1 ==============================
2 LLVM Language Reference Manual
3 ==============================
12 This document is a reference manual for the LLVM assembly language. LLVM
13 is a Static Single Assignment (SSA) based representation that provides
14 type safety, low-level operations, flexibility, and the capability of
15 representing 'all' high-level languages cleanly. It is the common code
16 representation used throughout all phases of the LLVM compilation
22 The LLVM code representation is designed to be used in three different
23 forms: as an in-memory compiler IR, as an on-disk bitcode representation
24 (suitable for fast loading by a Just-In-Time compiler), and as a human
25 readable assembly language representation. This allows LLVM to provide a
26 powerful intermediate representation for efficient compiler
27 transformations and analysis, while providing a natural means to debug
28 and visualize the transformations. The three different forms of LLVM are
29 all equivalent. This document describes the human readable
30 representation and notation.
32 The LLVM representation aims to be light-weight and low-level while
33 being expressive, typed, and extensible at the same time. It aims to be
34 a "universal IR" of sorts, by being at a low enough level that
35 high-level ideas may be cleanly mapped to it (similar to how
36 microprocessors are "universal IR's", allowing many source languages to
37 be mapped to them). By providing type information, LLVM can be used as
38 the target of optimizations: for example, through pointer analysis, it
39 can be proven that a C automatic variable is never accessed outside of
40 the current function, allowing it to be promoted to a simple SSA value
41 instead of a memory location.
48 It is important to note that this document describes 'well formed' LLVM
49 assembly language. There is a difference between what the parser accepts
50 and what is considered 'well formed'. For example, the following
51 instruction is syntactically okay, but not well formed:
57 because the definition of ``%x`` does not dominate all of its uses. The
58 LLVM infrastructure provides a verification pass that may be used to
59 verify that an LLVM module is well formed. This pass is automatically
60 run by the parser after parsing input assembly and by the optimizer
61 before it outputs bitcode. The violations pointed out by the verifier
62 pass indicate bugs in transformation passes or input to the parser.
69 LLVM identifiers come in two basic types: global and local. Global
70 identifiers (functions, global variables) begin with the ``'@'``
71 character. Local identifiers (register names, types) begin with the
72 ``'%'`` character. Additionally, there are three different formats for
73 identifiers, for different purposes:
75 #. Named values are represented as a string of characters with their
76 prefix. For example, ``%foo``, ``@DivisionByZero``,
77 ``%a.really.long.identifier``. The actual regular expression used is
78 '``[%@][-a-zA-Z$._][-a-zA-Z$._0-9]*``'. Identifiers that require other
79 characters in their names can be surrounded with quotes. Special
80 characters may be escaped using ``"\xx"`` where ``xx`` is the ASCII
81 code for the character in hexadecimal. In this way, any character can
82 be used in a name value, even quotes themselves. The ``"\01"`` prefix
83 can be used on global variables to suppress mangling.
84 #. Unnamed values are represented as an unsigned numeric value with
85 their prefix. For example, ``%12``, ``@2``, ``%44``.
86 #. Constants, which are described in the section Constants_ below.
88 LLVM requires that values start with a prefix for two reasons: Compilers
89 don't need to worry about name clashes with reserved words, and the set
90 of reserved words may be expanded in the future without penalty.
91 Additionally, unnamed identifiers allow a compiler to quickly come up
92 with a temporary variable without having to avoid symbol table
95 Reserved words in LLVM are very similar to reserved words in other
96 languages. There are keywords for different opcodes ('``add``',
97 '``bitcast``', '``ret``', etc...), for primitive type names ('``void``',
98 '``i32``', etc...), and others. These reserved words cannot conflict
99 with variable names, because none of them start with a prefix character
100 (``'%'`` or ``'@'``).
102 Here is an example of LLVM code to multiply the integer variable
109 %result = mul i32 %X, 8
111 After strength reduction:
115 %result = shl i32 %X, 3
121 %0 = add i32 %X, %X ; yields i32:%0
122 %1 = add i32 %0, %0 ; yields i32:%1
123 %result = add i32 %1, %1
125 This last way of multiplying ``%X`` by 8 illustrates several important
126 lexical features of LLVM:
128 #. Comments are delimited with a '``;``' and go until the end of line.
129 #. Unnamed temporaries are created when the result of a computation is
130 not assigned to a named value.
131 #. Unnamed temporaries are numbered sequentially (using a per-function
132 incrementing counter, starting with 0). Note that basic blocks and unnamed
133 function parameters are included in this numbering. For example, if the
134 entry basic block is not given a label name and all function parameters are
135 named, then it will get number 0.
137 It also shows a convention that we follow in this document. When
138 demonstrating instructions, we will follow an instruction with a comment
139 that defines the type and name of value produced.
147 LLVM programs are composed of ``Module``'s, each of which is a
148 translation unit of the input programs. Each module consists of
149 functions, global variables, and symbol table entries. Modules may be
150 combined together with the LLVM linker, which merges function (and
151 global variable) definitions, resolves forward declarations, and merges
152 symbol table entries. Here is an example of the "hello world" module:
156 ; Declare the string constant as a global constant.
157 @.str = private unnamed_addr constant [13 x i8] c"hello world\0A\00"
159 ; External declaration of the puts function
160 declare i32 @puts(i8* nocapture) nounwind
162 ; Definition of main function
163 define i32 @main() { ; i32()*
164 ; Convert [13 x i8]* to i8 *...
165 %cast210 = getelementptr [13 x i8], [13 x i8]* @.str, i64 0, i64 0
167 ; Call puts function to write out the string to stdout.
168 call i32 @puts(i8* %cast210)
173 !0 = !{i32 42, null, !"string"}
176 This example is made up of a :ref:`global variable <globalvars>` named
177 "``.str``", an external declaration of the "``puts``" function, a
178 :ref:`function definition <functionstructure>` for "``main``" and
179 :ref:`named metadata <namedmetadatastructure>` "``foo``".
181 In general, a module is made up of a list of global values (where both
182 functions and global variables are global values). Global values are
183 represented by a pointer to a memory location (in this case, a pointer
184 to an array of char, and a pointer to a function), and have one of the
185 following :ref:`linkage types <linkage>`.
192 All Global Variables and Functions have one of the following types of
196 Global values with "``private``" linkage are only directly
197 accessible by objects in the current module. In particular, linking
198 code into a module with an private global value may cause the
199 private to be renamed as necessary to avoid collisions. Because the
200 symbol is private to the module, all references can be updated. This
201 doesn't show up in any symbol table in the object file.
203 Similar to private, but the value shows as a local symbol
204 (``STB_LOCAL`` in the case of ELF) in the object file. This
205 corresponds to the notion of the '``static``' keyword in C.
206 ``available_externally``
207 Globals with "``available_externally``" linkage are never emitted 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.
535 LLVM IR allows you to specify both "identified" and "literal" :ref:`structure
536 types <t_struct>`. Literal types are uniqued structurally, but identified types
537 are never uniqued. An :ref:`opaque structural type <t_opaque>` can also be used
538 to forward declare a type that is not yet available.
540 An example of an identified structure specification is:
544 %mytype = type { %mytype*, i32 }
546 Prior to the LLVM 3.0 release, identified types were structurally uniqued. Only
547 literal types are uniqued in recent versions of LLVM.
554 Global variables define regions of memory allocated at compilation time
557 Global variable definitions must be initialized.
559 Global variables in other translation units can also be declared, in which
560 case they don't have an initializer.
562 Either global variable definitions or declarations may have an explicit section
563 to be placed in and may have an optional explicit alignment specified.
565 A variable may be defined as a global ``constant``, which indicates that
566 the contents of the variable will **never** be modified (enabling better
567 optimization, allowing the global data to be placed in the read-only
568 section of an executable, etc). Note that variables that need runtime
569 initialization cannot be marked ``constant`` as there is a store to the
572 LLVM explicitly allows *declarations* of global variables to be marked
573 constant, even if the final definition of the global is not. This
574 capability can be used to enable slightly better optimization of the
575 program, but requires the language definition to guarantee that
576 optimizations based on the 'constantness' are valid for the translation
577 units that do not include the definition.
579 As SSA values, global variables define pointer values that are in scope
580 (i.e. they dominate) all basic blocks in the program. Global variables
581 always define a pointer to their "content" type because they describe a
582 region of memory, and all memory objects in LLVM are accessed through
585 Global variables can be marked with ``unnamed_addr`` which indicates
586 that the address is not significant, only the content. Constants marked
587 like this can be merged with other constants if they have the same
588 initializer. Note that a constant with significant address *can* be
589 merged with a ``unnamed_addr`` constant, the result being a constant
590 whose address is significant.
592 If the ``local_unnamed_addr`` attribute is given, the address is known to
593 not be significant within the module.
595 A global variable may be declared to reside in a target-specific
596 numbered address space. For targets that support them, address spaces
597 may affect how optimizations are performed and/or what target
598 instructions are used to access the variable. The default address space
599 is zero. The address space qualifier must precede any other attributes.
601 LLVM allows an explicit section to be specified for globals. If the
602 target supports it, it will emit globals to the section specified.
603 Additionally, the global can placed in a comdat if the target has the necessary
606 By default, global initializers are optimized by assuming that global
607 variables defined within the module are not modified from their
608 initial values before the start of the global initializer. This is
609 true even for variables potentially accessible from outside the
610 module, including those with external linkage or appearing in
611 ``@llvm.used`` or dllexported variables. This assumption may be suppressed
612 by marking the variable with ``externally_initialized``.
614 An explicit alignment may be specified for a global, which must be a
615 power of 2. If not present, or if the alignment is set to zero, the
616 alignment of the global is set by the target to whatever it feels
617 convenient. If an explicit alignment is specified, the global is forced
618 to have exactly that alignment. Targets and optimizers are not allowed
619 to over-align the global if the global has an assigned section. In this
620 case, the extra alignment could be observable: for example, code could
621 assume that the globals are densely packed in their section and try to
622 iterate over them as an array, alignment padding would break this
623 iteration. The maximum alignment is ``1 << 29``.
625 Globals can also have a :ref:`DLL storage class <dllstorageclass>` and
626 an optional list of attached :ref:`metadata <metadata>`,
628 Variables and aliases can have a
629 :ref:`Thread Local Storage Model <tls_model>`.
633 @<GlobalVarName> = [Linkage] [Visibility] [DLLStorageClass] [ThreadLocal]
634 [(unnamed_addr|local_unnamed_addr)] [AddrSpace]
635 [ExternallyInitialized]
636 <global | constant> <Type> [<InitializerConstant>]
637 [, section "name"] [, comdat [($name)]]
638 [, align <Alignment>] (, !name !N)*
640 For example, the following defines a global in a numbered address space
641 with an initializer, section, and alignment:
645 @G = addrspace(5) constant float 1.0, section "foo", align 4
647 The following example just declares a global variable
651 @G = external global i32
653 The following example defines a thread-local global with the
654 ``initialexec`` TLS model:
658 @G = thread_local(initialexec) global i32 0, align 4
660 .. _functionstructure:
665 LLVM function definitions consist of the "``define``" keyword, an
666 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
667 style <visibility>`, an optional :ref:`DLL storage class <dllstorageclass>`,
668 an optional :ref:`calling convention <callingconv>`,
669 an optional ``unnamed_addr`` attribute, a return type, an optional
670 :ref:`parameter attribute <paramattrs>` for the return type, a function
671 name, a (possibly empty) argument list (each with optional :ref:`parameter
672 attributes <paramattrs>`), optional :ref:`function attributes <fnattrs>`,
673 an optional section, an optional alignment,
674 an optional :ref:`comdat <langref_comdats>`,
675 an optional :ref:`garbage collector name <gc>`, an optional :ref:`prefix <prefixdata>`,
676 an optional :ref:`prologue <prologuedata>`,
677 an optional :ref:`personality <personalityfn>`,
678 an optional list of attached :ref:`metadata <metadata>`,
679 an opening curly brace, a list of basic blocks, and a closing curly brace.
681 LLVM function declarations consist of the "``declare``" keyword, an
682 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility style
683 <visibility>`, an optional :ref:`DLL storage class <dllstorageclass>`, an
684 optional :ref:`calling convention <callingconv>`, an optional ``unnamed_addr``
685 or ``local_unnamed_addr`` attribute, a return type, an optional :ref:`parameter
686 attribute <paramattrs>` for the return type, a function name, a possibly
687 empty list of arguments, an optional alignment, an optional :ref:`garbage
688 collector name <gc>`, an optional :ref:`prefix <prefixdata>`, and an optional
689 :ref:`prologue <prologuedata>`.
691 A function definition contains a list of basic blocks, forming the CFG (Control
692 Flow Graph) for the function. Each basic block may optionally start with a label
693 (giving the basic block a symbol table entry), contains a list of instructions,
694 and ends with a :ref:`terminator <terminators>` instruction (such as a branch or
695 function return). If an explicit label is not provided, a block is assigned an
696 implicit numbered label, using the next value from the same counter as used for
697 unnamed temporaries (:ref:`see above<identifiers>`). For example, if a function
698 entry block does not have an explicit label, it will be assigned label "%0",
699 then the first unnamed temporary in that block will be "%1", etc.
701 The first basic block in a function is special in two ways: it is
702 immediately executed on entrance to the function, and it is not allowed
703 to have predecessor basic blocks (i.e. there can not be any branches to
704 the entry block of a function). Because the block can have no
705 predecessors, it also cannot have any :ref:`PHI nodes <i_phi>`.
707 LLVM allows an explicit section to be specified for functions. If the
708 target supports it, it will emit functions to the section specified.
709 Additionally, the function can be placed in a COMDAT.
711 An explicit alignment may be specified for a function. If not present,
712 or if the alignment is set to zero, the alignment of the function is set
713 by the target to whatever it feels convenient. If an explicit alignment
714 is specified, the function is forced to have at least that much
715 alignment. All alignments must be a power of 2.
717 If the ``unnamed_addr`` attribute is given, the address is known to not
718 be significant and two identical functions can be merged.
720 If the ``local_unnamed_addr`` attribute is given, the address is known to
721 not be significant within the module.
725 define [linkage] [visibility] [DLLStorageClass]
727 <ResultType> @<FunctionName> ([argument list])
728 [(unnamed_addr|local_unnamed_addr)] [fn Attrs] [section "name"]
729 [comdat [($name)]] [align N] [gc] [prefix Constant]
730 [prologue Constant] [personality Constant] (!name !N)* { ... }
732 The argument list is a comma separated sequence of arguments where each
733 argument is of the following form:
737 <type> [parameter Attrs] [name]
745 Aliases, unlike function or variables, don't create any new data. They
746 are just a new symbol and metadata for an existing position.
748 Aliases have a name and an aliasee that is either a global value or a
751 Aliases may have an optional :ref:`linkage type <linkage>`, an optional
752 :ref:`visibility style <visibility>`, an optional :ref:`DLL storage class
753 <dllstorageclass>` and an optional :ref:`tls model <tls_model>`.
757 @<Name> = [Linkage] [Visibility] [DLLStorageClass] [ThreadLocal] [(unnamed_addr|local_unnamed_addr)] alias <AliaseeTy>, <AliaseeTy>* @<Aliasee>
759 The linkage must be one of ``private``, ``internal``, ``linkonce``, ``weak``,
760 ``linkonce_odr``, ``weak_odr``, ``external``. Note that some system linkers
761 might not correctly handle dropping a weak symbol that is aliased.
763 Aliases that are not ``unnamed_addr`` are guaranteed to have the same address as
764 the aliasee expression. ``unnamed_addr`` ones are only guaranteed to point
767 If the ``local_unnamed_addr`` attribute is given, the address is known to
768 not be significant within the module.
770 Since aliases are only a second name, some restrictions apply, of which
771 some can only be checked when producing an object file:
773 * The expression defining the aliasee must be computable at assembly
774 time. Since it is just a name, no relocations can be used.
776 * No alias in the expression can be weak as the possibility of the
777 intermediate alias being overridden cannot be represented in an
780 * No global value in the expression can be a declaration, since that
781 would require a relocation, which is not possible.
788 IFuncs, like as aliases, don't create any new data or func. They are just a new
789 symbol that dynamic linker resolves at runtime by calling a resolver function.
791 IFuncs have a name and a resolver that is a function called by dynamic linker
792 that returns address of another function associated with the name.
794 IFunc may have an optional :ref:`linkage type <linkage>` and an optional
795 :ref:`visibility style <visibility>`.
799 @<Name> = [Linkage] [Visibility] ifunc <IFuncTy>, <ResolverTy>* @<Resolver>
807 Comdat IR provides access to COFF and ELF object file COMDAT functionality.
809 Comdats have a name which represents the COMDAT key. All global objects that
810 specify this key will only end up in the final object file if the linker chooses
811 that key over some other key. Aliases are placed in the same COMDAT that their
812 aliasee computes to, if any.
814 Comdats have a selection kind to provide input on how the linker should
815 choose between keys in two different object files.
819 $<Name> = comdat SelectionKind
821 The selection kind must be one of the following:
824 The linker may choose any COMDAT key, the choice is arbitrary.
826 The linker may choose any COMDAT key but the sections must contain the
829 The linker will choose the section containing the largest COMDAT key.
831 The linker requires that only section with this COMDAT key exist.
833 The linker may choose any COMDAT key but the sections must contain the
836 Note that the Mach-O platform doesn't support COMDATs and ELF only supports
837 ``any`` as a selection kind.
