or i32 4, %var ; yields i32:result = 4 | %var
7193 <result> = or i32 15, 40 ; yields i32:result = 47
7194 <result> = or i32 4, 8 ; yields i32:result = 12
7196 '``xor``' Instruction
7197 ^^^^^^^^^^^^^^^^^^^^^
7204 <result> = xor <ty> <op1>, <op2> ; yields ty:result
7209 The '``xor``' instruction returns the bitwise logical exclusive or of
7210 its two operands. The ``xor`` is used to implement the "one's
7211 complement" operation, which is the "~" operator in C.
7216 The two arguments to the '``xor``' instruction must be
7217 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
7218 arguments must have identical types.
7223 The truth table used for the '``xor``' instruction is:
7240 .. code-block:: text
7242 <result> = xor i32 4, %var ; yields i32:result = 4 ^ %var
7243 <result> = xor i32 15, 40 ; yields i32:result = 39
7244 <result> = xor i32 4, 8 ; yields i32:result = 12
7245 <result> = xor i32 %V, -1 ; yields i32:result = ~%V
7250 LLVM supports several instructions to represent vector operations in a
7251 target-independent manner. These instructions cover the element-access
7252 and vector-specific operations needed to process vectors effectively.
7253 While LLVM does directly support these vector operations, many
7254 sophisticated algorithms will want to use target-specific intrinsics to
7255 take full advantage of a specific target.
7257 .. _i_extractelement:
7259 '``extractelement``' Instruction
7260 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7267 <result> = extractelement <n x <ty>> <val>, <ty2> <idx> ; yields <ty>
7272 The '``extractelement``' instruction extracts a single scalar element
7273 from a vector at a specified index.
7278 The first operand of an '``extractelement``' instruction is a value of
7279 :ref:`vector <t_vector>` type. The second operand is an index indicating
7280 the position from which to extract the element. The index may be a
7281 variable of any integer type.
7286 The result is a scalar of the same type as the element type of ``val``.
7287 Its value is the value at position ``idx`` of ``val``. If ``idx``
7288 exceeds the length of ``val``, the results are undefined.
7293 .. code-block:: text
7295 <result> = extractelement <4 x i32> %vec, i32 0 ; yields i32
7297 .. _i_insertelement:
7299 '``insertelement``' Instruction
7300 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7307 <result> = insertelement <n x <ty>> <val>, <ty> <elt>, <ty2> <idx> ; yields <n x <ty>>
7312 The '``insertelement``' instruction inserts a scalar element into a
7313 vector at a specified index.
7318 The first operand of an '``insertelement``' instruction is a value of
7319 :ref:`vector <t_vector>` type. The second operand is a scalar value whose
7320 type must equal the element type of the first operand. The third operand
7321 is an index indicating the position at which to insert the value. The
7322 index may be a variable of any integer type.
7327 The result is a vector of the same type as ``val``. Its element values
7328 are those of ``val`` except at position ``idx``, where it gets the value
7329 ``elt``. If ``idx`` exceeds the length of ``val``, the results are
7335 .. code-block:: text
7337 <result> = insertelement <4 x i32> %vec, i32 1, i32 0 ; yields <4 x i32>
7339 .. _i_shufflevector:
7341 '``shufflevector``' Instruction
7342 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7349 <result> = shufflevector <n x <ty>> <v1>, <n x <ty>> <v2>, <m x i32> <mask> ; yields <m x <ty>>
7354 The '``shufflevector``' instruction constructs a permutation of elements
7355 from two input vectors, returning a vector with the same element type as
7356 the input and length that is the same as the shuffle mask.
7361 The first two operands of a '``shufflevector``' instruction are vectors
7362 with the same type. The third argument is a shuffle mask whose element
7363 type is always 'i32'. The result of the instruction is a vector whose
7364 length is the same as the shuffle mask and whose element type is the
7365 same as the element type of the first two operands.
7367 The shuffle mask operand is required to be a constant vector with either
7368 constant integer or undef values.
7373 The elements of the two input vectors are numbered from left to right
7374 across both of the vectors. The shuffle mask operand specifies, for each
7375 element of the result vector, which element of the two input vectors the
7376 result element gets. If the shuffle mask is undef, the result vector is
7377 undef. If any element of the mask operand is undef, that element of the
7378 result is undef. If the shuffle mask selects an undef element from one
7379 of the input vectors, the resulting element is undef.
7384 .. code-block:: text
7386 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
7387 <4 x i32> <i32 0, i32 4, i32 1, i32 5> ; yields <4 x i32>
7388 <result> = shufflevector <4 x i32> %v1, <4 x i32> undef,
7389 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32> - Identity shuffle.
7390 <result> = shufflevector <8 x i32> %v1, <8 x i32> undef,
7391 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32>
7392 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
7393 <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7 > ; yields <8 x i32>
7395 Aggregate Operations
7396 --------------------
7398 LLVM supports several instructions for working with
7399 :ref:`aggregate <t_aggregate>` values.
7403 '``extractvalue``' Instruction
7404 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7411 <result> = extractvalue <aggregate type> <val>, <idx>{, <idx>}*
7416 The '``extractvalue``' instruction extracts the value of a member field
7417 from an :ref:`aggregate <t_aggregate>` value.
7422 The first operand of an '``extractvalue``' instruction is a value of
7423 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The other operands are
7424 constant indices to specify which value to extract in a similar manner
7425 as indices in a '``getelementptr``' instruction.
7427 The major differences to ``getelementptr`` indexing are:
7429 - Since the value being indexed is not a pointer, the first index is
7430 omitted and assumed to be zero.
7431 - At least one index must be specified.
7432 - Not only struct indices but also array indices must be in bounds.
7437 The result is the value at the position in the aggregate specified by
7443 .. code-block:: text
7445 <result> = extractvalue {i32, float} %agg, 0 ; yields i32
7449 '``insertvalue``' Instruction
7450 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7457 <result> = insertvalue <aggregate type> <val>, <ty> <elt>, <idx>{, <idx>}* ; yields <aggregate type>
7462 The '``insertvalue``' instruction inserts a value into a member field in
7463 an :ref:`aggregate <t_aggregate>` value.
7468 The first operand of an '``insertvalue``' instruction is a value of
7469 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The second operand is
7470 a first-class value to insert. The following operands are constant
7471 indices indicating the position at which to insert the value in a
7472 similar manner as indices in a '``extractvalue``' instruction. The value
7473 to insert must have the same type as the value identified by the
7479 The result is an aggregate of the same type as ``val``. Its value is
7480 that of ``val`` except that the value at the position specified by the
7481 indices is that of ``elt``.
7486 .. code-block:: llvm
7488 %agg1 = insertvalue {i32, float} undef, i32 1, 0 ; yields {i32 1, float undef}
7489 %agg2 = insertvalue {i32, float} %agg1, float %val, 1 ; yields {i32 1, float %val}
7490 %agg3 = insertvalue {i32, {float}} undef, float %val, 1, 0 ; yields {i32 undef, {float %val}}
7494 Memory Access and Addressing Operations
7495 ---------------------------------------
7497 A key design point of an SSA-based representation is how it represents
7498 memory. In LLVM, no memory locations are in SSA form, which makes things
7499 very simple. This section describes how to read, write, and allocate
7504 '``alloca``' Instruction
7505 ^^^^^^^^^^^^^^^^^^^^^^^^
7512 <result> = alloca [inalloca] <type> [, <ty> <NumElements>] [, align <alignment>] [, addrspace(<num>)] ; yields type addrspace(num)*:result
7517 The '``alloca``' instruction allocates memory on the stack frame of the
7518 currently executing function, to be automatically released when this
7519 function returns to its caller. The object is always allocated in the
7520 address space for allocas indicated in the datalayout.
7525 The '``alloca``' instruction allocates ``sizeof(<type>)*NumElements``
7526 bytes of memory on the runtime stack, returning a pointer of the
7527 appropriate type to the program. If "NumElements" is specified, it is
7528 the number of elements allocated, otherwise "NumElements" is defaulted
7529 to be one. If a constant alignment is specified, the value result of the
7530 allocation is guaranteed to be aligned to at least that boundary. The
7531 alignment may not be greater than ``1 << 29``. If not specified, or if
7532 zero, the target can choose to align the allocation on any convenient
7533 boundary compatible with the type.
7535 '``type``' may be any sized type.
7540 Memory is allocated; a pointer is returned. The operation is undefined
7541 if there is insufficient stack space for the allocation. '``alloca``'d
7542 memory is automatically released when the function returns. The
7543 '``alloca``' instruction is commonly used to represent automatic
7544 variables that must have an address available. When the function returns
7545 (either with the ``ret`` or ``resume`` instructions), the memory is
7546 reclaimed. Allocating zero bytes is legal, but the result is undefined.
7547 The order in which memory is allocated (ie., which way the stack grows)
7553 .. code-block:: llvm
7555 %ptr = alloca i32 ; yields i32*:ptr
7556 %ptr = alloca i32, i32 4 ; yields i32*:ptr
7557 %ptr = alloca i32, i32 4, align 1024 ; yields i32*:ptr
7558 %ptr = alloca i32, align 1024 ; yields i32*:ptr
7562 '``load``' Instruction
7563 ^^^^^^^^^^^^^^^^^^^^^^
7570 <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>]
7571 <result> = load atomic [volatile] <ty>, <ty>* <pointer> [syncscope("<target-scope>")] <ordering>, align <alignment> [, !invariant.group !<index>]
7572 !<index> = !{ i32 1 }
7573 !<deref_bytes_node> = !{i64 <dereferenceable_bytes>}
7574 !<align_node> = !{ i64 <value_alignment> }
7579 The '``load``' instruction is used to read from memory.
7584 The argument to the ``load`` instruction specifies the memory address from which
7585 to load. The type specified must be a :ref:`first class <t_firstclass>` type of
7586 known size (i.e. not containing an :ref:`opaque structural type <t_opaque>`). If
7587 the ``load`` is marked as ``volatile``, then the optimizer is not allowed to
7588 modify the number or order of execution of this ``load`` with other
7589 :ref:`volatile operations <volatile>`.
7591 If the ``load`` is marked as ``atomic``, it takes an extra :ref:`ordering
7592 <ordering>` and optional ``syncscope("<target-scope>")`` argument. The
7593 ``release`` and ``acq_rel`` orderings are not valid on ``load`` instructions.
7594 Atomic loads produce :ref:`defined <memmodel>` results when they may see
7595 multiple atomic stores. The type of the pointee must be an integer, pointer, or
7596 floating-point type whose bit width is a power of two greater than or equal to
7597 eight and less than or equal to a target-specific size limit. ``align`` must be
7598 explicitly specified on atomic loads, and the load has undefined behavior if the
7599 alignment is not set to a value which is at least the size in bytes of the
7600 pointee. ``!nontemporal`` does not have any defined semantics for atomic loads.
7602 The optional constant ``align`` argument specifies the alignment of the
7603 operation (that is, the alignment of the memory address). A value of 0
7604 or an omitted ``align`` argument means that the operation has the ABI
7605 alignment for the target. It is the responsibility of the code emitter
7606 to ensure that the alignment information is correct. Overestimating the
7607 alignment results in undefined behavior. Underestimating the alignment
7608 may produce less efficient code. An alignment of 1 is always safe. The
7609 maximum possible alignment is ``1 << 29``. An alignment value higher
7610 than the size of the loaded type implies memory up to the alignment
7611 value bytes can be safely loaded without trapping in the default
7612 address space. Access of the high bytes can interfere with debugging
7613 tools, so should not be accessed if the function has the
7614 ``sanitize_thread`` or ``sanitize_address`` attributes.
7616 The optional ``!nontemporal`` metadata must reference a single
7617 metadata name ``<index>`` corresponding to a metadata node with one
7618 ``i32`` entry of value 1. The existence of the ``!nontemporal``
7619 metadata on the instruction tells the optimizer and code generator
7620 that this load is not expected to be reused in the cache. The code
7621 generator may select special instructions to save cache bandwidth, such
7622 as the ``MOVNT`` instruction on x86.
7624 The optional ``!invariant.load`` metadata must reference a single
7625 metadata name ``<index>`` corresponding to a metadata node with no
7626 entries. If a load instruction tagged with the ``!invariant.load``
7627 metadata is executed, the optimizer may assume the memory location
7628 referenced by the load contains the same value at all points in the
7629 program where the memory location is known to be dereferenceable.
7631 The optional ``!invariant.group`` metadata must reference a single metadata name
7632 ``<index>`` corresponding to a metadata node. See ``invariant.group`` metadata.
7634 The optional ``!nonnull`` metadata must reference a single
7635 metadata name ``<index>`` corresponding to a metadata node with no
7636 entries. The existence of the ``!nonnull`` metadata on the
7637 instruction tells the optimizer that the value loaded is known to
7638 never be null. This is analogous to the ``nonnull`` attribute
7639 on parameters and return values. This metadata can only be applied
7640 to loads of a pointer type.
7642 The optional ``!dereferenceable`` metadata must reference a single metadata
7643 name ``<deref_bytes_node>`` corresponding to a metadata node with one ``i64``
7644 entry. The existence of the ``!dereferenceable`` metadata on the instruction
7645 tells the optimizer that the value loaded is known to be dereferenceable.
7646 The number of bytes known to be dereferenceable is specified by the integer
7647 value in the metadata node. This is analogous to the ''dereferenceable''
7648 attribute on parameters and return values. This metadata can only be applied
7649 to loads of a pointer type.
7651 The optional ``!dereferenceable_or_null`` metadata must reference a single
7652 metadata name ``<deref_bytes_node>`` corresponding to a metadata node with one
7653 ``i64`` entry. The existence of the ``!dereferenceable_or_null`` metadata on the
7654 instruction tells the optimizer that the value loaded is known to be either
7655 dereferenceable or null.
7656 The number of bytes known to be dereferenceable is specified by the integer
7657 value in the metadata node. This is analogous to the ''dereferenceable_or_null''
7658 attribute on parameters and return values. This metadata can only be applied
7659 to loads of a pointer type.
7661 The optional ``!align`` metadata must reference a single metadata name
7662 ``<align_node>`` corresponding to a metadata node with one ``i64`` entry.
7663 The existence of the ``!align`` metadata on the instruction tells the
7664 optimizer that the value loaded is known to be aligned to a boundary specified
7665 by the integer value in the metadata node. The alignment must be a power of 2.
7666 This is analogous to the ''align'' attribute on parameters and return values.
7667 This metadata can only be applied to loads of a pointer type.
7672 The location of memory pointed to is loaded. If the value being loaded
7673 is of scalar type then the number of bytes read does not exceed the
7674 minimum number of bytes needed to hold all bits of the type. For
7675 example, loading an ``i24`` reads at most three bytes. When loading a
7676 value of a type like ``i20`` with a size that is not an integral number
7677 of bytes, the result is undefined if the value was not originally
7678 written using a store of the same type.
7683 .. code-block:: llvm
7685 %ptr = alloca i32 ; yields i32*:ptr
7686 store i32 3, i32* %ptr ; yields void
7687 %val = load i32, i32* %ptr ; yields i32:val = i32 3
7691 '``store``' Instruction
7692 ^^^^^^^^^^^^^^^^^^^^^^^
7699 store [volatile] <ty> <value>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>][, !invariant.group !<index>] ; yields void
7700 store atomic [volatile] <ty> <value>, <ty>* <pointer> [syncscope("<target-scope>")] <ordering>, align <alignment> [, !invariant.group !<index>] ; yields void
7705 The '``store``' instruction is used to write to memory.
7710 There are two arguments to the ``store`` instruction: a value to store and an
7711 address at which to store it. The type of the ``<pointer>`` operand must be a
7712 pointer to the :ref:`first class <t_firstclass>` type of the ``<value>``
7713 operand. If the ``store`` is marked as ``volatile``, then the optimizer is not
7714 allowed to modify the number or order of execution of this ``store`` with other
7715 :ref:`volatile operations <volatile>`. Only values of :ref:`first class
7716 <t_firstclass>` types of known size (i.e. not containing an :ref:`opaque
7717 structural type <t_opaque>`) can be stored.
7719 If the ``store`` is marked as ``atomic``, it takes an extra :ref:`ordering
7720 <ordering>` and optional ``syncscope("<target-scope>")`` argument. The
7721 ``acquire`` and ``acq_rel`` orderings aren't valid on ``store`` instructions.
7722 Atomic loads produce :ref:`defined <memmodel>` results when they may see
7723 multiple atomic stores. The type of the pointee must be an integer, pointer, or
7724 floating-point type whose bit width is a power of two greater than or equal to
7725 eight and less than or equal to a target-specific size limit. ``align`` must be
7726 explicitly specified on atomic stores, and the store has undefined behavior if
7727 the alignment is not set to a value which is at least the size in bytes of the
7728 pointee. ``!nontemporal`` does not have any defined semantics for atomic stores.
7730 The optional constant ``align`` argument specifies the alignment of the
7731 operation (that is, the alignment of the memory address). A value of 0
7732 or an omitted ``align`` argument means that the operation has the ABI
7733 alignment for the target. It is the responsibility of the code emitter
7734 to ensure that the alignment information is correct. Overestimating the
7735 alignment results in undefined behavior. Underestimating the
7736 alignment may produce less efficient code. An alignment of 1 is always
7737 safe. The maximum possible alignment is ``1 << 29``. An alignment
7738 value higher than the size of the stored type implies memory up to the
7739 alignment value bytes can be stored to without trapping in the default
7740 address space. Storing to the higher bytes however may result in data
7741 races if another thread can access the same address. Introducing a
7742 data race is not allowed. Storing to the extra bytes is not allowed
7743 even in situations where a data race is known to not exist if the
7744 function has the ``sanitize_address`` attribute.
7746 The optional ``!nontemporal`` metadata must reference a single metadata
7747 name ``<index>`` corresponding to a metadata node with one ``i32`` entry of
7748 value 1. The existence of the ``!nontemporal`` metadata on the instruction
7749 tells the optimizer and code generator that this load is not expected to
7750 be reused in the cache. The code generator may select special
7751 instructions to save cache bandwidth, such as the ``MOVNT`` instruction on
7754 The optional ``!invariant.group`` metadata must reference a
7755 single metadata name ``<index>``. See ``invariant.group`` metadata.
7760 The contents of memory are updated to contain ``<value>`` at the
7761 location specified by the ``<pointer>`` operand. If ``<value>`` is
7762 of scalar type then the number of bytes written does not exceed the
7763 minimum number of bytes needed to hold all bits of the type. For
7764 example, storing an ``i24`` writes at most three bytes. When writing a
7765 value of a type like ``i20`` with a size that is not an integral number
7766 of bytes, it is unspecified what happens to the extra bits that do not
7767 belong to the type, but they will typically be overwritten.
7772 .. code-block:: llvm
7774 %ptr = alloca i32 ; yields i32*:ptr
7775 store i32 3, i32* %ptr ; yields void
7776 %val = load i32, i32* %ptr ; yields i32:val = i32 3
7780 '``fence``' Instruction
7781 ^^^^^^^^^^^^^^^^^^^^^^^
7788 fence [syncscope("<target-scope>")] <ordering> ; yields void
7793 The '``fence``' instruction is used to introduce happens-before edges
7799 '``fence``' instructions take an :ref:`ordering <ordering>` argument which
7800 defines what *synchronizes-with* edges they add. They can only be given
7801 ``acquire``, ``release``, ``acq_rel``, and ``seq_cst`` orderings.
7806 A fence A which has (at least) ``release`` ordering semantics
7807 *synchronizes with* a fence B with (at least) ``acquire`` ordering
7808 semantics if and only if there exist atomic operations X and Y, both
7809 operating on some atomic object M, such that A is sequenced before X, X
7810 modifies M (either directly or through some side effect of a sequence
7811 headed by X), Y is sequenced before B, and Y observes M. This provides a
7812 *happens-before* dependency between A and B. Rather than an explicit
7813 ``fence``, one (but not both) of the atomic operations X or Y might
7814 provide a ``release`` or ``acquire`` (resp.) ordering constraint and
7815 still *synchronize-with* the explicit ``fence`` and establish the
7816 *happens-before* edge.
7818 A ``fence`` which has ``seq_cst`` ordering, in addition to having both
7819 ``acquire`` and ``release`` semantics specified above, participates in
7820 the global program order of other ``seq_cst`` operations and/or fences.
7822 A ``fence`` instruction can also take an optional
7823 ":ref:`syncscope <syncscope>`" argument.
7828 .. code-block:: text
7830 fence acquire ; yields void
7831 fence syncscope("singlethread") seq_cst ; yields void
7832 fence syncscope("agent") seq_cst ; yields void
7836 '``cmpxchg``' Instruction
7837 ^^^^^^^^^^^^^^^^^^^^^^^^^
7844 cmpxchg [weak] [volatile] <ty>* <pointer>, <ty> <cmp>, <ty> <new> [syncscope("<target-scope>")] <success ordering> <failure ordering> ; yields { ty, i1 }
7849 The '``cmpxchg``' instruction is used to atomically modify memory. It
7850 loads a value in memory and compares it to a given value. If they are
7851 equal, it tries to store a new value into the memory.
7856 There are three arguments to the '``cmpxchg``' instruction: an address
7857 to operate on, a value to compare to the value currently be at that
7858 address, and a new value to place at that address if the compared values
7859 are equal. The type of '<cmp>' must be an integer or pointer type whose
7860 bit width is a power of two greater than or equal to eight and less
7861 than or equal to a target-specific size limit. '<cmp>' and '<new>' must
7862 have the same type, and the type of '<pointer>' must be a pointer to
7863 that type. If the ``cmpxchg`` is marked as ``volatile``, then the
7864 optimizer is not allowed to modify the number or order of execution of
7865 this ``cmpxchg`` with other :ref:`volatile operations <volatile>`.
7867 The success and failure :ref:`ordering <ordering>` arguments specify how this
7868 ``cmpxchg`` synchronizes with other atomic operations. Both ordering parameters
7869 must be at least ``monotonic``, the ordering constraint on failure must be no
7870 stronger than that on success, and the failure ordering cannot be either
7871 ``release`` or ``acq_rel``.
7873 A ``cmpxchg`` instruction can also take an optional
7874 ":ref:`syncscope <syncscope>`" argument.
7876 The pointer passed into cmpxchg must have alignment greater than or
7877 equal to the size in memory of the operand.
7882 The contents of memory at the location specified by the '``<pointer>``' operand
7883 is read and compared to '``<cmp>``'; if the values are equal, '``<new>``' is
7884 written to the location. The original value at the location is returned,
7885 together with a flag indicating success (true) or failure (false).
7887 If the cmpxchg operation is marked as ``weak`` then a spurious failure is
7888 permitted: the operation may not write ``<new>`` even if the comparison
7891 If the cmpxchg operation is strong (the default), the i1 value is 1 if and only
7892 if the value loaded equals ``cmp``.
7894 A successful ``cmpxchg`` is a read-modify-write instruction for the purpose of
7895 identifying release sequences. A failed ``cmpxchg`` is equivalent to an atomic
7896 load with an ordering parameter determined the second ordering parameter.
7901 .. code-block:: llvm
7904 %orig = load atomic i32, i32* %ptr unordered, align 4 ; yields i32
7908 %cmp = phi i32 [ %orig, %entry ], [%value_loaded, %loop]
7909 %squared = mul i32 %cmp, %cmp
7910 %val_success = cmpxchg i32* %ptr, i32 %cmp, i32 %squared acq_rel monotonic ; yields { i32, i1 }
7911 %value_loaded = extractvalue { i32, i1 } %val_success, 0
7912 %success = extractvalue { i32, i1 } %val_success, 1
7913 br i1 %success, label %done, label %loop
7920 '``atomicrmw``' Instruction
7921 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7928 atomicrmw [volatile] <operation> <ty>* <pointer>, <ty> <value> [syncscope("<target-scope>")] <ordering> ; yields ty
7933 The '``atomicrmw``' instruction is used to atomically modify memory.
7938 There are three arguments to the '``atomicrmw``' instruction: an
7939 operation to apply, an address whose value to modify, an argument to the
7940 operation. The operation must be one of the following keywords:
7954 The type of '<value>' must be an integer type whose bit width is a power
7955 of two greater than or equal to eight and less than or equal to a
7956 target-specific size limit. The type of the '``<pointer>``' operand must
7957 be a pointer to that type. If the ``atomicrmw`` is marked as
7958 ``volatile``, then the optimizer is not allowed to modify the number or
7959 order of execution of this ``atomicrmw`` with other :ref:`volatile
7960 operations <volatile>`.
7962 A ``atomicrmw`` instruction can also take an optional
7963 ":ref:`syncscope <syncscope>`" argument.
7968 The contents of memory at the location specified by the '``<pointer>``'
7969 operand are atomically read, modified, and written back. The original
7970 value at the location is returned. The modification is specified by the
7973 - xchg: ``*ptr = val``
7974 - add: ``*ptr = *ptr + val``
7975 - sub: ``*ptr = *ptr - val``
7976 - and: ``*ptr = *ptr & val``
7977 - nand: ``*ptr = ~(*ptr & val)``
7978 - or: ``*ptr = *ptr | val``
7979 - xor: ``*ptr = *ptr ^ val``
7980 - max: ``*ptr = *ptr > val ? *ptr : val`` (using a signed comparison)
7981 - min: ``*ptr = *ptr < val ? *ptr : val`` (using a signed comparison)
7982 - umax: ``*ptr = *ptr > val ? *ptr : val`` (using an unsigned
7984 - umin: ``*ptr = *ptr < val ? *ptr : val`` (using an unsigned
7990 .. code-block:: llvm
7992 %old = atomicrmw add i32* %ptr, i32 1 acquire ; yields i32
7994 .. _i_getelementptr:
7996 '``getelementptr``' Instruction
7997 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8004 <result> = getelementptr <ty>, <ty>* <ptrval>{, [inrange] <ty> <idx>}*
8005 <result> = getelementptr inbounds <ty>, <ty>* <ptrval>{, [inrange] <ty> <idx>}*
8006 <result> = getelementptr <ty>, <ptr vector> <ptrval>, [inrange] <vector index type> <idx>
8011 The '``getelementptr``' instruction is used to get the address of a
8012 subelement of an :ref:`aggregate <t_aggregate>` data structure. It performs
8013 address calculation only and does not access memory. The instruction can also
8014 be used to calculate a vector of such addresses.