839 Here is an example of a COMDAT group where a function will only be selected if
840 the COMDAT key's section is the largest:
844 $foo = comdat largest
845 @foo = global i32 2, comdat($foo)
847 define void @bar() comdat($foo) {
851 As a syntactic sugar the ``$name`` can be omitted if the name is the same as
857 @foo = global i32 2, comdat
860 In a COFF object file, this will create a COMDAT section with selection kind
861 ``IMAGE_COMDAT_SELECT_LARGEST`` containing the contents of the ``@foo`` symbol
862 and another COMDAT section with selection kind
863 ``IMAGE_COMDAT_SELECT_ASSOCIATIVE`` which is associated with the first COMDAT
864 section and contains the contents of the ``@bar`` symbol.
866 There are some restrictions on the properties of the global object.
867 It, or an alias to it, must have the same name as the COMDAT group when
869 The contents and size of this object may be used during link-time to determine
870 which COMDAT groups get selected depending on the selection kind.
871 Because the name of the object must match the name of the COMDAT group, the
872 linkage of the global object must not be local; local symbols can get renamed
873 if a collision occurs in the symbol table.
875 The combined use of COMDATS and section attributes may yield surprising results.
882 @g1 = global i32 42, section "sec", comdat($foo)
883 @g2 = global i32 42, section "sec", comdat($bar)
885 From the object file perspective, this requires the creation of two sections
886 with the same name. This is necessary because both globals belong to different
887 COMDAT groups and COMDATs, at the object file level, are represented by
890 Note that certain IR constructs like global variables and functions may
891 create COMDATs in the object file in addition to any which are specified using
892 COMDAT IR. This arises when the code generator is configured to emit globals
893 in individual sections (e.g. when `-data-sections` or `-function-sections`
894 is supplied to `llc`).
896 .. _namedmetadatastructure:
901 Named metadata is a collection of metadata. :ref:`Metadata
902 nodes <metadata>` (but not metadata strings) are the only valid
903 operands for a named metadata.
905 #. Named metadata are represented as a string of characters with the
906 metadata prefix. The rules for metadata names are the same as for
907 identifiers, but quoted names are not allowed. ``"\xx"`` type escapes
908 are still valid, which allows any character to be part of a name.
912 ; Some unnamed metadata nodes, which are referenced by the named metadata.
917 !name = !{!0, !1, !2}
924 The return type and each parameter of a function type may have a set of
925 *parameter attributes* associated with them. Parameter attributes are
926 used to communicate additional information about the result or
927 parameters of a function. Parameter attributes are considered to be part
928 of the function, not of the function type, so functions with different
929 parameter attributes can have the same function type.
931 Parameter attributes are simple keywords that follow the type specified.
932 If multiple parameter attributes are needed, they are space separated.
937 declare i32 @printf(i8* noalias nocapture, ...)
938 declare i32 @atoi(i8 zeroext)
939 declare signext i8 @returns_signed_char()
941 Note that any attributes for the function result (``nounwind``,
942 ``readonly``) come immediately after the argument list.
944 Currently, only the following parameter attributes are defined:
947 This indicates to the code generator that the parameter or return
948 value should be zero-extended to the extent required by the target's
949 ABI by the caller (for a parameter) or the callee (for a return value).
951 This indicates to the code generator that the parameter or return
952 value should be sign-extended to the extent required by the target's
953 ABI (which is usually 32-bits) by the caller (for a parameter) or
954 the callee (for a return value).
956 This indicates that this parameter or return value should be treated
957 in a special target-dependent fashion while emitting code for
958 a function call or return (usually, by putting it in a register as
959 opposed to memory, though some targets use it to distinguish between
960 two different kinds of registers). Use of this attribute is
963 This indicates that the pointer parameter should really be passed by
964 value to the function. The attribute implies that a hidden copy of
965 the pointee is made between the caller and the callee, so the callee
966 is unable to modify the value in the caller. This attribute is only
967 valid on LLVM pointer arguments. It is generally used to pass
968 structs and arrays by value, but is also valid on pointers to
969 scalars. The copy is considered to belong to the caller not the
970 callee (for example, ``readonly`` functions should not write to
971 ``byval`` parameters). This is not a valid attribute for return
974 The byval attribute also supports specifying an alignment with the
975 align attribute. It indicates the alignment of the stack slot to
976 form and the known alignment of the pointer specified to the call
977 site. If the alignment is not specified, then the code generator
978 makes a target-specific assumption.
984 The ``inalloca`` argument attribute allows the caller to take the
985 address of outgoing stack arguments. An ``inalloca`` argument must
986 be a pointer to stack memory produced by an ``alloca`` instruction.
987 The alloca, or argument allocation, must also be tagged with the
988 inalloca keyword. Only the last argument may have the ``inalloca``
989 attribute, and that argument is guaranteed to be passed in memory.
991 An argument allocation may be used by a call at most once because
992 the call may deallocate it. The ``inalloca`` attribute cannot be
993 used in conjunction with other attributes that affect argument
994 storage, like ``inreg``, ``nest``, ``sret``, or ``byval``. The
995 ``inalloca`` attribute also disables LLVM's implicit lowering of
996 large aggregate return values, which means that frontend authors
997 must lower them with ``sret`` pointers.
999 When the call site is reached, the argument allocation must have
1000 been the most recent stack allocation that is still live, or the
1001 results are undefined. It is possible to allocate additional stack
1002 space after an argument allocation and before its call site, but it
1003 must be cleared off with :ref:`llvm.stackrestore
1004 <int_stackrestore>`.
1006 See :doc:`InAlloca` for more information on how to use this
1010 This indicates that the pointer parameter specifies the address of a
1011 structure that is the return value of the function in the source
1012 program. This pointer must be guaranteed by the caller to be valid:
1013 loads and stores to the structure may be assumed by the callee
1014 not to trap and to be properly aligned. This may only be applied to
1015 the first parameter. This is not a valid attribute for return
1019 This indicates that the pointer value may be assumed by the optimizer to
1020 have the specified alignment.
1022 Note that this attribute has additional semantics when combined with the
1023 ``byval`` attribute.
1028 This indicates that objects accessed via pointer values
1029 :ref:`based <pointeraliasing>` on the argument or return value are not also
1030 accessed, during the execution of the function, via pointer values not
1031 *based* on the argument or return value. The attribute on a return value
1032 also has additional semantics described below. The caller shares the
1033 responsibility with the callee for ensuring that these requirements are met.
1034 For further details, please see the discussion of the NoAlias response in
1035 :ref:`alias analysis <Must, May, or No>`.
1037 Note that this definition of ``noalias`` is intentionally similar
1038 to the definition of ``restrict`` in C99 for function arguments.
1040 For function return values, C99's ``restrict`` is not meaningful,
1041 while LLVM's ``noalias`` is. Furthermore, the semantics of the ``noalias``
1042 attribute on return values are stronger than the semantics of the attribute
1043 when used on function arguments. On function return values, the ``noalias``
1044 attribute indicates that the function acts like a system memory allocation
1045 function, returning a pointer to allocated storage disjoint from the
1046 storage for any other object accessible to the caller.
1049 This indicates that the callee does not make any copies of the
1050 pointer that outlive the callee itself. This is not a valid
1051 attribute for return values. Addresses used in volatile operations
1052 are considered to be captured.
1057 This indicates that the pointer parameter can be excised using the
1058 :ref:`trampoline intrinsics <int_trampoline>`. This is not a valid
1059 attribute for return values and can only be applied to one parameter.
1062 This indicates that the function always returns the argument as its return
1063 value. This is a hint to the optimizer and code generator used when
1064 generating the caller, allowing value propagation, tail call optimization,
1065 and omission of register saves and restores in some cases; it is not
1066 checked or enforced when generating the callee. The parameter and the
1067 function return type must be valid operands for the
1068 :ref:`bitcast instruction <i_bitcast>`. This is not a valid attribute for
1069 return values and can only be applied to one parameter.
1072 This indicates that the parameter or return pointer is not null. This
1073 attribute may only be applied to pointer typed parameters. This is not
1074 checked or enforced by LLVM, the caller must ensure that the pointer
1075 passed in is non-null, or the callee must ensure that the returned pointer
1078 ``dereferenceable(<n>)``
1079 This indicates that the parameter or return pointer is dereferenceable. This
1080 attribute may only be applied to pointer typed parameters. A pointer that
1081 is dereferenceable can be loaded from speculatively without a risk of
1082 trapping. The number of bytes known to be dereferenceable must be provided
1083 in parentheses. It is legal for the number of bytes to be less than the
1084 size of the pointee type. The ``nonnull`` attribute does not imply
1085 dereferenceability (consider a pointer to one element past the end of an
1086 array), however ``dereferenceable(<n>)`` does imply ``nonnull`` in
1087 ``addrspace(0)`` (which is the default address space).
1089 ``dereferenceable_or_null(<n>)``
1090 This indicates that the parameter or return value isn't both
1091 non-null and non-dereferenceable (up to ``<n>`` bytes) at the same
1092 time. All non-null pointers tagged with
1093 ``dereferenceable_or_null(<n>)`` are ``dereferenceable(<n>)``.
1094 For address space 0 ``dereferenceable_or_null(<n>)`` implies that
1095 a pointer is exactly one of ``dereferenceable(<n>)`` or ``null``,
1096 and in other address spaces ``dereferenceable_or_null(<n>)``
1097 implies that a pointer is at least one of ``dereferenceable(<n>)``
1098 or ``null`` (i.e. it may be both ``null`` and
1099 ``dereferenceable(<n>)``). This attribute may only be applied to
1100 pointer typed parameters.
1103 This indicates that the parameter is the self/context parameter. This is not
1104 a valid attribute for return values and can only be applied to one
1108 This attribute is motivated to model and optimize Swift error handling. It
1109 can be applied to a parameter with pointer to pointer type or a
1110 pointer-sized alloca. At the call site, the actual argument that corresponds
1111 to a ``swifterror`` parameter has to come from a ``swifterror`` alloca. A
1112 ``swifterror`` value (either the parameter or the alloca) can only be loaded
1113 and stored from, or used as a ``swifterror`` argument. This is not a valid
1114 attribute for return values and can only be applied to one parameter.
1116 These constraints allow the calling convention to optimize access to
1117 ``swifterror`` variables by associating them with a specific register at
1118 call boundaries rather than placing them in memory. Since this does change
1119 the calling convention, a function which uses the ``swifterror`` attribute
1120 on a parameter is not ABI-compatible with one which does not.
1122 These constraints also allow LLVM to assume that a ``swifterror`` argument
1123 does not alias any other memory visible within a function and that a
1124 ``swifterror`` alloca passed as an argument does not escape.
1128 Garbage Collector Strategy Names
1129 --------------------------------
1131 Each function may specify a garbage collector strategy name, which is simply a
1134 .. code-block:: llvm
1136 define void @f() gc "name" { ... }
1138 The supported values of *name* includes those :ref:`built in to LLVM
1139 <builtin-gc-strategies>` and any provided by loaded plugins. Specifying a GC
1140 strategy will cause the compiler to alter its output in order to support the
1141 named garbage collection algorithm. Note that LLVM itself does not contain a
1142 garbage collector, this functionality is restricted to generating machine code
1143 which can interoperate with a collector provided externally.
1150 Prefix data is data associated with a function which the code
1151 generator will emit immediately before the function's entrypoint.
1152 The purpose of this feature is to allow frontends to associate
1153 language-specific runtime metadata with specific functions and make it
1154 available through the function pointer while still allowing the
1155 function pointer to be called.
1157 To access the data for a given function, a program may bitcast the
1158 function pointer to a pointer to the constant's type and dereference
1159 index -1. This implies that the IR symbol points just past the end of
1160 the prefix data. For instance, take the example of a function annotated
1161 with a single ``i32``,
1163 .. code-block:: llvm
1165 define void @f() prefix i32 123 { ... }
1167 The prefix data can be referenced as,
1169 .. code-block:: llvm
1171 %0 = bitcast void* () @f to i32*
1172 %a = getelementptr inbounds i32, i32* %0, i32 -1
1173 %b = load i32, i32* %a
1175 Prefix data is laid out as if it were an initializer for a global variable
1176 of the prefix data's type. The function will be placed such that the
1177 beginning of the prefix data is aligned. This means that if the size
1178 of the prefix data is not a multiple of the alignment size, the
1179 function's entrypoint will not be aligned. If alignment of the
1180 function's entrypoint is desired, padding must be added to the prefix
1183 A function may have prefix data but no body. This has similar semantics
1184 to the ``available_externally`` linkage in that the data may be used by the
1185 optimizers but will not be emitted in the object file.
1192 The ``prologue`` attribute allows arbitrary code (encoded as bytes) to
1193 be inserted prior to the function body. This can be used for enabling
1194 function hot-patching and instrumentation.
1196 To maintain the semantics of ordinary function calls, the prologue data must
1197 have a particular format. Specifically, it must begin with a sequence of
1198 bytes which decode to a sequence of machine instructions, valid for the
1199 module's target, which transfer control to the point immediately succeeding
1200 the prologue data, without performing any other visible action. This allows
1201 the inliner and other passes to reason about the semantics of the function
1202 definition without needing to reason about the prologue data. Obviously this
1203 makes the format of the prologue data highly target dependent.
1205 A trivial example of valid prologue data for the x86 architecture is ``i8 144``,
1206 which encodes the ``nop`` instruction:
1208 .. code-block:: text
1210 define void @f() prologue i8 144 { ... }
1212 Generally prologue data can be formed by encoding a relative branch instruction
1213 which skips the metadata, as in this example of valid prologue data for the
1214 x86_64 architecture, where the first two bytes encode ``jmp .+10``:
1216 .. code-block:: text
1218 %0 = type <{ i8, i8, i8* }>
1220 define void @f() prologue %0 <{ i8 235, i8 8, i8* @md}> { ... }
1222 A function may have prologue data but no body. This has similar semantics
1223 to the ``available_externally`` linkage in that the data may be used by the
1224 optimizers but will not be emitted in the object file.
1228 Personality Function
1229 --------------------
1231 The ``personality`` attribute permits functions to specify what function
1232 to use for exception handling.
1239 Attribute groups are groups of attributes that are referenced by objects within
1240 the IR. They are important for keeping ``.ll`` files readable, because a lot of
1241 functions will use the same set of attributes. In the degenerative case of a
1242 ``.ll`` file that corresponds to a single ``.c`` file, the single attribute
1243 group will capture the important command line flags used to build that file.
1245 An attribute group is a module-level object. To use an attribute group, an
1246 object references the attribute group's ID (e.g. ``#37``). An object may refer
1247 to more than one attribute group. In that situation, the attributes from the
1248 different groups are merged.
1250 Here is an example of attribute groups for a function that should always be
1251 inlined, has a stack alignment of 4, and which shouldn't use SSE instructions:
1253 .. code-block:: llvm
1255 ; Target-independent attributes:
1256 attributes #0 = { alwaysinline alignstack=4 }
1258 ; Target-dependent attributes:
1259 attributes #1 = { "no-sse" }
1261 ; Function @f has attributes: alwaysinline, alignstack=4, and "no-sse".
1262 define void @f() #0 #1 { ... }
1269 Function attributes are set to communicate additional information about
1270 a function. Function attributes are considered to be part of the
1271 function, not of the function type, so functions with different function
1272 attributes can have the same function type.
1274 Function attributes are simple keywords that follow the type specified.
1275 If multiple attributes are needed, they are space separated. For
1278 .. code-block:: llvm
1280 define void @f() noinline { ... }
1281 define void @f() alwaysinline { ... }
1282 define void @f() alwaysinline optsize { ... }
1283 define void @f() optsize { ... }
1286 This attribute indicates that, when emitting the prologue and
1287 epilogue, the backend should forcibly align the stack pointer.
1288 Specify the desired alignment, which must be a power of two, in
1290 ``allocsize(<EltSizeParam>[, <NumEltsParam>])``
1291 This attribute indicates that the annotated function will always return at
1292 least a given number of bytes (or null). Its arguments are zero-indexed
1293 parameter numbers; if one argument is provided, then it's assumed that at
1294 least ``CallSite.Args[EltSizeParam]`` bytes will be available at the
1295 returned pointer. If two are provided, then it's assumed that
1296 ``CallSite.Args[EltSizeParam] * CallSite.Args[NumEltsParam]`` bytes are
1297 available. The referenced parameters must be integer types. No assumptions
1298 are made about the contents of the returned block of memory.
1300 This attribute indicates that the inliner should attempt to inline
1301 this function into callers whenever possible, ignoring any active
1302 inlining size threshold for this caller.
1304 This indicates that the callee function at a call site should be
1305 recognized as a built-in function, even though the function's declaration
1306 uses the ``nobuiltin`` attribute. This is only valid at call sites for
1307 direct calls to functions that are declared with the ``nobuiltin``
1310 This attribute indicates that this function is rarely called. When
1311 computing edge weights, basic blocks post-dominated by a cold
1312 function call are also considered to be cold; and, thus, given low
1315 In some parallel execution models, there exist operations that cannot be
1316 made control-dependent on any additional values. We call such operations
1317 ``convergent``, and mark them with this attribute.