8019 The first argument is always a type used as the basis for the calculations.
8020 The second argument is always a pointer or a vector of pointers, and is the
8021 base address to start from. The remaining arguments are indices
8022 that indicate which of the elements of the aggregate object are indexed.
8023 The interpretation of each index is dependent on the type being indexed
8024 into. The first index always indexes the pointer value given as the
8025 second argument, the second index indexes a value of the type pointed to
8026 (not necessarily the value directly pointed to, since the first index
8027 can be non-zero), etc. The first type indexed into must be a pointer
8028 value, subsequent types can be arrays, vectors, and structs. Note that
8029 subsequent types being indexed into can never be pointers, since that
8030 would require loading the pointer before continuing calculation.
8032 The type of each index argument depends on the type it is indexing into.
8033 When indexing into a (optionally packed) structure, only ``i32`` integer
8034 **constants** are allowed (when using a vector of indices they must all
8035 be the **same** ``i32`` integer constant). When indexing into an array,
8036 pointer or vector, integers of any width are allowed, and they are not
8037 required to be constant. These integers are treated as signed values
8040 For example, let's consider a C code fragment and how it gets compiled
8056 int *foo(struct ST *s) {
8057 return &s[1].Z.B[5][13];
8060 The LLVM code generated by Clang is:
8062 .. code-block:: llvm
8064 %struct.RT = type { i8, [10 x [20 x i32]], i8 }
8065 %struct.ST = type { i32, double, %struct.RT }
8067 define i32* @foo(%struct.ST* %s) nounwind uwtable readnone optsize ssp {
8069 %arrayidx = getelementptr inbounds %struct.ST, %struct.ST* %s, i64 1, i32 2, i32 1, i64 5, i64 13
8076 In the example above, the first index is indexing into the
8077 '``%struct.ST*``' type, which is a pointer, yielding a '``%struct.ST``'
8078 = '``{ i32, double, %struct.RT }``' type, a structure. The second index
8079 indexes into the third element of the structure, yielding a
8080 '``%struct.RT``' = '``{ i8 , [10 x [20 x i32]], i8 }``' type, another
8081 structure. The third index indexes into the second element of the
8082 structure, yielding a '``[10 x [20 x i32]]``' type, an array. The two
8083 dimensions of the array are subscripted into, yielding an '``i32``'
8084 type. The '``getelementptr``' instruction returns a pointer to this
8085 element, thus computing a value of '``i32*``' type.
8087 Note that it is perfectly legal to index partially through a structure,
8088 returning a pointer to an inner element. Because of this, the LLVM code
8089 for the given testcase is equivalent to:
8091 .. code-block:: llvm
8093 define i32* @foo(%struct.ST* %s) {
8094 %t1 = getelementptr %struct.ST, %struct.ST* %s, i32 1 ; yields %struct.ST*:%t1
8095 %t2 = getelementptr %struct.ST, %struct.ST* %t1, i32 0, i32 2 ; yields %struct.RT*:%t2
8096 %t3 = getelementptr %struct.RT, %struct.RT* %t2, i32 0, i32 1 ; yields [10 x [20 x i32]]*:%t3
8097 %t4 = getelementptr [10 x [20 x i32]], [10 x [20 x i32]]* %t3, i32 0, i32 5 ; yields [20 x i32]*:%t4
8098 %t5 = getelementptr [20 x i32], [20 x i32]* %t4, i32 0, i32 13 ; yields i32*:%t5
8102 If the ``inbounds`` keyword is present, the result value of the
8103 ``getelementptr`` is a :ref:`poison value <poisonvalues>` if the base
8104 pointer is not an *in bounds* address of an allocated object, or if any
8105 of the addresses that would be formed by successive addition of the
8106 offsets implied by the indices to the base address with infinitely
8107 precise signed arithmetic are not an *in bounds* address of that
8108 allocated object. The *in bounds* addresses for an allocated object are
8109 all the addresses that point into the object, plus the address one byte
8110 past the end. The only *in bounds* address for a null pointer in the
8111 default address-space is the null pointer itself. In cases where the
8112 base is a vector of pointers the ``inbounds`` keyword applies to each
8113 of the computations element-wise.
8115 If the ``inbounds`` keyword is not present, the offsets are added to the
8116 base address with silently-wrapping two's complement arithmetic. If the
8117 offsets have a different width from the pointer, they are sign-extended
8118 or truncated to the width of the pointer. The result value of the
8119 ``getelementptr`` may be outside the object pointed to by the base
8120 pointer. The result value may not necessarily be used to access memory
8121 though, even if it happens to point into allocated storage. See the
8122 :ref:`Pointer Aliasing Rules <pointeraliasing>` section for more
8125 If the ``inrange`` keyword is present before any index, loading from or
8126 storing to any pointer derived from the ``getelementptr`` has undefined
8127 behavior if the load or store would access memory outside of the bounds of
8128 the element selected by the index marked as ``inrange``. The result of a
8129 pointer comparison or ``ptrtoint`` (including ``ptrtoint``-like operations
8130 involving memory) involving a pointer derived from a ``getelementptr`` with
8131 the ``inrange`` keyword is undefined, with the exception of comparisons
8132 in the case where both operands are in the range of the element selected
8133 by the ``inrange`` keyword, inclusive of the address one past the end of
8134 that element. Note that the ``inrange`` keyword is currently only allowed
8135 in constant ``getelementptr`` expressions.
8137 The getelementptr instruction is often confusing. For some more insight
8138 into how it works, see :doc:`the getelementptr FAQ <GetElementPtr>`.
8143 .. code-block:: llvm
8145 ; yields [12 x i8]*:aptr
8146 %aptr = getelementptr {i32, [12 x i8]}, {i32, [12 x i8]}* %saptr, i64 0, i32 1
8148 %vptr = getelementptr {i32, <2 x i8>}, {i32, <2 x i8>}* %svptr, i64 0, i32 1, i32 1
8150 %eptr = getelementptr [12 x i8], [12 x i8]* %aptr, i64 0, i32 1
8152 %iptr = getelementptr [10 x i32], [10 x i32]* @arr, i16 0, i16 0
8157 The ``getelementptr`` returns a vector of pointers, instead of a single address,
8158 when one or more of its arguments is a vector. In such cases, all vector
8159 arguments should have the same number of elements, and every scalar argument
8160 will be effectively broadcast into a vector during address calculation.
8162 .. code-block:: llvm
8164 ; All arguments are vectors:
8165 ; A[i] = ptrs[i] + offsets[i]*sizeof(i8)
8166 %A = getelementptr i8, <4 x i8*> %ptrs, <4 x i64> %offsets
8168 ; Add the same scalar offset to each pointer of a vector:
8169 ; A[i] = ptrs[i] + offset*sizeof(i8)
8170 %A = getelementptr i8, <4 x i8*> %ptrs, i64 %offset
8172 ; Add distinct offsets to the same pointer:
8173 ; A[i] = ptr + offsets[i]*sizeof(i8)
8174 %A = getelementptr i8, i8* %ptr, <4 x i64> %offsets
8176 ; In all cases described above the type of the result is <4 x i8*>
8178 The two following instructions are equivalent:
8180 .. code-block:: llvm
8182 getelementptr %struct.ST, <4 x %struct.ST*> %s, <4 x i64> %ind1,
8183 <4 x i32> <i32 2, i32 2, i32 2, i32 2>,
8184 <4 x i32> <i32 1, i32 1, i32 1, i32 1>,
8186 <4 x i64> <i64 13, i64 13, i64 13, i64 13>
8188 getelementptr %struct.ST, <4 x %struct.ST*> %s, <4 x i64> %ind1,
8189 i32 2, i32 1, <4 x i32> %ind4, i64 13
8191 Let's look at the C code, where the vector version of ``getelementptr``
8196 // Let's assume that we vectorize the following loop:
8197 double *A, *B; int *C;
8198 for (int i = 0; i < size; ++i) {
8202 .. code-block:: llvm
8204 ; get pointers for 8 elements from array B
8205 %ptrs = getelementptr double, double* %B, <8 x i32> %C
8206 ; load 8 elements from array B into A
8207 %A = call <8 x double> @llvm.masked.gather.v8f64.v8p0f64(<8 x double*> %ptrs,
8208 i32 8, <8 x i1> %mask, <8 x double> %passthru)
8210 Conversion Operations
8211 ---------------------
8213 The instructions in this category are the conversion instructions
8214 (casting) which all take a single operand and a type. They perform
8215 various bit conversions on the operand.
8219 '``trunc .. to``' Instruction
8220 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8227 <result> = trunc <ty> <value> to <ty2> ; yields ty2
8232 The '``trunc``' instruction truncates its operand to the type ``ty2``.
8237 The '``trunc``' instruction takes a value to trunc, and a type to trunc
8238 it to. Both types must be of :ref:`integer <t_integer>` types, or vectors
8239 of the same number of integers. The bit size of the ``value`` must be
8240 larger than the bit size of the destination type, ``ty2``. Equal sized
8241 types are not allowed.
8246 The '``trunc``' instruction truncates the high order bits in ``value``
8247 and converts the remaining bits to ``ty2``. Since the source size must
8248 be larger than the destination size, ``trunc`` cannot be a *no-op cast*.
8249 It will always truncate bits.
8254 .. code-block:: llvm
8256 %X = trunc i32 257 to i8 ; yields i8:1
8257 %Y = trunc i32 123 to i1 ; yields i1:true
8258 %Z = trunc i32 122 to i1 ; yields i1:false
8259 %W = trunc <2 x i16> <i16 8, i16 7> to <2 x i8> ; yields <i8 8, i8 7>
8263 '``zext .. to``' Instruction
8264 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8271 <result> = zext <ty> <value> to <ty2> ; yields ty2
8276 The '``zext``' instruction zero extends its operand to type ``ty2``.
8281 The '``zext``' instruction takes a value to cast, and a type to cast it
8282 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
8283 the same number of integers. The bit size of the ``value`` must be
8284 smaller than the bit size of the destination type, ``ty2``.
8289 The ``zext`` fills the high order bits of the ``value`` with zero bits
8290 until it reaches the size of the destination type, ``ty2``.
8292 When zero extending from i1, the result will always be either 0 or 1.
8297 .. code-block:: llvm
8299 %X = zext i32 257 to i64 ; yields i64:257
8300 %Y = zext i1 true to i32 ; yields i32:1
8301 %Z = zext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
8305 '``sext .. to``' Instruction
8306 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8313 <result> = sext <ty> <value> to <ty2> ; yields ty2
8318 The '``sext``' sign extends ``value`` to the type ``ty2``.
8323 The '``sext``' instruction takes a value to cast, and a type to cast it
8324 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
8325 the same number of integers. The bit size of the ``value`` must be
8326 smaller than the bit size of the destination type, ``ty2``.
8331 The '``sext``' instruction performs a sign extension by copying the sign
8332 bit (highest order bit) of the ``value`` until it reaches the bit size
8333 of the type ``ty2``.
8335 When sign extending from i1, the extension always results in -1 or 0.
8340 .. code-block:: llvm
8342 %X = sext i8 -1 to i16 ; yields i16 :65535
8343 %Y = sext i1 true to i32 ; yields i32:-1
8344 %Z = sext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
8346 '``fptrunc .. to``' Instruction
8347 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8354 <result> = fptrunc <ty> <value> to <ty2> ; yields ty2
8359 The '``fptrunc``' instruction truncates ``value`` to type ``ty2``.
8364 The '``fptrunc``' instruction takes a :ref:`floating-point <t_floating>`
8365 value to cast and a :ref:`floating-point <t_floating>` type to cast it to.
8366 The size of ``value`` must be larger than the size of ``ty2``. This
8367 implies that ``fptrunc`` cannot be used to make a *no-op cast*.
8372 The '``fptrunc``' instruction casts a ``value`` from a larger
8373 :ref:`floating-point <t_floating>` type to a smaller :ref:`floating-point
8374 <t_floating>` type. If the value cannot fit (i.e. overflows) within the
8375 destination type, ``ty2``, then the results are undefined. If the cast produces
8376 an inexact result, how rounding is performed (e.g. truncation, also known as
8377 round to zero) is undefined.
8382 .. code-block:: llvm
8384 %X = fptrunc double 123.0 to float ; yields float:123.0
8385 %Y = fptrunc double 1.0E+300 to float ; yields undefined
8387 '``fpext .. to``' Instruction
8388 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8395 <result> = fpext <ty> <value> to <ty2> ; yields ty2
8400 The '``fpext``' extends a floating-point ``value`` to a larger floating-point
8406 The '``fpext``' instruction takes a :ref:`floating-point <t_floating>`
8407 ``value`` to cast, and a :ref:`floating-point <t_floating>` type to cast it
8408 to. The source type must be smaller than the destination type.
8413 The '``fpext``' instruction extends the ``value`` from a smaller
8414 :ref:`floating-point <t_floating>` type to a larger :ref:`floating-point
8415 <t_floating>` type. The ``fpext`` cannot be used to make a
8416 *no-op cast* because it always changes bits. Use ``bitcast`` to make a
8417 *no-op cast* for a floating-point cast.
8422 .. code-block:: llvm
8424 %X = fpext float 3.125 to double ; yields double:3.125000e+00
8425 %Y = fpext double %X to fp128 ; yields fp128:0xL00000000000000004000900000000000
8427 '``fptoui .. to``' Instruction
8428 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8435 <result> = fptoui <ty> <value> to <ty2> ; yields ty2
8440 The '``fptoui``' converts a floating-point ``value`` to its unsigned
8441 integer equivalent of type ``ty2``.
8446 The '``fptoui``' instruction takes a value to cast, which must be a
8447 scalar or vector :ref:`floating-point <t_floating>` value, and a type to
8448 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
8449 ``ty`` is a vector floating-point type, ``ty2`` must be a vector integer
8450 type with the same number of elements as ``ty``
8455 The '``fptoui``' instruction converts its :ref:`floating-point
8456 <t_floating>` operand into the nearest (rounding towards zero)
8457 unsigned integer value. If the value cannot fit in ``ty2``, the results
8463 .. code-block:: llvm
8465 %X = fptoui double 123.0 to i32 ; yields i32:123
8466 %Y = fptoui float 1.0E+300 to i1 ; yields undefined:1
8467 %Z = fptoui float 1.04E+17 to i8 ; yields undefined:1
8469 '``fptosi .. to``' Instruction
8470 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8477 <result> = fptosi <ty> <value> to <ty2> ; yields ty2
8482 The '``fptosi``' instruction converts :ref:`floating-point <t_floating>`
8483 ``value`` to type ``ty2``.
8488 The '``fptosi``' instruction takes a value to cast, which must be a
8489 scalar or vector :ref:`floating-point <t_floating>` value, and a type to
8490 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
8491 ``ty`` is a vector floating-point type, ``ty2`` must be a vector integer
8492 type with the same number of elements as ``ty``
8497 The '``fptosi``' instruction converts its :ref:`floating-point
8498 <t_floating>` operand into the nearest (rounding towards zero)
8499 signed integer value. If the value cannot fit in ``ty2``, the results
8505 .. code-block:: llvm
8507 %X = fptosi double -123.0 to i32 ; yields i32:-123
8508 %Y = fptosi float 1.0E-247 to i1 ; yields undefined:1
8509 %Z = fptosi float 1.04E+17 to i8 ; yields undefined:1
8511 '``uitofp .. to``' Instruction
8512 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8519 <result> = uitofp <ty> <value> to <ty2> ; yields ty2
8524 The '``uitofp``' instruction regards ``value`` as an unsigned integer
8525 and converts that value to the ``ty2`` type.
8530 The '``uitofp``' instruction takes a value to cast, which must be a
8531 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
8532 ``ty2``, which must be an :ref:`floating-point <t_floating>` type. If
8533 ``ty`` is a vector integer type, ``ty2`` must be a vector floating-point
8534 type with the same number of elements as ``ty``
8539 The '``uitofp``' instruction interprets its operand as an unsigned
8540 integer quantity and converts it to the corresponding floating-point
8541 value. If the value cannot fit in the floating-point value, the results
8547 .. code-block:: llvm
8549 %X = uitofp i32 257 to float ; yields float:257.0
8550 %Y = uitofp i8 -1 to double ; yields double:255.0
8552 '``sitofp .. to``' Instruction
8553 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8560 <result> = sitofp <ty> <value> to <ty2> ; yields ty2
8565 The '``sitofp``' instruction regards ``value`` as a signed integer and
8566 converts that value to the ``ty2`` type.
8571 The '``sitofp``' instruction takes a value to cast, which must be a
8572 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
8573 ``ty2``, which must be an :ref:`floating-point <t_floating>` type. If
8574 ``ty`` is a vector integer type, ``ty2`` must be a vector floating-point
8575 type with the same number of elements as ``ty``
8580 The '``sitofp``' instruction interprets its operand as a signed integer
8581 quantity and converts it to the corresponding floating-point value. If
8582 the value cannot fit in the floating-point value, the results are
8588 .. code-block:: llvm
8590 %X = sitofp i32 257 to float ; yields float:257.0
8591 %Y = sitofp i8 -1 to double ; yields double:-1.0
8595 '``ptrtoint .. to``' Instruction
8596 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8603 <result> = ptrtoint <ty> <value> to <ty2> ; yields ty2
8608 The '``ptrtoint``' instruction converts the pointer or a vector of
8609 pointers ``value`` to the integer (or vector of integers) type ``ty2``.
8614 The '``ptrtoint``' instruction takes a ``value`` to cast, which must be
8615 a value of type :ref:`pointer <t_pointer>` or a vector of pointers, and a
8616 type to cast it to ``ty2``, which must be an :ref:`integer <t_integer>` or
8617 a vector of integers type.
8622 The '``ptrtoint``' instruction converts ``value`` to integer type
8623 ``ty2`` by interpreting the pointer value as an integer and either
8624 truncating or zero extending that value to the size of the integer type.
8625 If ``value`` is smaller than ``ty2`` then a zero extension is done. If
8626 ``value`` is larger than ``ty2`` then a truncation is done. If they are
8627 the same size, then nothing is done (*no-op cast*) other than a type
8633 .. code-block:: llvm
8635 %X = ptrtoint i32* %P to i8 ; yields truncation on 32-bit architecture
8636 %Y = ptrtoint i32* %P to i64 ; yields zero extension on 32-bit architecture
8637 %Z = ptrtoint <4 x i32*> %P to <4 x i64>; yields vector zero extension for a vector of addresses on 32-bit architecture
8641 '``inttoptr .. to``' Instruction
8642 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8649 <result> = inttoptr <ty> <value> to <ty2> ; yields ty2
8654 The '``inttoptr``' instruction converts an integer ``value`` to a
8655 pointer type, ``ty2``.
8660 The '``inttoptr``' instruction takes an :ref:`integer <t_integer>` value to
8661 cast, and a type to cast it to, which must be a :ref:`pointer <t_pointer>`
8667 The '``inttoptr``' instruction converts ``value`` to type ``ty2`` by
8668 applying either a zero extension or a truncation depending on the size
8669 of the integer ``value``. If ``value`` is larger than the size of a
8670 pointer then a truncation is done. If ``value`` is smaller than the size
8671 of a pointer then a zero extension is done. If they are the same size,
8672 nothing is done (*no-op cast*).
8677 .. code-block:: llvm
8679 %X = inttoptr i32 255 to i32* ; yields zero extension on 64-bit architecture
8680 %Y = inttoptr i32 255 to i32* ; yields no-op on 32-bit architecture
8681 %Z = inttoptr i64 0 to i32* ; yields truncation on 32-bit architecture
8682 %Z = inttoptr <4 x i32> %G to <4 x i8*>; yields truncation of vector G to four pointers
8686 '``bitcast .. to``' Instruction
8687 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8694 <result> = bitcast <ty> <value> to <ty2> ; yields ty2
8699 The '``bitcast``' instruction converts ``value`` to type ``ty2`` without
8705 The '``bitcast``' instruction takes a value to cast, which must be a
8706 non-aggregate first class value, and a type to cast it to, which must
8707 also be a non-aggregate :ref:`first class <t_firstclass>` type. The
8708 bit sizes of ``value`` and the destination type, ``ty2``, must be
8709 identical. If the source type is a pointer, the destination type must
8710 also be a pointer of the same size. This instruction supports bitwise
8711 conversion of vectors to integers and to vectors of other types (as
8712 long as they have the same size).
8717 The '``bitcast``' instruction converts ``value`` to type ``ty2``. It
8718 is always a *no-op cast* because no bits change with this
8719 conversion. The conversion is done as if the ``value`` had been stored
8720 to memory and read back as type ``ty2``. Pointer (or vector of
8721 pointers) types may only be converted to other pointer (or vector of
8722 pointers) types with the same address space through this instruction.
8723 To convert pointers to other types, use the :ref:`inttoptr <i_inttoptr>`
8724 or :ref:`ptrtoint <i_ptrtoint>` instructions first.
8729 .. code-block:: text
8731 %X = bitcast i8 255 to i8 ; yields i8 :-1
8732 %Y = bitcast i32* %x to sint* ; yields sint*:%x
8733 %Z = bitcast <2 x int> %V to i64; ; yields i64: %V
8734 %Z = bitcast <2 x i32*> %V to <2 x i64*> ; yields <2 x i64*>
8736 .. _i_addrspacecast:
8738 '``addrspacecast .. to``' Instruction
8739 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8746 <result> = addrspacecast <pty> <ptrval> to <pty2> ; yields pty2
8751 The '``addrspacecast``' instruction converts ``ptrval`` from ``pty`` in
8752 address space ``n`` to type ``pty2`` in address space ``m``.
8757 The '``addrspacecast``' instruction takes a pointer or vector of pointer value
8758 to cast and a pointer type to cast it to, which must have a different
8764 The '``addrspacecast``' instruction converts the pointer value
8765 ``ptrval`` to type ``pty2``. It can be a *no-op cast* or a complex
8766 value modification, depending on the target and the address space
8767 pair. Pointer conversions within the same address space must be
8768 performed with the ``bitcast`` instruction. Note that if the address space
8769 conversion is legal then both result and operand refer to the same memory
8775 .. code-block:: llvm
8777 %X = addrspacecast i32* %x to i32 addrspace(1)* ; yields i32 addrspace(1)*:%x
8778 %Y = addrspacecast i32 addrspace(1)* %y to i64 addrspace(2)* ; yields i64 addrspace(2)*:%y
8779 %Z = addrspacecast <4 x i32*> %z to <4 x float addrspace(3)*> ; yields <4 x float addrspace(3)*>:%z
8786 The instructions in this category are the "miscellaneous" instructions,
8787 which defy better classification.
8791 '``icmp``' Instruction
8792 ^^^^^^^^^^^^^^^^^^^^^^
8799 <result> = icmp <cond> <ty> <op1>, <op2> ; yields i1 or <N x i1>:result
8804 The '``icmp``' instruction returns a boolean value or a vector of
8805 boolean values based on comparison of its two integer, integer vector,
8806 pointer, or pointer vector operands.
8811 The '``icmp``' instruction takes three operands. The first operand is
8812 the condition code indicating the kind of comparison to perform. It is
8813 not a value, just a keyword. The possible condition codes are:
8816 #. ``ne``: not equal
8817 #. ``ugt``: unsigned greater than
8818 #. ``uge``: unsigned greater or equal
8819 #. ``ult``: unsigned less than
8820 #. ``ule``: unsigned less or equal
8821 #. ``sgt``: signed greater than
8822 #. ``sge``: signed greater or equal
8823 #. ``slt``: signed less than
8824 #. ``sle``: signed less or equal
8826 The remaining two arguments must be :ref:`integer <t_integer>` or
8827 :ref:`pointer <t_pointer>` or integer :ref:`vector <t_vector>` typed. They
8828 must also be identical types.
8833 The '``icmp``' compares ``op1`` and ``op2`` according to the condition
8834 code given as ``cond``. The comparison performed always yields either an
8835 :ref:`i1 <t_integer>` or vector of ``i1`` result, as follows:
8837 #. ``eq``: yields ``true`` if the operands are equal, ``false``
8838 otherwise. No sign interpretation is necessary or performed.
8839 #. ``ne``: yields ``true`` if the operands are unequal, ``false``
8840 otherwise. No sign interpretation is necessary or performed.
8841 #. ``ugt``: interprets the operands as unsigned values and yields
8842 ``true`` if ``op1`` is greater than ``op2``.
8843 #. ``uge``: interprets the operands as unsigned values and yields
8844 ``true`` if ``op1`` is greater than or equal to ``op2``.
8845 #. ``ult``: interprets the operands as unsigned values and yields
8846 ``true`` if ``op1`` is less than ``op2``.
8847 #. ``ule``: interprets the operands as unsigned values and yields
8848 ``true`` if ``op1`` is less than or equal to ``op2``.
8849 #. ``sgt``: interprets the operands as signed values and yields ``true``
8850 if ``op1`` is greater than ``op2``.
8851 #. ``sge``: interprets the operands as signed values and yields ``true``
8852 if ``op1`` is greater than or equal to ``op2``.
8853 #. ``slt``: interprets the operands as signed values and yields ``true``
8854 if ``op1`` is less than ``op2``.
8855 #. ``sle``: interprets the operands as signed values and yields ``true``
8856 if ``op1`` is less than or equal to ``op2``.
8858 If the operands are :ref:`pointer <t_pointer>` typed, the pointer values
8859 are compared as if they were integers.
8861 If the operands are integer vectors, then they are compared element by
8862 element. The result is an ``i1`` vector with the same number of elements
8863 as the values being compared. Otherwise, the result is an ``i1``.
8868 .. code-block:: text
8870 <result> = icmp eq i32 4, 5 ; yields: result=false
8871 <result> = icmp ne float* %X, %X ; yields: result=false
8872 <result> = icmp ult i16 4, 5 ; yields: result=true
8873 <result> = icmp sgt i16 4, 5 ; yields: result=false
8874 <result> = icmp ule i16 -4, 5 ; yields: result=false
8875 <result> = icmp sge i16 4, 5 ; yields: result=false
8879 '``fcmp``' Instruction
8880 ^^^^^^^^^^^^^^^^^^^^^^
8887 <result> = fcmp [fast-math flags]* <cond> <ty> <op1>, <op2> ; yields i1 or <N x i1>:result
8892 The '``fcmp``' instruction returns a boolean value or vector of boolean
8893 values based on comparison of its operands.
8895 If the operands are floating-point scalars, then the result type is a
8896 boolean (:ref:`i1 <t_integer>`).
8898 If the operands are floating-point vectors, then the result type is a
8899 vector of boolean with the same number of elements as the operands being
8905 The '``fcmp``' instruction takes three operands. The first operand is
8906 the condition code indicating the kind of comparison to perform. It is
8907 not a value, just a keyword. The possible condition codes are:
8909 #. ``false``: no comparison, always returns false
8910 #. ``oeq``: ordered and equal
8911 #. ``ogt``: ordered and greater than
8912 #. ``oge``: ordered and greater than or equal
8913 #. ``olt``: ordered and less than
8914 #. ``ole``: ordered and less than or equal
8915 #. ``one``: ordered and not equal
8916 #. ``ord``: ordered (no nans)
8917 #. ``ueq``: unordered or equal
8918 #. ``ugt``: unordered or greater than
8919 #. ``uge``: unordered or greater than or equal
8920 #. ``ult``: unordered or less than
8921 #. ``ule``: unordered or less than or equal
8922 #. ``une``: unordered or not equal
8923 #. ``uno``: unordered (either nans)
8924 #. ``true``: no comparison, always returns true
8926 *Ordered* means that neither operand is a QNAN while *unordered* means
8927 that either operand may be a QNAN.
8929 Each of ``val1`` and ``val2`` arguments must be either a :ref:`floating-point
8930 <t_floating>` type or a :ref:`vector <t_vector>` of floating-point type.
8931 They must have identical types.
8936 The '``fcmp``' instruction compares ``op1`` and ``op2`` according to the
8937 condition code given as ``cond``. If the operands are vectors, then the
8938 vectors are compared element by element. Each comparison performed
8939 always yields an :ref:`i1 <t_integer>` result, as follows:
8941 #. ``false``: always yields ``false``, regardless of operands.
8942 #. ``oeq``: yields ``true`` if both operands are not a QNAN and ``op1``
8943 is equal to ``op2``.
8944 #. ``ogt``: yields ``true`` if both operands are not a QNAN and ``op1``
8945 is greater than ``op2``.
8946 #. ``oge``: yields ``true`` if both operands are not a QNAN and ``op1``
8947 is greater than or equal to ``op2``.
8948 #. ``olt``: yields ``true`` if both operands are not a QNAN and ``op1``
8949 is less than ``op2``.
8950 #. ``ole``: yields ``true`` if both operands are not a QNAN and ``op1``
8951 is less than or equal to ``op2``.
8952 #. ``one``: yields ``true`` if both operands are not a QNAN and ``op1``
8953 is not equal to ``op2``.
8954 #. ``ord``: yields ``true`` if both operands are not a QNAN.
8955 #. ``ueq``: yields ``true`` if either operand is a QNAN or ``op1`` is
8957 #. ``ugt``: yields ``true`` if either operand is a QNAN or ``op1`` is
8958 greater than ``op2``.
8959 #. ``uge``: yields ``true`` if either operand is a QNAN or ``op1`` is
8960 greater than or equal to ``op2``.
8961 #. ``ult``: yields ``true`` if either operand is a QNAN or ``op1`` is
8963 #. ``ule``: yields ``true`` if either operand is a QNAN or ``op1`` is
8964 less than or equal to ``op2``.
8965 #. ``une``: yields ``true`` if either operand is a QNAN or ``op1`` is
8966 not equal to ``op2``.
8967 #. ``uno``: yields ``true`` if either operand is a QNAN.
8968 #. ``true``: always yields ``true``, regardless of operands.
8970 The ``fcmp`` instruction can also optionally take any number of
8971 :ref:`fast-math flags <fastmath>`, which are optimization hints to enable
8972 otherwise unsafe floating-point optimizations.
8974 Any set of fast-math flags are legal on an ``fcmp`` instruction, but the
8975 only flags that have any effect on its semantics are those that allow
8976 assumptions to be made about the values of input arguments; namely
8977 ``nnan``, ``ninf``, and ``nsz``. See :ref:`fastmath` for more information.
8982 .. code-block:: text
8984 <result> = fcmp oeq float 4.0, 5.0 ; yields: result=false
8985 <result> = fcmp one float 4.0, 5.0 ; yields: result=true
8986 <result> = fcmp olt float 4.0, 5.0 ; yields: result=true
8987 <result> = fcmp ueq double 1.0, 2.0 ; yields: result=false
8991 '``phi``' Instruction
8992 ^^^^^^^^^^^^^^^^^^^^^
8999 <result> = phi <ty> [ <val0>, <label0>], ...
9004 The '``phi``' instruction is used to implement the φ node in the SSA
9005 graph representing the function.
9010 The type of the incoming values is specified with the first type field.
9011 After this, the '``phi``' instruction takes a list of pairs as
9012 arguments, with one pair for each predecessor basic block of the current
9013 block. Only values of :ref:`first class <t_firstclass>` type may be used as
9014 the value arguments to the PHI node. Only labels may be used as the
9017 There must be no non-phi instructions between the start of a basic block
9018 and the PHI instructions: i.e. PHI instructions must be first in a basic
9021 For the purposes of the SSA form, the use of each incoming value is
9022 deemed to occur on the edge from the corresponding predecessor block to
9023 the current block (but after any definition of an '``invoke``'
9024 instruction's return value on the same edge).
9029 At runtime, the '``phi``' instruction logically takes on the value
9030 specified by the pair corresponding to the predecessor basic block that
9031 executed just prior to the current block.
9036 .. code-block:: llvm
9038 Loop: ; Infinite loop that counts from 0 on up...
9039 %indvar = phi i32 [ 0, %LoopHeader ], [ %nextindvar, %Loop ]
9040 %nextindvar = add i32 %indvar, 1
9045 '``select``' Instruction
9046 ^^^^^^^^^^^^^^^^^^^^^^^^
9053 <result> = select selty <cond>, <ty> <val1>, <ty> <val2> ; yields ty
9055 selty is either i1 or {<N x i1>}
9060 The '``select``' instruction is used to choose one value based on a
9061 condition, without IR-level branching.
9066 The '``select``' instruction requires an 'i1' value or a vector of 'i1'
9067 values indicating the condition, and two values of the same :ref:`first
9068 class <t_firstclass>` type.
9073 If the condition is an i1 and it evaluates to 1, the instruction returns
9074 the first value argument; otherwise, it returns the second value
9077 If the condition is a vector of i1, then the value arguments must be
9078 vectors of the same size, and the selection is done element by element.
9080 If the condition is an i1 and the value arguments are vectors of the
9081 same size, then an entire vector is selected.
9086 .. code-block:: llvm
9088 %X = select i1 true, i8 17, i8 42 ; yields i8:17
9092 '``call``' Instruction
9093 ^^^^^^^^^^^^^^^^^^^^^^
9100 <result> = [tail | musttail | notail ] call [fast-math flags] [cconv] [ret attrs] <ty>|<fnty> <fnptrval>(<function args>) [fn attrs]
9106 The '``call``' instruction represents a simple function call.
9111 This instruction requires several arguments:
9113 #. The optional ``tail`` and ``musttail`` markers indicate that the optimizers
9114 should perform tail call optimization. The ``tail`` marker is a hint that
9115 `can be ignored <CodeGenerator.html#sibcallopt>`_. The ``musttail`` marker
9116 means that the call must be tail call optimized in order for the program to
9117 be correct. The ``musttail`` marker provides these guarantees:
9119 #. The call will not cause unbounded stack growth if it is part of a
9120 recursive cycle in the call graph.
9121 #. Arguments with the :ref:`inalloca <attr_inalloca>` attribute are
9124 Both markers imply that the callee does not access allocas from the caller.
9125 The ``tail`` marker additionally implies that the callee does not access
9126 varargs from the caller, while ``musttail`` implies that varargs from the
9127 caller are passed to the callee. Calls marked ``musttail`` must obey the
9128 following additional rules:
9130 - The call must immediately precede a :ref:`ret <i_ret>` instruction,
9131 or a pointer bitcast followed by a ret instruction.
9132 - The ret instruction must return the (possibly bitcasted) value
9133 produced by the call or void.
9134 - The caller and callee prototypes must match. Pointer types of
9135 parameters or return types may differ in pointee type, but not
9137 - The calling conventions of the caller and callee must match.
9138 - All ABI-impacting function attributes, such as sret, byval, inreg,
9139 returned, and inalloca, must match.
9140 - The callee must be varargs iff the caller is varargs. Bitcasting a
9141 non-varargs function to the appropriate varargs type is legal so
9142 long as the non-varargs prefixes obey the other rules.
9144 Tail call optimization for calls marked ``tail`` is guaranteed to occur if
9145 the following conditions are met:
9147 - Caller and callee both have the calling convention ``fastcc``.
9148 - The call is in tail position (ret immediately follows call and ret
9149 uses value of call or is void).
9150 - Option ``-tailcallopt`` is enabled, or
9151 ``llvm::GuaranteedTailCallOpt`` is ``true``.
9152 - `Platform-specific constraints are
9153 met. <CodeGenerator.html#tailcallopt>`_
9155 #. The optional ``notail`` marker indicates that the optimizers should not add
9156 ``tail`` or ``musttail`` markers to the call. It is used to prevent tail
9157 call optimization from being performed on the call.
9159 #. The optional ``fast-math flags`` marker indicates that the call has one or more
9160 :ref:`fast-math flags <fastmath>`, which are optimization hints to enable
9161 otherwise unsafe floating-point optimizations. Fast-math flags are only valid
9162 for calls that return a floating-point scalar or vector type.
9164 #. The optional "cconv" marker indicates which :ref:`calling
9165 convention <callingconv>` the call should use. If none is
9166 specified, the call defaults to using C calling conventions. The
9167 calling convention of the call must match the calling convention of
9168 the target function, or else the behavior is undefined.
9169 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
9170 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
9172 #. '``ty``': the type of the call instruction itself which is also the
9173 type of the return value. Functions that return no value are marked
9175 #. '``fnty``': shall be the signature of the function being called. The
9176 argument types must match the types implied by this signature. This
9177 type can be omitted if the function is not varargs.
9178 #. '``fnptrval``': An LLVM value containing a pointer to a function to
9179 be called. In most cases, this is a direct function call, but
9180 indirect ``call``'s are just as possible, calling an arbitrary pointer
9182 #. '``function args``': argument list whose types match the function
9183 signature argument types and parameter attributes. All arguments must
9184 be of :ref:`first class <t_firstclass>` type. If the function signature
9185 indicates the function accepts a variable number of arguments, the
9186 extra arguments can be specified.
9187 #. The optional :ref:`function attributes <fnattrs>` list.
9188 #. The optional :ref:`operand bundles <opbundles>` list.
9193 The '``call``' instruction is used to cause control flow to transfer to
9194 a specified function, with its incoming arguments bound to the specified
9195 values. Upon a '``ret``' instruction in the called function, control
9196 flow continues with the instruction after the function call, and the
9197 return value of the function is bound to the result argument.
9202 .. code-block:: llvm
9204 %retval = call i32 @test(i32 %argc)
9205 call i32 (i8*, ...)* @printf(i8* %msg, i32 12, i8 42) ; yields i32
9206 %X = tail call i32 @foo() ; yields i32
9207 %Y = tail call fastcc i32 @foo() ; yields i32
9208 call void %foo(i8 97 signext)
9210 %struct.A = type { i32, i8 }
9211 %r = call %struct.A @foo() ; yields { i32, i8 }
9212 %gr = extractvalue %struct.A %r, 0 ; yields i32
9213 %gr1 = extractvalue %struct.A %r, 1 ; yields i8
9214 %Z = call void @foo() noreturn ; indicates that %foo never returns normally
9215 %ZZ = call zeroext i32 @bar() ; Return value is %zero extended
9217 llvm treats calls to some functions with names and arguments that match
9218 the standard C99 library as being the C99 library functions, and may
9219 perform optimizations or generate code for them under that assumption.
9220 This is something we'd like to change in the future to provide better
9221 support for freestanding environments and non-C-based languages.
9225 '``va_arg``' Instruction
9226 ^^^^^^^^^^^^^^^^^^^^^^^^
9233 <resultval> = va_arg <va_list*> <arglist>, <argty>
9238 The '``va_arg``' instruction is used to access arguments passed through
9239 the "variable argument" area of a function call. It is used to implement
9240 the ``va_arg`` macro in C.
9245 This instruction takes a ``va_list*`` value and the type of the
9246 argument. It returns a value of the specified argument type and
9247 increments the ``va_list`` to point to the next argument. The actual
9248 type of ``va_list`` is target specific.
9253 The '``va_arg``' instruction loads an argument of the specified type
9254 from the specified ``va_list`` and causes the ``va_list`` to point to
9255 the next argument. For more information, see the variable argument
9256 handling :ref:`Intrinsic Functions <int_varargs>`.
9258 It is legal for this instruction to be called in a function which does
9259 not take a variable number of arguments, for example, the ``vfprintf``
9262 ``va_arg`` is an LLVM instruction instead of an :ref:`intrinsic
9263 function <intrinsics>` because it takes a type as an argument.
9268 See the :ref:`variable argument processing <int_varargs>` section.
9270 Note that the code generator does not yet fully support va\_arg on many
9271 targets. Also, it does not currently support va\_arg with aggregate
9272 types on any target.
9276 '``landingpad``' Instruction
9277 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9284 <resultval> = landingpad <resultty> <clause>+
9285 <resultval> = landingpad <resultty> cleanup <clause>*
9287 <clause> := catch <type> <value>
9288 <clause> := filter <array constant type> <array constant>
9293 The '``landingpad``' instruction is used by `LLVM's exception handling
9294 system <ExceptionHandling.html#overview>`_ to specify that a basic block
9295 is a landing pad --- one where the exception lands, and corresponds to the
9296 code found in the ``catch`` portion of a ``try``/``catch`` sequence. It
9297 defines values supplied by the :ref:`personality function <personalityfn>` upon
9298 re-entry to the function. The ``resultval`` has the type ``resultty``.
9304 ``cleanup`` flag indicates that the landing pad block is a cleanup.
9306 A ``clause`` begins with the clause type --- ``catch`` or ``filter`` --- and
9307 contains the global variable representing the "type" that may be caught
9308 or filtered respectively. Unlike the ``catch`` clause, the ``filter``
9309 clause takes an array constant as its argument. Use
9310 "``[0 x i8**] undef``" for a filter which cannot throw. The
9311 '``landingpad``' instruction must contain *at least* one ``clause`` or
9312 the ``cleanup`` flag.
9317 The '``landingpad``' instruction defines the values which are set by the
9318 :ref:`personality function <personalityfn>` upon re-entry to the function, and
9319 therefore the "result type" of the ``landingpad`` instruction. As with
9320 calling conventions, how the personality function results are
9321 represented in LLVM IR is target specific.
9323 The clauses are applied in order from top to bottom. If two
9324 ``landingpad`` instructions are merged together through inlining, the
9325 clauses from the calling function are appended to the list of clauses.
9326 When the call stack is being unwound due to an exception being thrown,
9327 the exception is compared against each ``clause`` in turn. If it doesn't
9328 match any of the clauses, and the ``cleanup`` flag is not set, then
9329 unwinding continues further up the call stack.
9331 The ``landingpad`` instruction has several restrictions:
9333 - A landing pad block is a basic block which is the unwind destination
9334 of an '``invoke``' instruction.
9335 - A landing pad block must have a '``landingpad``' instruction as its
9336 first non-PHI instruction.
9337 - There can be only one '``landingpad``' instruction within the landing
9339 - A basic block that is not a landing pad block may not include a
9340 '``landingpad``' instruction.
9345 .. code-block:: llvm
9347 ;; A landing pad which can catch an integer.
9348 %res = landingpad { i8*, i32 }
9350 ;; A landing pad that is a cleanup.
9351 %res = landingpad { i8*, i32 }
9353 ;; A landing pad which can catch an integer and can only throw a double.
9354 %res = landingpad { i8*, i32 }
9356 filter [1 x i8**] [@_ZTId]
9360 '``catchpad``' Instruction
9361 ^^^^^^^^^^^^^^^^^^^^^^^^^^
9368 <resultval> = catchpad within <catchswitch> [<args>*]
9373 The '``catchpad``' instruction is used by `LLVM's exception handling
9374 system <ExceptionHandling.html#overview>`_ to specify that a basic block
9375 begins a catch handler --- one where a personality routine attempts to transfer
9376 control to catch an exception.
9381 The ``catchswitch`` operand must always be a token produced by a
9382 :ref:`catchswitch <i_catchswitch>` instruction in a predecessor block. This
9383 ensures that each ``catchpad`` has exactly one predecessor block, and it always
9384 terminates in a ``catchswitch``.
9386 The ``args`` correspond to whatever information the personality routine
9387 requires to know if this is an appropriate handler for the exception. Control
9388 will transfer to the ``catchpad`` if this is the first appropriate handler for
9391 The ``resultval`` has the type :ref:`token <t_token>` and is used to match the
9392 ``catchpad`` to corresponding :ref:`catchrets <i_catchret>` and other nested EH
9398 When the call stack is being unwound due to an exception being thrown, the
9399 exception is compared against the ``args``. If it doesn't match, control will
9400 not reach the ``catchpad`` instruction. The representation of ``args`` is
9401 entirely target and personality function-specific.
9403 Like the :ref:`landingpad <i_landingpad>` instruction, the ``catchpad``
9404 instruction must be the first non-phi of its parent basic block.
9406 The meaning of the tokens produced and consumed by ``catchpad`` and other "pad"
9407 instructions is described in the
9408 `Windows exception handling documentation\ <ExceptionHandling.html#wineh>`_.
9410 When a ``catchpad`` has been "entered" but not yet "exited" (as
9411 described in the `EH documentation\ <ExceptionHandling.html#wineh-constraints>`_),
9412 it is undefined behavior to execute a :ref:`call <i_call>` or :ref:`invoke <i_invoke>`
9413 that does not carry an appropriate :ref:`"funclet" bundle <ob_funclet>`.
9418 .. code-block:: text
9421 %cs = catchswitch within none [label %handler0] unwind to caller
9422 ;; A catch block which can catch an integer.
9424 %tok = catchpad within %cs [i8** @_ZTIi]
9428 '``cleanuppad``' Instruction
9429 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9436 <resultval> = cleanuppad within <parent> [<args>*]
9441 The '``cleanuppad``' instruction is used by `LLVM's exception handling
9442 system <ExceptionHandling.html#overview>`_ to specify that a basic block
9443 is a cleanup block --- one where a personality routine attempts to
9444 transfer control to run cleanup actions.
9445 The ``args`` correspond to whatever additional
9446 information the :ref:`personality function <personalityfn>` requires to
9447 execute the cleanup.
9448 The ``resultval`` has the type :ref:`token <t_token>` and is used to
9449 match the ``cleanuppad`` to corresponding :ref:`cleanuprets <i_cleanupret>`.
9450 The ``parent`` argument is the token of the funclet that contains the
9451 ``cleanuppad`` instruction. If the ``cleanuppad`` is not inside a funclet,
9452 this operand may be the token ``none``.
9457 The instruction takes a list of arbitrary values which are interpreted
9458 by the :ref:`personality function <personalityfn>`.
9463 When the call stack is being unwound due to an exception being thrown,
9464 the :ref:`personality function <personalityfn>` transfers control to the
9465 ``cleanuppad`` with the aid of the personality-specific arguments.
9466 As with calling conventions, how the personality function results are
9467 represented in LLVM IR is target specific.
9469 The ``cleanuppad`` instruction has several restrictions:
9471 - A cleanup block is a basic block which is the unwind destination of
9472 an exceptional instruction.
9473 - A cleanup block must have a '``cleanuppad``' instruction as its
9474 first non-PHI instruction.
9475 - There can be only one '``cleanuppad``' instruction within the
9477 - A basic block that is not a cleanup block may not include a
9478 '``cleanuppad``' instruction.
9480 When a ``cleanuppad`` has been "entered" but not yet "exited" (as
9481 described in the `EH documentation\ <ExceptionHandling.html#wineh-constraints>`_),
9482 it is undefined behavior to execute a :ref:`call <i_call>` or :ref:`invoke <i_invoke>`
9483 that does not carry an appropriate :ref:`"funclet" bundle <ob_funclet>`.
9488 .. code-block:: text
9490 %tok = cleanuppad within %cs []
9497 LLVM supports the notion of an "intrinsic function". These functions
9498 have well known names and semantics and are required to follow certain
9499 restrictions. Overall, these intrinsics represent an extension mechanism
9500 for the LLVM language that does not require changing all of the
9501 transformations in LLVM when adding to the language (or the bitcode
9502 reader/writer, the parser, etc...).
9504 Intrinsic function names must all start with an "``llvm.``" prefix. This
9505 prefix is reserved in LLVM for intrinsic names; thus, function names may
9506 not begin with this prefix. Intrinsic functions must always be external
9507 functions: you cannot define the body of intrinsic functions. Intrinsic
9508 functions may only be used in call or invoke instructions: it is illegal
9509 to take the address of an intrinsic function. Additionally, because
9510 intrinsic functions are part of the LLVM language, it is required if any
9511 are added that they be documented here.
9513 Some intrinsic functions can be overloaded, i.e., the intrinsic
9514 represents a family of functions that perform the same operation but on
9515 different data types. Because LLVM can represent over 8 million
9516 different integer types, overloading is used commonly to allow an
9517 intrinsic function to operate on any integer type. One or more of the
9518 argument types or the result type can be overloaded to accept any
9519 integer type. Argument types may also be defined as exactly matching a
9520 previous argument's type or the result type. This allows an intrinsic
9521 function which accepts multiple arguments, but needs all of them to be
9522 of the same type, to only be overloaded with respect to a single
9523 argument or the result.
9525 Overloaded intrinsics will have the names of its overloaded argument
9526 types encoded into its function name, each preceded by a period. Only
9527 those types which are overloaded result in a name suffix. Arguments
9528 whose type is matched against another type do not. For example, the
9529 ``llvm.ctpop`` function can take an integer of any width and returns an
9530 integer of exactly the same integer width. This leads to a family of
9531 functions such as ``i8 @llvm.ctpop.i8(i8 %val)`` and
9532 ``i29 @llvm.ctpop.i29(i29 %val)``. Only one type, the return type, is
9533 overloaded, and only one type suffix is required. Because the argument's
9534 type is matched against the return type, it does not require its own
9537 To learn how to add an intrinsic function, please see the `Extending
9538 LLVM Guide <ExtendingLLVM.html>`_.
9542 Variable Argument Handling Intrinsics
9543 -------------------------------------
9545 Variable argument support is defined in LLVM with the
9546 :ref:`va_arg <i_va_arg>` instruction and these three intrinsic
9547 functions. These functions are related to the similarly named macros
9548 defined in the ``<stdarg.h>`` header file.
9550 All of these functions operate on arguments that use a target-specific
9551 value type "``va_list``". The LLVM assembly language reference manual
9552 does not define what this type is, so all transformations should be
9553 prepared to handle these functions regardless of the type used.
9555 This example shows how the :ref:`va_arg <i_va_arg>` instruction and the
9556 variable argument handling intrinsic functions are used.
9558 .. code-block:: llvm
9560 ; This struct is different for every platform. For most platforms,
9561 ; it is merely an i8*.
9562 %struct.va_list = type { i8* }
9564 ; For Unix x86_64 platforms, va_list is the following struct:
9565 ; %struct.va_list = type { i32, i32, i8*, i8* }
9567 define i32 @test(i32 %X, ...) {
9568 ; Initialize variable argument processing
9569 %ap = alloca %struct.va_list
9570 %ap2 = bitcast %struct.va_list* %ap to i8*
9571 call void @llvm.va_start(i8* %ap2)
9573 ; Read a single integer argument
9574 %tmp = va_arg i8* %ap2, i32
9576 ; Demonstrate usage of llvm.va_copy and llvm.va_end
9578 %aq2 = bitcast i8** %aq to i8*
9579 call void @llvm.va_copy(i8* %aq2, i8* %ap2)
9580 call void @llvm.va_end(i8* %aq2)
9582 ; Stop processing of arguments.
9583 call void @llvm.va_end(i8* %ap2)
9587 declare void @llvm.va_start(i8*)
9588 declare void @llvm.va_copy(i8*, i8*)
9589 declare void @llvm.va_end(i8*)
9593 '``llvm.va_start``' Intrinsic
9594 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9601 declare void @llvm.va_start(i8* <arglist>)
9606 The '``llvm.va_start``' intrinsic initializes ``*<arglist>`` for
9607 subsequent use by ``va_arg``.
9612 The argument is a pointer to a ``va_list`` element to initialize.
9617 The '``llvm.va_start``' intrinsic works just like the ``va_start`` macro
9618 available in C. In a target-dependent way, it initializes the
9619 ``va_list`` element to which the argument points, so that the next call
9620 to ``va_arg`` will produce the first variable argument passed to the
9621 function. Unlike the C ``va_start`` macro, this intrinsic does not need
9622 to know the last argument of the function as the compiler can figure
9625 '``llvm.va_end``' Intrinsic
9626 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9633 declare void @llvm.va_end(i8* <arglist>)
9638 The '``llvm.va_end``' intrinsic destroys ``*<arglist>``, which has been
9639 initialized previously with ``llvm.va_start`` or ``llvm.va_copy``.
9644 The argument is a pointer to a ``va_list`` to destroy.
9649 The '``llvm.va_end``' intrinsic works just like the ``va_end`` macro
9650 available in C. In a target-dependent way, it destroys the ``va_list``
9651 element to which the argument points. Calls to
9652 :ref:`llvm.va_start <int_va_start>` and
9653 :ref:`llvm.va_copy <int_va_copy>` must be matched exactly with calls to
9658 '``llvm.va_copy``' Intrinsic
9659 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9666 declare void @llvm.va_copy(i8* <destarglist>, i8* <srcarglist>)
9671 The '``llvm.va_copy``' intrinsic copies the current argument position
9672 from the source argument list to the destination argument list.