1319 The ``convergent`` attribute may appear on functions or call/invoke
1320 instructions. When it appears on a function, it indicates that calls to
1321 this function should not be made control-dependent on additional values.
1322 For example, the intrinsic ``llvm.nvvm.barrier0`` is ``convergent``, so
1323 calls to this intrinsic cannot be made control-dependent on additional
1326 When it appears on a call/invoke, the ``convergent`` attribute indicates
1327 that we should treat the call as though we're calling a convergent
1328 function. This is particularly useful on indirect calls; without this we
1329 may treat such calls as though the target is non-convergent.
1331 The optimizer may remove the ``convergent`` attribute on functions when it
1332 can prove that the function does not execute any convergent operations.
1333 Similarly, the optimizer may remove ``convergent`` on calls/invokes when it
1334 can prove that the call/invoke cannot call a convergent function.
1335 ``inaccessiblememonly``
1336 This attribute indicates that the function may only access memory that
1337 is not accessible by the module being compiled. This is a weaker form
1339 ``inaccessiblemem_or_argmemonly``
1340 This attribute indicates that the function may only access memory that is
1341 either not accessible by the module being compiled, or is pointed to
1342 by its pointer arguments. This is a weaker form of ``argmemonly``
1344 This attribute indicates that the source code contained a hint that
1345 inlining this function is desirable (such as the "inline" keyword in
1346 C/C++). It is just a hint; it imposes no requirements on the
1349 This attribute indicates that the function should be added to a
1350 jump-instruction table at code-generation time, and that all address-taken
1351 references to this function should be replaced with a reference to the
1352 appropriate jump-instruction-table function pointer. Note that this creates
1353 a new pointer for the original function, which means that code that depends
1354 on function-pointer identity can break. So, any function annotated with
1355 ``jumptable`` must also be ``unnamed_addr``.
1357 This attribute suggests that optimization passes and code generator
1358 passes make choices that keep the code size of this function as small
1359 as possible and perform optimizations that may sacrifice runtime
1360 performance in order to minimize the size of the generated code.
1362 This attribute disables prologue / epilogue emission for the
1363 function. This can have very system-specific consequences.
1365 This indicates that the callee function at a call site is not recognized as
1366 a built-in function. LLVM will retain the original call and not replace it
1367 with equivalent code based on the semantics of the built-in function, unless
1368 the call site uses the ``builtin`` attribute. This is valid at call sites
1369 and on function declarations and definitions.
1371 This attribute indicates that calls to the function cannot be
1372 duplicated. A call to a ``noduplicate`` function may be moved
1373 within its parent function, but may not be duplicated within
1374 its parent function.
1376 A function containing a ``noduplicate`` call may still
1377 be an inlining candidate, provided that the call is not
1378 duplicated by inlining. That implies that the function has
1379 internal linkage and only has one call site, so the original
1380 call is dead after inlining.
1382 This attributes disables implicit floating point instructions.
1384 This attribute indicates that the inliner should never inline this
1385 function in any situation. This attribute may not be used together
1386 with the ``alwaysinline`` attribute.
1388 This attribute suppresses lazy symbol binding for the function. This
1389 may make calls to the function faster, at the cost of extra program
1390 startup time if the function is not called during program startup.
1392 This attribute indicates that the code generator should not use a
1393 red zone, even if the target-specific ABI normally permits it.
1395 This function attribute indicates that the function never returns
1396 normally. This produces undefined behavior at runtime if the
1397 function ever does dynamically return.
1399 This function attribute indicates that the function does not call itself
1400 either directly or indirectly down any possible call path. This produces
1401 undefined behavior at runtime if the function ever does recurse.
1403 This function attribute indicates that the function never raises an
1404 exception. If the function does raise an exception, its runtime
1405 behavior is undefined. However, functions marked nounwind may still
1406 trap or generate asynchronous exceptions. Exception handling schemes
1407 that are recognized by LLVM to handle asynchronous exceptions, such
1408 as SEH, will still provide their implementation defined semantics.
1410 This function attribute indicates that most optimization passes will skip
1411 this function, with the exception of interprocedural optimization passes.
1412 Code generation defaults to the "fast" instruction selector.
1413 This attribute cannot be used together with the ``alwaysinline``
1414 attribute; this attribute is also incompatible
1415 with the ``minsize`` attribute and the ``optsize`` attribute.
1417 This attribute requires the ``noinline`` attribute to be specified on
1418 the function as well, so the function is never inlined into any caller.
1419 Only functions with the ``alwaysinline`` attribute are valid
1420 candidates for inlining into the body of this function.
1422 This attribute suggests that optimization passes and code generator
1423 passes make choices that keep the code size of this function low,
1424 and otherwise do optimizations specifically to reduce code size as
1425 long as they do not significantly impact runtime performance.
1426 ``"patchable-function"``
1427 This attribute tells the code generator that the code
1428 generated for this function needs to follow certain conventions that
1429 make it possible for a runtime function to patch over it later.
1430 The exact effect of this attribute depends on its string value,
1431 for which there currently is one legal possibility:
1433 * ``"prologue-short-redirect"`` - This style of patchable
1434 function is intended to support patching a function prologue to
1435 redirect control away from the function in a thread safe
1436 manner. It guarantees that the first instruction of the
1437 function will be large enough to accommodate a short jump
1438 instruction, and will be sufficiently aligned to allow being
1439 fully changed via an atomic compare-and-swap instruction.
1440 While the first requirement can be satisfied by inserting large
1441 enough NOP, LLVM can and will try to re-purpose an existing
1442 instruction (i.e. one that would have to be emitted anyway) as
1443 the patchable instruction larger than a short jump.
1445 ``"prologue-short-redirect"`` is currently only supported on
1448 This attribute by itself does not imply restrictions on
1449 inter-procedural optimizations. All of the semantic effects the
1450 patching may have to be separately conveyed via the linkage type.
1452 On a function, this attribute indicates that the function computes its
1453 result (or decides to unwind an exception) based strictly on its arguments,
1454 without dereferencing any pointer arguments or otherwise accessing
1455 any mutable state (e.g. memory, control registers, etc) visible to
1456 caller functions. It does not write through any pointer arguments
1457 (including ``byval`` arguments) and never changes any state visible
1458 to callers. This means that it cannot unwind exceptions by calling
1459 the ``C++`` exception throwing methods.
1461 On an argument, this attribute indicates that the function does not
1462 dereference that pointer argument, even though it may read or write the
1463 memory that the pointer points to if accessed through other pointers.
1465 On a function, this attribute indicates that the function does not write
1466 through any pointer arguments (including ``byval`` arguments) or otherwise
1467 modify any state (e.g. memory, control registers, etc) visible to
1468 caller functions. It may dereference pointer arguments and read
1469 state that may be set in the caller. A readonly function always
1470 returns the same value (or unwinds an exception identically) when
1471 called with the same set of arguments and global state. It cannot
1472 unwind an exception by calling the ``C++`` exception throwing
1475 On an argument, this attribute indicates that the function does not write
1476 through this pointer argument, even though it may write to the memory that
1477 the pointer points to.
1479 On a function, this attribute indicates that the function may write to but
1480 does not read from memory.
1482 On an argument, this attribute indicates that the function may write to but
1483 does not read through this pointer argument (even though it may read from
1484 the memory that the pointer points to).
1486 This attribute indicates that the only memory accesses inside function are
1487 loads and stores from objects pointed to by its pointer-typed arguments,
1488 with arbitrary offsets. Or in other words, all memory operations in the
1489 function can refer to memory only using pointers based on its function
1491 Note that ``argmemonly`` can be used together with ``readonly`` attribute
1492 in order to specify that function reads only from its arguments.
1494 This attribute indicates that this function can return twice. The C
1495 ``setjmp`` is an example of such a function. The compiler disables
1496 some optimizations (like tail calls) in the caller of these
1499 This attribute indicates that
1500 `SafeStack <http://clang.llvm.org/docs/SafeStack.html>`_
1501 protection is enabled for this function.
1503 If a function that has a ``safestack`` attribute is inlined into a
1504 function that doesn't have a ``safestack`` attribute or which has an
1505 ``ssp``, ``sspstrong`` or ``sspreq`` attribute, then the resulting
1506 function will have a ``safestack`` attribute.
1507 ``sanitize_address``
1508 This attribute indicates that AddressSanitizer checks
1509 (dynamic address safety analysis) are enabled for this function.
1511 This attribute indicates that MemorySanitizer checks (dynamic detection
1512 of accesses to uninitialized memory) are enabled for this function.
1514 This attribute indicates that ThreadSanitizer checks
1515 (dynamic thread safety analysis) are enabled for this function.
1517 This attribute indicates that the function should emit a stack
1518 smashing protector. It is in the form of a "canary" --- a random value
1519 placed on the stack before the local variables that's checked upon
1520 return from the function to see if it has been overwritten. A
1521 heuristic is used to determine if a function needs stack protectors
1522 or not. The heuristic used will enable protectors for functions with:
1524 - Character arrays larger than ``ssp-buffer-size`` (default 8).
1525 - Aggregates containing character arrays larger than ``ssp-buffer-size``.
1526 - Calls to alloca() with variable sizes or constant sizes greater than
1527 ``ssp-buffer-size``.
1529 Variables that are identified as requiring a protector will be arranged
1530 on the stack such that they are adjacent to the stack protector guard.
1532 If a function that has an ``ssp`` attribute is inlined into a
1533 function that doesn't have an ``ssp`` attribute, then the resulting
1534 function will have an ``ssp`` attribute.
1536 This attribute indicates that the function should *always* emit a
1537 stack smashing protector. This overrides the ``ssp`` function
1540 Variables that are identified as requiring a protector will be arranged
1541 on the stack such that they are adjacent to the stack protector guard.
1542 The specific layout rules are:
1544 #. Large arrays and structures containing large arrays
1545 (``>= ssp-buffer-size``) are closest to the stack protector.
1546 #. Small arrays and structures containing small arrays
1547 (``< ssp-buffer-size``) are 2nd closest to the protector.
1548 #. Variables that have had their address taken are 3rd closest to the
1551 If a function that has an ``sspreq`` attribute is inlined into a
1552 function that doesn't have an ``sspreq`` attribute or which has an
1553 ``ssp`` or ``sspstrong`` attribute, then the resulting function will have
1554 an ``sspreq`` attribute.
1556 This attribute indicates that the function should emit a stack smashing
1557 protector. This attribute causes a strong heuristic to be used when
1558 determining if a function needs stack protectors. The strong heuristic
1559 will enable protectors for functions with:
1561 - Arrays of any size and type
1562 - Aggregates containing an array of any size and type.
1563 - Calls to alloca().
1564 - Local variables that have had their address taken.
1566 Variables that are identified as requiring a protector will be arranged
1567 on the stack such that they are adjacent to the stack protector guard.
1568 The specific layout rules are:
1570 #. Large arrays and structures containing large arrays
1571 (``>= ssp-buffer-size``) are closest to the stack protector.
1572 #. Small arrays and structures containing small arrays
1573 (``< ssp-buffer-size``) are 2nd closest to the protector.
1574 #. Variables that have had their address taken are 3rd closest to the
1577 This overrides the ``ssp`` function attribute.
1579 If a function that has an ``sspstrong`` attribute is inlined into a
1580 function that doesn't have an ``sspstrong`` attribute, then the
1581 resulting function will have an ``sspstrong`` attribute.
1583 This attribute indicates that the function will delegate to some other
1584 function with a tail call. The prototype of a thunk should not be used for
1585 optimization purposes. The caller is expected to cast the thunk prototype to
1586 match the thunk target prototype.
1588 This attribute indicates that the ABI being targeted requires that
1589 an unwind table entry be produced for this function even if we can
1590 show that no exceptions passes by it. This is normally the case for
1591 the ELF x86-64 abi, but it can be disabled for some compilation
1600 Note: operand bundles are a work in progress, and they should be
1601 considered experimental at this time.
1603 Operand bundles are tagged sets of SSA values that can be associated
1604 with certain LLVM instructions (currently only ``call`` s and
1605 ``invoke`` s). In a way they are like metadata, but dropping them is
1606 incorrect and will change program semantics.
1610 operand bundle set ::= '[' operand bundle (, operand bundle )* ']'
1611 operand bundle ::= tag '(' [ bundle operand ] (, bundle operand )* ')'
1612 bundle operand ::= SSA value
1613 tag ::= string constant
1615 Operand bundles are **not** part of a function's signature, and a
1616 given function may be called from multiple places with different kinds
1617 of operand bundles. This reflects the fact that the operand bundles
1618 are conceptually a part of the ``call`` (or ``invoke``), not the
1619 callee being dispatched to.
1621 Operand bundles are a generic mechanism intended to support
1622 runtime-introspection-like functionality for managed languages. While
1623 the exact semantics of an operand bundle depend on the bundle tag,
1624 there are certain limitations to how much the presence of an operand
1625 bundle can influence the semantics of a program. These restrictions
1626 are described as the semantics of an "unknown" operand bundle. As
1627 long as the behavior of an operand bundle is describable within these
1628 restrictions, LLVM does not need to have special knowledge of the
1629 operand bundle to not miscompile programs containing it.
1631 - The bundle operands for an unknown operand bundle escape in unknown
1632 ways before control is transferred to the callee or invokee.
1633 - Calls and invokes with operand bundles have unknown read / write
1634 effect on the heap on entry and exit (even if the call target is
1635 ``readnone`` or ``readonly``), unless they're overridden with
1636 callsite specific attributes.
1637 - An operand bundle at a call site cannot change the implementation
1638 of the called function. Inter-procedural optimizations work as
1639 usual as long as they take into account the first two properties.
1641 More specific types of operand bundles are described below.
1643 .. _deopt_opbundles:
1645 Deoptimization Operand Bundles
1646 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1648 Deoptimization operand bundles are characterized by the ``"deopt"``
1649 operand bundle tag. These operand bundles represent an alternate
1650 "safe" continuation for the call site they're attached to, and can be
1651 used by a suitable runtime to deoptimize the compiled frame at the
1652 specified call site. There can be at most one ``"deopt"`` operand
1653 bundle attached to a call site. Exact details of deoptimization is
1654 out of scope for the language reference, but it usually involves
1655 rewriting a compiled frame into a set of interpreted frames.
1657 From the compiler's perspective, deoptimization operand bundles make
1658 the call sites they're attached to at least ``readonly``. They read
1659 through all of their pointer typed operands (even if they're not
1660 otherwise escaped) and the entire visible heap. Deoptimization
1661 operand bundles do not capture their operands except during
1662 deoptimization, in which case control will not be returned to the
1665 The inliner knows how to inline through calls that have deoptimization
1666 operand bundles. Just like inlining through a normal call site
1667 involves composing the normal and exceptional continuations, inlining
1668 through a call site with a deoptimization operand bundle needs to
1669 appropriately compose the "safe" deoptimization continuation. The
1670 inliner does this by prepending the parent's deoptimization
1671 continuation to every deoptimization continuation in the inlined body.
1672 E.g. inlining ``@f`` into ``@g`` in the following example
1674 .. code-block:: llvm
1677 call void @x() ;; no deopt state
1678 call void @y() [ "deopt"(i32 10) ]
1679 call void @y() [ "deopt"(i32 10), "unknown"(i8* null) ]
1684 call void @f() [ "deopt"(i32 20) ]
1690 .. code-block:: llvm
1693 call void @x() ;; still no deopt state
1694 call void @y() [ "deopt"(i32 20, i32 10) ]
1695 call void @y() [ "deopt"(i32 20, i32 10), "unknown"(i8* null) ]
1699 It is the frontend's responsibility to structure or encode the
1700 deoptimization state in a way that syntactically prepending the
1701 caller's deoptimization state to the callee's deoptimization state is
1702 semantically equivalent to composing the caller's deoptimization
1703 continuation after the callee's deoptimization continuation.
1707 Funclet Operand Bundles
1708 ^^^^^^^^^^^^^^^^^^^^^^^
1710 Funclet operand bundles are characterized by the ``"funclet"``
1711 operand bundle tag. These operand bundles indicate that a call site
1712 is within a particular funclet. There can be at most one
1713 ``"funclet"`` operand bundle attached to a call site and it must have
1714 exactly one bundle operand.
1716 If any funclet EH pads have been "entered" but not "exited" (per the
1717 `description in the EH doc\ <ExceptionHandling.html#wineh-constraints>`_),
1718 it is undefined behavior to execute a ``call`` or ``invoke`` which:
1720 * does not have a ``"funclet"`` bundle and is not a ``call`` to a nounwind
1722 * has a ``"funclet"`` bundle whose operand is not the most-recently-entered
1723 not-yet-exited funclet EH pad.
1725 Similarly, if no funclet EH pads have been entered-but-not-yet-exited,
1726 executing a ``call`` or ``invoke`` with a ``"funclet"`` bundle is undefined behavior.