9677 The first argument is a pointer to a ``va_list`` element to initialize.
9678 The second argument is a pointer to a ``va_list`` element to copy from.
9683 The '``llvm.va_copy``' intrinsic works just like the ``va_copy`` macro
9684 available in C. In a target-dependent way, it copies the source
9685 ``va_list`` element into the destination ``va_list`` element. This
9686 intrinsic is necessary because the `` llvm.va_start`` intrinsic may be
9687 arbitrarily complex and require, for example, memory allocation.
9689 Accurate Garbage Collection Intrinsics
9690 --------------------------------------
9692 LLVM's support for `Accurate Garbage Collection <GarbageCollection.html>`_
9693 (GC) requires the frontend to generate code containing appropriate intrinsic
9694 calls and select an appropriate GC strategy which knows how to lower these
9695 intrinsics in a manner which is appropriate for the target collector.
9697 These intrinsics allow identification of :ref:`GC roots on the
9698 stack <int_gcroot>`, as well as garbage collector implementations that
9699 require :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers.
9700 Frontends for type-safe garbage collected languages should generate
9701 these intrinsics to make use of the LLVM garbage collectors. For more
9702 details, see `Garbage Collection with LLVM <GarbageCollection.html>`_.
9704 Experimental Statepoint Intrinsics
9705 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9707 LLVM provides an second experimental set of intrinsics for describing garbage
9708 collection safepoints in compiled code. These intrinsics are an alternative
9709 to the ``llvm.gcroot`` intrinsics, but are compatible with the ones for
9710 :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers. The
9711 differences in approach are covered in the `Garbage Collection with LLVM
9712 <GarbageCollection.html>`_ documentation. The intrinsics themselves are
9713 described in :doc:`Statepoints`.
9717 '``llvm.gcroot``' Intrinsic
9718 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9725 declare void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
9730 The '``llvm.gcroot``' intrinsic declares the existence of a GC root to
9731 the code generator, and allows some metadata to be associated with it.
9736 The first argument specifies the address of a stack object that contains
9737 the root pointer. The second pointer (which must be either a constant or
9738 a global value address) contains the meta-data to be associated with the
9744 At runtime, a call to this intrinsic stores a null pointer into the
9745 "ptrloc" location. At compile-time, the code generator generates
9746 information to allow the runtime to find the pointer at GC safe points.
9747 The '``llvm.gcroot``' intrinsic may only be used in a function which
9748 :ref:`specifies a GC algorithm <gc>`.
9752 '``llvm.gcread``' Intrinsic
9753 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9760 declare i8* @llvm.gcread(i8* %ObjPtr, i8** %Ptr)
9765 The '``llvm.gcread``' intrinsic identifies reads of references from heap
9766 locations, allowing garbage collector implementations that require read
9772 The second argument is the address to read from, which should be an
9773 address allocated from the garbage collector. The first object is a
9774 pointer to the start of the referenced object, if needed by the language
9775 runtime (otherwise null).
9780 The '``llvm.gcread``' intrinsic has the same semantics as a load
9781 instruction, but may be replaced with substantially more complex code by
9782 the garbage collector runtime, as needed. The '``llvm.gcread``'
9783 intrinsic may only be used in a function which :ref:`specifies a GC
9788 '``llvm.gcwrite``' Intrinsic
9789 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9796 declare void @llvm.gcwrite(i8* %P1, i8* %Obj, i8** %P2)
9801 The '``llvm.gcwrite``' intrinsic identifies writes of references to heap
9802 locations, allowing garbage collector implementations that require write
9803 barriers (such as generational or reference counting collectors).
9808 The first argument is the reference to store, the second is the start of
9809 the object to store it to, and the third is the address of the field of
9810 Obj to store to. If the runtime does not require a pointer to the
9811 object, Obj may be null.
9816 The '``llvm.gcwrite``' intrinsic has the same semantics as a store
9817 instruction, but may be replaced with substantially more complex code by
9818 the garbage collector runtime, as needed. The '``llvm.gcwrite``'
9819 intrinsic may only be used in a function which :ref:`specifies a GC
9822 Code Generator Intrinsics
9823 -------------------------
9825 These intrinsics are provided by LLVM to expose special features that
9826 may only be implemented with code generator support.
9828 '``llvm.returnaddress``' Intrinsic
9829 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9836 declare i8* @llvm.returnaddress(i32 <level>)
9841 The '``llvm.returnaddress``' intrinsic attempts to compute a
9842 target-specific value indicating the return address of the current
9843 function or one of its callers.
9848 The argument to this intrinsic indicates which function to return the
9849 address for. Zero indicates the calling function, one indicates its
9850 caller, etc. The argument is **required** to be a constant integer
9856 The '``llvm.returnaddress``' intrinsic either returns a pointer
9857 indicating the return address of the specified call frame, or zero if it
9858 cannot be identified. The value returned by this intrinsic is likely to
9859 be incorrect or 0 for arguments other than zero, so it should only be
9860 used for debugging purposes.
9862 Note that calling this intrinsic does not prevent function inlining or
9863 other aggressive transformations, so the value returned may not be that
9864 of the obvious source-language caller.
9866 '``llvm.addressofreturnaddress``' Intrinsic
9867 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9874 declare i8* @llvm.addressofreturnaddress()
9879 The '``llvm.addressofreturnaddress``' intrinsic returns a target-specific
9880 pointer to the place in the stack frame where the return address of the
9881 current function is stored.
9886 Note that calling this intrinsic does not prevent function inlining or
9887 other aggressive transformations, so the value returned may not be that
9888 of the obvious source-language caller.
9890 This intrinsic is only implemented for x86.
9892 '``llvm.frameaddress``' Intrinsic
9893 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9900 declare i8* @llvm.frameaddress(i32 <level>)
9905 The '``llvm.frameaddress``' intrinsic attempts to return the
9906 target-specific frame pointer value for the specified stack frame.
9911 The argument to this intrinsic indicates which function to return the
9912 frame pointer for. Zero indicates the calling function, one indicates
9913 its caller, etc. The argument is **required** to be a constant integer
9919 The '``llvm.frameaddress``' intrinsic either returns a pointer
9920 indicating the frame address of the specified call frame, or zero if it
9921 cannot be identified. The value returned by this intrinsic is likely to
9922 be incorrect or 0 for arguments other than zero, so it should only be
9923 used for debugging purposes.
9925 Note that calling this intrinsic does not prevent function inlining or
9926 other aggressive transformations, so the value returned may not be that
9927 of the obvious source-language caller.
9929 '``llvm.localescape``' and '``llvm.localrecover``' Intrinsics
9930 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9937 declare void @llvm.localescape(...)
9938 declare i8* @llvm.localrecover(i8* %func, i8* %fp, i32 %idx)
9943 The '``llvm.localescape``' intrinsic escapes offsets of a collection of static
9944 allocas, and the '``llvm.localrecover``' intrinsic applies those offsets to a
9945 live frame pointer to recover the address of the allocation. The offset is
9946 computed during frame layout of the caller of ``llvm.localescape``.
9951 All arguments to '``llvm.localescape``' must be pointers to static allocas or
9952 casts of static allocas. Each function can only call '``llvm.localescape``'
9953 once, and it can only do so from the entry block.
9955 The ``func`` argument to '``llvm.localrecover``' must be a constant
9956 bitcasted pointer to a function defined in the current module. The code
9957 generator cannot determine the frame allocation offset of functions defined in
9960 The ``fp`` argument to '``llvm.localrecover``' must be a frame pointer of a
9961 call frame that is currently live. The return value of '``llvm.localaddress``'
9962 is one way to produce such a value, but various runtimes also expose a suitable
9963 pointer in platform-specific ways.
9965 The ``idx`` argument to '``llvm.localrecover``' indicates which alloca passed to
9966 '``llvm.localescape``' to recover. It is zero-indexed.
9971 These intrinsics allow a group of functions to share access to a set of local
9972 stack allocations of a one parent function. The parent function may call the
9973 '``llvm.localescape``' intrinsic once from the function entry block, and the
9974 child functions can use '``llvm.localrecover``' to access the escaped allocas.
9975 The '``llvm.localescape``' intrinsic blocks inlining, as inlining changes where
9976 the escaped allocas are allocated, which would break attempts to use
9977 '``llvm.localrecover``'.
9979 .. _int_read_register:
9980 .. _int_write_register:
9982 '``llvm.read_register``' and '``llvm.write_register``' Intrinsics
9983 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9990 declare i32 @llvm.read_register.i32(metadata)
9991 declare i64 @llvm.read_register.i64(metadata)
9992 declare void @llvm.write_register.i32(metadata, i32 @value)
9993 declare void @llvm.write_register.i64(metadata, i64 @value)
9999 The '``llvm.read_register``' and '``llvm.write_register``' intrinsics
10000 provides access to the named register. The register must be valid on
10001 the architecture being compiled to. The type needs to be compatible
10002 with the register being read.
10007 The '``llvm.read_register``' intrinsic returns the current value of the
10008 register, where possible. The '``llvm.write_register``' intrinsic sets
10009 the current value of the register, where possible.
10011 This is useful to implement named register global variables that need
10012 to always be mapped to a specific register, as is common practice on
10013 bare-metal programs including OS kernels.
10015 The compiler doesn't check for register availability or use of the used
10016 register in surrounding code, including inline assembly. Because of that,
10017 allocatable registers are not supported.
10019 Warning: So far it only works with the stack pointer on selected
10020 architectures (ARM, AArch64, PowerPC and x86_64). Significant amount of
10021 work is needed to support other registers and even more so, allocatable
10026 '``llvm.stacksave``' Intrinsic
10027 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10034 declare i8* @llvm.stacksave()
10039 The '``llvm.stacksave``' intrinsic is used to remember the current state
10040 of the function stack, for use with
10041 :ref:`llvm.stackrestore <int_stackrestore>`. This is useful for
10042 implementing language features like scoped automatic variable sized
10048 This intrinsic returns a opaque pointer value that can be passed to
10049 :ref:`llvm.stackrestore <int_stackrestore>`. When an
10050 ``llvm.stackrestore`` intrinsic is executed with a value saved from
10051 ``llvm.stacksave``, it effectively restores the state of the stack to
10052 the state it was in when the ``llvm.stacksave`` intrinsic executed. In
10053 practice, this pops any :ref:`alloca <i_alloca>` blocks from the stack that
10054 were allocated after the ``llvm.stacksave`` was executed.
10056 .. _int_stackrestore:
10058 '``llvm.stackrestore``' Intrinsic
10059 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10066 declare void @llvm.stackrestore(i8* %ptr)
10071 The '``llvm.stackrestore``' intrinsic is used to restore the state of
10072 the function stack to the state it was in when the corresponding
10073 :ref:`llvm.stacksave <int_stacksave>` intrinsic executed. This is
10074 useful for implementing language features like scoped automatic variable
10075 sized arrays in C99.
10080 See the description for :ref:`llvm.stacksave <int_stacksave>`.
10082 .. _int_get_dynamic_area_offset:
10084 '``llvm.get.dynamic.area.offset``' Intrinsic
10085 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10092 declare i32 @llvm.get.dynamic.area.offset.i32()
10093 declare i64 @llvm.get.dynamic.area.offset.i64()
10098 The '``llvm.get.dynamic.area.offset.*``' intrinsic family is used to
10099 get the offset from native stack pointer to the address of the most
10100 recent dynamic alloca on the caller's stack. These intrinsics are
10101 intendend for use in combination with
10102 :ref:`llvm.stacksave <int_stacksave>` to get a
10103 pointer to the most recent dynamic alloca. This is useful, for example,
10104 for AddressSanitizer's stack unpoisoning routines.
10109 These intrinsics return a non-negative integer value that can be used to
10110 get the address of the most recent dynamic alloca, allocated by :ref:`alloca <i_alloca>`
10111 on the caller's stack. In particular, for targets where stack grows downwards,
10112 adding this offset to the native stack pointer would get the address of the most
10113 recent dynamic alloca. For targets where stack grows upwards, the situation is a bit more
10114 complicated, because subtracting this value from stack pointer would get the address
10115 one past the end of the most recent dynamic alloca.
10117 Although for most targets `llvm.get.dynamic.area.offset <int_get_dynamic_area_offset>`
10118 returns just a zero, for others, such as PowerPC and PowerPC64, it returns a
10119 compile-time-known constant value.
10121 The return value type of :ref:`llvm.get.dynamic.area.offset <int_get_dynamic_area_offset>`
10122 must match the target's default address space's (address space 0) pointer type.
10124 '``llvm.prefetch``' Intrinsic
10125 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10132 declare void @llvm.prefetch(i8* <address>, i32 <rw>, i32 <locality>, i32 <cache type>)
10137 The '``llvm.prefetch``' intrinsic is a hint to the code generator to
10138 insert a prefetch instruction if supported; otherwise, it is a noop.
10139 Prefetches have no effect on the behavior of the program but can change
10140 its performance characteristics.
10145 ``address`` is the address to be prefetched, ``rw`` is the specifier
10146 determining if the fetch should be for a read (0) or write (1), and
10147 ``locality`` is a temporal locality specifier ranging from (0) - no
10148 locality, to (3) - extremely local keep in cache. The ``cache type``
10149 specifies whether the prefetch is performed on the data (1) or
10150 instruction (0) cache. The ``rw``, ``locality`` and ``cache type``
10151 arguments must be constant integers.
10156 This intrinsic does not modify the behavior of the program. In
10157 particular, prefetches cannot trap and do not produce a value. On
10158 targets that support this intrinsic, the prefetch can provide hints to
10159 the processor cache for better performance.
10161 '``llvm.pcmarker``' Intrinsic
10162 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10169 declare void @llvm.pcmarker(i32 <id>)
10174 The '``llvm.pcmarker``' intrinsic is a method to export a Program
10175 Counter (PC) in a region of code to simulators and other tools. The
10176 method is target specific, but it is expected that the marker will use
10177 exported symbols to transmit the PC of the marker. The marker makes no
10178 guarantees that it will remain with any specific instruction after
10179 optimizations. It is possible that the presence of a marker will inhibit
10180 optimizations. The intended use is to be inserted after optimizations to
10181 allow correlations of simulation runs.
10186 ``id`` is a numerical id identifying the marker.
10191 This intrinsic does not modify the behavior of the program. Backends
10192 that do not support this intrinsic may ignore it.
10194 '``llvm.readcyclecounter``' Intrinsic
10195 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10202 declare i64 @llvm.readcyclecounter()
10207 The '``llvm.readcyclecounter``' intrinsic provides access to the cycle
10208 counter register (or similar low latency, high accuracy clocks) on those
10209 targets that support it. On X86, it should map to RDTSC. On Alpha, it
10210 should map to RPCC. As the backing counters overflow quickly (on the
10211 order of 9 seconds on alpha), this should only be used for small
10217 When directly supported, reading the cycle counter should not modify any
10218 memory. Implementations are allowed to either return a application
10219 specific value or a system wide value. On backends without support, this
10220 is lowered to a constant 0.
10222 Note that runtime support may be conditional on the privilege-level code is
10223 running at and the host platform.
10225 '``llvm.clear_cache``' Intrinsic
10226 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10233 declare void @llvm.clear_cache(i8*, i8*)
10238 The '``llvm.clear_cache``' intrinsic ensures visibility of modifications
10239 in the specified range to the execution unit of the processor. On
10240 targets with non-unified instruction and data cache, the implementation
10241 flushes the instruction cache.
10246 On platforms with coherent instruction and data caches (e.g. x86), this
10247 intrinsic is a nop. On platforms with non-coherent instruction and data
10248 cache (e.g. ARM, MIPS), the intrinsic is lowered either to appropriate
10249 instructions or a system call, if cache flushing requires special
10252 The default behavior is to emit a call to ``__clear_cache`` from the run
10255 This instrinsic does *not* empty the instruction pipeline. Modifications
10256 of the current function are outside the scope of the intrinsic.
10258 '``llvm.instrprof.increment``' Intrinsic
10259 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10266 declare void @llvm.instrprof.increment(i8* <name>, i64 <hash>,
10267 i32 <num-counters>, i32 <index>)
10272 The '``llvm.instrprof.increment``' intrinsic can be emitted by a
10273 frontend for use with instrumentation based profiling. These will be
10274 lowered by the ``-instrprof`` pass to generate execution counts of a
10275 program at runtime.
10280 The first argument is a pointer to a global variable containing the
10281 name of the entity being instrumented. This should generally be the
10282 (mangled) function name for a set of counters.
10284 The second argument is a hash value that can be used by the consumer
10285 of the profile data to detect changes to the instrumented source, and
10286 the third is the number of counters associated with ``name``. It is an
10287 error if ``hash`` or ``num-counters`` differ between two instances of
10288 ``instrprof.increment`` that refer to the same name.
10290 The last argument refers to which of the counters for ``name`` should
10291 be incremented. It should be a value between 0 and ``num-counters``.
10296 This intrinsic represents an increment of a profiling counter. It will
10297 cause the ``-instrprof`` pass to generate the appropriate data
10298 structures and the code to increment the appropriate value, in a
10299 format that can be written out by a compiler runtime and consumed via
10300 the ``llvm-profdata`` tool.
10302 '``llvm.instrprof.increment.step``' Intrinsic
10303 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10310 declare void @llvm.instrprof.increment.step(i8* <name>, i64 <hash>,
10311 i32 <num-counters>,
10312 i32 <index>, i64 <step>)
10317 The '``llvm.instrprof.increment.step``' intrinsic is an extension to
10318 the '``llvm.instrprof.increment``' intrinsic with an additional fifth
10319 argument to specify the step of the increment.
10323 The first four arguments are the same as '``llvm.instrprof.increment``'
10326 The last argument specifies the value of the increment of the counter variable.
10330 See description of '``llvm.instrprof.increment``' instrinsic.
10333 '``llvm.instrprof.value.profile``' Intrinsic
10334 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10341 declare void @llvm.instrprof.value.profile(i8* <name>, i64 <hash>,
10342 i64 <value>, i32 <value_kind>,
10348 The '``llvm.instrprof.value.profile``' intrinsic can be emitted by a
10349 frontend for use with instrumentation based profiling. This will be
10350 lowered by the ``-instrprof`` pass to find out the target values,
10351 instrumented expressions take in a program at runtime.
10356 The first argument is a pointer to a global variable containing the
10357 name of the entity being instrumented. ``name`` should generally be the
10358 (mangled) function name for a set of counters.
10360 The second argument is a hash value that can be used by the consumer
10361 of the profile data to detect changes to the instrumented source. It
10362 is an error if ``hash`` differs between two instances of
10363 ``llvm.instrprof.*`` that refer to the same name.
10365 The third argument is the value of the expression being profiled. The profiled
10366 expression's value should be representable as an unsigned 64-bit value. The
10367 fourth argument represents the kind of value profiling that is being done. The
10368 supported value profiling kinds are enumerated through the
10369 ``InstrProfValueKind`` type declared in the
10370 ``<include/llvm/ProfileData/InstrProf.h>`` header file. The last argument is the
10371 index of the instrumented expression within ``name``. It should be >= 0.
10376 This intrinsic represents the point where a call to a runtime routine
10377 should be inserted for value profiling of target expressions. ``-instrprof``
10378 pass will generate the appropriate data structures and replace the
10379 ``llvm.instrprof.value.profile`` intrinsic with the call to the profile
10380 runtime library with proper arguments.
10382 '``llvm.thread.pointer``' Intrinsic
10383 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10390 declare i8* @llvm.thread.pointer()
10395 The '``llvm.thread.pointer``' intrinsic returns the value of the thread
10401 The '``llvm.thread.pointer``' intrinsic returns a pointer to the TLS area
10402 for the current thread. The exact semantics of this value are target
10403 specific: it may point to the start of TLS area, to the end, or somewhere
10404 in the middle. Depending on the target, this intrinsic may read a register,
10405 call a helper function, read from an alternate memory space, or perform
10406 other operations necessary to locate the TLS area. Not all targets support
10409 Standard C Library Intrinsics
10410 -----------------------------
10412 LLVM provides intrinsics for a few important standard C library
10413 functions. These intrinsics allow source-language front-ends to pass
10414 information about the alignment of the pointer arguments to the code
10415 generator, providing opportunity for more efficient code generation.
10419 '``llvm.memcpy``' Intrinsic
10420 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10425 This is an overloaded intrinsic. You can use ``llvm.memcpy`` on any
10426 integer bit width and for different address spaces. Not all targets
10427 support all bit widths however.
10431 declare void @llvm.memcpy.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
10432 i32 <len>, i1 <isvolatile>)
10433 declare void @llvm.memcpy.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
10434 i64 <len>, i1 <isvolatile>)
10439 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
10440 source location to the destination location.
10442 Note that, unlike the standard libc function, the ``llvm.memcpy.*``
10443 intrinsics do not return a value, takes extra isvolatile
10444 arguments and the pointers can be in specified address spaces.
10449 The first argument is a pointer to the destination, the second is a
10450 pointer to the source. The third argument is an integer argument
10451 specifying the number of bytes to copy, and the fourth is a
10452 boolean indicating a volatile access.
10454 The :ref:`align <attr_align>` parameter attribute can be provided
10455 for the first and second arguments.
10457 If the ``isvolatile`` parameter is ``true``, the ``llvm.memcpy`` call is
10458 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
10459 very cleanly specified and it is unwise to depend on it.
10464 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
10465 source location to the destination location, which are not allowed to
10466 overlap. It copies "len" bytes of memory over. If the argument is known
10467 to be aligned to some boundary, this can be specified as the fourth
10468 argument, otherwise it should be set to 0 or 1 (both meaning no alignment).
10472 '``llvm.memmove``' Intrinsic
10473 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10478 This is an overloaded intrinsic. You can use llvm.memmove on any integer
10479 bit width and for different address space. Not all targets support all
10480 bit widths however.
10484 declare void @llvm.memmove.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
10485 i32 <len>, i1 <isvolatile>)
10486 declare void @llvm.memmove.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
10487 i64 <len>, i1 <isvolatile>)
10492 The '``llvm.memmove.*``' intrinsics move a block of memory from the
10493 source location to the destination location. It is similar to the
10494 '``llvm.memcpy``' intrinsic but allows the two memory locations to
10497 Note that, unlike the standard libc function, the ``llvm.memmove.*``
10498 intrinsics do not return a value, takes an extra isvolatile
10499 argument and the pointers can be in specified address spaces.
10504 The first argument is a pointer to the destination, the second is a
10505 pointer to the source. The third argument is an integer argument
10506 specifying the number of bytes to copy, and the fourth is a
10507 boolean indicating a volatile access.
10509 The :ref:`align <attr_align>` parameter attribute can be provided
10510 for the first and second arguments.
10512 If the ``isvolatile`` parameter is ``true``, the ``llvm.memmove`` call
10513 is a :ref:`volatile operation <volatile>`. The detailed access behavior is
10514 not very cleanly specified and it is unwise to depend on it.
10519 The '``llvm.memmove.*``' intrinsics copy a block of memory from the
10520 source location to the destination location, which may overlap. It
10521 copies "len" bytes of memory over. If the argument is known to be
10522 aligned to some boundary, this can be specified as the fourth argument,
10523 otherwise it should be set to 0 or 1 (both meaning no alignment).
10527 '``llvm.memset.*``' Intrinsics
10528 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10533 This is an overloaded intrinsic. You can use llvm.memset on any integer
10534 bit width and for different address spaces. However, not all targets
10535 support all bit widths.
10539 declare void @llvm.memset.p0i8.i32(i8* <dest>, i8 <val>,
10540 i32 <len>, i1 <isvolatile>)
10541 declare void @llvm.memset.p0i8.i64(i8* <dest>, i8 <val>,
10542 i64 <len>, i1 <isvolatile>)
10547 The '``llvm.memset.*``' intrinsics fill a block of memory with a
10548 particular byte value.
10550 Note that, unlike the standard libc function, the ``llvm.memset``
10551 intrinsic does not return a value and takes an extra volatile
10552 argument. Also, the destination can be in an arbitrary address space.
10557 The first argument is a pointer to the destination to fill, the second
10558 is the byte value with which to fill it, the third argument is an
10559 integer argument specifying the number of bytes to fill, and the fourth
10560 is a boolean indicating a volatile access.
10562 The :ref:`align <attr_align>` parameter attribute can be provided
10563 for the first arguments.
10565 If the ``isvolatile`` parameter is ``true``, the ``llvm.memset`` call is
10566 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
10567 very cleanly specified and it is unwise to depend on it.
10572 The '``llvm.memset.*``' intrinsics fill "len" bytes of memory starting
10573 at the destination location.
10575 '``llvm.sqrt.*``' Intrinsic
10576 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10581 This is an overloaded intrinsic. You can use ``llvm.sqrt`` on any
10582 floating-point or vector of floating-point type. Not all targets support
10587 declare float @llvm.sqrt.f32(float %Val)
10588 declare double @llvm.sqrt.f64(double %Val)
10589 declare x86_fp80 @llvm.sqrt.f80(x86_fp80 %Val)
10590 declare fp128 @llvm.sqrt.f128(fp128 %Val)
10591 declare ppc_fp128 @llvm.sqrt.ppcf128(ppc_fp128 %Val)
10596 The '``llvm.sqrt``' intrinsics return the square root of the specified value.
10601 The argument and return value are floating-point numbers of the same type.
10606 Return the same value as a corresponding libm '``sqrt``' function but without
10607 trapping or setting ``errno``. For types specified by IEEE-754, the result
10608 matches a conforming libm implementation.
10610 When specified with the fast-math-flag 'afn', the result may be approximated
10611 using a less accurate calculation.