1728 GC Transition Operand Bundles
1729 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1731 GC transition operand bundles are characterized by the
1732 ``"gc-transition"`` operand bundle tag. These operand bundles mark a
1733 call as a transition between a function with one GC strategy to a
1734 function with a different GC strategy. If coordinating the transition
1735 between GC strategies requires additional code generation at the call
1736 site, these bundles may contain any values that are needed by the
1737 generated code. For more details, see :ref:`GC Transitions
1738 <gc_transition_args>`.
1742 Module-Level Inline Assembly
1743 ----------------------------
1745 Modules may contain "module-level inline asm" blocks, which corresponds
1746 to the GCC "file scope inline asm" blocks. These blocks are internally
1747 concatenated by LLVM and treated as a single unit, but may be separated
1748 in the ``.ll`` file if desired. The syntax is very simple:
1750 .. code-block:: llvm
1752 module asm "inline asm code goes here"
1753 module asm "more can go here"
1755 The strings can contain any character by escaping non-printable
1756 characters. The escape sequence used is simply "\\xx" where "xx" is the
1757 two digit hex code for the number.
1759 Note that the assembly string *must* be parseable by LLVM's integrated assembler
1760 (unless it is disabled), even when emitting a ``.s`` file.
1762 .. _langref_datalayout:
1767 A module may specify a target specific data layout string that specifies
1768 how data is to be laid out in memory. The syntax for the data layout is
1771 .. code-block:: llvm
1773 target datalayout = "layout specification"
1775 The *layout specification* consists of a list of specifications
1776 separated by the minus sign character ('-'). Each specification starts
1777 with a letter and may include other information after the letter to
1778 define some aspect of the data layout. The specifications accepted are
1782 Specifies that the target lays out data in big-endian form. That is,
1783 the bits with the most significance have the lowest address
1786 Specifies that the target lays out data in little-endian form. That
1787 is, the bits with the least significance have the lowest address
1790 Specifies the natural alignment of the stack in bits. Alignment
1791 promotion of stack variables is limited to the natural stack
1792 alignment to avoid dynamic stack realignment. The stack alignment
1793 must be a multiple of 8-bits. If omitted, the natural stack
1794 alignment defaults to "unspecified", which does not prevent any
1795 alignment promotions.
1796 ``p[n]:<size>:<abi>:<pref>``
1797 This specifies the *size* of a pointer and its ``<abi>`` and
1798 ``<pref>``\erred alignments for address space ``n``. All sizes are in
1799 bits. The address space, ``n``, is optional, and if not specified,
1800 denotes the default address space 0. The value of ``n`` must be
1801 in the range [1,2^23).
1802 ``i<size>:<abi>:<pref>``
1803 This specifies the alignment for an integer type of a given bit
1804 ``<size>``. The value of ``<size>`` must be in the range [1,2^23).
1805 ``v<size>:<abi>:<pref>``
1806 This specifies the alignment for a vector type of a given bit
1808 ``f<size>:<abi>:<pref>``
1809 This specifies the alignment for a floating point type of a given bit
1810 ``<size>``. Only values of ``<size>`` that are supported by the target
1811 will work. 32 (float) and 64 (double) are supported on all targets; 80
1812 or 128 (different flavors of long double) are also supported on some
1815 This specifies the alignment for an object of aggregate type.
1817 If present, specifies that llvm names are mangled in the output. The
1820 * ``e``: ELF mangling: Private symbols get a ``.L`` prefix.
1821 * ``m``: Mips mangling: Private symbols get a ``$`` prefix.
1822 * ``o``: Mach-O mangling: Private symbols get ``L`` prefix. Other
1823 symbols get a ``_`` prefix.
1824 * ``w``: Windows COFF prefix: Similar to Mach-O, but stdcall and fastcall
1825 functions also get a suffix based on the frame size.
1826 * ``x``: Windows x86 COFF prefix: Similar to Windows COFF, but use a ``_``
1827 prefix for ``__cdecl`` functions.
1828 ``n<size1>:<size2>:<size3>...``
1829 This specifies a set of native integer widths for the target CPU in
1830 bits. For example, it might contain ``n32`` for 32-bit PowerPC,
1831 ``n32:64`` for PowerPC 64, or ``n8:16:32:64`` for X86-64. Elements of
1832 this set are considered to support most general arithmetic operations
1835 On every specification that takes a ``<abi>:<pref>``, specifying the
1836 ``<pref>`` alignment is optional. If omitted, the preceding ``:``
1837 should be omitted too and ``<pref>`` will be equal to ``<abi>``.
1839 When constructing the data layout for a given target, LLVM starts with a
1840 default set of specifications which are then (possibly) overridden by
1841 the specifications in the ``datalayout`` keyword. The default
1842 specifications are given in this list:
1844 - ``E`` - big endian
1845 - ``p:64:64:64`` - 64-bit pointers with 64-bit alignment.
1846 - ``p[n]:64:64:64`` - Other address spaces are assumed to be the
1847 same as the default address space.
1848 - ``S0`` - natural stack alignment is unspecified
1849 - ``i1:8:8`` - i1 is 8-bit (byte) aligned
1850 - ``i8:8:8`` - i8 is 8-bit (byte) aligned
1851 - ``i16:16:16`` - i16 is 16-bit aligned
1852 - ``i32:32:32`` - i32 is 32-bit aligned
1853 - ``i64:32:64`` - i64 has ABI alignment of 32-bits but preferred
1854 alignment of 64-bits
1855 - ``f16:16:16`` - half is 16-bit aligned
1856 - ``f32:32:32`` - float is 32-bit aligned
1857 - ``f64:64:64`` - double is 64-bit aligned
1858 - ``f128:128:128`` - quad is 128-bit aligned
1859 - ``v64:64:64`` - 64-bit vector is 64-bit aligned
1860 - ``v128:128:128`` - 128-bit vector is 128-bit aligned
1861 - ``a:0:64`` - aggregates are 64-bit aligned
1863 When LLVM is determining the alignment for a given type, it uses the
1866 #. If the type sought is an exact match for one of the specifications,
1867 that specification is used.
1868 #. If no match is found, and the type sought is an integer type, then
1869 the smallest integer type that is larger than the bitwidth of the
1870 sought type is used. If none of the specifications are larger than
1871 the bitwidth then the largest integer type is used. For example,
1872 given the default specifications above, the i7 type will use the
1873 alignment of i8 (next largest) while both i65 and i256 will use the
1874 alignment of i64 (largest specified).
1875 #. If no match is found, and the type sought is a vector type, then the
1876 largest vector type that is smaller than the sought vector type will
1877 be used as a fall back. This happens because <128 x double> can be
1878 implemented in terms of 64 <2 x double>, for example.
1880 The function of the data layout string may not be what you expect.
1881 Notably, this is not a specification from the frontend of what alignment
1882 the code generator should use.
1884 Instead, if specified, the target data layout is required to match what
1885 the ultimate *code generator* expects. This string is used by the
1886 mid-level optimizers to improve code, and this only works if it matches
1887 what the ultimate code generator uses. There is no way to generate IR
1888 that does not embed this target-specific detail into the IR. If you
1889 don't specify the string, the default specifications will be used to
1890 generate a Data Layout and the optimization phases will operate
1891 accordingly and introduce target specificity into the IR with respect to
1892 these default specifications.
1899 A module may specify a target triple string that describes the target
1900 host. The syntax for the target triple is simply:
1902 .. code-block:: llvm
1904 target triple = "x86_64-apple-macosx10.7.0"
1906 The *target triple* string consists of a series of identifiers delimited
1907 by the minus sign character ('-'). The canonical forms are:
1911 ARCHITECTURE-VENDOR-OPERATING_SYSTEM
1912 ARCHITECTURE-VENDOR-OPERATING_SYSTEM-ENVIRONMENT
1914 This information is passed along to the backend so that it generates
1915 code for the proper architecture. It's possible to override this on the
1916 command line with the ``-mtriple`` command line option.
1918 .. _pointeraliasing:
1920 Pointer Aliasing Rules
1921 ----------------------
1923 Any memory access must be done through a pointer value associated with
1924 an address range of the memory access, otherwise the behavior is
1925 undefined. Pointer values are associated with address ranges according
1926 to the following rules:
1928 - A pointer value is associated with the addresses associated with any
1929 value it is *based* on.
1930 - An address of a global variable is associated with the address range
1931 of the variable's storage.
1932 - The result value of an allocation instruction is associated with the
1933 address range of the allocated storage.
1934 - A null pointer in the default address-space is associated with no
1936 - An integer constant other than zero or a pointer value returned from
1937 a function not defined within LLVM may be associated with address
1938 ranges allocated through mechanisms other than those provided by
1939 LLVM. Such ranges shall not overlap with any ranges of addresses
1940 allocated by mechanisms provided by LLVM.
1942 A pointer value is *based* on another pointer value according to the
1945 - A pointer value formed from a ``getelementptr`` operation is *based*
1946 on the first value operand of the ``getelementptr``.
1947 - The result value of a ``bitcast`` is *based* on the operand of the
1949 - A pointer value formed by an ``inttoptr`` is *based* on all pointer
1950 values that contribute (directly or indirectly) to the computation of
1951 the pointer's value.
1952 - The "*based* on" relationship is transitive.
1954 Note that this definition of *"based"* is intentionally similar to the
1955 definition of *"based"* in C99, though it is slightly weaker.
1957 LLVM IR does not associate types with memory. The result type of a
1958 ``load`` merely indicates the size and alignment of the memory from
1959 which to load, as well as the interpretation of the value. The first
1960 operand type of a ``store`` similarly only indicates the size and
1961 alignment of the store.
1963 Consequently, type-based alias analysis, aka TBAA, aka
1964 ``-fstrict-aliasing``, is not applicable to general unadorned LLVM IR.
1965 :ref:`Metadata <metadata>` may be used to encode additional information
1966 which specialized optimization passes may use to implement type-based
1971 Volatile Memory Accesses
1972 ------------------------
1974 Certain memory accesses, such as :ref:`load <i_load>`'s,
1975 :ref:`store <i_store>`'s, and :ref:`llvm.memcpy <int_memcpy>`'s may be
1976 marked ``volatile``. The optimizers must not change the number of
1977 volatile operations or change their order of execution relative to other
1978 volatile operations. The optimizers *may* change the order of volatile
1979 operations relative to non-volatile operations. This is not Java's
1980 "volatile" and has no cross-thread synchronization behavior.
1982 IR-level volatile loads and stores cannot safely be optimized into
1983 llvm.memcpy or llvm.memmove intrinsics even when those intrinsics are
1984 flagged volatile. Likewise, the backend should never split or merge
1985 target-legal volatile load/store instructions.
1987 .. admonition:: Rationale
1989 Platforms may rely on volatile loads and stores of natively supported
1990 data width to be executed as single instruction. For example, in C
1991 this holds for an l-value of volatile primitive type with native
1992 hardware support, but not necessarily for aggregate types. The
1993 frontend upholds these expectations, which are intentionally
1994 unspecified in the IR. The rules above ensure that IR transformations
1995 do not violate the frontend's contract with the language.
1999 Memory Model for Concurrent Operations
2000 --------------------------------------
2002 The LLVM IR does not define any way to start parallel threads of
2003 execution or to register signal handlers. Nonetheless, there are
2004 platform-specific ways to create them, and we define LLVM IR's behavior
2005 in their presence. This model is inspired by the C++0x memory model.
2007 For a more informal introduction to this model, see the :doc:`Atomics`.
2009 We define a *happens-before* partial order as the least partial order
2012 - Is a superset of single-thread program order, and
2013 - When a *synchronizes-with* ``b``, includes an edge from ``a`` to
2014 ``b``. *Synchronizes-with* pairs are introduced by platform-specific
2015 techniques, like pthread locks, thread creation, thread joining,
2016 etc., and by atomic instructions. (See also :ref:`Atomic Memory Ordering
2017 Constraints <ordering>`).
2019 Note that program order does not introduce *happens-before* edges
2020 between a thread and signals executing inside that thread.
2022 Every (defined) read operation (load instructions, memcpy, atomic
2023 loads/read-modify-writes, etc.) R reads a series of bytes written by
2024 (defined) write operations (store instructions, atomic
2025 stores/read-modify-writes, memcpy, etc.). For the purposes of this
2026 section, initialized globals are considered to have a write of the
2027 initializer which is atomic and happens before any other read or write
2028 of the memory in question. For each byte of a read R, R\ :sub:`byte`
2029 may see any write to the same byte, except:
2031 - If write\ :sub:`1` happens before write\ :sub:`2`, and
2032 write\ :sub:`2` happens before R\ :sub:`byte`, then
2033 R\ :sub:`byte` does not see write\ :sub:`1`.
2034 - If R\ :sub:`byte` happens before write\ :sub:`3`, then
2035 R\ :sub:`byte` does not see write\ :sub:`3`.
2037 Given that definition, R\ :sub:`byte` is defined as follows:
2039 - If R is volatile, the result is target-dependent. (Volatile is
2040 supposed to give guarantees which can support ``sig_atomic_t`` in
2041 C/C++, and may be used for accesses to addresses that do not behave
2042 like normal memory. It does not generally provide cross-thread
2044 - Otherwise, if there is no write to the same byte that happens before
2045 R\ :sub:`byte`, R\ :sub:`byte` returns ``undef`` for that byte.
2046 - Otherwise, if R\ :sub:`byte` may see exactly one write,
2047 R\ :sub:`byte` returns the value written by that write.
2048 - Otherwise, if R is atomic, and all the writes R\ :sub:`byte` may
2049 see are atomic, it chooses one of the values written. See the :ref:`Atomic
2050 Memory Ordering Constraints <ordering>` section for additional
2051 constraints on how the choice is made.
2052 - Otherwise R\ :sub:`byte` returns ``undef``.
2054 R returns the value composed of the series of bytes it read. This
2055 implies that some bytes within the value may be ``undef`` **without**
2056 the entire value being ``undef``. Note that this only defines the
2057 semantics of the operation; it doesn't mean that targets will emit more
2058 than one instruction to read the series of bytes.
2060 Note that in cases where none of the atomic intrinsics are used, this
2061 model places only one restriction on IR transformations on top of what
2062 is required for single-threaded execution: introducing a store to a byte
2063 which might not otherwise be stored is not allowed in general.
2064 (Specifically, in the case where another thread might write to and read
2065 from an address, introducing a store can change a load that may see
2066 exactly one write into a load that may see multiple writes.)
2070 Atomic Memory Ordering Constraints
2071 ----------------------------------
2073 Atomic instructions (:ref:`cmpxchg <i_cmpxchg>`,
2074 :ref:`atomicrmw <i_atomicrmw>`, :ref:`fence <i_fence>`,
2075 :ref:`atomic load <i_load>`, and :ref:`atomic store <i_store>`) take
2076 ordering parameters that determine which other atomic instructions on
2077 the same address they *synchronize with*. These semantics are borrowed
2078 from Java and C++0x, but are somewhat more colloquial. If these
2079 descriptions aren't precise enough, check those specs (see spec
2080 references in the :doc:`atomics guide <Atomics>`).
2081 :ref:`fence <i_fence>` instructions treat these orderings somewhat
2082 differently since they don't take an address. See that instruction's
2083 documentation for details.
2085 For a simpler introduction to the ordering constraints, see the
2089 The set of values that can be read is governed by the happens-before
2090 partial order. A value cannot be read unless some operation wrote
2091 it. This is intended to provide a guarantee strong enough to model
2092 Java's non-volatile shared variables. This ordering cannot be
2093 specified for read-modify-write operations; it is not strong enough
2094 to make them atomic in any interesting way.
2096 In addition to the guarantees of ``unordered``, there is a single
2097 total order for modifications by ``monotonic`` operations on each
2098 address. All modification orders must be compatible with the
2099 happens-before order. There is no guarantee that the modification
2100 orders can be combined to a global total order for the whole program
2101 (and this often will not be possible). The read in an atomic
2102 read-modify-write operation (:ref:`cmpxchg <i_cmpxchg>` and
2103 :ref:`atomicrmw <i_atomicrmw>`) reads the value in the modification
2104 order immediately before the value it writes. If one atomic read
2105 happens before another atomic read of the same address, the later
2106 read must see the same value or a later value in the address's
2107 modification order. This disallows reordering of ``monotonic`` (or
2108 stronger) operations on the same address. If an address is written
2109 ``monotonic``-ally by one thread, and other threads ``monotonic``-ally
2110 read that address repeatedly, the other threads must eventually see
2111 the write. This corresponds to the C++0x/C1x
2112 ``memory_order_relaxed``.
2114 In addition to the guarantees of ``monotonic``, a
2115 *synchronizes-with* edge may be formed with a ``release`` operation.
2116 This is intended to model C++'s ``memory_order_acquire``.
2118 In addition to the guarantees of ``monotonic``, if this operation
2119 writes a value which is subsequently read by an ``acquire``
2120 operation, it *synchronizes-with* that operation. (This isn't a
2121 complete description; see the C++0x definition of a release
2122 sequence.) This corresponds to the C++0x/C1x
2123 ``memory_order_release``.
2124 ``acq_rel`` (acquire+release)
2125 Acts as both an ``acquire`` and ``release`` operation on its
2126 address. This corresponds to the C++0x/C1x ``memory_order_acq_rel``.