10613 '``llvm.powi.*``' Intrinsic
10614 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10619 This is an overloaded intrinsic. You can use ``llvm.powi`` on any
10620 floating-point or vector of floating-point type. Not all targets support
10625 declare float @llvm.powi.f32(float %Val, i32 %power)
10626 declare double @llvm.powi.f64(double %Val, i32 %power)
10627 declare x86_fp80 @llvm.powi.f80(x86_fp80 %Val, i32 %power)
10628 declare fp128 @llvm.powi.f128(fp128 %Val, i32 %power)
10629 declare ppc_fp128 @llvm.powi.ppcf128(ppc_fp128 %Val, i32 %power)
10634 The '``llvm.powi.*``' intrinsics return the first operand raised to the
10635 specified (positive or negative) power. The order of evaluation of
10636 multiplications is not defined. When a vector of floating-point type is
10637 used, the second argument remains a scalar integer value.
10642 The second argument is an integer power, and the first is a value to
10643 raise to that power.
10648 This function returns the first value raised to the second power with an
10649 unspecified sequence of rounding operations.
10651 '``llvm.sin.*``' Intrinsic
10652 ^^^^^^^^^^^^^^^^^^^^^^^^^^
10657 This is an overloaded intrinsic. You can use ``llvm.sin`` on any
10658 floating-point or vector of floating-point type. Not all targets support
10663 declare float @llvm.sin.f32(float %Val)
10664 declare double @llvm.sin.f64(double %Val)
10665 declare x86_fp80 @llvm.sin.f80(x86_fp80 %Val)
10666 declare fp128 @llvm.sin.f128(fp128 %Val)
10667 declare ppc_fp128 @llvm.sin.ppcf128(ppc_fp128 %Val)
10672 The '``llvm.sin.*``' intrinsics return the sine of the operand.
10677 The argument and return value are floating-point numbers of the same type.
10682 Return the same value as a corresponding libm '``sin``' function but without
10683 trapping or setting ``errno``.
10685 When specified with the fast-math-flag 'afn', the result may be approximated
10686 using a less accurate calculation.
10688 '``llvm.cos.*``' Intrinsic
10689 ^^^^^^^^^^^^^^^^^^^^^^^^^^
10694 This is an overloaded intrinsic. You can use ``llvm.cos`` on any
10695 floating-point or vector of floating-point type. Not all targets support
10700 declare float @llvm.cos.f32(float %Val)
10701 declare double @llvm.cos.f64(double %Val)
10702 declare x86_fp80 @llvm.cos.f80(x86_fp80 %Val)
10703 declare fp128 @llvm.cos.f128(fp128 %Val)
10704 declare ppc_fp128 @llvm.cos.ppcf128(ppc_fp128 %Val)
10709 The '``llvm.cos.*``' intrinsics return the cosine of the operand.
10714 The argument and return value are floating-point numbers of the same type.
10719 Return the same value as a corresponding libm '``cos``' function but without
10720 trapping or setting ``errno``.
10722 When specified with the fast-math-flag 'afn', the result may be approximated
10723 using a less accurate calculation.
10725 '``llvm.pow.*``' Intrinsic
10726 ^^^^^^^^^^^^^^^^^^^^^^^^^^
10731 This is an overloaded intrinsic. You can use ``llvm.pow`` on any
10732 floating-point or vector of floating-point type. Not all targets support
10737 declare float @llvm.pow.f32(float %Val, float %Power)
10738 declare double @llvm.pow.f64(double %Val, double %Power)
10739 declare x86_fp80 @llvm.pow.f80(x86_fp80 %Val, x86_fp80 %Power)
10740 declare fp128 @llvm.pow.f128(fp128 %Val, fp128 %Power)
10741 declare ppc_fp128 @llvm.pow.ppcf128(ppc_fp128 %Val, ppc_fp128 Power)
10746 The '``llvm.pow.*``' intrinsics return the first operand raised to the
10747 specified (positive or negative) power.
10752 The arguments and return value are floating-point numbers of the same type.
10757 Return the same value as a corresponding libm '``pow``' function but without
10758 trapping or setting ``errno``.
10760 When specified with the fast-math-flag 'afn', the result may be approximated
10761 using a less accurate calculation.
10763 '``llvm.exp.*``' Intrinsic
10764 ^^^^^^^^^^^^^^^^^^^^^^^^^^
10769 This is an overloaded intrinsic. You can use ``llvm.exp`` on any
10770 floating-point or vector of floating-point type. Not all targets support
10775 declare float @llvm.exp.f32(float %Val)
10776 declare double @llvm.exp.f64(double %Val)
10777 declare x86_fp80 @llvm.exp.f80(x86_fp80 %Val)
10778 declare fp128 @llvm.exp.f128(fp128 %Val)
10779 declare ppc_fp128 @llvm.exp.ppcf128(ppc_fp128 %Val)
10784 The '``llvm.exp.*``' intrinsics compute the base-e exponential of the specified
10790 The argument and return value are floating-point numbers of the same type.
10795 Return the same value as a corresponding libm '``exp``' function but without
10796 trapping or setting ``errno``.
10798 When specified with the fast-math-flag 'afn', the result may be approximated
10799 using a less accurate calculation.
10801 '``llvm.exp2.*``' Intrinsic
10802 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10807 This is an overloaded intrinsic. You can use ``llvm.exp2`` on any
10808 floating-point or vector of floating-point type. Not all targets support
10813 declare float @llvm.exp2.f32(float %Val)
10814 declare double @llvm.exp2.f64(double %Val)
10815 declare x86_fp80 @llvm.exp2.f80(x86_fp80 %Val)
10816 declare fp128 @llvm.exp2.f128(fp128 %Val)
10817 declare ppc_fp128 @llvm.exp2.ppcf128(ppc_fp128 %Val)
10822 The '``llvm.exp2.*``' intrinsics compute the base-2 exponential of the
10828 The argument and return value are floating-point numbers of the same type.
10833 Return the same value as a corresponding libm '``exp2``' function but without
10834 trapping or setting ``errno``.
10836 When specified with the fast-math-flag 'afn', the result may be approximated
10837 using a less accurate calculation.
10839 '``llvm.log.*``' Intrinsic
10840 ^^^^^^^^^^^^^^^^^^^^^^^^^^
10845 This is an overloaded intrinsic. You can use ``llvm.log`` on any
10846 floating-point or vector of floating-point type. Not all targets support
10851 declare float @llvm.log.f32(float %Val)
10852 declare double @llvm.log.f64(double %Val)
10853 declare x86_fp80 @llvm.log.f80(x86_fp80 %Val)
10854 declare fp128 @llvm.log.f128(fp128 %Val)
10855 declare ppc_fp128 @llvm.log.ppcf128(ppc_fp128 %Val)
10860 The '``llvm.log.*``' intrinsics compute the base-e logarithm of the specified
10866 The argument and return value are floating-point numbers of the same type.
10871 Return the same value as a corresponding libm '``log``' function but without
10872 trapping or setting ``errno``.
10874 When specified with the fast-math-flag 'afn', the result may be approximated
10875 using a less accurate calculation.
10877 '``llvm.log10.*``' Intrinsic
10878 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10883 This is an overloaded intrinsic. You can use ``llvm.log10`` on any
10884 floating-point or vector of floating-point type. Not all targets support
10889 declare float @llvm.log10.f32(float %Val)
10890 declare double @llvm.log10.f64(double %Val)
10891 declare x86_fp80 @llvm.log10.f80(x86_fp80 %Val)
10892 declare fp128 @llvm.log10.f128(fp128 %Val)
10893 declare ppc_fp128 @llvm.log10.ppcf128(ppc_fp128 %Val)
10898 The '``llvm.log10.*``' intrinsics compute the base-10 logarithm of the
10904 The argument and return value are floating-point numbers of the same type.
10909 Return the same value as a corresponding libm '``log10``' function but without
10910 trapping or setting ``errno``.
10912 When specified with the fast-math-flag 'afn', the result may be approximated
10913 using a less accurate calculation.
10915 '``llvm.log2.*``' Intrinsic
10916 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10921 This is an overloaded intrinsic. You can use ``llvm.log2`` on any
10922 floating-point or vector of floating-point type. Not all targets support
10927 declare float @llvm.log2.f32(float %Val)
10928 declare double @llvm.log2.f64(double %Val)
10929 declare x86_fp80 @llvm.log2.f80(x86_fp80 %Val)
10930 declare fp128 @llvm.log2.f128(fp128 %Val)
10931 declare ppc_fp128 @llvm.log2.ppcf128(ppc_fp128 %Val)
10936 The '``llvm.log2.*``' intrinsics compute the base-2 logarithm of the specified
10942 The argument and return value are floating-point numbers of the same type.
10947 Return the same value as a corresponding libm '``log2``' function but without
10948 trapping or setting ``errno``.
10950 When specified with the fast-math-flag 'afn', the result may be approximated
10951 using a less accurate calculation.
10953 '``llvm.fma.*``' Intrinsic
10954 ^^^^^^^^^^^^^^^^^^^^^^^^^^
10959 This is an overloaded intrinsic. You can use ``llvm.fma`` on any
10960 floating-point or vector of floating-point type. Not all targets support
10965 declare float @llvm.fma.f32(float %a, float %b, float %c)
10966 declare double @llvm.fma.f64(double %a, double %b, double %c)
10967 declare x86_fp80 @llvm.fma.f80(x86_fp80 %a, x86_fp80 %b, x86_fp80 %c)
10968 declare fp128 @llvm.fma.f128(fp128 %a, fp128 %b, fp128 %c)
10969 declare ppc_fp128 @llvm.fma.ppcf128(ppc_fp128 %a, ppc_fp128 %b, ppc_fp128 %c)
10974 The '``llvm.fma.*``' intrinsics perform the fused multiply-add operation.
10979 The arguments and return value are floating-point numbers of the same type.
10984 Return the same value as a corresponding libm '``fma``' function but without
10985 trapping or setting ``errno``.
10987 When specified with the fast-math-flag 'afn', the result may be approximated
10988 using a less accurate calculation.
10990 '``llvm.fabs.*``' Intrinsic
10991 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10996 This is an overloaded intrinsic. You can use ``llvm.fabs`` on any
10997 floating-point or vector of floating-point type. Not all targets support
11002 declare float @llvm.fabs.f32(float %Val)
11003 declare double @llvm.fabs.f64(double %Val)
11004 declare x86_fp80 @llvm.fabs.f80(x86_fp80 %Val)
11005 declare fp128 @llvm.fabs.f128(fp128 %Val)
11006 declare ppc_fp128 @llvm.fabs.ppcf128(ppc_fp128 %Val)
11011 The '``llvm.fabs.*``' intrinsics return the absolute value of the
11017 The argument and return value are floating-point numbers of the same
11023 This function returns the same values as the libm ``fabs`` functions
11024 would, and handles error conditions in the same way.
11026 '``llvm.minnum.*``' Intrinsic
11027 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11032 This is an overloaded intrinsic. You can use ``llvm.minnum`` on any
11033 floating-point or vector of floating-point type. Not all targets support
11038 declare float @llvm.minnum.f32(float %Val0, float %Val1)
11039 declare double @llvm.minnum.f64(double %Val0, double %Val1)
11040 declare x86_fp80 @llvm.minnum.f80(x86_fp80 %Val0, x86_fp80 %Val1)
11041 declare fp128 @llvm.minnum.f128(fp128 %Val0, fp128 %Val1)
11042 declare ppc_fp128 @llvm.minnum.ppcf128(ppc_fp128 %Val0, ppc_fp128 %Val1)
11047 The '``llvm.minnum.*``' intrinsics return the minimum of the two
11054 The arguments and return value are floating-point numbers of the same
11060 Follows the IEEE-754 semantics for minNum, which also match for libm's
11063 If either operand is a NaN, returns the other non-NaN operand. Returns
11064 NaN only if both operands are NaN. If the operands compare equal,
11065 returns a value that compares equal to both operands. This means that
11066 fmin(+/-0.0, +/-0.0) could return either -0.0 or 0.0.
11068 '``llvm.maxnum.*``' Intrinsic
11069 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11074 This is an overloaded intrinsic. You can use ``llvm.maxnum`` on any
11075 floating-point or vector of floating-point type. Not all targets support
11080 declare float @llvm.maxnum.f32(float %Val0, float %Val1l)
11081 declare double @llvm.maxnum.f64(double %Val0, double %Val1)
11082 declare x86_fp80 @llvm.maxnum.f80(x86_fp80 %Val0, x86_fp80 %Val1)
11083 declare fp128 @llvm.maxnum.f128(fp128 %Val0, fp128 %Val1)
11084 declare ppc_fp128 @llvm.maxnum.ppcf128(ppc_fp128 %Val0, ppc_fp128 %Val1)
11089 The '``llvm.maxnum.*``' intrinsics return the maximum of the two
11096 The arguments and return value are floating-point numbers of the same
11101 Follows the IEEE-754 semantics for maxNum, which also match for libm's
11104 If either operand is a NaN, returns the other non-NaN operand. Returns
11105 NaN only if both operands are NaN. If the operands compare equal,
11106 returns a value that compares equal to both operands. This means that
11107 fmax(+/-0.0, +/-0.0) could return either -0.0 or 0.0.
11109 '``llvm.copysign.*``' Intrinsic
11110 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11115 This is an overloaded intrinsic. You can use ``llvm.copysign`` on any
11116 floating-point or vector of floating-point type. Not all targets support
11121 declare float @llvm.copysign.f32(float %Mag, float %Sgn)
11122 declare double @llvm.copysign.f64(double %Mag, double %Sgn)
11123 declare x86_fp80 @llvm.copysign.f80(x86_fp80 %Mag, x86_fp80 %Sgn)
11124 declare fp128 @llvm.copysign.f128(fp128 %Mag, fp128 %Sgn)
11125 declare ppc_fp128 @llvm.copysign.ppcf128(ppc_fp128 %Mag, ppc_fp128 %Sgn)
11130 The '``llvm.copysign.*``' intrinsics return a value with the magnitude of the
11131 first operand and the sign of the second operand.
11136 The arguments and return value are floating-point numbers of the same
11142 This function returns the same values as the libm ``copysign``
11143 functions would, and handles error conditions in the same way.
11145 '``llvm.floor.*``' Intrinsic
11146 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11151 This is an overloaded intrinsic. You can use ``llvm.floor`` on any
11152 floating-point or vector of floating-point type. Not all targets support
11157 declare float @llvm.floor.f32(float %Val)
11158 declare double @llvm.floor.f64(double %Val)
11159 declare x86_fp80 @llvm.floor.f80(x86_fp80 %Val)
11160 declare fp128 @llvm.floor.f128(fp128 %Val)
11161 declare ppc_fp128 @llvm.floor.ppcf128(ppc_fp128 %Val)
11166 The '``llvm.floor.*``' intrinsics return the floor of the operand.
11171 The argument and return value are floating-point numbers of the same
11177 This function returns the same values as the libm ``floor`` functions
11178 would, and handles error conditions in the same way.
11180 '``llvm.ceil.*``' Intrinsic
11181 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
11186 This is an overloaded intrinsic. You can use ``llvm.ceil`` on any
11187 floating-point or vector of floating-point type. Not all targets support
11192 declare float @llvm.ceil.f32(float %Val)
11193 declare double @llvm.ceil.f64(double %Val)
11194 declare x86_fp80 @llvm.ceil.f80(x86_fp80 %Val)
11195 declare fp128 @llvm.ceil.f128(fp128 %Val)
11196 declare ppc_fp128 @llvm.ceil.ppcf128(ppc_fp128 %Val)
11201 The '``llvm.ceil.*``' intrinsics return the ceiling of the operand.
11206 The argument and return value are floating-point numbers of the same
11212 This function returns the same values as the libm ``ceil`` functions
11213 would, and handles error conditions in the same way.
11215 '``llvm.trunc.*``' Intrinsic
11216 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11221 This is an overloaded intrinsic. You can use ``llvm.trunc`` on any
11222 floating-point or vector of floating-point type. Not all targets support
11227 declare float @llvm.trunc.f32(float %Val)
11228 declare double @llvm.trunc.f64(double %Val)
11229 declare x86_fp80 @llvm.trunc.f80(x86_fp80 %Val)
11230 declare fp128 @llvm.trunc.f128(fp128 %Val)
11231 declare ppc_fp128 @llvm.trunc.ppcf128(ppc_fp128 %Val)
11236 The '``llvm.trunc.*``' intrinsics returns the operand rounded to the
11237 nearest integer not larger in magnitude than the operand.
11242 The argument and return value are floating-point numbers of the same
11248 This function returns the same values as the libm ``trunc`` functions
11249 would, and handles error conditions in the same way.
11251 '``llvm.rint.*``' Intrinsic
11252 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
11257 This is an overloaded intrinsic. You can use ``llvm.rint`` on any
11258 floating-point or vector of floating-point type. Not all targets support
11263 declare float @llvm.rint.f32(float %Val)
11264 declare double @llvm.rint.f64(double %Val)
11265 declare x86_fp80 @llvm.rint.f80(x86_fp80 %Val)
11266 declare fp128 @llvm.rint.f128(fp128 %Val)
11267 declare ppc_fp128 @llvm.rint.ppcf128(ppc_fp128 %Val)
11272 The '``llvm.rint.*``' intrinsics returns the operand rounded to the
11273 nearest integer. It may raise an inexact floating-point exception if the
11274 operand isn't an integer.
11279 The argument and return value are floating-point numbers of the same
11285 This function returns the same values as the libm ``rint`` functions
11286 would, and handles error conditions in the same way.
11288 '``llvm.nearbyint.*``' Intrinsic
11289 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11294 This is an overloaded intrinsic. You can use ``llvm.nearbyint`` on any
11295 floating-point or vector of floating-point type. Not all targets support
11300 declare float @llvm.nearbyint.f32(float %Val)
11301 declare double @llvm.nearbyint.f64(double %Val)
11302 declare x86_fp80 @llvm.nearbyint.f80(x86_fp80 %Val)
11303 declare fp128 @llvm.nearbyint.f128(fp128 %Val)
11304 declare ppc_fp128 @llvm.nearbyint.ppcf128(ppc_fp128 %Val)
11309 The '``llvm.nearbyint.*``' intrinsics returns the operand rounded to the
11315 The argument and return value are floating-point numbers of the same
11321 This function returns the same values as the libm ``nearbyint``
11322 functions would, and handles error conditions in the same way.
11324 '``llvm.round.*``' Intrinsic
11325 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11330 This is an overloaded intrinsic. You can use ``llvm.round`` on any
11331 floating-point or vector of floating-point type. Not all targets support
11336 declare float @llvm.round.f32(float %Val)
11337 declare double @llvm.round.f64(double %Val)
11338 declare x86_fp80 @llvm.round.f80(x86_fp80 %Val)
11339 declare fp128 @llvm.round.f128(fp128 %Val)
11340 declare ppc_fp128 @llvm.round.ppcf128(ppc_fp128 %Val)
11345 The '``llvm.round.*``' intrinsics returns the operand rounded to the
11351 The argument and return value are floating-point numbers of the same
11357 This function returns the same values as the libm ``round``
11358 functions would, and handles error conditions in the same way.
11360 Bit Manipulation Intrinsics
11361 ---------------------------
11363 LLVM provides intrinsics for a few important bit manipulation
11364 operations. These allow efficient code generation for some algorithms.
11366 '``llvm.bitreverse.*``' Intrinsics
11367 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11372 This is an overloaded intrinsic function. You can use bitreverse on any
11377 declare i16 @llvm.bitreverse.i16(i16 <id>)
11378 declare i32 @llvm.bitreverse.i32(i32 <id>)
11379 declare i64 @llvm.bitreverse.i64(i64 <id>)
11384 The '``llvm.bitreverse``' family of intrinsics is used to reverse the
11385 bitpattern of an integer value; for example ``0b10110110`` becomes
11391 The ``llvm.bitreverse.iN`` intrinsic returns an iN value that has bit
11392 ``M`` in the input moved to bit ``N-M`` in the output.
11394 '``llvm.bswap.*``' Intrinsics
11395 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11400 This is an overloaded intrinsic function. You can use bswap on any
11401 integer type that is an even number of bytes (i.e. BitWidth % 16 == 0).
11405 declare i16 @llvm.bswap.i16(i16 <id>)
11406 declare i32 @llvm.bswap.i32(i32 <id>)
11407 declare i64 @llvm.bswap.i64(i64 <id>)
11412 The '``llvm.bswap``' family of intrinsics is used to byte swap integer
11413 values with an even number of bytes (positive multiple of 16 bits).
11414 These are useful for performing operations on data that is not in the
11415 target's native byte order.
11420 The ``llvm.bswap.i16`` intrinsic returns an i16 value that has the high
11421 and low byte of the input i16 swapped. Similarly, the ``llvm.bswap.i32``
11422 intrinsic returns an i32 value that has the four bytes of the input i32
11423 swapped, so that if the input bytes are numbered 0, 1, 2, 3 then the
11424 returned i32 will have its bytes in 3, 2, 1, 0 order. The
11425 ``llvm.bswap.i48``, ``llvm.bswap.i64`` and other intrinsics extend this
11426 concept to additional even-byte lengths (6 bytes, 8 bytes and more,
11429 '``llvm.ctpop.*``' Intrinsic
11430 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11435 This is an overloaded intrinsic. You can use llvm.ctpop on any integer
11436 bit width, or on any vector with integer elements. Not all targets
11437 support all bit widths or vector types, however.
11441 declare i8 @llvm.ctpop.i8(i8 <src>)
11442 declare i16 @llvm.ctpop.i16(i16 <src>)
11443 declare i32 @llvm.ctpop.i32(i32 <src>)
11444 declare i64 @llvm.ctpop.i64(i64 <src>)
11445 declare i256 @llvm.ctpop.i256(i256 <src>)
11446 declare <2 x i32> @llvm.ctpop.v2i32(<2 x i32> <src>)
11451 The '``llvm.ctpop``' family of intrinsics counts the number of bits set
11457 The only argument is the value to be counted. The argument may be of any
11458 integer type, or a vector with integer elements. The return type must
11459 match the argument type.
11464 The '``llvm.ctpop``' intrinsic counts the 1's in a variable, or within
11465 each element of a vector.
11467 '``llvm.ctlz.*``' Intrinsic
11468 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
11473 This is an overloaded intrinsic. You can use ``llvm.ctlz`` on any
11474 integer bit width, or any vector whose elements are integers. Not all
11475 targets support all bit widths or vector types, however.
11479 declare i8 @llvm.ctlz.i8 (i8 <src>, i1 <is_zero_undef>)
11480 declare i16 @llvm.ctlz.i16 (i16 <src>, i1 <is_zero_undef>)
11481 declare i32 @llvm.ctlz.i32 (i32 <src>, i1 <is_zero_undef>)
11482 declare i64 @llvm.ctlz.i64 (i64 <src>, i1 <is_zero_undef>)
11483 declare i256 @llvm.ctlz.i256(i256 <src>, i1 <is_zero_undef>)
11484 declare <2 x i32> @llvm.ctlz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
11489 The '``llvm.ctlz``' family of intrinsic functions counts the number of
11490 leading zeros in a variable.
11495 The first argument is the value to be counted. This argument may be of
11496 any integer type, or a vector with integer element type. The return
11497 type must match the first argument type.
11499 The second argument must be a constant and is a flag to indicate whether
11500 the intrinsic should ensure that a zero as the first argument produces a
11501 defined result. Historically some architectures did not provide a
11502 defined result for zero values as efficiently, and many algorithms are
11503 now predicated on avoiding zero-value inputs.
11508 The '``llvm.ctlz``' intrinsic counts the leading (most significant)
11509 zeros in a variable, or within each element of the vector. If
11510 ``src == 0`` then the result is the size in bits of the type of ``src``
11511 if ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
11512 ``llvm.ctlz(i32 2) = 30``.
11514 '``llvm.cttz.*``' Intrinsic
11515 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
11520 This is an overloaded intrinsic. You can use ``llvm.cttz`` on any
11521 integer bit width, or any vector of integer elements. Not all targets
11522 support all bit widths or vector types, however.
11526 declare i8 @llvm.cttz.i8 (i8 <src>, i1 <is_zero_undef>)
11527 declare i16 @llvm.cttz.i16 (i16 <src>, i1 <is_zero_undef>)
11528 declare i32 @llvm.cttz.i32 (i32 <src>, i1 <is_zero_undef>)
11529 declare i64 @llvm.cttz.i64 (i64 <src>, i1 <is_zero_undef>)
11530 declare i256 @llvm.cttz.i256(i256 <src>, i1 <is_zero_undef>)
11531 declare <2 x i32> @llvm.cttz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
11536 The '``llvm.cttz``' family of intrinsic functions counts the number of
11542 The first argument is the value to be counted. This argument may be of
11543 any integer type, or a vector with integer element type. The return
11544 type must match the first argument type.
11546 The second argument must be a constant and is a flag to indicate whether
11547 the intrinsic should ensure that a zero as the first argument produces a
11548 defined result. Historically some architectures did not provide a
11549 defined result for zero values as efficiently, and many algorithms are
11550 now predicated on avoiding zero-value inputs.
11555 The '``llvm.cttz``' intrinsic counts the trailing (least significant)
11556 zeros in a variable, or within each element of a vector. If ``src == 0``
11557 then the result is the size in bits of the type of ``src`` if
11558 ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
11559 ``llvm.cttz(2) = 1``.
11563 Arithmetic with Overflow Intrinsics
11564 -----------------------------------
11566 LLVM provides intrinsics for fast arithmetic overflow checking.
11568 Each of these intrinsics returns a two-element struct. The first
11569 element of this struct contains the result of the corresponding
11570 arithmetic operation modulo 2\ :sup:`n`\ , where n is the bit width of
11571 the result. Therefore, for example, the first element of the struct
11572 returned by ``llvm.sadd.with.overflow.i32`` is always the same as the
11573 result of a 32-bit ``add`` instruction with the same operands, where
11574 the ``add`` is *not* modified by an ``nsw`` or ``nuw`` flag.
11576 The second element of the result is an ``i1`` that is 1 if the
11577 arithmetic operation overflowed and 0 otherwise. An operation
11578 overflows if, for any values of its operands ``A`` and ``B`` and for
11579 any ``N`` larger than the operands' width, ``ext(A op B) to iN`` is
11580 not equal to ``(ext(A) to iN) op (ext(B) to iN)`` where ``ext`` is
11581 ``sext`` for signed overflow and ``zext`` for unsigned overflow, and
11582 ``op`` is the underlying arithmetic operation.
11584 The behavior of these intrinsics is well-defined for all argument
11587 '``llvm.sadd.with.overflow.*``' Intrinsics
11588 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11593 This is an overloaded intrinsic. You can use ``llvm.sadd.with.overflow``
11594 on any integer bit width.
11598 declare {i16, i1} @llvm.sadd.with.overflow.i16(i16 %a, i16 %b)
11599 declare {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
11600 declare {i64, i1} @llvm.sadd.with.overflow.i64(i64 %a, i64 %b)
11605 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
11606 a signed addition of the two arguments, and indicate whether an overflow
11607 occurred during the signed summation.
11612 The arguments (%a and %b) and the first element of the result structure
11613 may be of integer types of any bit width, but they must have the same
11614 bit width. The second element of the result structure must be of type
11615 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
11621 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
11622 a signed addition of the two variables. They return a structure --- the
11623 first element of which is the signed summation, and the second element
11624 of which is a bit specifying if the signed summation resulted in an
11630 .. code-block:: llvm
11632 %res = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
11633 %sum = extractvalue {i32, i1} %res, 0
11634 %obit = extractvalue {i32, i1} %res, 1
11635 br i1 %obit, label %overflow, label %normal
11637 '``llvm.uadd.with.overflow.*``' Intrinsics
11638 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11643 This is an overloaded intrinsic. You can use ``llvm.uadd.with.overflow``
11644 on any integer bit width.
11648 declare {i16, i1} @llvm.uadd.with.overflow.i16(i16 %a, i16 %b)
11649 declare {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
11650 declare {i64, i1} @llvm.uadd.with.overflow.i64(i64 %a, i64 %b)
11655 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
11656 an unsigned addition of the two arguments, and indicate whether a carry
11657 occurred during the unsigned summation.
11662 The arguments (%a and %b) and the first element of the result structure
11663 may be of integer types of any bit width, but they must have the same
11664 bit width. The second element of the result structure must be of type
11665 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
11671 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
11672 an unsigned addition of the two arguments. They return a structure --- the
11673 first element of which is the sum, and the second element of which is a
11674 bit specifying if the unsigned summation resulted in a carry.
11679 .. code-block:: llvm
11681 %res = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
11682 %sum = extractvalue {i32, i1} %res, 0
11683 %obit = extractvalue {i32, i1} %res, 1
11684 br i1 %obit, label %carry, label %normal
11686 '``llvm.ssub.with.overflow.*``' Intrinsics
11687 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11692 This is an overloaded intrinsic. You can use ``llvm.ssub.with.overflow``
11693 on any integer bit width.
11697 declare {i16, i1} @llvm.ssub.with.overflow.i16(i16 %a, i16 %b)
11698 declare {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
11699 declare {i64, i1} @llvm.ssub.with.overflow.i64(i64 %a, i64 %b)
11704 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
11705 a signed subtraction of the two arguments, and indicate whether an
11706 overflow occurred during the signed subtraction.
11711 The arguments (%a and %b) and the first element of the result structure
11712 may be of integer types of any bit width, but they must have the same
11713 bit width. The second element of the result structure must be of type
11714 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
11720 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
11721 a signed subtraction of the two arguments. They return a structure --- the
11722 first element of which is the subtraction, and the second element of
11723 which is a bit specifying if the signed subtraction resulted in an
11729 .. code-block:: llvm
11731 %res = call {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
11732 %sum = extractvalue {i32, i1} %res, 0
11733 %obit = extractvalue {i32, i1} %res, 1
11734 br i1 %obit, label %overflow, label %normal
11736 '``llvm.usub.with.overflow.*``' Intrinsics
11737 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11742 This is an overloaded intrinsic. You can use ``llvm.usub.with.overflow``
11743 on any integer bit width.
11747 declare {i16, i1} @llvm.usub.with.overflow.i16(i16 %a, i16 %b)
11748 declare {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
11749 declare {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
11754 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
11755 an unsigned subtraction of the two arguments, and indicate whether an
11756 overflow occurred during the unsigned subtraction.
11761 The arguments (%a and %b) and the first element of the result structure
11762 may be of integer types of any bit width, but they must have the same
11763 bit width. The second element of the result structure must be of type
11764 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
11770 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
11771 an unsigned subtraction of the two arguments. They return a structure ---
11772 the first element of which is the subtraction, and the second element of
11773 which is a bit specifying if the unsigned subtraction resulted in an
11779 .. code-block:: llvm
11781 %res = call {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
11782 %sum = extractvalue {i32, i1} %res, 0
11783 %obit = extractvalue {i32, i1} %res, 1
11784 br i1 %obit, label %overflow, label %normal
11786 '``llvm.smul.with.overflow.*``' Intrinsics
11787 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11792 This is an overloaded intrinsic. You can use ``llvm.smul.with.overflow``
11793 on any integer bit width.
11797 declare {i16, i1} @llvm.smul.with.overflow.i16(i16 %a, i16 %b)
11798 declare {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
11799 declare {i64, i1} @llvm.smul.with.overflow.i64(i64 %a, i64 %b)
11804 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
11805 a signed multiplication of the two arguments, and indicate whether an
11806 overflow occurred during the signed multiplication.
11811 The arguments (%a and %b) and the first element of the result structure
11812 may be of integer types of any bit width, but they must have the same
11813 bit width. The second element of the result structure must be of type
11814 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
11820 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
11821 a signed multiplication of the two arguments. They return a structure ---
11822 the first element of which is the multiplication, and the second element
11823 of which is a bit specifying if the signed multiplication resulted in an
11829 .. code-block:: llvm
11831 %res = call {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
11832 %sum = extractvalue {i32, i1} %res, 0
11833 %obit = extractvalue {i32, i1} %res, 1
11834 br i1 %obit, label %overflow, label %normal
11836 '``llvm.umul.with.overflow.*``' Intrinsics
11837 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11842 This is an overloaded intrinsic. You can use ``llvm.umul.with.overflow``
11843 on any integer bit width.
11847 declare {i16, i1} @llvm.umul.with.overflow.i16(i16 %a, i16 %b)
11848 declare {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
11849 declare {i64, i1} @llvm.umul.with.overflow.i64(i64 %a, i64 %b)
11854 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
11855 a unsigned multiplication of the two arguments, and indicate whether an
11856 overflow occurred during the unsigned multiplication.
11861 The arguments (%a and %b) and the first element of the result structure
11862 may be of integer types of any bit width, but they must have the same
11863 bit width. The second element of the result structure must be of type
11864 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
11870 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
11871 an unsigned multiplication of the two arguments. They return a structure ---
11872 the first element of which is the multiplication, and the second
11873 element of which is a bit specifying if the unsigned multiplication
11874 resulted in an overflow.
11879 .. code-block:: llvm
11881 %res = call {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
11882 %sum = extractvalue {i32, i1} %res, 0
11883 %obit = extractvalue {i32, i1} %res, 1
11884 br i1 %obit, label %overflow, label %normal
11886 Specialised Arithmetic Intrinsics
11887 ---------------------------------
11889 '``llvm.canonicalize.*``' Intrinsic
11890 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11897 declare float @llvm.canonicalize.f32(float %a)
11898 declare double @llvm.canonicalize.f64(double %b)
11903 The '``llvm.canonicalize.*``' intrinsic returns the platform specific canonical
11904 encoding of a floating-point number. This canonicalization is useful for
11905 implementing certain numeric primitives such as frexp. The canonical encoding is
11906 defined by IEEE-754-2008 to be:
11910 2.1.8 canonical encoding: The preferred encoding of a floating-point
11911 representation in a format. Applied to declets, significands of finite
11912 numbers, infinities, and NaNs, especially in decimal formats.
11914 This operation can also be considered equivalent to the IEEE-754-2008
11915 conversion of a floating-point value to the same format. NaNs are handled
11916 according to section 6.2.
11918 Examples of non-canonical encodings:
11920 - x87 pseudo denormals, pseudo NaNs, pseudo Infinity, Unnormals. These are
11921 converted to a canonical representation per hardware-specific protocol.
11922 - Many normal decimal floating-point numbers have non-canonical alternative
11924 - Some machines, like GPUs or ARMv7 NEON, do not support subnormal values.
11925 These are treated as non-canonical encodings of zero and will be flushed to
11926 a zero of the same sign by this operation.
11928 Note that per IEEE-754-2008 6.2, systems that support signaling NaNs with
11929 default exception handling must signal an invalid exception, and produce a
11932 This function should always be implementable as multiplication by 1.0, provided
11933 that the compiler does not constant fold the operation. Likewise, division by
11934 1.0 and ``llvm.minnum(x, x)`` are possible implementations. Addition with
11935 -0.0 is also sufficient provided that the rounding mode is not -Infinity.
11937 ``@llvm.canonicalize`` must preserve the equality relation. That is:
11939 - ``(@llvm.canonicalize(x) == x)`` is equivalent to ``(x == x)``
11940 - ``(@llvm.canonicalize(x) == @llvm.canonicalize(y))`` is equivalent to
11943 Additionally, the sign of zero must be conserved:
11944 ``@llvm.canonicalize(-0.0) = -0.0`` and ``@llvm.canonicalize(+0.0) = +0.0``
11946 The payload bits of a NaN must be conserved, with two exceptions.
11947 First, environments which use only a single canonical representation of NaN
11948 must perform said canonicalization. Second, SNaNs must be quieted per the
11951 The canonicalization operation may be optimized away if:
11953 - The input is known to be canonical. For example, it was produced by a
11954 floating-point operation that is required by the standard to be canonical.
11955 - The result is consumed only by (or fused with) other floating-point
11956 operations. That is, the bits of the floating-point value are not examined.
11958 '``llvm.fmuladd.*``' Intrinsic
11959 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11966 declare float @llvm.fmuladd.f32(float %a, float %b, float %c)
11967 declare double @llvm.fmuladd.f64(double %a, double %b, double %c)
11972 The '``llvm.fmuladd.*``' intrinsic functions represent multiply-add
11973 expressions that can be fused if the code generator determines that (a) the
11974 target instruction set has support for a fused operation, and (b) that the
11975 fused operation is more efficient than the equivalent, separate pair of mul
11976 and add instructions.
11981 The '``llvm.fmuladd.*``' intrinsics each take three arguments: two
11982 multiplicands, a and b, and an addend c.
11991 %0 = call float @llvm.fmuladd.f32(%a, %b, %c)
11993 is equivalent to the expression a \* b + c, except that rounding will
11994 not be performed between the multiplication and addition steps if the
11995 code generator fuses the operations. Fusion is not guaranteed, even if
11996 the target platform supports it. If a fused multiply-add is required the
11997 corresponding llvm.fma.\* intrinsic function should be used
11998 instead. This never sets errno, just as '``llvm.fma.*``'.
12003 .. code-block:: llvm
12005 %r2 = call float @llvm.fmuladd.f32(float %a, float %b, float %c) ; yields float:r2 = (a * b) + c
12008 Experimental Vector Reduction Intrinsics
12009 ----------------------------------------
12011 Horizontal reductions of vectors can be expressed using the following
12012 intrinsics. Each one takes a vector operand as an input and applies its
12013 respective operation across all elements of the vector, returning a single
12014 scalar result of the same element type.
12017 '``llvm.experimental.vector.reduce.add.*``' Intrinsic
12018 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12025 declare i32 @llvm.experimental.vector.reduce.add.i32.v4i32(<4 x i32> %a)
12026 declare i64 @llvm.experimental.vector.reduce.add.i64.v2i64(<2 x i64> %a)
12031 The '``llvm.experimental.vector.reduce.add.*``' intrinsics do an integer ``ADD``
12032 reduction of a vector, returning the result as a scalar. The return type matches
12033 the element-type of the vector input.
12037 The argument to this intrinsic must be a vector of integer values.
12039 '``llvm.experimental.vector.reduce.fadd.*``' Intrinsic
12040 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12047 declare float @llvm.experimental.vector.reduce.fadd.f32.v4f32(float %acc, <4 x float> %a)
12048 declare double @llvm.experimental.vector.reduce.fadd.f64.v2f64(double %acc, <2 x double> %a)
12053 The '``llvm.experimental.vector.reduce.fadd.*``' intrinsics do a floating-point
12054 ``ADD`` reduction of a vector, returning the result as a scalar. The return type
12055 matches the element-type of the vector input.
12057 If the intrinsic call has fast-math flags, then the reduction will not preserve
12058 the associativity of an equivalent scalarized counterpart. If it does not have
12059 fast-math flags, then the reduction will be *ordered*, implying that the
12060 operation respects the associativity of a scalarized reduction.
12065 The first argument to this intrinsic is a scalar accumulator value, which is
12066 only used when there are no fast-math flags attached. This argument may be undef
12067 when fast-math flags are used.
12069 The second argument must be a vector of floating-point values.
12074 .. code-block:: llvm
12076 %fast = call fast float @llvm.experimental.vector.reduce.fadd.f32.v4f32(float undef, <4 x float> %input) ; fast reduction
12077 %ord = call float @llvm.experimental.vector.reduce.fadd.f32.v4f32(float %acc, <4 x float> %input) ; ordered reduction
12080 '``llvm.experimental.vector.reduce.mul.*``' Intrinsic
12081 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12088 declare i32 @llvm.experimental.vector.reduce.mul.i32.v4i32(<4 x i32> %a)
12089 declare i64 @llvm.experimental.vector.reduce.mul.i64.v2i64(<2 x i64> %a)
12094 The '``llvm.experimental.vector.reduce.mul.*``' intrinsics do an integer ``MUL``
12095 reduction of a vector, returning the result as a scalar. The return type matches
12096 the element-type of the vector input.
12100 The argument to this intrinsic must be a vector of integer values.
12102 '``llvm.experimental.vector.reduce.fmul.*``' Intrinsic
12103 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12110 declare float @llvm.experimental.vector.reduce.fmul.f32.v4f32(float %acc, <4 x float> %a)
12111 declare double @llvm.experimental.vector.reduce.fmul.f64.v2f64(double %acc, <2 x double> %a)
12116 The '``llvm.experimental.vector.reduce.fmul.*``' intrinsics do a floating-point
12117 ``MUL`` reduction of a vector, returning the result as a scalar. The return type
12118 matches the element-type of the vector input.
12120 If the intrinsic call has fast-math flags, then the reduction will not preserve
12121 the associativity of an equivalent scalarized counterpart. If it does not have
12122 fast-math flags, then the reduction will be *ordered*, implying that the
12123 operation respects the associativity of a scalarized reduction.
12128 The first argument to this intrinsic is a scalar accumulator value, which is
12129 only used when there are no fast-math flags attached. This argument may be undef
12130 when fast-math flags are used.
12132 The second argument must be a vector of floating-point values.
12137 .. code-block:: llvm
12139 %fast = call fast float @llvm.experimental.vector.reduce.fmul.f32.v4f32(float undef, <4 x float> %input) ; fast reduction
12140 %ord = call float @llvm.experimental.vector.reduce.fmul.f32.v4f32(float %acc, <4 x float> %input) ; ordered reduction
12142 '``llvm.experimental.vector.reduce.and.*``' Intrinsic
12143 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12150 declare i32 @llvm.experimental.vector.reduce.and.i32.v4i32(<4 x i32> %a)
12155 The '``llvm.experimental.vector.reduce.and.*``' intrinsics do a bitwise ``AND``
12156 reduction of a vector, returning the result as a scalar. The return type matches
12157 the element-type of the vector input.
12161 The argument to this intrinsic must be a vector of integer values.
12163 '``llvm.experimental.vector.reduce.or.*``' Intrinsic
12164 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12171 declare i32 @llvm.experimental.vector.reduce.or.i32.v4i32(<4 x i32> %a)
12176 The '``llvm.experimental.vector.reduce.or.*``' intrinsics do a bitwise ``OR`` reduction
12177 of a vector, returning the result as a scalar. The return type matches the
12178 element-type of the vector input.
12182 The argument to this intrinsic must be a vector of integer values.
12184 '``llvm.experimental.vector.reduce.xor.*``' Intrinsic
12185 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12192 declare i32 @llvm.experimental.vector.reduce.xor.i32.v4i32(<4 x i32> %a)
12197 The '``llvm.experimental.vector.reduce.xor.*``' intrinsics do a bitwise ``XOR``
12198 reduction of a vector, returning the result as a scalar. The return type matches
12199 the element-type of the vector input.
12203 The argument to this intrinsic must be a vector of integer values.
12205 '``llvm.experimental.vector.reduce.smax.*``' Intrinsic
12206 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12213 declare i32 @llvm.experimental.vector.reduce.smax.i32.v4i32(<4 x i32> %a)
12218 The '``llvm.experimental.vector.reduce.smax.*``' intrinsics do a signed integer
12219 ``MAX`` reduction of a vector, returning the result as a scalar. The return type
12220 matches the element-type of the vector input.
12224 The argument to this intrinsic must be a vector of integer values.
12226 '``llvm.experimental.vector.reduce.smin.*``' Intrinsic
12227 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12234 declare i32 @llvm.experimental.vector.reduce.smin.i32.v4i32(<4 x i32> %a)
12239 The '``llvm.experimental.vector.reduce.smin.*``' intrinsics do a signed integer
12240 ``MIN`` reduction of a vector, returning the result as a scalar. The return type
12241 matches the element-type of the vector input.
12245 The argument to this intrinsic must be a vector of integer values.
12247 '``llvm.experimental.vector.reduce.umax.*``' Intrinsic
12248 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12255 declare i32 @llvm.experimental.vector.reduce.umax.i32.v4i32(<4 x i32> %a)
12260 The '``llvm.experimental.vector.reduce.umax.*``' intrinsics do an unsigned
12261 integer ``MAX`` reduction of a vector, returning the result as a scalar. The
12262 return type matches the element-type of the vector input.
12266 The argument to this intrinsic must be a vector of integer values.
12268 '``llvm.experimental.vector.reduce.umin.*``' Intrinsic
12269 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12276 declare i32 @llvm.experimental.vector.reduce.umin.i32.v4i32(<4 x i32> %a)
12281 The '``llvm.experimental.vector.reduce.umin.*``' intrinsics do an unsigned
12282 integer ``MIN`` reduction of a vector, returning the result as a scalar. The
12283 return type matches the element-type of the vector input.
12287 The argument to this intrinsic must be a vector of integer values.
12289 '``llvm.experimental.vector.reduce.fmax.*``' Intrinsic
12290 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12297 declare float @llvm.experimental.vector.reduce.fmax.f32.v4f32(<4 x float> %a)
12298 declare double @llvm.experimental.vector.reduce.fmax.f64.v2f64(<2 x double> %a)
12303 The '``llvm.experimental.vector.reduce.fmax.*``' intrinsics do a floating-point
12304 ``MAX`` reduction of a vector, returning the result as a scalar. The return type
12305 matches the element-type of the vector input.
12307 If the intrinsic call has the ``nnan`` fast-math flag then the operation can
12308 assume that NaNs are not present in the input vector.
12312 The argument to this intrinsic must be a vector of floating-point values.
12314 '``llvm.experimental.vector.reduce.fmin.*``' Intrinsic
12315 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12322 declare float @llvm.experimental.vector.reduce.fmin.f32.v4f32(<4 x float> %a)
12323 declare double @llvm.experimental.vector.reduce.fmin.f64.v2f64(<2 x double> %a)
12328 The '``llvm.experimental.vector.reduce.fmin.*``' intrinsics do a floating-point
12329 ``MIN`` reduction of a vector, returning the result as a scalar. The return type
12330 matches the element-type of the vector input.
12332 If the intrinsic call has the ``nnan`` fast-math flag then the operation can
12333 assume that NaNs are not present in the input vector.
12337 The argument to this intrinsic must be a vector of floating-point values.
12339 Half Precision Floating-Point Intrinsics
12340 ----------------------------------------
12342 For most target platforms, half precision floating-point is a
12343 storage-only format. This means that it is a dense encoding (in memory)
12344 but does not support computation in the format.
12346 This means that code must first load the half-precision floating-point
12347 value as an i16, then convert it to float with
12348 :ref:`llvm.convert.from.fp16 <int_convert_from_fp16>`. Computation can
12349 then be performed on the float value (including extending to double
12350 etc). To store the value back to memory, it is first converted to float
12351 if needed, then converted to i16 with
12352 :ref:`llvm.convert.to.fp16 <int_convert_to_fp16>`, then storing as an
12355 .. _int_convert_to_fp16:
12357 '``llvm.convert.to.fp16``' Intrinsic
12358 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12365 declare i16 @llvm.convert.to.fp16.f32(float %a)
12366 declare i16 @llvm.convert.to.fp16.f64(double %a)
12371 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion from a
12372 conventional floating-point type to half precision floating-point format.
12377 The intrinsic function contains single argument - the value to be
12383 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion from a
12384 conventional floating-point format to half precision floating-point format. The
12385 return value is an ``i16`` which contains the converted number.
12390 .. code-block:: llvm
12392 %res = call i16 @llvm.convert.to.fp16.f32(float %a)
12393 store i16 %res, i16* @x, align 2
12395 .. _int_convert_from_fp16:
12397 '``llvm.convert.from.fp16``' Intrinsic
12398 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12405 declare float @llvm.convert.from.fp16.f32(i16 %a)
12406 declare double @llvm.convert.from.fp16.f64(i16 %a)
12411 The '``llvm.convert.from.fp16``' intrinsic function performs a
12412 conversion from half precision floating-point format to single precision
12413 floating-point format.
12418 The intrinsic function contains single argument - the value to be
12424 The '``llvm.convert.from.fp16``' intrinsic function performs a
12425 conversion from half single precision floating-point format to single
12426 precision floating-point format. The input half-float value is
12427 represented by an ``i16`` value.
12432 .. code-block:: llvm
12434 %a = load i16, i16* @x, align 2
12435 %res = call float @llvm.convert.from.fp16(i16 %a)
12437 .. _dbg_intrinsics:
12439 Debugger Intrinsics
12440 -------------------
12442 The LLVM debugger intrinsics (which all start with ``llvm.dbg.``
12443 prefix), are described in the `LLVM Source Level
12444 Debugging <SourceLevelDebugging.html#format-common-intrinsics>`_
12447 Exception Handling Intrinsics
12448 -----------------------------
12450 The LLVM exception handling intrinsics (which all start with
12451 ``llvm.eh.`` prefix), are described in the `LLVM Exception
12452 Handling <ExceptionHandling.html#format-common-intrinsics>`_ document.
12454 .. _int_trampoline:
12456 Trampoline Intrinsics
12457 ---------------------
12459 These intrinsics make it possible to excise one parameter, marked with
12460 the :ref:`nest <nest>` attribute, from a function. The result is a
12461 callable function pointer lacking the nest parameter - the caller does
12462 not need to provide a value for it. Instead, the value to use is stored
12463 in advance in a "trampoline", a block of memory usually allocated on the
12464 stack, which also contains code to splice the nest value into the
12465 argument list. This is used to implement the GCC nested function address
12468 For example, if the function is ``i32 f(i8* nest %c, i32 %x, i32 %y)``
12469 then the resulting function pointer has signature ``i32 (i32, i32)*``.
12470 It can be created as follows:
12472 .. code-block:: llvm
12474 %tramp = alloca [10 x i8], align 4 ; size and alignment only correct for X86
12475 %tramp1 = getelementptr [10 x i8], [10 x i8]* %tramp, i32 0, i32 0
12476 call i8* @llvm.init.trampoline(i8* %tramp1, i8* bitcast (i32 (i8*, i32, i32)* @f to i8*), i8* %nval)
12477 %p = call i8* @llvm.adjust.trampoline(i8* %tramp1)
12478 %fp = bitcast i8* %p to i32 (i32, i32)*
12480 The call ``%val = call i32 %fp(i32 %x, i32 %y)`` is then equivalent to
12481 ``%val = call i32 %f(i8* %nval, i32 %x, i32 %y)``.
12485 '``llvm.init.trampoline``' Intrinsic
12486 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12493 declare void @llvm.init.trampoline(i8* <tramp>, i8* <func>, i8* <nval>)
12498 This fills the memory pointed to by ``tramp`` with executable code,
12499 turning it into a trampoline.
12504 The ``llvm.init.trampoline`` intrinsic takes three arguments, all
12505 pointers. The ``tramp`` argument must point to a sufficiently large and
12506 sufficiently aligned block of memory; this memory is written to by the
12507 intrinsic. Note that the size and the alignment are target-specific -
12508 LLVM currently provides no portable way of determining them, so a
12509 front-end that generates this intrinsic needs to have some
12510 target-specific knowledge. The ``func`` argument must hold a function
12511 bitcast to an ``i8*``.
12516 The block of memory pointed to by ``tramp`` is filled with target
12517 dependent code, turning it into a function. Then ``tramp`` needs to be
12518 passed to :ref:`llvm.adjust.trampoline <int_at>` to get a pointer which can
12519 be :ref:`bitcast (to a new function) and called <int_trampoline>`. The new
12520 function's signature is the same as that of ``func`` with any arguments
12521 marked with the ``nest`` attribute removed. At most one such ``nest``
12522 argument is allowed, and it must be of pointer type. Calling the new
12523 function is equivalent to calling ``func`` with the same argument list,
12524 but with ``nval`` used for the missing ``nest`` argument. If, after
12525 calling ``llvm.init.trampoline``, the memory pointed to by ``tramp`` is
12526 modified, then the effect of any later call to the returned function
12527 pointer is undefined.
12531 '``llvm.adjust.trampoline``' Intrinsic
12532 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12539 declare i8* @llvm.adjust.trampoline(i8* <tramp>)
12544 This performs any required machine-specific adjustment to the address of
12545 a trampoline (passed as ``tramp``).
12550 ``tramp`` must point to a block of memory which already has trampoline
12551 code filled in by a previous call to
12552 :ref:`llvm.init.trampoline <int_it>`.
12557 On some architectures the address of the code to be executed needs to be
12558 different than the address where the trampoline is actually stored. This
12559 intrinsic returns the executable address corresponding to ``tramp``
12560 after performing the required machine specific adjustments. The pointer
12561 returned can then be :ref:`bitcast and executed <int_trampoline>`.
12563 .. _int_mload_mstore:
12565 Masked Vector Load and Store Intrinsics
12566 ---------------------------------------
12568 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.