2127 ``seq_cst`` (sequentially consistent)
2128 In addition to the guarantees of ``acq_rel`` (``acquire`` for an
2129 operation that only reads, ``release`` for an operation that only
2130 writes), there is a global total order on all
2131 sequentially-consistent operations on all addresses, which is
2132 consistent with the *happens-before* partial order and with the
2133 modification orders of all the affected addresses. Each
2134 sequentially-consistent read sees the last preceding write to the
2135 same address in this global order. This corresponds to the C++0x/C1x
2136 ``memory_order_seq_cst`` and Java volatile.
2140 If an atomic operation is marked ``singlethread``, it only *synchronizes
2141 with* or participates in modification and seq\_cst total orderings with
2142 other operations running in the same thread (for example, in signal
2150 LLVM IR floating-point binary ops (:ref:`fadd <i_fadd>`,
2151 :ref:`fsub <i_fsub>`, :ref:`fmul <i_fmul>`, :ref:`fdiv <i_fdiv>`,
2152 :ref:`frem <i_frem>`, :ref:`fcmp <i_fcmp>`) have the following flags that can
2153 be set to enable otherwise unsafe floating point operations
2156 No NaNs - Allow optimizations to assume the arguments and result are not
2157 NaN. Such optimizations are required to retain defined behavior over
2158 NaNs, but the value of the result is undefined.
2161 No Infs - Allow optimizations to assume the arguments and result are not
2162 +/-Inf. Such optimizations are required to retain defined behavior over
2163 +/-Inf, but the value of the result is undefined.
2166 No Signed Zeros - Allow optimizations to treat the sign of a zero
2167 argument or result as insignificant.
2170 Allow Reciprocal - Allow optimizations to use the reciprocal of an
2171 argument rather than perform division.
2174 Fast - Allow algebraically equivalent transformations that may
2175 dramatically change results in floating point (e.g. reassociate). This
2176 flag implies all the others.
2180 Use-list Order Directives
2181 -------------------------
2183 Use-list directives encode the in-memory order of each use-list, allowing the
2184 order to be recreated. ``<order-indexes>`` is a comma-separated list of
2185 indexes that are assigned to the referenced value's uses. The referenced
2186 value's use-list is immediately sorted by these indexes.
2188 Use-list directives may appear at function scope or global scope. They are not
2189 instructions, and have no effect on the semantics of the IR. When they're at
2190 function scope, they must appear after the terminator of the final basic block.
2192 If basic blocks have their address taken via ``blockaddress()`` expressions,
2193 ``uselistorder_bb`` can be used to reorder their use-lists from outside their
2200 uselistorder <ty> <value>, { <order-indexes> }
2201 uselistorder_bb @function, %block { <order-indexes> }
2207 define void @foo(i32 %arg1, i32 %arg2) {
2209 ; ... instructions ...
2211 ; ... instructions ...
2213 ; At function scope.
2214 uselistorder i32 %arg1, { 1, 0, 2 }
2215 uselistorder label %bb, { 1, 0 }
2219 uselistorder i32* @global, { 1, 2, 0 }
2220 uselistorder i32 7, { 1, 0 }
2221 uselistorder i32 (i32) @bar, { 1, 0 }
2222 uselistorder_bb @foo, %bb, { 5, 1, 3, 2, 0, 4 }
2224 .. _source_filename:
2229 The *source filename* string is set to the original module identifier,
2230 which will be the name of the compiled source file when compiling from
2231 source through the clang front end, for example. It is then preserved through
2234 This is currently necessary to generate a consistent unique global
2235 identifier for local functions used in profile data, which prepends the
2236 source file name to the local function name.
2238 The syntax for the source file name is simply:
2240 .. code-block:: text
2242 source_filename = "/path/to/source.c"
2249 The LLVM type system is one of the most important features of the
2250 intermediate representation. Being typed enables a number of
2251 optimizations to be performed on the intermediate representation
2252 directly, without having to do extra analyses on the side before the
2253 transformation. A strong type system makes it easier to read the
2254 generated code and enables novel analyses and transformations that are
2255 not feasible to perform on normal three address code representations.
2265 The void type does not represent any value and has no size.
2283 The function type can be thought of as a function signature. It consists of a
2284 return type and a list of formal parameter types. The return type of a function
2285 type is a void type or first class type --- except for :ref:`label <t_label>`
2286 and :ref:`metadata <t_metadata>` types.
2292 <returntype> (<parameter list>)
2294 ...where '``<parameter list>``' is a comma-separated list of type
2295 specifiers. Optionally, the parameter list may include a type ``...``, which
2296 indicates that the function takes a variable number of arguments. Variable
2297 argument functions can access their arguments with the :ref:`variable argument
2298 handling intrinsic <int_varargs>` functions. '``<returntype>``' is any type
2299 except :ref:`label <t_label>` and :ref:`metadata <t_metadata>`.
2303 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2304 | ``i32 (i32)`` | function taking an ``i32``, returning an ``i32`` |
2305 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2306 | ``float (i16, i32 *) *`` | :ref:`Pointer <t_pointer>` to a function that takes an ``i16`` and a :ref:`pointer <t_pointer>` to ``i32``, returning ``float``. |
2307 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2308 | ``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. |
2309 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2310 | ``{i32, i32} (i32)`` | A function taking an ``i32``, returning a :ref:`structure <t_struct>` containing two ``i32`` values |
2311 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2318 The :ref:`first class <t_firstclass>` types are perhaps the most important.
2319 Values of these types are the only ones which can be produced by
2327 These are the types that are valid in registers from CodeGen's perspective.
2336 The integer type is a very simple type that simply specifies an
2337 arbitrary bit width for the integer type desired. Any bit width from 1
2338 bit to 2\ :sup:`23`\ -1 (about 8 million) can be specified.
2346 The number of bits the integer will occupy is specified by the ``N``
2352 +----------------+------------------------------------------------+
2353 | ``i1`` | a single-bit integer. |
2354 +----------------+------------------------------------------------+
2355 | ``i32`` | a 32-bit integer. |
2356 +----------------+------------------------------------------------+
2357 | ``i1942652`` | a really big integer of over 1 million bits. |
2358 +----------------+------------------------------------------------+
2362 Floating Point Types
2363 """"""""""""""""""""
2372 - 16-bit floating point value
2375 - 32-bit floating point value
2378 - 64-bit floating point value
2381 - 128-bit floating point value (112-bit mantissa)
2384 - 80-bit floating point value (X87)
2387 - 128-bit floating point value (two 64-bits)
2394 The x86_mmx type represents a value held in an MMX register on an x86
2395 machine. The operations allowed on it are quite limited: parameters and
2396 return values, load and store, and bitcast. User-specified MMX
2397 instructions are represented as intrinsic or asm calls with arguments
2398 and/or results of this type. There are no arrays, vectors or constants
2415 The pointer type is used to specify memory locations. Pointers are
2416 commonly used to reference objects in memory.
2418 Pointer types may have an optional address space attribute defining the
2419 numbered address space where the pointed-to object resides. The default
2420 address space is number zero. The semantics of non-zero address spaces
2421 are target-specific.
2423 Note that LLVM does not permit pointers to void (``void*``) nor does it
2424 permit pointers to labels (``label*``). Use ``i8*`` instead.
2434 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2435 | ``[4 x i32]*`` | A :ref:`pointer <t_pointer>` to :ref:`array <t_array>` of four ``i32`` values. |
2436 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2437 | ``i32 (i32*) *`` | A :ref:`pointer <t_pointer>` to a :ref:`function <t_function>` that takes an ``i32*``, returning an ``i32``. |
2438 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2439 | ``i32 addrspace(5)*`` | A :ref:`pointer <t_pointer>` to an ``i32`` value that resides in address space #5. |
2440 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2449 A vector type is a simple derived type that represents a vector of
2450 elements. Vector types are used when multiple primitive data are
2451 operated in parallel using a single instruction (SIMD). A vector type
2452 requires a size (number of elements) and an underlying primitive data
2453 type. Vector types are considered :ref:`first class <t_firstclass>`.
2459 < <# elements> x <elementtype> >
2461 The number of elements is a constant integer value larger than 0;
2462 elementtype may be any integer, floating point or pointer type. Vectors
2463 of size zero are not allowed.
2467 +-------------------+--------------------------------------------------+
2468 | ``<4 x i32>`` | Vector of 4 32-bit integer values. |
2469 +-------------------+--------------------------------------------------+
2470 | ``<8 x float>`` | Vector of 8 32-bit floating-point values. |
2471 +-------------------+--------------------------------------------------+
2472 | ``<2 x i64>`` | Vector of 2 64-bit integer values. |
2473 +-------------------+--------------------------------------------------+
2474 | ``<4 x i64*>`` | Vector of 4 pointers to 64-bit integer values. |
2475 +-------------------+--------------------------------------------------+
2484 The label type represents code labels.
2499 The token type is used when a value is associated with an instruction
2500 but all uses of the value must not attempt to introspect or obscure it.
2501 As such, it is not appropriate to have a :ref:`phi <i_phi>` or
2502 :ref:`select <i_select>` of type token.
2519 The metadata type represents embedded metadata. No derived types may be
2520 created from metadata except for :ref:`function <t_function>` arguments.
2533 Aggregate Types are a subset of derived types that can contain multiple
2534 member types. :ref:`Arrays <t_array>` and :ref:`structs <t_struct>` are
2535 aggregate types. :ref:`Vectors <t_vector>` are not considered to be
2545 The array type is a very simple derived type that arranges elements
2546 sequentially in memory. The array type requires a size (number of
2547 elements) and an underlying data type.
2553 [<# elements> x <elementtype>]
2555 The number of elements is a constant integer value; ``elementtype`` may
2556 be any type with a size.
2560 +------------------+--------------------------------------+
2561 | ``[40 x i32]`` | Array of 40 32-bit integer values. |
2562 +------------------+--------------------------------------+
2563 | ``[41 x i32]`` | Array of 41 32-bit integer values. |
2564 +------------------+--------------------------------------+
2565 | ``[4 x i8]`` | Array of 4 8-bit integer values. |
2566 +------------------+--------------------------------------+
2568 Here are some examples of multidimensional arrays:
2570 +-----------------------------+----------------------------------------------------------+
2571 | ``[3 x [4 x i32]]`` | 3x4 array of 32-bit integer values. |
2572 +-----------------------------+----------------------------------------------------------+
2573 | ``[12 x [10 x float]]`` | 12x10 array of single precision floating point values. |
2574 +-----------------------------+----------------------------------------------------------+
2575 | ``[2 x [3 x [4 x i16]]]`` | 2x3x4 array of 16-bit integer values. |
2576 +-----------------------------+----------------------------------------------------------+
2578 There is no restriction on indexing beyond the end of the array implied
2579 by a static type (though there are restrictions on indexing beyond the
2580 bounds of an allocated object in some cases). This means that
2581 single-dimension 'variable sized array' addressing can be implemented in
2582 LLVM with a zero length array type. An implementation of 'pascal style
2583 arrays' in LLVM could use the type "``{ i32, [0 x float]}``", for
2593 The structure type is used to represent a collection of data members
2594 together in memory. The elements of a structure may be any type that has
2597 Structures in memory are accessed using '``load``' and '``store``' by
2598 getting a pointer to a field with the '``getelementptr``' instruction.
2599 Structures in registers are accessed using the '``extractvalue``' and
2600 '``insertvalue``' instructions.
2602 Structures may optionally be "packed" structures, which indicate that
2603 the alignment of the struct is one byte, and that there is no padding
2604 between the elements. In non-packed structs, padding between field types
2605 is inserted as defined by the DataLayout string in the module, which is
2606 required to match what the underlying code generator expects.
2608 Structures can either be "literal" or "identified". A literal structure
2609 is defined inline with other types (e.g. ``{i32, i32}*``) whereas
2610 identified types are always defined at the top level with a name.
2611 Literal types are uniqued by their contents and can never be recursive
2612 or opaque since there is no way to write one. Identified types can be
2613 recursive, can be opaqued, and are never uniqued.
2619 %T1 = type { <type list> } ; Identified normal struct type
2620 %T2 = type <{ <type list> }> ; Identified packed struct type
2624 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2625 | ``{ i32, i32, i32 }`` | A triple of three ``i32`` values |
2626 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2627 | ``{ 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``. |
2628 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2629 | ``<{ i8, i32 }>`` | A packed struct known to be 5 bytes in size. |
2630 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2634 Opaque Structure Types
2635 """"""""""""""""""""""
2639 Opaque structure types are used to represent named structure types that
2640 do not have a body specified. This corresponds (for example) to the C
2641 notion of a forward declared structure.
2652 +--------------+-------------------+
2653 | ``opaque`` | An opaque type. |
2654 +--------------+-------------------+
2661 LLVM has several different basic types of constants. This section
2662 describes them all and their syntax.
2667 **Boolean constants**
2668 The two strings '``true``' and '``false``' are both valid constants
2670 **Integer constants**
2671 Standard integers (such as '4') are constants of the
2672 :ref:`integer <t_integer>` type. Negative numbers may be used with
2674 **Floating point constants**
2675 Floating point constants use standard decimal notation (e.g.
2676 123.421), exponential notation (e.g. 1.23421e+2), or a more precise
2677 hexadecimal notation (see below). The assembler requires the exact
2678 decimal value of a floating-point constant. For example, the
2679 assembler accepts 1.25 but rejects 1.3 because 1.3 is a repeating
2680 decimal in binary. Floating point constants must have a :ref:`floating
2681 point <t_floating>` type.
2682 **Null pointer constants**
2683 The identifier '``null``' is recognized as a null pointer constant
2684 and must be of :ref:`pointer type <t_pointer>`.
2686 The identifier '``none``' is recognized as an empty token constant
2687 and must be of :ref:`token type <t_token>`.
2689 The one non-intuitive notation for constants is the hexadecimal form of
2690 floating point constants. For example, the form
2691 '``double 0x432ff973cafa8000``' is equivalent to (but harder to read
2692 than) '``double 4.5e+15``'. The only time hexadecimal floating point
2693 constants are required (and the only time that they are generated by the
2694 disassembler) is when a floating point constant must be emitted but it
2695 cannot be represented as a decimal floating point number in a reasonable
2696 number of digits. For example, NaN's, infinities, and other special
2697 values are represented in their IEEE hexadecimal format so that assembly
2698 and disassembly do not cause any bits to change in the constants.
2700 When using the hexadecimal form, constants of types half, float, and
2701 double are represented using the 16-digit form shown above (which
2702 matches the IEEE754 representation for double); half and float values
2703 must, however, be exactly representable as IEEE 754 half and single
2704 precision, respectively. Hexadecimal format is always used for long
2705 double, and there are three forms of long double. The 80-bit format used
2706 by x86 is represented as ``0xK`` followed by 20 hexadecimal digits. The
2707 128-bit format used by PowerPC (two adjacent doubles) is represented by
2708 ``0xM`` followed by 32 hexadecimal digits. The IEEE 128-bit format is
2709 represented by ``0xL`` followed by 32 hexadecimal digits. Long doubles
2710 will only work if they match the long double format on your target.
2711 The IEEE 16-bit format (half precision) is represented by ``0xH``
2712 followed by 4 hexadecimal digits. All hexadecimal formats are big-endian
2713 (sign bit at the left).
2715 There are no constants of type x86_mmx.
2717 .. _complexconstants:
2722 Complex constants are a (potentially recursive) combination of simple
2723 constants and smaller complex constants.
2725 **Structure constants**
2726 Structure constants are represented with notation similar to
2727 structure type definitions (a comma separated list of elements,
2728 surrounded by braces (``{}``)). For example:
2729 "``{ i32 4, float 17.0, i32* @G }``", where "``@G``" is declared as
2730 "``@G = external global i32``". Structure constants must have
2731 :ref:`structure type <t_struct>`, and the number and types of elements
2732 must match those specified by the type.
2734 Array constants are represented with notation similar to array type
2735 definitions (a comma separated list of elements, surrounded by
2736 square brackets (``[]``)). For example:
2737 "``[ i32 42, i32 11, i32 74 ]``". Array constants must have
2738 :ref:`array type <t_array>`, and the number and types of elements must
2739 match those specified by the type. As a special case, character array
2740 constants may also be represented as a double-quoted string using the ``c``
2741 prefix. For example: "``c"Hello World\0A\00"``".
2742 **Vector constants**
2743 Vector constants are represented with notation similar to vector
2744 type definitions (a comma separated list of elements, surrounded by
2745 less-than/greater-than's (``<>``)). For example:
2746 "``< i32 42, i32 11, i32 74, i32 100 >``". Vector constants
2747 must have :ref:`vector type <t_vector>`, and the number and types of
2748 elements must match those specified by the type.
2749 **Zero initialization**
2750 The string '``zeroinitializer``' can be used to zero initialize a
2751 value to zero of *any* type, including scalar and
2752 :ref:`aggregate <t_aggregate>` types. This is often used to avoid
2753 having to print large zero initializers (e.g. for large arrays) and
2754 is always exactly equivalent to using explicit zero initializers.
2756 A metadata node is a constant tuple without types. For example:
2757 "``!{!0, !{!2, !0}, !"test"}``". Metadata can reference constant values,
2758 for example: "``!{!0, i32 0, i8* @global, i64 (i64)* @function, !"str"}``".
2759 Unlike other typed constants that are meant to be interpreted as part of
2760 the instruction stream, metadata is a place to attach additional
2761 information such as debug info.
2763 Global Variable and Function Addresses
2764 --------------------------------------
2766 The addresses of :ref:`global variables <globalvars>` and
2767 :ref:`functions <functionstructure>` are always implicitly valid
2768 (link-time) constants. These constants are explicitly referenced when
2769 the :ref:`identifier for the global <identifiers>` is used and always have
2770 :ref:`pointer <t_pointer>` type. For example, the following is a legal LLVM
2773 .. code-block:: llvm
2777 @Z = global [2 x i32*] [ i32* @X, i32* @Y ]
2784 The string '``undef``' can be used anywhere a constant is expected, and
2785 indicates that the user of the value may receive an unspecified
2786 bit-pattern. Undefined values may be of any type (other than '``label``'
2787 or '``void``') and be used anywhere a constant is permitted.
2789 Undefined values are useful because they indicate to the compiler that
2790 the program is well defined no matter what value is used. This gives the
2791 compiler more freedom to optimize. Here are some examples of
2792 (potentially surprising) transformations that are valid (in pseudo IR):
2794 .. code-block:: llvm
2804 This is safe because all of the output bits are affected by the undef
2805 bits. Any output bit can have a zero or one depending on the input bits.
2807 .. code-block:: llvm
2818 These logical operations have bits that are not always affected by the
2819 input. For example, if ``%X`` has a zero bit, then the output of the
2820 '``and``' operation will always be a zero for that bit, no matter what
2821 the corresponding bit from the '``undef``' is. As such, it is unsafe to
2822 optimize or assume that the result of the '``and``' is '``undef``'.
2823 However, it is safe to assume that all bits of the '``undef``' could be
2824 0, and optimize the '``and``' to 0. Likewise, it is safe to assume that
2825 all the bits of the '``undef``' operand to the '``or``' could be set,
2826 allowing the '``or``' to be folded to -1.
2828 .. code-block:: llvm
2830 %A = select undef, %X, %Y
2831 %B = select undef, 42, %Y
2832 %C = select %X, %Y, undef
2842 This set of examples shows that undefined '``select``' (and conditional
2843 branch) conditions can go *either way*, but they have to come from one
2844 of the two operands. In the ``%A`` example, if ``%X`` and ``%Y`` were
2845 both known to have a clear low bit, then ``%A`` would have to have a
2846 cleared low bit. However, in the ``%C`` example, the optimizer is
2847 allowed to assume that the '``undef``' operand could be the same as
2848 ``%Y``, allowing the whole '``select``' to be eliminated.
2850 .. code-block:: text
2852 %A = xor undef, undef
2869 This example points out that two '``undef``' operands are not
2870 necessarily the same. This can be surprising to people (and also matches
2871 C semantics) where they assume that "``X^X``" is always zero, even if
2872 ``X`` is undefined. This isn't true for a number of reasons, but the
2873 short answer is that an '``undef``' "variable" can arbitrarily change
2874 its value over its "live range". This is true because the variable
2875 doesn't actually *have a live range*. Instead, the value is logically
2876 read from arbitrary registers that happen to be around when needed, so
2877 the value is not necessarily consistent over time. In fact, ``%A`` and
2878 ``%C`` need to have the same semantics or the core LLVM "replace all
2879 uses with" concept would not hold.
2881 .. code-block:: llvm
2889 These examples show the crucial difference between an *undefined value*
2890 and *undefined behavior*. An undefined value (like '``undef``') is
2891 allowed to have an arbitrary bit-pattern. This means that the ``%A``
2892 operation can be constant folded to '``undef``', because the '``undef``'
2893 could be an SNaN, and ``fdiv`` is not (currently) defined on SNaN's.
2894 However, in the second example, we can make a more aggressive
2895 assumption: because the ``undef`` is allowed to be an arbitrary value,
2896 we are allowed to assume that it could be zero. Since a divide by zero
2897 has *undefined behavior*, we are allowed to assume that the operation
2898 does not execute at all. This allows us to delete the divide and all
2899 code after it. Because the undefined operation "can't happen", the
2900 optimizer can assume that it occurs in dead code.
2902 .. code-block:: text
2904 a: store undef -> %X
2905 b: store %X -> undef
2910 These examples reiterate the ``fdiv`` example: a store *of* an undefined
2911 value can be assumed to not have any effect; we can assume that the
2912 value is overwritten with bits that happen to match what was already
2913 there. However, a store *to* an undefined location could clobber
2914 arbitrary memory, therefore, it has undefined behavior.
2921 Poison values are similar to :ref:`undef values <undefvalues>`, however
2922 they also represent the fact that an instruction or constant expression
2923 that cannot evoke side effects has nevertheless detected a condition
2924 that results in undefined behavior.
2926 There is currently no way of representing a poison value in the IR; they
2927 only exist when produced by operations such as :ref:`add <i_add>` with
2930 Poison value behavior is defined in terms of value *dependence*:
2932 - Values other than :ref:`phi <i_phi>` nodes depend on their operands.
2933 - :ref:`Phi <i_phi>` nodes depend on the operand corresponding to
2934 their dynamic predecessor basic block.
2935 - Function arguments depend on the corresponding actual argument values
2936 in the dynamic callers of their functions.
2937 - :ref:`Call <i_call>` instructions depend on the :ref:`ret <i_ret>`
2938 instructions that dynamically transfer control back to them.
2939 - :ref:`Invoke <i_invoke>` instructions depend on the
2940 :ref:`ret <i_ret>`, :ref:`resume <i_resume>`, or exception-throwing
2941 call instructions that dynamically transfer control back to them.
2942 - Non-volatile loads and stores depend on the most recent stores to all
2943 of the referenced memory addresses, following the order in the IR
2944 (including loads and stores implied by intrinsics such as
2945 :ref:`@llvm.memcpy <int_memcpy>`.)
2946 - An instruction with externally visible side effects depends on the
2947 most recent preceding instruction with externally visible side
2948 effects, following the order in the IR. (This includes :ref:`volatile
2949 operations <volatile>`.)
2950 - An instruction *control-depends* on a :ref:`terminator
2951 instruction <terminators>` if the terminator instruction has
2952 multiple successors and the instruction is always executed when
2953 control transfers to one of the successors, and may not be executed
2954 when control is transferred to another.
2955 - Additionally, an instruction also *control-depends* on a terminator
2956 instruction if the set of instructions it otherwise depends on would
2957 be different if the terminator had transferred control to a different
2959 - Dependence is transitive.
2961 Poison values have the same behavior as :ref:`undef values <undefvalues>`,
2962 with the additional effect that any instruction that has a *dependence*
2963 on a poison value has undefined behavior.
2965 Here are some examples:
2967 .. code-block:: llvm
2970 %poison = sub nuw i32 0, 1 ; Results in a poison value.
2971 %still_poison = and i32 %poison, 0 ; 0, but also poison.
2972 %poison_yet_again = getelementptr i32, i32* @h, i32 %still_poison
2973 store i32 0, i32* %poison_yet_again ; memory at @h[0] is poisoned
2975 store i32 %poison, i32* @g ; Poison value stored to memory.
2976 %poison2 = load i32, i32* @g ; Poison value loaded back from memory.
2978 store volatile i32 %poison, i32* @g ; External observation; undefined behavior.
2980 %narrowaddr = bitcast i32* @g to i16*
2981 %wideaddr = bitcast i32* @g to i64*
2982 %poison3 = load i16, i16* %narrowaddr ; Returns a poison value.
2983 %poison4 = load i64, i64* %wideaddr ; Returns a poison value.
2985 %cmp = icmp slt i32 %poison, 0 ; Returns a poison value.
2986 br i1 %cmp, label %true, label %end ; Branch to either destination.
2989 store volatile i32 0, i32* @g ; This is control-dependent on %cmp, so
2990 ; it has undefined behavior.
2994 %p = phi i32 [ 0, %entry ], [ 1, %true ]
2995 ; Both edges into this PHI are
2996 ; control-dependent on %cmp, so this
2997 ; always results in a poison value.
2999 store volatile i32 0, i32* @g ; This would depend on the store in %true
3000 ; if %cmp is true, or the store in %entry
3001 ; otherwise, so this is undefined behavior.
3003 br i1 %cmp, label %second_true, label %second_end
3004 ; The same branch again, but this time the
3005 ; true block doesn't have side effects.
3012 store volatile i32 0, i32* @g ; This time, the instruction always depends
3013 ; on the store in %end. Also, it is
3014 ; control-equivalent to %end, so this is
3015 ; well-defined (ignoring earlier undefined
3016 ; behavior in this example).
3020 Addresses of Basic Blocks
3021 -------------------------
3023 ``blockaddress(@function, %block)``
3025 The '``blockaddress``' constant computes the address of the specified
3026 basic block in the specified function, and always has an ``i8*`` type.
3027 Taking the address of the entry block is illegal.
3029 This value only has defined behavior when used as an operand to the
3030 ':ref:`indirectbr <i_indirectbr>`' instruction, or for comparisons
3031 against null. Pointer equality tests between labels addresses results in
3032 undefined behavior --- though, again, comparison against null is ok, and
3033 no label is equal to the null pointer. This may be passed around as an
3034 opaque pointer sized value as long as the bits are not inspected. This
3035 allows ``ptrtoint`` and arithmetic to be performed on these values so
3036 long as the original value is reconstituted before the ``indirectbr``
3039 Finally, some targets may provide defined semantics when using the value
3040 as the operand to an inline assembly, but that is target specific.
3044 Constant Expressions
3045 --------------------
3047 Constant expressions are used to allow expressions involving other
3048 constants to be used as constants. Constant expressions may be of any
3049 :ref:`first class <t_firstclass>` type and may involve any LLVM operation
3050 that does not have side effects (e.g. load and call are not supported).
3051 The following is the syntax for constant expressions:
3053 ``trunc (CST to TYPE)``
3054 Truncate a constant to another type. The bit size of CST must be
3055 larger than the bit size of TYPE. Both types must be integers.
3056 ``zext (CST to TYPE)``
3057 Zero extend a constant to another type. The bit size of CST must be
3058 smaller than the bit size of TYPE. Both types must be integers.
3059 ``sext (CST to TYPE)``
3060 Sign extend a constant to another type. The bit size of CST must be
3061 smaller than the bit size of TYPE. Both types must be integers.
3062 ``fptrunc (CST to TYPE)``
3063 Truncate a floating point constant to another floating point type.
3064 The size of CST must be larger than the size of TYPE. Both types
3065 must be floating point.
3066 ``fpext (CST to TYPE)``
3067 Floating point extend a constant to another type. The size of CST
3068 must be smaller or equal to the size of TYPE. Both types must be
3070 ``fptoui (CST to TYPE)``
3071 Convert a floating point constant to the corresponding unsigned
3072 integer constant. TYPE must be a scalar or vector integer type. CST
3073 must be of scalar or vector floating point type. Both CST and TYPE
3074 must be scalars, or vectors of the same number of elements. If the
3075 value won't fit in the integer type, the results are undefined.
3076 ``fptosi (CST to TYPE)``
3077 Convert a floating point constant to the corresponding signed
3078 integer constant. TYPE must be a scalar or vector integer type. CST
3079 must be of scalar or vector floating point type. Both CST and TYPE
3080 must be scalars, or vectors of the same number of elements. If the
3081 value won't fit in the integer type, the results are undefined.
3082 ``uitofp (CST to TYPE)``
3083 Convert an unsigned integer constant to the corresponding floating
3084 point constant. TYPE must be a scalar or vector floating point type.
3085 CST must be of scalar or vector integer type. Both CST and TYPE must
3086 be scalars, or vectors of the same number of elements. If the value
3087 won't fit in the floating point type, the results are undefined.
3088 ``sitofp (CST to TYPE)``
3089 Convert a signed integer constant to the corresponding floating
3090 point constant. TYPE must be a scalar or vector floating point type.
3091 CST must be of scalar or vector integer type. Both CST and TYPE must
3092 be scalars, or vectors of the same number of elements. If the value
3093 won't fit in the floating point type, the results are undefined.
3094 ``ptrtoint (CST to TYPE)``
3095 Convert a pointer typed constant to the corresponding integer
3096 constant. ``TYPE`` must be an integer type. ``CST`` must be of
3097 pointer type. The ``CST`` value is zero extended, truncated, or
3098 unchanged to make it fit in ``TYPE``.
3099 ``inttoptr (CST to TYPE)``
3100 Convert an integer constant to a pointer constant. TYPE must be a
3101 pointer type. CST must be of integer type. The CST value is zero
3102 extended, truncated, or unchanged to make it fit in a pointer size.
3103 This one is *really* dangerous!
3104 ``bitcast (CST to TYPE)``
3105 Convert a constant, CST, to another TYPE. The constraints of the
3106 operands are the same as those for the :ref:`bitcast
3107 instruction <i_bitcast>`.
3108 ``addrspacecast (CST to TYPE)``
3109 Convert a constant pointer or constant vector of pointer, CST, to another
3110 TYPE in a different address space. The constraints of the operands are the
3111 same as those for the :ref:`addrspacecast instruction <i_addrspacecast>`.
3112 ``getelementptr (TY, CSTPTR, IDX0, IDX1, ...)``, ``getelementptr inbounds (TY, CSTPTR, IDX0, IDX1, ...)``
3113 Perform the :ref:`getelementptr operation <i_getelementptr>` on
3114 constants. As with the :ref:`getelementptr <i_getelementptr>`
3115 instruction, the index list may have zero or more indexes, which are
3116 required to make sense for the type of "pointer to TY".
3117 ``select (COND, VAL1, VAL2)``
3118 Perform the :ref:`select operation <i_select>` on constants.
3119 ``icmp COND (VAL1, VAL2)``
3120 Performs the :ref:`icmp operation <i_icmp>` on constants.
3121 ``fcmp COND (VAL1, VAL2)``
3122 Performs the :ref:`fcmp operation <i_fcmp>` on constants.
3123 ``extractelement (VAL, IDX)``
3124 Perform the :ref:`extractelement operation <i_extractelement>` on
3126 ``insertelement (VAL, ELT, IDX)``
3127 Perform the :ref:`insertelement operation <i_insertelement>` on
3129 ``shufflevector (VEC1, VEC2, IDXMASK)``
3130 Perform the :ref:`shufflevector operation <i_shufflevector>` on
3132 ``extractvalue (VAL, IDX0, IDX1, ...)``
3133 Perform the :ref:`extractvalue operation <i_extractvalue>` on
3134 constants. The index list is interpreted in a similar manner as
3135 indices in a ':ref:`getelementptr <i_getelementptr>`' operation. At
3136 least one index value must be specified.
3137 ``insertvalue (VAL, ELT, IDX0, IDX1, ...)``
3138 Perform the :ref:`insertvalue operation <i_insertvalue>` on constants.
3139 The index list is interpreted in a similar manner as indices in a
3140 ':ref:`getelementptr <i_getelementptr>`' operation. At least one index
3141 value must be specified.
3142 ``OPCODE (LHS, RHS)``
3143 Perform the specified operation of the LHS and RHS constants. OPCODE
3144 may be any of the :ref:`binary <binaryops>` or :ref:`bitwise
3145 binary <bitwiseops>` operations. The constraints on operands are
3146 the same as those for the corresponding instruction (e.g. no bitwise
3147 operations on floating point values are allowed).
3154 Inline Assembler Expressions
3155 ----------------------------
3157 LLVM supports inline assembler expressions (as opposed to :ref:`Module-Level
3158 Inline Assembly <moduleasm>`) through the use of a special value. This value
3159 represents the inline assembler as a template string (containing the
3160 instructions to emit), a list of operand constraints (stored as a string), a
3161 flag that indicates whether or not the inline asm expression has side effects,
3162 and a flag indicating whether the function containing the asm needs to align its
3163 stack conservatively.
3165 The template string supports argument substitution of the operands using "``$``"
3166 followed by a number, to indicate substitution of the given register/memory
3167 location, as specified by the constraint string. "``${NUM:MODIFIER}``" may also
3168 be used, where ``MODIFIER`` is a target-specific annotation for how to print the
3169 operand (See :ref:`inline-asm-modifiers`).
3171 A literal "``$``" may be included by using "``$$``" in the template. To include
3172 other special characters into the output, the usual "``\XX``" escapes may be
3173 used, just as in other strings. Note that after template substitution, the
3174 resulting assembly string is parsed by LLVM's integrated assembler unless it is
3175 disabled -- even when emitting a ``.s`` file -- and thus must contain assembly
3176 syntax known to LLVM.
3178 LLVM's support for inline asm is modeled closely on the requirements of Clang's
3179 GCC-compatible inline-asm support. Thus, the feature-set and the constraint and
3180 modifier codes listed here are similar or identical to those in GCC's inline asm
3181 support. However, to be clear, the syntax of the template and constraint strings
3182 described here is *not* the same as the syntax accepted by GCC and Clang, and,
3183 while most constraint letters are passed through as-is by Clang, some get
3184 translated to other codes when converting from the C source to the LLVM
3187 An example inline assembler expression is:
3189 .. code-block:: llvm
3191 i32 (i32) asm "bswap $0", "=r,r"
3193 Inline assembler expressions may **only** be used as the callee operand
3194 of a :ref:`call <i_call>` or an :ref:`invoke <i_invoke>` instruction.