12572 '``llvm.masked.load.*``' Intrinsics
12573 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12577 This is an overloaded intrinsic. The loaded data is a vector of any integer, floating-point or pointer data type.
12581 declare <16 x float> @llvm.masked.load.v16f32.p0v16f32 (<16 x float>* <ptr>, i32 <alignment>, <16 x i1> <mask>, <16 x float> <passthru>)
12582 declare <2 x double> @llvm.masked.load.v2f64.p0v2f64 (<2 x double>* <ptr>, i32 <alignment>, <2 x i1> <mask>, <2 x double> <passthru>)
12583 ;; The data is a vector of pointers to double
12584 declare <8 x double*> @llvm.masked.load.v8p0f64.p0v8p0f64 (<8 x double*>* <ptr>, i32 <alignment>, <8 x i1> <mask>, <8 x double*> <passthru>)
12585 ;; The data is a vector of function pointers
12586 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>)
12591 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.
12597 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.
12603 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.
12604 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.
12609 %res = call <16 x float> @llvm.masked.load.v16f32.p0v16f32 (<16 x float>* %ptr, i32 4, <16 x i1>%mask, <16 x float> %passthru)
12611 ;; The result of the two following instructions is identical aside from potential memory access exception
12612 %loadlal = load <16 x float>, <16 x float>* %ptr, align 4
12613 %res = select <16 x i1> %mask, <16 x float> %loadlal, <16 x float> %passthru
12617 '``llvm.masked.store.*``' Intrinsics
12618 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12622 This is an overloaded intrinsic. The data stored in memory is a vector of any integer, floating-point or pointer data type.
12626 declare void @llvm.masked.store.v8i32.p0v8i32 (<8 x i32> <value>, <8 x i32>* <ptr>, i32 <alignment>, <8 x i1> <mask>)
12627 declare void @llvm.masked.store.v16f32.p0v16f32 (<16 x float> <value>, <16 x float>* <ptr>, i32 <alignment>, <16 x i1> <mask>)
12628 ;; The data is a vector of pointers to double
12629 declare void @llvm.masked.store.v8p0f64.p0v8p0f64 (<8 x double*> <value>, <8 x double*>* <ptr>, i32 <alignment>, <8 x i1> <mask>)
12630 ;; The data is a vector of function pointers
12631 declare void @llvm.masked.store.v4p0f_i32f.p0v4p0f_i32f (<4 x i32 ()*> <value>, <4 x i32 ()*>* <ptr>, i32 <alignment>, <4 x i1> <mask>)
12636 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.
12641 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.
12647 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.
12648 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.
12652 call void @llvm.masked.store.v16f32.p0v16f32(<16 x float> %value, <16 x float>* %ptr, i32 4, <16 x i1> %mask)
12654 ;; The result of the following instructions is identical aside from potential data races and memory access exceptions
12655 %oldval = load <16 x float>, <16 x float>* %ptr, align 4
12656 %res = select <16 x i1> %mask, <16 x float> %value, <16 x float> %oldval
12657 store <16 x float> %res, <16 x float>* %ptr, align 4
12660 Masked Vector Gather and Scatter Intrinsics
12661 -------------------------------------------
12663 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.
12667 '``llvm.masked.gather.*``' Intrinsics
12668 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12672 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.
12676 declare <16 x float> @llvm.masked.gather.v16f32.v16p0f32 (<16 x float*> <ptrs>, i32 <alignment>, <16 x i1> <mask>, <16 x float> <passthru>)
12677 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>)
12678 declare <8 x float*> @llvm.masked.gather.v8p0f32.v8p0p0f32 (<8 x float**> <ptrs>, i32 <alignment>, <8 x i1> <mask>, <8 x float*> <passthru>)
12683 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.
12689 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.
12695 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.
12696 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.
12701 %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)
12703 ;; The gather with all-true mask is equivalent to the following instruction sequence
12704 %ptr0 = extractelement <4 x double*> %ptrs, i32 0
12705 %ptr1 = extractelement <4 x double*> %ptrs, i32 1
12706 %ptr2 = extractelement <4 x double*> %ptrs, i32 2
12707 %ptr3 = extractelement <4 x double*> %ptrs, i32 3
12709 %val0 = load double, double* %ptr0, align 8
12710 %val1 = load double, double* %ptr1, align 8
12711 %val2 = load double, double* %ptr2, align 8
12712 %val3 = load double, double* %ptr3, align 8
12714 %vec0 = insertelement <4 x double>undef, %val0, 0
12715 %vec01 = insertelement <4 x double>%vec0, %val1, 1
12716 %vec012 = insertelement <4 x double>%vec01, %val2, 2
12717 %vec0123 = insertelement <4 x double>%vec012, %val3, 3
12721 '``llvm.masked.scatter.*``' Intrinsics
12722 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12726 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.
12730 declare void @llvm.masked.scatter.v8i32.v8p0i32 (<8 x i32> <value>, <8 x i32*> <ptrs>, i32 <alignment>, <8 x i1> <mask>)
12731 declare void @llvm.masked.scatter.v16f32.v16p1f32 (<16 x float> <value>, <16 x float addrspace(1)*> <ptrs>, i32 <alignment>, <16 x i1> <mask>)
12732 declare void @llvm.masked.scatter.v4p0f64.v4p0p0f64 (<4 x double*> <value>, <4 x double**> <ptrs>, i32 <alignment>, <4 x i1> <mask>)
12737 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.
12742 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.
12748 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.
12752 ;; This instruction unconditionally stores data vector in multiple addresses
12753 call @llvm.masked.scatter.v8i32.v8p0i32 (<8 x i32> %value, <8 x i32*> %ptrs, i32 4, <8 x i1> <true, true, .. true>)
12755 ;; It is equivalent to a list of scalar stores
12756 %val0 = extractelement <8 x i32> %value, i32 0
12757 %val1 = extractelement <8 x i32> %value, i32 1
12759 %val7 = extractelement <8 x i32> %value, i32 7
12760 %ptr0 = extractelement <8 x i32*> %ptrs, i32 0
12761 %ptr1 = extractelement <8 x i32*> %ptrs, i32 1
12763 %ptr7 = extractelement <8 x i32*> %ptrs, i32 7
12764 ;; Note: the order of the following stores is important when they overlap:
12765 store i32 %val0, i32* %ptr0, align 4
12766 store i32 %val1, i32* %ptr1, align 4
12768 store i32 %val7, i32* %ptr7, align 4
12774 This class of intrinsics provides information about the lifetime of
12775 memory objects and ranges where variables are immutable.
12779 '``llvm.lifetime.start``' Intrinsic
12780 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12787 declare void @llvm.lifetime.start(i64 <size>, i8* nocapture <ptr>)
12792 The '``llvm.lifetime.start``' intrinsic specifies the start of a memory
12798 The first argument is a constant integer representing the size of the
12799 object, or -1 if it is variable sized. The second argument is a pointer
12805 This intrinsic indicates that before this point in the code, the value
12806 of the memory pointed to by ``ptr`` is dead. This means that it is known
12807 to never be used and has an undefined value. A load from the pointer
12808 that precedes this intrinsic can be replaced with ``'undef'``.
12812 '``llvm.lifetime.end``' Intrinsic
12813 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12820 declare void @llvm.lifetime.end(i64 <size>, i8* nocapture <ptr>)
12825 The '``llvm.lifetime.end``' intrinsic specifies the end of a memory
12831 The first argument is a constant integer representing the size of the
12832 object, or -1 if it is variable sized. The second argument is a pointer
12838 This intrinsic indicates that after this point in the code, the value of
12839 the memory pointed to by ``ptr`` is dead. This means that it is known to
12840 never be used and has an undefined value. Any stores into the memory
12841 object following this intrinsic may be removed as dead.
12843 '``llvm.invariant.start``' Intrinsic
12844 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12848 This is an overloaded intrinsic. The memory object can belong to any address space.
12852 declare {}* @llvm.invariant.start.p0i8(i64 <size>, i8* nocapture <ptr>)
12857 The '``llvm.invariant.start``' intrinsic specifies that the contents of
12858 a memory object will not change.
12863 The first argument is a constant integer representing the size of the
12864 object, or -1 if it is variable sized. The second argument is a pointer
12870 This intrinsic indicates that until an ``llvm.invariant.end`` that uses
12871 the return value, the referenced memory location is constant and
12874 '``llvm.invariant.end``' Intrinsic
12875 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12879 This is an overloaded intrinsic. The memory object can belong to any address space.
12883 declare void @llvm.invariant.end.p0i8({}* <start>, i64 <size>, i8* nocapture <ptr>)
12888 The '``llvm.invariant.end``' intrinsic specifies that the contents of a
12889 memory object are mutable.
12894 The first argument is the matching ``llvm.invariant.start`` intrinsic.
12895 The second argument is a constant integer representing the size of the
12896 object, or -1 if it is variable sized and the third argument is a
12897 pointer to the object.
12902 This intrinsic indicates that the memory is mutable again.
12904 '``llvm.invariant.group.barrier``' Intrinsic
12905 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12909 This is an overloaded intrinsic. The memory object can belong to any address
12910 space. The returned pointer must belong to the same address space as the
12915 declare i8* @llvm.invariant.group.barrier.p0i8(i8* <ptr>)
12920 The '``llvm.invariant.group.barrier``' intrinsic can be used when an invariant
12921 established by invariant.group metadata no longer holds, to obtain a new pointer
12922 value that does not carry the invariant information.
12928 The ``llvm.invariant.group.barrier`` takes only one argument, which is
12929 the pointer to the memory for which the ``invariant.group`` no longer holds.
12934 Returns another pointer that aliases its argument but which is considered different
12935 for the purposes of ``load``/``store`` ``invariant.group`` metadata.
12939 Constrained Floating-Point Intrinsics
12940 -------------------------------------
12942 These intrinsics are used to provide special handling of floating-point
12943 operations when specific rounding mode or floating-point exception behavior is
12944 required. By default, LLVM optimization passes assume that the rounding mode is
12945 round-to-nearest and that floating-point exceptions will not be monitored.
12946 Constrained FP intrinsics are used to support non-default rounding modes and
12947 accurately preserve exception behavior without compromising LLVM's ability to
12948 optimize FP code when the default behavior is used.
12950 Each of these intrinsics corresponds to a normal floating-point operation. The
12951 first two arguments and the return value are the same as the corresponding FP
12954 The third argument is a metadata argument specifying the rounding mode to be
12955 assumed. This argument must be one of the following strings:
12965 If this argument is "round.dynamic" optimization passes must assume that the
12966 rounding mode is unknown and may change at runtime. No transformations that
12967 depend on rounding mode may be performed in this case.
12969 The other possible values for the rounding mode argument correspond to the
12970 similarly named IEEE rounding modes. If the argument is any of these values
12971 optimization passes may perform transformations as long as they are consistent
12972 with the specified rounding mode.
12974 For example, 'x-0'->'x' is not a valid transformation if the rounding mode is
12975 "round.downward" or "round.dynamic" because if the value of 'x' is +0 then
12976 'x-0' should evaluate to '-0' when rounding downward. However, this
12977 transformation is legal for all other rounding modes.
12979 For values other than "round.dynamic" optimization passes may assume that the
12980 actual runtime rounding mode (as defined in a target-specific manner) matches
12981 the specified rounding mode, but this is not guaranteed. Using a specific
12982 non-dynamic rounding mode which does not match the actual rounding mode at
12983 runtime results in undefined behavior.
12985 The fourth argument to the constrained floating-point intrinsics specifies the
12986 required exception behavior. This argument must be one of the following
12995 If this argument is "fpexcept.ignore" optimization passes may assume that the
12996 exception status flags will not be read and that floating-point exceptions will
12997 be masked. This allows transformations to be performed that may change the
12998 exception semantics of the original code. For example, FP operations may be
12999 speculatively executed in this case whereas they must not be for either of the
13000 other possible values of this argument.
13002 If the exception behavior argument is "fpexcept.maytrap" optimization passes
13003 must avoid transformations that may raise exceptions that would not have been
13004 raised by the original code (such as speculatively executing FP operations), but
13005 passes are not required to preserve all exceptions that are implied by the
13006 original code. For example, exceptions may be potentially hidden by constant
13009 If the exception behavior argument is "fpexcept.strict" all transformations must
13010 strictly preserve the floating-point exception semantics of the original code.
13011 Any FP exception that would have been raised by the original code must be raised
13012 by the transformed code, and the transformed code must not raise any FP
13013 exceptions that would not have been raised by the original code. This is the
13014 exception behavior argument that will be used if the code being compiled reads
13015 the FP exception status flags, but this mode can also be used with code that
13016 unmasks FP exceptions.
13018 The number and order of floating-point exceptions is NOT guaranteed. For
13019 example, a series of FP operations that each may raise exceptions may be
13020 vectorized into a single instruction that raises each unique exception a single
13024 '``llvm.experimental.constrained.fadd``' Intrinsic
13025 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13033 @llvm.experimental.constrained.fadd(<type> <op1>, <type> <op2>,
13034 metadata <rounding mode>,
13035 metadata <exception behavior>)
13040 The '``llvm.experimental.constrained.fadd``' intrinsic returns the sum of its
13047 The first two arguments to the '``llvm.experimental.constrained.fadd``'
13048 intrinsic must be :ref:`floating-point <t_floating>` or :ref:`vector <t_vector>`
13049 of floating-point values. Both arguments must have identical types.
13051 The third and fourth arguments specify the rounding mode and exception
13052 behavior as described above.
13057 The value produced is the floating-point sum of the two value operands and has
13058 the same type as the operands.
13061 '``llvm.experimental.constrained.fsub``' Intrinsic
13062 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13070 @llvm.experimental.constrained.fsub(<type> <op1>, <type> <op2>,
13071 metadata <rounding mode>,
13072 metadata <exception behavior>)
13077 The '``llvm.experimental.constrained.fsub``' intrinsic returns the difference
13078 of its two operands.
13084 The first two arguments to the '``llvm.experimental.constrained.fsub``'
13085 intrinsic must be :ref:`floating-point <t_floating>` or :ref:`vector <t_vector>`
13086 of floating-point values. Both arguments must have identical types.
13088 The third and fourth arguments specify the rounding mode and exception
13089 behavior as described above.
13094 The value produced is the floating-point difference of the two value operands
13095 and has the same type as the operands.
13098 '``llvm.experimental.constrained.fmul``' Intrinsic
13099 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13107 @llvm.experimental.constrained.fmul(<type> <op1>, <type> <op2>,
13108 metadata <rounding mode>,
13109 metadata <exception behavior>)
13114 The '``llvm.experimental.constrained.fmul``' intrinsic returns the product of
13121 The first two arguments to the '``llvm.experimental.constrained.fmul``'
13122 intrinsic must be :ref:`floating-point <t_floating>` or :ref:`vector <t_vector>`
13123 of floating-point values. Both arguments must have identical types.
13125 The third and fourth arguments specify the rounding mode and exception
13126 behavior as described above.
13131 The value produced is the floating-point product of the two value operands and
13132 has the same type as the operands.
13135 '``llvm.experimental.constrained.fdiv``' Intrinsic
13136 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13144 @llvm.experimental.constrained.fdiv(<type> <op1>, <type> <op2>,
13145 metadata <rounding mode>,
13146 metadata <exception behavior>)
13151 The '``llvm.experimental.constrained.fdiv``' intrinsic returns the quotient of
13158 The first two arguments to the '``llvm.experimental.constrained.fdiv``'
13159 intrinsic must be :ref:`floating-point <t_floating>` or :ref:`vector <t_vector>`
13160 of floating-point values. Both arguments must have identical types.
13162 The third and fourth arguments specify the rounding mode and exception
13163 behavior as described above.
13168 The value produced is the floating-point quotient of the two value operands and
13169 has the same type as the operands.
13172 '``llvm.experimental.constrained.frem``' Intrinsic
13173 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13181 @llvm.experimental.constrained.frem(<type> <op1>, <type> <op2>,
13182 metadata <rounding mode>,
13183 metadata <exception behavior>)
13188 The '``llvm.experimental.constrained.frem``' intrinsic returns the remainder
13189 from the division of its two operands.
13195 The first two arguments to the '``llvm.experimental.constrained.frem``'
13196 intrinsic must be :ref:`floating-point <t_floating>` or :ref:`vector <t_vector>`
13197 of floating-point values. Both arguments must have identical types.
13199 The third and fourth arguments specify the rounding mode and exception
13200 behavior as described above. The rounding mode argument has no effect, since
13201 the result of frem is never rounded, but the argument is included for
13202 consistency with the other constrained floating-point intrinsics.
13207 The value produced is the floating-point remainder from the division of the two
13208 value operands and has the same type as the operands. The remainder has the
13209 same sign as the dividend.
13211 '``llvm.experimental.constrained.fma``' Intrinsic
13212 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13220 @llvm.experimental.constrained.fma(<type> <op1>, <type> <op2>, <type> <op3>,
13221 metadata <rounding mode>,
13222 metadata <exception behavior>)
13227 The '``llvm.experimental.constrained.fma``' intrinsic returns the result of a
13228 fused-multiply-add operation on its operands.
13233 The first three arguments to the '``llvm.experimental.constrained.fma``'
13234 intrinsic must be :ref:`floating-point <t_floating>` or :ref:`vector
13235 <t_vector>` of floating-point values. All arguments must have identical types.
13237 The fourth and fifth arguments specify the rounding mode and exception behavior
13238 as described above.
13243 The result produced is the product of the first two operands added to the third
13244 operand computed with infinite precision, and then rounded to the target
13247 Constrained libm-equivalent Intrinsics
13248 --------------------------------------
13250 In addition to the basic floating-point operations for which constrained
13251 intrinsics are described above, there are constrained versions of various
13252 operations which provide equivalent behavior to a corresponding libm function.
13253 These intrinsics allow the precise behavior of these operations with respect to
13254 rounding mode and exception behavior to be controlled.
13256 As with the basic constrained floating-point intrinsics, the rounding mode
13257 and exception behavior arguments only control the behavior of the optimizer.
13258 They do not change the runtime floating-point environment.
13261 '``llvm.experimental.constrained.sqrt``' Intrinsic
13262 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13270 @llvm.experimental.constrained.sqrt(<type> <op1>,
13271 metadata <rounding mode>,
13272 metadata <exception behavior>)
13277 The '``llvm.experimental.constrained.sqrt``' intrinsic returns the square root
13278 of the specified value, returning the same value as the libm '``sqrt``'
13279 functions would, but without setting ``errno``.
13284 The first argument and the return type are floating-point numbers of the same
13287 The second and third arguments specify the rounding mode and exception
13288 behavior as described above.
13293 This function returns the nonnegative square root of the specified value.
13294 If the value is less than negative zero, a floating-point exception occurs
13295 and the return value is architecture specific.
13298 '``llvm.experimental.constrained.pow``' Intrinsic
13299 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13307 @llvm.experimental.constrained.pow(<type> <op1>, <type> <op2>,
13308 metadata <rounding mode>,
13309 metadata <exception behavior>)
13314 The '``llvm.experimental.constrained.pow``' intrinsic returns the first operand
13315 raised to the (positive or negative) power specified by the second operand.
13320 The first two arguments and the return value are floating-point numbers of the
13321 same type. The second argument specifies the power to which the first argument
13324 The third and fourth arguments specify the rounding mode and exception
13325 behavior as described above.
13330 This function returns the first value raised to the second power,
13331 returning the same values as the libm ``pow`` functions would, and
13332 handles error conditions in the same way.
13335 '``llvm.experimental.constrained.powi``' Intrinsic
13336 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13344 @llvm.experimental.constrained.powi(<type> <op1>, i32 <op2>,
13345 metadata <rounding mode>,
13346 metadata <exception behavior>)
13351 The '``llvm.experimental.constrained.powi``' intrinsic returns the first operand
13352 raised to the (positive or negative) power specified by the second operand. The
13353 order of evaluation of multiplications is not defined. When a vector of
13354 floating-point type is used, the second argument remains a scalar integer value.
13360 The first argument and the return value are floating-point numbers of the same
13361 type. The second argument is a 32-bit signed integer specifying the power to
13362 which the first argument should be raised.
13364 The third and fourth arguments specify the rounding mode and exception
13365 behavior as described above.
13370 This function returns the first value raised to the second power with an
13371 unspecified sequence of rounding operations.
13374 '``llvm.experimental.constrained.sin``' Intrinsic
13375 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13383 @llvm.experimental.constrained.sin(<type> <op1>,
13384 metadata <rounding mode>,
13385 metadata <exception behavior>)
13390 The '``llvm.experimental.constrained.sin``' intrinsic returns the sine of the
13396 The first argument and the return type are floating-point numbers of the same
13399 The second and third arguments specify the rounding mode and exception
13400 behavior as described above.
13405 This function returns the sine of the specified operand, returning the
13406 same values as the libm ``sin`` functions would, and handles error
13407 conditions in the same way.
13410 '``llvm.experimental.constrained.cos``' Intrinsic
13411 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13419 @llvm.experimental.constrained.cos(<type> <op1>,
13420 metadata <rounding mode>,
13421 metadata <exception behavior>)
13426 The '``llvm.experimental.constrained.cos``' intrinsic returns the cosine of the
13432 The first argument and the return type are floating-point numbers of the same
13435 The second and third arguments specify the rounding mode and exception
13436 behavior as described above.
13441 This function returns the cosine of the specified operand, returning the
13442 same values as the libm ``cos`` functions would, and handles error
13443 conditions in the same way.
13446 '``llvm.experimental.constrained.exp``' Intrinsic
13447 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13455 @llvm.experimental.constrained.exp(<type> <op1>,
13456 metadata <rounding mode>,
13457 metadata <exception behavior>)
13462 The '``llvm.experimental.constrained.exp``' intrinsic computes the base-e
13463 exponential of the specified value.
13468 The first argument and the return value are floating-point numbers of the same
13471 The second and third arguments specify the rounding mode and exception
13472 behavior as described above.
13477 This function returns the same values as the libm ``exp`` functions
13478 would, and handles error conditions in the same way.
13481 '``llvm.experimental.constrained.exp2``' Intrinsic
13482 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13490 @llvm.experimental.constrained.exp2(<type> <op1>,
13491 metadata <rounding mode>,
13492 metadata <exception behavior>)
13497 The '``llvm.experimental.constrained.exp2``' intrinsic computes the base-2
13498 exponential of the specified value.
13504 The first argument and the return value are floating-point numbers of the same
13507 The second and third arguments specify the rounding mode and exception
13508 behavior as described above.
13513 This function returns the same values as the libm ``exp2`` functions
13514 would, and handles error conditions in the same way.
13517 '``llvm.experimental.constrained.log``' Intrinsic
13518 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13526 @llvm.experimental.constrained.log(<type> <op1>,
13527 metadata <rounding mode>,
13528 metadata <exception behavior>)
13533 The '``llvm.experimental.constrained.log``' intrinsic computes the base-e
13534 logarithm of the specified value.
13539 The first argument and the return value are floating-point numbers of the same
13542 The second and third arguments specify the rounding mode and exception
13543 behavior as described above.
13549 This function returns the same values as the libm ``log`` functions
13550 would, and handles error conditions in the same way.
13553 '``llvm.experimental.constrained.log10``' Intrinsic
13554 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13562 @llvm.experimental.constrained.log10(<type> <op1>,
13563 metadata <rounding mode>,
13564 metadata <exception behavior>)
13569 The '``llvm.experimental.constrained.log10``' intrinsic computes the base-10
13570 logarithm of the specified value.
13575 The first argument and the return value are floating-point numbers of the same
13578 The second and third arguments specify the rounding mode and exception
13579 behavior as described above.
13584 This function returns the same values as the libm ``log10`` functions
13585 would, and handles error conditions in the same way.
13588 '``llvm.experimental.constrained.log2``' Intrinsic
13589 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13597 @llvm.experimental.constrained.log2(<type> <op1>,
13598 metadata <rounding mode>,
13599 metadata <exception behavior>)
13604 The '``llvm.experimental.constrained.log2``' intrinsic computes the base-2
13605 logarithm of the specified value.
13610 The first argument and the return value are floating-point numbers of the same
13613 The second and third arguments specify the rounding mode and exception
13614 behavior as described above.
13619 This function returns the same values as the libm ``log2`` functions
13620 would, and handles error conditions in the same way.
13623 '``llvm.experimental.constrained.rint``' Intrinsic
13624 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13632 @llvm.experimental.constrained.rint(<type> <op1>,
13633 metadata <rounding mode>,
13634 metadata <exception behavior>)
13639 The '``llvm.experimental.constrained.rint``' intrinsic returns the first
13640 operand rounded to the nearest integer. It may raise an inexact floating-point
13641 exception if the operand is not an integer.
13646 The first argument and the return value are floating-point numbers of the same
13649 The second and third arguments specify the rounding mode and exception
13650 behavior as described above.
13655 This function returns the same values as the libm ``rint`` functions
13656 would, and handles error conditions in the same way. The rounding mode is
13657 described, not determined, by the rounding mode argument. The actual rounding
13658 mode is determined by the runtime floating-point environment. The rounding
13659 mode argument is only intended as information to the compiler.
13662 '``llvm.experimental.constrained.nearbyint``' Intrinsic
13663 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13671 @llvm.experimental.constrained.nearbyint(<type> <op1>,
13672 metadata <rounding mode>,
13673 metadata <exception behavior>)
13678 The '``llvm.experimental.constrained.nearbyint``' intrinsic returns the first
13679 operand rounded to the nearest integer. It will not raise an inexact
13680 floating-point exception if the operand is not an integer.
13686 The first argument and the return value are floating-point numbers of the same
13689 The second and third arguments specify the rounding mode and exception
13690 behavior as described above.
13695 This function returns the same values as the libm ``nearbyint`` functions
13696 would, and handles error conditions in the same way. The rounding mode is
13697 described, not determined, by the rounding mode argument. The actual rounding
13698 mode is determined by the runtime floating-point environment. The rounding
13699 mode argument is only intended as information to the compiler.