3195 Thus, typically we have:
3197 .. code-block:: llvm
3199 %X = call i32 asm "bswap $0", "=r,r"(i32 %Y)
3201 Inline asms with side effects not visible in the constraint list must be
3202 marked as having side effects. This is done through the use of the
3203 '``sideeffect``' keyword, like so:
3205 .. code-block:: llvm
3207 call void asm sideeffect "eieio", ""()
3209 In some cases inline asms will contain code that will not work unless
3210 the stack is aligned in some way, such as calls or SSE instructions on
3211 x86, yet will not contain code that does that alignment within the asm.
3212 The compiler should make conservative assumptions about what the asm
3213 might contain and should generate its usual stack alignment code in the
3214 prologue if the '``alignstack``' keyword is present:
3216 .. code-block:: llvm
3218 call void asm alignstack "eieio", ""()
3220 Inline asms also support using non-standard assembly dialects. The
3221 assumed dialect is ATT. When the '``inteldialect``' keyword is present,
3222 the inline asm is using the Intel dialect. Currently, ATT and Intel are
3223 the only supported dialects. An example is:
3225 .. code-block:: llvm
3227 call void asm inteldialect "eieio", ""()
3229 If multiple keywords appear the '``sideeffect``' keyword must come
3230 first, the '``alignstack``' keyword second and the '``inteldialect``'
3233 Inline Asm Constraint String
3234 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3236 The constraint list is a comma-separated string, each element containing one or
3237 more constraint codes.
3239 For each element in the constraint list an appropriate register or memory
3240 operand will be chosen, and it will be made available to assembly template
3241 string expansion as ``$0`` for the first constraint in the list, ``$1`` for the
3244 There are three different types of constraints, which are distinguished by a
3245 prefix symbol in front of the constraint code: Output, Input, and Clobber. The
3246 constraints must always be given in that order: outputs first, then inputs, then
3247 clobbers. They cannot be intermingled.
3249 There are also three different categories of constraint codes:
3251 - Register constraint. This is either a register class, or a fixed physical
3252 register. This kind of constraint will allocate a register, and if necessary,
3253 bitcast the argument or result to the appropriate type.
3254 - Memory constraint. This kind of constraint is for use with an instruction
3255 taking a memory operand. Different constraints allow for different addressing
3256 modes used by the target.
3257 - Immediate value constraint. This kind of constraint is for an integer or other
3258 immediate value which can be rendered directly into an instruction. The
3259 various target-specific constraints allow the selection of a value in the
3260 proper range for the instruction you wish to use it with.
3265 Output constraints are specified by an "``=``" prefix (e.g. "``=r``"). This
3266 indicates that the assembly will write to this operand, and the operand will
3267 then be made available as a return value of the ``asm`` expression. Output
3268 constraints do not consume an argument from the call instruction. (Except, see
3269 below about indirect outputs).
3271 Normally, it is expected that no output locations are written to by the assembly
3272 expression until *all* of the inputs have been read. As such, LLVM may assign
3273 the same register to an output and an input. If this is not safe (e.g. if the
3274 assembly contains two instructions, where the first writes to one output, and
3275 the second reads an input and writes to a second output), then the "``&``"
3276 modifier must be used (e.g. "``=&r``") to specify that the output is an
3277 "early-clobber" output. Marking an output as "early-clobber" ensures that LLVM
3278 will not use the same register for any inputs (other than an input tied to this
3284 Input constraints do not have a prefix -- just the constraint codes. Each input
3285 constraint will consume one argument from the call instruction. It is not
3286 permitted for the asm to write to any input register or memory location (unless
3287 that input is tied to an output). Note also that multiple inputs may all be
3288 assigned to the same register, if LLVM can determine that they necessarily all
3289 contain the same value.
3291 Instead of providing a Constraint Code, input constraints may also "tie"
3292 themselves to an output constraint, by providing an integer as the constraint
3293 string. Tied inputs still consume an argument from the call instruction, and
3294 take up a position in the asm template numbering as is usual -- they will simply
3295 be constrained to always use the same register as the output they've been tied
3296 to. For example, a constraint string of "``=r,0``" says to assign a register for
3297 output, and use that register as an input as well (it being the 0'th
3300 It is permitted to tie an input to an "early-clobber" output. In that case, no
3301 *other* input may share the same register as the input tied to the early-clobber
3302 (even when the other input has the same value).
3304 You may only tie an input to an output which has a register constraint, not a
3305 memory constraint. Only a single input may be tied to an output.
3307 There is also an "interesting" feature which deserves a bit of explanation: if a
3308 register class constraint allocates a register which is too small for the value
3309 type operand provided as input, the input value will be split into multiple
3310 registers, and all of them passed to the inline asm.
3312 However, this feature is often not as useful as you might think.
3314 Firstly, the registers are *not* guaranteed to be consecutive. So, on those
3315 architectures that have instructions which operate on multiple consecutive
3316 instructions, this is not an appropriate way to support them. (e.g. the 32-bit
3317 SparcV8 has a 64-bit load, which instruction takes a single 32-bit register. The
3318 hardware then loads into both the named register, and the next register. This
3319 feature of inline asm would not be useful to support that.)
3321 A few of the targets provide a template string modifier allowing explicit access
3322 to the second register of a two-register operand (e.g. MIPS ``L``, ``M``, and
3323 ``D``). On such an architecture, you can actually access the second allocated
3324 register (yet, still, not any subsequent ones). But, in that case, you're still
3325 probably better off simply splitting the value into two separate operands, for
3326 clarity. (e.g. see the description of the ``A`` constraint on X86, which,
3327 despite existing only for use with this feature, is not really a good idea to
3330 Indirect inputs and outputs
3331 """""""""""""""""""""""""""
3333 Indirect output or input constraints can be specified by the "``*``" modifier
3334 (which goes after the "``=``" in case of an output). This indicates that the asm
3335 will write to or read from the contents of an *address* provided as an input
3336 argument. (Note that in this way, indirect outputs act more like an *input* than
3337 an output: just like an input, they consume an argument of the call expression,
3338 rather than producing a return value. An indirect output constraint is an
3339 "output" only in that the asm is expected to write to the contents of the input
3340 memory location, instead of just read from it).
3342 This is most typically used for memory constraint, e.g. "``=*m``", to pass the
3343 address of a variable as a value.
3345 It is also possible to use an indirect *register* constraint, but only on output
3346 (e.g. "``=*r``"). This will cause LLVM to allocate a register for an output
3347 value normally, and then, separately emit a store to the address provided as
3348 input, after the provided inline asm. (It's not clear what value this
3349 functionality provides, compared to writing the store explicitly after the asm
3350 statement, and it can only produce worse code, since it bypasses many
3351 optimization passes. I would recommend not using it.)
3357 A clobber constraint is indicated by a "``~``" prefix. A clobber does not
3358 consume an input operand, nor generate an output. Clobbers cannot use any of the
3359 general constraint code letters -- they may use only explicit register
3360 constraints, e.g. "``~{eax}``". The one exception is that a clobber string of
3361 "``~{memory}``" indicates that the assembly writes to arbitrary undeclared
3362 memory locations -- not only the memory pointed to by a declared indirect
3368 After a potential prefix comes constraint code, or codes.
3370 A Constraint Code is either a single letter (e.g. "``r``"), a "``^``" character
3371 followed by two letters (e.g. "``^wc``"), or "``{``" register-name "``}``"
3374 The one and two letter constraint codes are typically chosen to be the same as
3375 GCC's constraint codes.
3377 A single constraint may include one or more than constraint code in it, leaving
3378 it up to LLVM to choose which one to use. This is included mainly for
3379 compatibility with the translation of GCC inline asm coming from clang.
3381 There are two ways to specify alternatives, and either or both may be used in an
3382 inline asm constraint list:
3384 1) Append the codes to each other, making a constraint code set. E.g. "``im``"
3385 or "``{eax}m``". This means "choose any of the options in the set". The
3386 choice of constraint is made independently for each constraint in the
3389 2) Use "``|``" between constraint code sets, creating alternatives. Every
3390 constraint in the constraint list must have the same number of alternative
3391 sets. With this syntax, the same alternative in *all* of the items in the
3392 constraint list will be chosen together.
3394 Putting those together, you might have a two operand constraint string like
3395 ``"rm|r,ri|rm"``. This indicates that if operand 0 is ``r`` or ``m``, then
3396 operand 1 may be one of ``r`` or ``i``. If operand 0 is ``r``, then operand 1
3397 may be one of ``r`` or ``m``. But, operand 0 and 1 cannot both be of type m.
3399 However, the use of either of the alternatives features is *NOT* recommended, as
3400 LLVM is not able to make an intelligent choice about which one to use. (At the
3401 point it currently needs to choose, not enough information is available to do so
3402 in a smart way.) Thus, it simply tries to make a choice that's most likely to
3403 compile, not one that will be optimal performance. (e.g., given "``rm``", it'll
3404 always choose to use memory, not registers). And, if given multiple registers,
3405 or multiple register classes, it will simply choose the first one. (In fact, it
3406 doesn't currently even ensure explicitly specified physical registers are
3407 unique, so specifying multiple physical registers as alternatives, like
3408 ``{r11}{r12},{r11}{r12}``, will assign r11 to both operands, not at all what was
3411 Supported Constraint Code List
3412 """"""""""""""""""""""""""""""
3414 The constraint codes are, in general, expected to behave the same way they do in
3415 GCC. LLVM's support is often implemented on an 'as-needed' basis, to support C
3416 inline asm code which was supported by GCC. A mismatch in behavior between LLVM
3417 and GCC likely indicates a bug in LLVM.
3419 Some constraint codes are typically supported by all targets:
3421 - ``r``: A register in the target's general purpose register class.
3422 - ``m``: A memory address operand. It is target-specific what addressing modes
3423 are supported, typical examples are register, or register + register offset,
3424 or register + immediate offset (of some target-specific size).
3425 - ``i``: An integer constant (of target-specific width). Allows either a simple
3426 immediate, or a relocatable value.
3427 - ``n``: An integer constant -- *not* including relocatable values.
3428 - ``s``: An integer constant, but allowing *only* relocatable values.
3429 - ``X``: Allows an operand of any kind, no constraint whatsoever. Typically
3430 useful to pass a label for an asm branch or call.
3432 .. FIXME: but that surely isn't actually okay to jump out of an asm
3433 block without telling llvm about the control transfer???)
3435 - ``{register-name}``: Requires exactly the named physical register.
3437 Other constraints are target-specific:
3441 - ``z``: An immediate integer 0. Outputs ``WZR`` or ``XZR``, as appropriate.
3442 - ``I``: An immediate integer valid for an ``ADD`` or ``SUB`` instruction,
3443 i.e. 0 to 4095 with optional shift by 12.
3444 - ``J``: An immediate integer that, when negated, is valid for an ``ADD`` or
3445 ``SUB`` instruction, i.e. -1 to -4095 with optional left shift by 12.
3446 - ``K``: An immediate integer that is valid for the 'bitmask immediate 32' of a
3447 logical instruction like ``AND``, ``EOR``, or ``ORR`` with a 32-bit register.
3448 - ``L``: An immediate integer that is valid for the 'bitmask immediate 64' of a
3449 logical instruction like ``AND``, ``EOR``, or ``ORR`` with a 64-bit register.
3450 - ``M``: An immediate integer for use with the ``MOV`` assembly alias on a
3451 32-bit register. This is a superset of ``K``: in addition to the bitmask
3452 immediate, also allows immediate integers which can be loaded with a single
3453 ``MOVZ`` or ``MOVL`` instruction.
3454 - ``N``: An immediate integer for use with the ``MOV`` assembly alias on a
3455 64-bit register. This is a superset of ``L``.
3456 - ``Q``: Memory address operand must be in a single register (no
3457 offsets). (However, LLVM currently does this for the ``m`` constraint as
3459 - ``r``: A 32 or 64-bit integer register (W* or X*).
3460 - ``w``: A 32, 64, or 128-bit floating-point/SIMD register.
3461 - ``x``: A lower 128-bit floating-point/SIMD register (``V0`` to ``V15``).
3465 - ``r``: A 32 or 64-bit integer register.
3466 - ``[0-9]v``: The 32-bit VGPR register, number 0-9.
3467 - ``[0-9]s``: The 32-bit SGPR register, number 0-9.
3472 - ``Q``, ``Um``, ``Un``, ``Uq``, ``Us``, ``Ut``, ``Uv``, ``Uy``: Memory address
3473 operand. Treated the same as operand ``m``, at the moment.
3475 ARM and ARM's Thumb2 mode:
3477 - ``j``: An immediate integer between 0 and 65535 (valid for ``MOVW``)
3478 - ``I``: An immediate integer valid for a data-processing instruction.
3479 - ``J``: An immediate integer between -4095 and 4095.
3480 - ``K``: An immediate integer whose bitwise inverse is valid for a
3481 data-processing instruction. (Can be used with template modifier "``B``" to
3482 print the inverted value).
3483 - ``L``: An immediate integer whose negation is valid for a data-processing
3484 instruction. (Can be used with template modifier "``n``" to print the negated
3486 - ``M``: A power of two or a integer between 0 and 32.
3487 - ``N``: Invalid immediate constraint.
3488 - ``O``: Invalid immediate constraint.
3489 - ``r``: A general-purpose 32-bit integer register (``r0-r15``).
3490 - ``l``: In Thumb2 mode, low 32-bit GPR registers (``r0-r7``). In ARM mode, same
3492 - ``h``: In Thumb2 mode, a high 32-bit GPR register (``r8-r15``). In ARM mode,
3494 - ``w``: A 32, 64, or 128-bit floating-point/SIMD register: ``s0-s31``,
3495 ``d0-d31``, or ``q0-q15``.
3496 - ``x``: A 32, 64, or 128-bit floating-point/SIMD register: ``s0-s15``,
3497 ``d0-d7``, or ``q0-q3``.
3498 - ``t``: A floating-point/SIMD register, only supports 32-bit values:
3503 - ``I``: An immediate integer between 0 and 255.
3504 - ``J``: An immediate integer between -255 and -1.
3505 - ``K``: An immediate integer between 0 and 255, with optional left-shift by
3507 - ``L``: An immediate integer between -7 and 7.
3508 - ``M``: An immediate integer which is a multiple of 4 between 0 and 1020.
3509 - ``N``: An immediate integer between 0 and 31.
3510 - ``O``: An immediate integer which is a multiple of 4 between -508 and 508.
3511 - ``r``: A low 32-bit GPR register (``r0-r7``).
3512 - ``l``: A low 32-bit GPR register (``r0-r7``).
3513 - ``h``: A high GPR register (``r0-r7``).
3514 - ``w``: A 32, 64, or 128-bit floating-point/SIMD register: ``s0-s31``,
3515 ``d0-d31``, or ``q0-q15``.
3516 - ``x``: A 32, 64, or 128-bit floating-point/SIMD register: ``s0-s15``,
3517 ``d0-d7``, or ``q0-q3``.
3518 - ``t``: A floating-point/SIMD register, only supports 32-bit values:
3524 - ``o``, ``v``: A memory address operand, treated the same as constraint ``m``,
3526 - ``r``: A 32 or 64-bit register.
3530 - ``r``: An 8 or 16-bit register.
3534 - ``I``: An immediate signed 16-bit integer.
3535 - ``J``: An immediate integer zero.
3536 - ``K``: An immediate unsigned 16-bit integer.
3537 - ``L``: An immediate 32-bit integer, where the lower 16 bits are 0.
3538 - ``N``: An immediate integer between -65535 and -1.
3539 - ``O``: An immediate signed 15-bit integer.
3540 - ``P``: An immediate integer between 1 and 65535.
3541 - ``m``: A memory address operand. In MIPS-SE mode, allows a base address
3542 register plus 16-bit immediate offset. In MIPS mode, just a base register.
3543 - ``R``: A memory address operand. In MIPS-SE mode, allows a base address
3544 register plus a 9-bit signed offset. In MIPS mode, the same as constraint
3546 - ``ZC``: A memory address operand, suitable for use in a ``pref``, ``ll``, or
3547 ``sc`` instruction on the given subtarget (details vary).
3548 - ``r``, ``d``, ``y``: A 32 or 64-bit GPR register.
3549 - ``f``: A 32 or 64-bit FPU register (``F0-F31``), or a 128-bit MSA register
3550 (``W0-W31``). In the case of MSA registers, it is recommended to use the ``w``
3551 argument modifier for compatibility with GCC.
3552 - ``c``: A 32-bit or 64-bit GPR register suitable for indirect jump (always
3554 - ``l``: The ``lo`` register, 32 or 64-bit.
3559 - ``b``: A 1-bit integer register.
3560 - ``c`` or ``h``: A 16-bit integer register.
3561 - ``r``: A 32-bit integer register.
3562 - ``l`` or ``N``: A 64-bit integer register.
3563 - ``f``: A 32-bit float register.
3564 - ``d``: A 64-bit float register.
3569 - ``I``: An immediate signed 16-bit integer.