13705 This class of intrinsics is designed to be generic and has no specific
13708 '``llvm.var.annotation``' Intrinsic
13709 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13716 declare void @llvm.var.annotation(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
13721 The '``llvm.var.annotation``' intrinsic.
13726 The first argument is a pointer to a value, the second is a pointer to a
13727 global string, the third is a pointer to a global string which is the
13728 source file name, and the last argument is the line number.
13733 This intrinsic allows annotation of local variables with arbitrary
13734 strings. This can be useful for special purpose optimizations that want
13735 to look for these annotations. These have no other defined use; they are
13736 ignored by code generation and optimization.
13738 '``llvm.ptr.annotation.*``' Intrinsic
13739 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13744 This is an overloaded intrinsic. You can use '``llvm.ptr.annotation``' on a
13745 pointer to an integer of any width. *NOTE* you must specify an address space for
13746 the pointer. The identifier for the default address space is the integer
13751 declare i8* @llvm.ptr.annotation.p<address space>i8(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
13752 declare i16* @llvm.ptr.annotation.p<address space>i16(i16* <val>, i8* <str>, i8* <str>, i32 <int>)
13753 declare i32* @llvm.ptr.annotation.p<address space>i32(i32* <val>, i8* <str>, i8* <str>, i32 <int>)
13754 declare i64* @llvm.ptr.annotation.p<address space>i64(i64* <val>, i8* <str>, i8* <str>, i32 <int>)
13755 declare i256* @llvm.ptr.annotation.p<address space>i256(i256* <val>, i8* <str>, i8* <str>, i32 <int>)
13760 The '``llvm.ptr.annotation``' intrinsic.
13765 The first argument is a pointer to an integer value of arbitrary bitwidth
13766 (result of some expression), the second is a pointer to a global string, the
13767 third is a pointer to a global string which is the source file name, and the
13768 last argument is the line number. It returns the value of the first argument.
13773 This intrinsic allows annotation of a pointer to an integer with arbitrary
13774 strings. This can be useful for special purpose optimizations that want to look
13775 for these annotations. These have no other defined use; they are ignored by code
13776 generation and optimization.
13778 '``llvm.annotation.*``' Intrinsic
13779 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13784 This is an overloaded intrinsic. You can use '``llvm.annotation``' on
13785 any integer bit width.
13789 declare i8 @llvm.annotation.i8(i8 <val>, i8* <str>, i8* <str>, i32 <int>)
13790 declare i16 @llvm.annotation.i16(i16 <val>, i8* <str>, i8* <str>, i32 <int>)
13791 declare i32 @llvm.annotation.i32(i32 <val>, i8* <str>, i8* <str>, i32 <int>)
13792 declare i64 @llvm.annotation.i64(i64 <val>, i8* <str>, i8* <str>, i32 <int>)
13793 declare i256 @llvm.annotation.i256(i256 <val>, i8* <str>, i8* <str>, i32 <int>)
13798 The '``llvm.annotation``' intrinsic.
13803 The first argument is an integer value (result of some expression), the
13804 second is a pointer to a global string, the third is a pointer to a
13805 global string which is the source file name, and the last argument is
13806 the line number. It returns the value of the first argument.
13811 This intrinsic allows annotations to be put on arbitrary expressions
13812 with arbitrary strings. This can be useful for special purpose
13813 optimizations that want to look for these annotations. These have no
13814 other defined use; they are ignored by code generation and optimization.
13816 '``llvm.codeview.annotation``' Intrinsic
13817 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13822 This annotation emits a label at its program point and an associated
13823 ``S_ANNOTATION`` codeview record with some additional string metadata. This is
13824 used to implement MSVC's ``__annotation`` intrinsic. It is marked
13825 ``noduplicate``, so calls to this intrinsic prevent inlining and should be
13826 considered expensive.
13830 declare void @llvm.codeview.annotation(metadata)
13835 The argument should be an MDTuple containing any number of MDStrings.
13837 '``llvm.trap``' Intrinsic
13838 ^^^^^^^^^^^^^^^^^^^^^^^^^
13845 declare void @llvm.trap() noreturn nounwind
13850 The '``llvm.trap``' intrinsic.
13860 This intrinsic is lowered to the target dependent trap instruction. If
13861 the target does not have a trap instruction, this intrinsic will be
13862 lowered to a call of the ``abort()`` function.
13864 '``llvm.debugtrap``' Intrinsic
13865 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13872 declare void @llvm.debugtrap() nounwind
13877 The '``llvm.debugtrap``' intrinsic.
13887 This intrinsic is lowered to code which is intended to cause an
13888 execution trap with the intention of requesting the attention of a
13891 '``llvm.stackprotector``' Intrinsic
13892 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13899 declare void @llvm.stackprotector(i8* <guard>, i8** <slot>)
13904 The ``llvm.stackprotector`` intrinsic takes the ``guard`` and stores it
13905 onto the stack at ``slot``. The stack slot is adjusted to ensure that it
13906 is placed on the stack before local variables.
13911 The ``llvm.stackprotector`` intrinsic requires two pointer arguments.
13912 The first argument is the value loaded from the stack guard
13913 ``@__stack_chk_guard``. The second variable is an ``alloca`` that has
13914 enough space to hold the value of the guard.
13919 This intrinsic causes the prologue/epilogue inserter to force the position of
13920 the ``AllocaInst`` stack slot to be before local variables on the stack. This is
13921 to ensure that if a local variable on the stack is overwritten, it will destroy
13922 the value of the guard. When the function exits, the guard on the stack is
13923 checked against the original guard by ``llvm.stackprotectorcheck``. If they are
13924 different, then ``llvm.stackprotectorcheck`` causes the program to abort by
13925 calling the ``__stack_chk_fail()`` function.
13927 '``llvm.stackguard``' Intrinsic
13928 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13935 declare i8* @llvm.stackguard()
13940 The ``llvm.stackguard`` intrinsic returns the system stack guard value.
13942 It should not be generated by frontends, since it is only for internal usage.
13943 The reason why we create this intrinsic is that we still support IR form Stack
13944 Protector in FastISel.
13954 On some platforms, the value returned by this intrinsic remains unchanged
13955 between loads in the same thread. On other platforms, it returns the same
13956 global variable value, if any, e.g. ``@__stack_chk_guard``.
13958 Currently some platforms have IR-level customized stack guard loading (e.g.
13959 X86 Linux) that is not handled by ``llvm.stackguard()``, while they should be
13962 '``llvm.objectsize``' Intrinsic
13963 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
13970 declare i32 @llvm.objectsize.i32(i8* <object>, i1 <min>, i1 <nullunknown>)
13971 declare i64 @llvm.objectsize.i64(i8* <object>, i1 <min>, i1 <nullunknown>)
13976 The ``llvm.objectsize`` intrinsic is designed to provide information to
13977 the optimizers to determine at compile time whether a) an operation
13978 (like memcpy) will overflow a buffer that corresponds to an object, or
13979 b) that a runtime check for overflow isn't necessary. An object in this
13980 context means an allocation of a specific class, structure, array, or
13986 The ``llvm.objectsize`` intrinsic takes three arguments. The first argument is
13987 a pointer to or into the ``object``. The second argument determines whether
13988 ``llvm.objectsize`` returns 0 (if true) or -1 (if false) when the object size
13989 is unknown. The third argument controls how ``llvm.objectsize`` acts when
13990 ``null`` is used as its pointer argument. If it's true and the pointer is in
13991 address space 0, ``null`` is treated as an opaque value with an unknown number
13992 of bytes. Otherwise, ``llvm.objectsize`` reports 0 bytes available when given
13995 The second and third arguments only accept constants.
14000 The ``llvm.objectsize`` intrinsic is lowered to a constant representing
14001 the size of the object concerned. If the size cannot be determined at
14002 compile time, ``llvm.objectsize`` returns ``i32/i64 -1 or 0`` (depending
14003 on the ``min`` argument).
14005 '``llvm.expect``' Intrinsic
14006 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
14011 This is an overloaded intrinsic. You can use ``llvm.expect`` on any
14016 declare i1 @llvm.expect.i1(i1 <val>, i1 <expected_val>)
14017 declare i32 @llvm.expect.i32(i32 <val>, i32 <expected_val>)
14018 declare i64 @llvm.expect.i64(i64 <val>, i64 <expected_val>)
14023 The ``llvm.expect`` intrinsic provides information about expected (the
14024 most probable) value of ``val``, which can be used by optimizers.
14029 The ``llvm.expect`` intrinsic takes two arguments. The first argument is
14030 a value. The second argument is an expected value, this needs to be a
14031 constant value, variables are not allowed.
14036 This intrinsic is lowered to the ``val``.
14040 '``llvm.assume``' Intrinsic
14041 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
14048 declare void @llvm.assume(i1 %cond)
14053 The ``llvm.assume`` allows the optimizer to assume that the provided
14054 condition is true. This information can then be used in simplifying other parts
14060 The condition which the optimizer may assume is always true.
14065 The intrinsic allows the optimizer to assume that the provided condition is
14066 always true whenever the control flow reaches the intrinsic call. No code is
14067 generated for this intrinsic, and instructions that contribute only to the
14068 provided condition are not used for code generation. If the condition is
14069 violated during execution, the behavior is undefined.
14071 Note that the optimizer might limit the transformations performed on values
14072 used by the ``llvm.assume`` intrinsic in order to preserve the instructions
14073 only used to form the intrinsic's input argument. This might prove undesirable
14074 if the extra information provided by the ``llvm.assume`` intrinsic does not cause
14075 sufficient overall improvement in code quality. For this reason,
14076 ``llvm.assume`` should not be used to document basic mathematical invariants
14077 that the optimizer can otherwise deduce or facts that are of little use to the
14082 '``llvm.ssa_copy``' Intrinsic
14083 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
14090 declare type @llvm.ssa_copy(type %operand) returned(1) readnone
14095 The first argument is an operand which is used as the returned value.
14100 The ``llvm.ssa_copy`` intrinsic can be used to attach information to
14101 operations by copying them and giving them new names. For example,
14102 the PredicateInfo utility uses it to build Extended SSA form, and
14103 attach various forms of information to operands that dominate specific
14104 uses. It is not meant for general use, only for building temporary
14105 renaming forms that require value splits at certain points.
14109 '``llvm.type.test``' Intrinsic
14110 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
14117 declare i1 @llvm.type.test(i8* %ptr, metadata %type) nounwind readnone
14123 The first argument is a pointer to be tested. The second argument is a
14124 metadata object representing a :doc:`type identifier <TypeMetadata>`.
14129 The ``llvm.type.test`` intrinsic tests whether the given pointer is associated
14130 with the given type identifier.
14132 '``llvm.type.checked.load``' Intrinsic
14133 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
14140 declare {i8*, i1} @llvm.type.checked.load(i8* %ptr, i32 %offset, metadata %type) argmemonly nounwind readonly
14146 The first argument is a pointer from which to load a function pointer. The
14147 second argument is the byte offset from which to load the function pointer. The
14148 third argument is a metadata object representing a :doc:`type identifier
14154 The ``llvm.type.checked.load`` intrinsic safely loads a function pointer from a
14155 virtual table pointer using type metadata. This intrinsic is used to implement
14156 control flow integrity in conjunction with virtual call optimization. The
14157 virtual call optimization pass will optimize away ``llvm.type.checked.load``
14158 intrinsics associated with devirtualized calls, thereby removing the type
14159 check in cases where it is not needed to enforce the control flow integrity
14162 If the given pointer is associated with a type metadata identifier, this
14163 function returns true as the second element of its return value. (Note that
14164 the function may also return true if the given pointer is not associated
14165 with a type metadata identifier.) If the function's return value's second
14166 element is true, the following rules apply to the first element:
14168 - If the given pointer is associated with the given type metadata identifier,
14169 it is the function pointer loaded from the given byte offset from the given
14172 - If the given pointer is not associated with the given type metadata
14173 identifier, it is one of the following (the choice of which is unspecified):
14175 1. The function pointer that would have been loaded from an arbitrarily chosen
14176 (through an unspecified mechanism) pointer associated with the type
14179 2. If the function has a non-void return type, a pointer to a function that
14180 returns an unspecified value without causing side effects.
14182 If the function's return value's second element is false, the value of the
14183 first element is undefined.
14186 '``llvm.donothing``' Intrinsic
14187 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
14194 declare void @llvm.donothing() nounwind readnone
14199 The ``llvm.donothing`` intrinsic doesn't perform any operation. It's one of only
14200 three intrinsics (besides ``llvm.experimental.patchpoint`` and
14201 ``llvm.experimental.gc.statepoint``) that can be called with an invoke
14212 This intrinsic does nothing, and it's removed by optimizers and ignored
14215 '``llvm.experimental.deoptimize``' Intrinsic
14216 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
14223 declare type @llvm.experimental.deoptimize(...) [ "deopt"(...) ]
14228 This intrinsic, together with :ref:`deoptimization operand bundles
14229 <deopt_opbundles>`, allow frontends to express transfer of control and
14230 frame-local state from the currently executing (typically more specialized,
14231 hence faster) version of a function into another (typically more generic, hence
14234 In languages with a fully integrated managed runtime like Java and JavaScript
14235 this intrinsic can be used to implement "uncommon trap" or "side exit" like
14236 functionality. In unmanaged languages like C and C++, this intrinsic can be
14237 used to represent the slow paths of specialized functions.
14243 The intrinsic takes an arbitrary number of arguments, whose meaning is
14244 decided by the :ref:`lowering strategy<deoptimize_lowering>`.
14249 The ``@llvm.experimental.deoptimize`` intrinsic executes an attached
14250 deoptimization continuation (denoted using a :ref:`deoptimization
14251 operand bundle <deopt_opbundles>`) and returns the value returned by
14252 the deoptimization continuation. Defining the semantic properties of
14253 the continuation itself is out of scope of the language reference --
14254 as far as LLVM is concerned, the deoptimization continuation can
14255 invoke arbitrary side effects, including reading from and writing to
14258 Deoptimization continuations expressed using ``"deopt"`` operand bundles always
14259 continue execution to the end of the physical frame containing them, so all
14260 calls to ``@llvm.experimental.deoptimize`` must be in "tail position":
14262 - ``@llvm.experimental.deoptimize`` cannot be invoked.
14263 - The call must immediately precede a :ref:`ret <i_ret>` instruction.
14264 - The ``ret`` instruction must return the value produced by the
14265 ``@llvm.experimental.deoptimize`` call if there is one, or void.
14267 Note that the above restrictions imply that the return type for a call to
14268 ``@llvm.experimental.deoptimize`` will match the return type of its immediate
14271 The inliner composes the ``"deopt"`` continuations of the caller into the
14272 ``"deopt"`` continuations present in the inlinee, and also updates calls to this
14273 intrinsic to return directly from the frame of the function it inlined into.
14275 All declarations of ``@llvm.experimental.deoptimize`` must share the
14276 same calling convention.
14278 .. _deoptimize_lowering:
14283 Calls to ``@llvm.experimental.deoptimize`` are lowered to calls to the
14284 symbol ``__llvm_deoptimize`` (it is the frontend's responsibility to
14285 ensure that this symbol is defined). The call arguments to
14286 ``@llvm.experimental.deoptimize`` are lowered as if they were formal
14287 arguments of the specified types, and not as varargs.
14290 '``llvm.experimental.guard``' Intrinsic
14291 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
14298 declare void @llvm.experimental.guard(i1, ...) [ "deopt"(...) ]
14303 This intrinsic, together with :ref:`deoptimization operand bundles
14304 <deopt_opbundles>`, allows frontends to express guards or checks on
14305 optimistic assumptions made during compilation. The semantics of
14306 ``@llvm.experimental.guard`` is defined in terms of
14307 ``@llvm.experimental.deoptimize`` -- its body is defined to be
14310 .. code-block:: text
14312 define void @llvm.experimental.guard(i1 %pred, <args...>) {
14313 %realPred = and i1 %pred, undef
14314 br i1 %realPred, label %continue, label %leave [, !make.implicit !{}]
14317 call void @llvm.experimental.deoptimize(<args...>) [ "deopt"() ]
14325 with the optional ``[, !make.implicit !{}]`` present if and only if it
14326 is present on the call site. For more details on ``!make.implicit``,
14327 see :doc:`FaultMaps`.
14329 In words, ``@llvm.experimental.guard`` executes the attached
14330 ``"deopt"`` continuation if (but **not** only if) its first argument
14331 is ``false``. Since the optimizer is allowed to replace the ``undef``
14332 with an arbitrary value, it can optimize guard to fail "spuriously",
14333 i.e. without the original condition being false (hence the "not only
14334 if"); and this allows for "check widening" type optimizations.
14336 ``@llvm.experimental.guard`` cannot be invoked.
14339 '``llvm.load.relative``' Intrinsic
14340 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
14347 declare i8* @llvm.load.relative.iN(i8* %ptr, iN %offset) argmemonly nounwind readonly
14352 This intrinsic loads a 32-bit value from the address ``%ptr + %offset``,
14353 adds ``%ptr`` to that value and returns it. The constant folder specifically
14354 recognizes the form of this intrinsic and the constant initializers it may
14355 load from; if a loaded constant initializer is known to have the form
14356 ``i32 trunc(x - %ptr)``, the intrinsic call is folded to ``x``.
14358 LLVM provides that the calculation of such a constant initializer will
14359 not overflow at link time under the medium code model if ``x`` is an
14360 ``unnamed_addr`` function. However, it does not provide this guarantee for
14361 a constant initializer folded into a function body. This intrinsic can be
14362 used to avoid the possibility of overflows when loading from such a constant.
14364 '``llvm.sideeffect``' Intrinsic
14365 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
14372 declare void @llvm.sideeffect() inaccessiblememonly nounwind
14377 The ``llvm.sideeffect`` intrinsic doesn't perform any operation. Optimizers
14378 treat it as having side effects, so it can be inserted into a loop to
14379 indicate that the loop shouldn't be assumed to terminate (which could
14380 potentially lead to the loop being optimized away entirely), even if it's
14381 an infinite loop with no other side effects.
14391 This intrinsic actually does nothing, but optimizers must assume that it
14392 has externally observable side effects.
14394 Stack Map Intrinsics
14395 --------------------
14397 LLVM provides experimental intrinsics to support runtime patching
14398 mechanisms commonly desired in dynamic language JITs. These intrinsics
14399 are described in :doc:`StackMaps`.
14401 Element Wise Atomic Memory Intrinsics
14402 -------------------------------------
14404 These intrinsics are similar to the standard library memory intrinsics except
14405 that they perform memory transfer as a sequence of atomic memory accesses.
14407 .. _int_memcpy_element_unordered_atomic:
14409 '``llvm.memcpy.element.unordered.atomic``' Intrinsic
14410 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
14415 This is an overloaded intrinsic. You can use ``llvm.memcpy.element.unordered.atomic`` on
14416 any integer bit width and for different address spaces. Not all targets
14417 support all bit widths however.
14421 declare void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i32(i8* <dest>,
14424 i32 <element_size>)
14425 declare void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* <dest>,
14428 i32 <element_size>)
14433 The '``llvm.memcpy.element.unordered.atomic.*``' intrinsic is a specialization of the
14434 '``llvm.memcpy.*``' intrinsic. It differs in that the ``dest`` and ``src`` are treated
14435 as arrays with elements that are exactly ``element_size`` bytes, and the copy between
14436 buffers uses a sequence of :ref:`unordered atomic <ordering>` load/store operations
14437 that are a positive integer multiple of the ``element_size`` in size.
14442 The first three arguments are the same as they are in the :ref:`@llvm.memcpy <int_memcpy>`
14443 intrinsic, with the added constraint that ``len`` is required to be a positive integer
14444 multiple of the ``element_size``. If ``len`` is not a positive integer multiple of
14445 ``element_size``, then the behaviour of the intrinsic is undefined.
14447 ``element_size`` must be a compile-time constant positive power of two no greater than
14448 target-specific atomic access size limit.
14450 For each of the input pointers ``align`` parameter attribute must be specified. It
14451 must be a power of two no less than the ``element_size``. Caller guarantees that
14452 both the source and destination pointers are aligned to that boundary.
14457 The '``llvm.memcpy.element.unordered.atomic.*``' intrinsic copies ``len`` bytes of
14458 memory from the source location to the destination location. These locations are not
14459 allowed to overlap. The memory copy is performed as a sequence of load/store operations
14460 where each access is guaranteed to be a multiple of ``element_size`` bytes wide and
14461 aligned at an ``element_size`` boundary.
14463 The order of the copy is unspecified. The same value may be read from the source
14464 buffer many times, but only one write is issued to the destination buffer per
14465 element. It is well defined to have concurrent reads and writes to both source and
14466 destination provided those reads and writes are unordered atomic when specified.
14468 This intrinsic does not provide any additional ordering guarantees over those
14469 provided by a set of unordered loads from the source location and stores to the
14475 In the most general case call to the '``llvm.memcpy.element.unordered.atomic.*``' is
14476 lowered to a call to the symbol ``__llvm_memcpy_element_unordered_atomic_*``. Where '*'
14477 is replaced with an actual element size.
14479 Optimizer is allowed to inline memory copy when it's profitable to do so.
14481 '``llvm.memmove.element.unordered.atomic``' Intrinsic
14482 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
14487 This is an overloaded intrinsic. You can use
14488 ``llvm.memmove.element.unordered.atomic`` on any integer bit width and for
14489 different address spaces. Not all targets support all bit widths however.
14493 declare void @llvm.memmove.element.unordered.atomic.p0i8.p0i8.i32(i8* <dest>,
14496 i32 <element_size>)
14497 declare void @llvm.memmove.element.unordered.atomic.p0i8.p0i8.i64(i8* <dest>,
14500 i32 <element_size>)
14505 The '``llvm.memmove.element.unordered.atomic.*``' intrinsic is a specialization
14506 of the '``llvm.memmove.*``' intrinsic. It differs in that the ``dest`` and
14507 ``src`` are treated as arrays with elements that are exactly ``element_size``
14508 bytes, and the copy between buffers uses a sequence of
14509 :ref:`unordered atomic <ordering>` load/store operations that are a positive
14510 integer multiple of the ``element_size`` in size.
14515 The first three arguments are the same as they are in the
14516 :ref:`@llvm.memmove <int_memmove>` intrinsic, with the added constraint that
14517 ``len`` is required to be a positive integer multiple of the ``element_size``.
14518 If ``len`` is not a positive integer multiple of ``element_size``, then the
14519 behaviour of the intrinsic is undefined.
14521 ``element_size`` must be a compile-time constant positive power of two no
14522 greater than a target-specific atomic access size limit.
14524 For each of the input pointers the ``align`` parameter attribute must be
14525 specified. It must be a power of two no less than the ``element_size``. Caller
14526 guarantees that both the source and destination pointers are aligned to that
14532 The '``llvm.memmove.element.unordered.atomic.*``' intrinsic copies ``len`` bytes
14533 of memory from the source location to the destination location. These locations
14534 are allowed to overlap. The memory copy is performed as a sequence of load/store
14535 operations where each access is guaranteed to be a multiple of ``element_size``
14536 bytes wide and aligned at an ``element_size`` boundary.
14538 The order of the copy is unspecified. The same value may be read from the source
14539 buffer many times, but only one write is issued to the destination buffer per
14540 element. It is well defined to have concurrent reads and writes to both source
14541 and destination provided those reads and writes are unordered atomic when
14544 This intrinsic does not provide any additional ordering guarantees over those
14545 provided by a set of unordered loads from the source location and stores to the
14551 In the most general case call to the
14552 '``llvm.memmove.element.unordered.atomic.*``' is lowered to a call to the symbol
14553 ``__llvm_memmove_element_unordered_atomic_*``. Where '*' is replaced with an
14554 actual element size.
14556 The optimizer is allowed to inline the memory copy when it's profitable to do so.
14558 .. _int_memset_element_unordered_atomic:
14560 '``llvm.memset.element.unordered.atomic``' Intrinsic
14561 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
14566 This is an overloaded intrinsic. You can use ``llvm.memset.element.unordered.atomic`` on
14567 any integer bit width and for different address spaces. Not all targets
14568 support all bit widths however.
14572 declare void @llvm.memset.element.unordered.atomic.p0i8.i32(i8* <dest>,
14575 i32 <element_size>)
14576 declare void @llvm.memset.element.unordered.atomic.p0i8.i64(i8* <dest>,
14579 i32 <element_size>)
14584 The '``llvm.memset.element.unordered.atomic.*``' intrinsic is a specialization of the
14585 '``llvm.memset.*``' intrinsic. It differs in that the ``dest`` is treated as an array
14586 with elements that are exactly ``element_size`` bytes, and the assignment to that array
14587 uses uses a sequence of :ref:`unordered atomic <ordering>` store operations
14588 that are a positive integer multiple of the ``element_size`` in size.
14593 The first three arguments are the same as they are in the :ref:`@llvm.memset <int_memset>`
14594 intrinsic, with the added constraint that ``len`` is required to be a positive integer
14595 multiple of the ``element_size``. If ``len`` is not a positive integer multiple of
14596 ``element_size``, then the behaviour of the intrinsic is undefined.
14598 ``element_size`` must be a compile-time constant positive power of two no greater than
14599 target-specific atomic access size limit.
14601 The ``dest`` input pointer must have the ``align`` parameter attribute specified. It
14602 must be a power of two no less than the ``element_size``. Caller guarantees that
14603 the destination pointer is aligned to that boundary.
14608 The '``llvm.memset.element.unordered.atomic.*``' intrinsic sets the ``len`` bytes of
14609 memory starting at the destination location to the given ``value``. The memory is
14610 set with a sequence of store operations where each access is guaranteed to be a
14611 multiple of ``element_size`` bytes wide and aligned at an ``element_size`` boundary.
14613 The order of the assignment is unspecified. Only one write is issued to the
14614 destination buffer per element. It is well defined to have concurrent reads and
14615 writes to the destination provided those reads and writes are unordered atomic
14618 This intrinsic does not provide any additional ordering guarantees over those
14619 provided by a set of unordered stores to the destination.
14624 In the most general case call to the '``llvm.memset.element.unordered.atomic.*``' is
14625 lowered to a call to the symbol ``__llvm_memset_element_unordered_atomic_*``. Where '*'
14626 is replaced with an actual element size.
14628 The optimizer is allowed to inline the memory assignment when it's profitable to do so.