3570 - ``J``: An immediate unsigned 16-bit integer, shifted left 16 bits.
3571 - ``K``: An immediate unsigned 16-bit integer.
3572 - ``L``: An immediate signed 16-bit integer, shifted left 16 bits.
3573 - ``M``: An immediate integer greater than 31.
3574 - ``N``: An immediate integer that is an exact power of 2.
3575 - ``O``: The immediate integer constant 0.
3576 - ``P``: An immediate integer constant whose negation is a signed 16-bit
3578 - ``es``, ``o``, ``Q``, ``Z``, ``Zy``: A memory address operand, currently
3579 treated the same as ``m``.
3580 - ``r``: A 32 or 64-bit integer register.
3581 - ``b``: A 32 or 64-bit integer register, excluding ``R0`` (that is:
3583 - ``f``: A 32 or 64-bit float register (``F0-F31``), or when QPX is enabled, a
3584 128 or 256-bit QPX register (``Q0-Q31``; aliases the ``F`` registers).
3585 - ``v``: For ``4 x f32`` or ``4 x f64`` types, when QPX is enabled, a
3586 128 or 256-bit QPX register (``Q0-Q31``), otherwise a 128-bit
3587 altivec vector register (``V0-V31``).
3589 .. FIXME: is this a bug that v accepts QPX registers? I think this
3590 is supposed to only use the altivec vector registers?
3592 - ``y``: Condition register (``CR0-CR7``).
3593 - ``wc``: An individual CR bit in a CR register.
3594 - ``wa``, ``wd``, ``wf``: Any 128-bit VSX vector register, from the full VSX
3595 register set (overlapping both the floating-point and vector register files).
3596 - ``ws``: A 32 or 64-bit floating point register, from the full VSX register
3601 - ``I``: An immediate 13-bit signed integer.
3602 - ``r``: A 32-bit integer register.
3606 - ``I``: An immediate unsigned 8-bit integer.
3607 - ``J``: An immediate unsigned 12-bit integer.
3608 - ``K``: An immediate signed 16-bit integer.
3609 - ``L``: An immediate signed 20-bit integer.
3610 - ``M``: An immediate integer 0x7fffffff.
3611 - ``Q``: A memory address operand with a base address and a 12-bit immediate
3612 unsigned displacement.
3613 - ``R``: A memory address operand with a base address, a 12-bit immediate
3614 unsigned displacement, and an index register.
3615 - ``S``: A memory address operand with a base address and a 20-bit immediate
3616 signed displacement.
3617 - ``T``: A memory address operand with a base address, a 20-bit immediate
3618 signed displacement, and an index register.
3619 - ``r`` or ``d``: A 32, 64, or 128-bit integer register.
3620 - ``a``: A 32, 64, or 128-bit integer address register (excludes R0, which in an
3621 address context evaluates as zero).
3622 - ``h``: A 32-bit value in the high part of a 64bit data register
3624 - ``f``: A 32, 64, or 128-bit floating point register.
3628 - ``I``: An immediate integer between 0 and 31.
3629 - ``J``: An immediate integer between 0 and 64.
3630 - ``K``: An immediate signed 8-bit integer.
3631 - ``L``: An immediate integer, 0xff or 0xffff or (in 64-bit mode only)
3633 - ``M``: An immediate integer between 0 and 3.
3634 - ``N``: An immediate unsigned 8-bit integer.
3635 - ``O``: An immediate integer between 0 and 127.
3636 - ``e``: An immediate 32-bit signed integer.
3637 - ``Z``: An immediate 32-bit unsigned integer.
3638 - ``o``, ``v``: Treated the same as ``m``, at the moment.
3639 - ``q``: An 8, 16, 32, or 64-bit register which can be accessed as an 8-bit
3640 ``l`` integer register. On X86-32, this is the ``a``, ``b``, ``c``, and ``d``
3641 registers, and on X86-64, it is all of the integer registers.
3642 - ``Q``: An 8, 16, 32, or 64-bit register which can be accessed as an 8-bit
3643 ``h`` integer register. This is the ``a``, ``b``, ``c``, and ``d`` registers.
3644 - ``r`` or ``l``: An 8, 16, 32, or 64-bit integer register.
3645 - ``R``: An 8, 16, 32, or 64-bit "legacy" integer register -- one which has
3646 existed since i386, and can be accessed without the REX prefix.
3647 - ``f``: A 32, 64, or 80-bit '387 FPU stack pseudo-register.
3648 - ``y``: A 64-bit MMX register, if MMX is enabled.
3649 - ``x``: If SSE is enabled: a 32 or 64-bit scalar operand, or 128-bit vector
3650 operand in a SSE register. If AVX is also enabled, can also be a 256-bit
3651 vector operand in an AVX register. If AVX-512 is also enabled, can also be a
3652 512-bit vector operand in an AVX512 register, Otherwise, an error.
3653 - ``Y``: The same as ``x``, if *SSE2* is enabled, otherwise an error.
3654 - ``A``: Special case: allocates EAX first, then EDX, for a single operand (in
3655 32-bit mode, a 64-bit integer operand will get split into two registers). It
3656 is not recommended to use this constraint, as in 64-bit mode, the 64-bit
3657 operand will get allocated only to RAX -- if two 32-bit operands are needed,
3658 you're better off splitting it yourself, before passing it to the asm
3663 - ``r``: A 32-bit integer register.
3666 .. _inline-asm-modifiers:
3668 Asm template argument modifiers
3669 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3671 In the asm template string, modifiers can be used on the operand reference, like
3674 The modifiers are, in general, expected to behave the same way they do in
3675 GCC. LLVM's support is often implemented on an 'as-needed' basis, to support C
3676 inline asm code which was supported by GCC. A mismatch in behavior between LLVM
3677 and GCC likely indicates a bug in LLVM.
3681 - ``c``: Print an immediate integer constant unadorned, without
3682 the target-specific immediate punctuation (e.g. no ``$`` prefix).
3683 - ``n``: Negate and print immediate integer constant unadorned, without the
3684 target-specific immediate punctuation (e.g. no ``$`` prefix).
3685 - ``l``: Print as an unadorned label, without the target-specific label
3686 punctuation (e.g. no ``$`` prefix).
3690 - ``w``: Print a GPR register with a ``w*`` name instead of ``x*`` name. E.g.,
3691 instead of ``x30``, print ``w30``.
3692 - ``x``: Print a GPR register with a ``x*`` name. (this is the default, anyhow).
3693 - ``b``, ``h``, ``s``, ``d``, ``q``: Print a floating-point/SIMD register with a
3694 ``b*``, ``h*``, ``s*``, ``d*``, or ``q*`` name, rather than the default of
3703 - ``a``: Print an operand as an address (with ``[`` and ``]`` surrounding a
3707 - ``y``: Print a VFP single-precision register as an indexed double (e.g. print
3708 as ``d4[1]`` instead of ``s9``)
3709 - ``B``: Bitwise invert and print an immediate integer constant without ``#``
3711 - ``L``: Print the low 16-bits of an immediate integer constant.
3712 - ``M``: Print as a register set suitable for ldm/stm. Also prints *all*
3713 register operands subsequent to the specified one (!), so use carefully.
3714 - ``Q``: Print the low-order register of a register-pair, or the low-order
3715 register of a two-register operand.
3716 - ``R``: Print the high-order register of a register-pair, or the high-order
3717 register of a two-register operand.
3718 - ``H``: Print the second register of a register-pair. (On a big-endian system,
3719 ``H`` is equivalent to ``Q``, and on little-endian system, ``H`` is equivalent
3722 .. FIXME: H doesn't currently support printing the second register
3723 of a two-register operand.
3725 - ``e``: Print the low doubleword register of a NEON quad register.
3726 - ``f``: Print the high doubleword register of a NEON quad register.
3727 - ``m``: Print the base register of a memory operand without the ``[`` and ``]``
3732 - ``L``: Print the second register of a two-register operand. Requires that it
3733 has been allocated consecutively to the first.
3735 .. FIXME: why is it restricted to consecutive ones? And there's
3736 nothing that ensures that happens, is there?
3738 - ``I``: Print the letter 'i' if the operand is an integer constant, otherwise
3739 nothing. Used to print 'addi' vs 'add' instructions.
3743 No additional modifiers.
3747 - ``X``: Print an immediate integer as hexadecimal
3748 - ``x``: Print the low 16 bits of an immediate integer as hexadecimal.
3749 - ``d``: Print an immediate integer as decimal.
3750 - ``m``: Subtract one and print an immediate integer as decimal.
3751 - ``z``: Print $0 if an immediate zero, otherwise print normally.
3752 - ``L``: Print the low-order register of a two-register operand, or prints the
3753 address of the low-order word of a double-word memory operand.
3755 .. FIXME: L seems to be missing memory operand support.
3757 - ``M``: Print the high-order register of a two-register operand, or prints the
3758 address of the high-order word of a double-word memory operand.
3760 .. FIXME: M seems to be missing memory operand support.
3762 - ``D``: Print the second register of a two-register operand, or prints the
3763 second word of a double-word memory operand. (On a big-endian system, ``D`` is
3764 equivalent to ``L``, and on little-endian system, ``D`` is equivalent to
3766 - ``w``: No effect. Provided for compatibility with GCC which requires this
3767 modifier in order to print MSA registers (``W0-W31``) with the ``f``
3776 - ``L``: Print the second register of a two-register operand. Requires that it
3777 has been allocated consecutively to the first.
3779 .. FIXME: why is it restricted to consecutive ones? And there's
3780 nothing that ensures that happens, is there?
3782 - ``I``: Print the letter 'i' if the operand is an integer constant, otherwise
3783 nothing. Used to print 'addi' vs 'add' instructions.
3784 - ``y``: For a memory operand, prints formatter for a two-register X-form
3785 instruction. (Currently always prints ``r0,OPERAND``).
3786 - ``U``: Prints 'u' if the memory operand is an update form, and nothing
3787 otherwise. (NOTE: LLVM does not support update form, so this will currently
3788 always print nothing)
3789 - ``X``: Prints 'x' if the memory operand is an indexed form. (NOTE: LLVM does
3790 not support indexed form, so this will currently always print nothing)
3798 SystemZ implements only ``n``, and does *not* support any of the other
3799 target-independent modifiers.
3803 - ``c``: Print an unadorned integer or symbol name. (The latter is
3804 target-specific behavior for this typically target-independent modifier).
3805 - ``A``: Print a register name with a '``*``' before it.
3806 - ``b``: Print an 8-bit register name (e.g. ``al``); do nothing on a memory
3808 - ``h``: Print the upper 8-bit register name (e.g. ``ah``); do nothing on a
3810 - ``w``: Print the 16-bit register name (e.g. ``ax``); do nothing on a memory
3812 - ``k``: Print the 32-bit register name (e.g. ``eax``); do nothing on a memory
3814 - ``q``: Print the 64-bit register name (e.g. ``rax``), if 64-bit registers are
3815 available, otherwise the 32-bit register name; do nothing on a memory operand.
3816 - ``n``: Negate and print an unadorned integer, or, for operands other than an
3817 immediate integer (e.g. a relocatable symbol expression), print a '-' before
3818 the operand. (The behavior for relocatable symbol expressions is a
3819 target-specific behavior for this typically target-independent modifier)
3820 - ``H``: Print a memory reference with additional offset +8.
3821 - ``P``: Print a memory reference or operand for use as the argument of a call
3822 instruction. (E.g. omit ``(rip)``, even though it's PC-relative.)
3826 No additional modifiers.
3832 The call instructions that wrap inline asm nodes may have a
3833 "``!srcloc``" MDNode attached to it that contains a list of constant
3834 integers. If present, the code generator will use the integer as the
3835 location cookie value when report errors through the ``LLVMContext``
3836 error reporting mechanisms. This allows a front-end to correlate backend
3837 errors that occur with inline asm back to the source code that produced
3840 .. code-block:: llvm
3842 call void asm sideeffect "something bad", ""(), !srcloc !42
3844 !42 = !{ i32 1234567 }
3846 It is up to the front-end to make sense of the magic numbers it places
3847 in the IR. If the MDNode contains multiple constants, the code generator
3848 will use the one that corresponds to the line of the asm that the error
3856 LLVM IR allows metadata to be attached to instructions in the program
3857 that can convey extra information about the code to the optimizers and
3858 code generator. One example application of metadata is source-level
3859 debug information. There are two metadata primitives: strings and nodes.
3861 Metadata does not have a type, and is not a value. If referenced from a
3862 ``call`` instruction, it uses the ``metadata`` type.
3864 All metadata are identified in syntax by a exclamation point ('``!``').
3866 .. _metadata-string:
3868 Metadata Nodes and Metadata Strings
3869 -----------------------------------
3871 A metadata string is a string surrounded by double quotes. It can
3872 contain any character by escaping non-printable characters with
3873 "``\xx``" where "``xx``" is the two digit hex code. For example:
3876 Metadata nodes are represented with notation similar to structure
3877 constants (a comma separated list of elements, surrounded by braces and
3878 preceded by an exclamation point). Metadata nodes can have any values as
3879 their operand. For example:
3881 .. code-block:: llvm
3883 !{ !"test\00", i32 10}
3885 Metadata nodes that aren't uniqued use the ``distinct`` keyword. For example:
3887 .. code-block:: text
3889 !0 = distinct !{!"test\00", i32 10}
3891 ``distinct`` nodes are useful when nodes shouldn't be merged based on their
3892 content. They can also occur when transformations cause uniquing collisions
3893 when metadata operands change.
3895 A :ref:`named metadata <namedmetadatastructure>` is a collection of
3896 metadata nodes, which can be looked up in the module symbol table. For
3899 .. code-block:: llvm
3903 Metadata can be used as function arguments. Here ``llvm.dbg.value``
3904 function is using two metadata arguments:
3906 .. code-block:: llvm
3908 call void @llvm.dbg.value(metadata !24, i64 0, metadata !25)
3910 Metadata can be attached to an instruction. Here metadata ``!21`` is attached
3911 to the ``add`` instruction using the ``!dbg`` identifier:
3913 .. code-block:: llvm
3915 %indvar.next = add i64 %indvar, 1, !dbg !21
3917 Metadata can also be attached to a function definition. Here metadata ``!22``
3918 is attached to the ``foo`` function using the ``!dbg`` identifier:
3920 .. code-block:: llvm
3922 define void @foo() !dbg !22 {
3926 More information about specific metadata nodes recognized by the
3927 optimizers and code generator is found below.
3929 .. _specialized-metadata:
3931 Specialized Metadata Nodes
3932 ^^^^^^^^^^^^^^^^^^^^^^^^^^
3934 Specialized metadata nodes are custom data structures in metadata (as opposed
3935 to generic tuples). Their fields are labelled, and can be specified in any
3938 These aren't inherently debug info centric, but currently all the specialized
3939 metadata nodes are related to debug info.
3946 ``DICompileUnit`` nodes represent a compile unit. The ``enums:``,
3947 ``retainedTypes:``, ``subprograms:``, ``globals:``, ``imports:`` and ``macros:``
3948 fields are tuples containing the debug info to be emitted along with the compile
3949 unit, regardless of code optimizations (some nodes are only emitted if there are
3950 references to them from instructions).
3952 .. code-block:: text
3954 !0 = !DICompileUnit(language: DW_LANG_C99, file: !1, producer: "clang",
3955 isOptimized: true, flags: "-O2", runtimeVersion: 2,
3956 splitDebugFilename: "abc.debug", emissionKind: FullDebug,
3957 enums: !2, retainedTypes: !3, subprograms: !4,
3958 globals: !5, imports: !6, macros: !7, dwoId: 0x0abcd)
3960 Compile unit descriptors provide the root scope for objects declared in a
3961 specific compilation unit. File descriptors are defined using this scope.
3962 These descriptors are collected by a named metadata ``!llvm.dbg.cu``. They
3963 keep track of subprograms, global variables, type information, and imported
3964 entities (declarations and namespaces).
3971 ``DIFile`` nodes represent files. The ``filename:`` can include slashes.
3973 .. code-block:: llvm
3975 !0 = !DIFile(filename: "path/to/file", directory: "/path/to/dir")
3977 Files are sometimes used in ``scope:`` fields, and are the only valid target
3978 for ``file:`` fields.
3985 ``DIBasicType`` nodes represent primitive types, such as ``int``, ``bool`` and
3986 ``float``. ``tag:`` defaults to ``DW_TAG_base_type``.
3988 .. code-block:: text
3990 !0 = !DIBasicType(name: "unsigned char", size: 8, align: 8,
3991 encoding: DW_ATE_unsigned_char)
3992 !1 = !DIBasicType(tag: DW_TAG_unspecified_type, name: "decltype(nullptr)")
3994 The ``encoding:`` describes the details of the type. Usually it's one of the
3997 .. code-block:: text
4003 DW_ATE_signed_char = 6
4005 DW_ATE_unsigned_char = 8
4007 .. _DISubroutineType:
4012 ``DISubroutineType`` nodes represent subroutine types. Their ``types:`` field
4013 refers to a tuple; the first operand is the return type, while the rest are the
4014 types of the formal arguments in order. If the first operand is ``null``, that
4015 represents a function with no return value (such as ``void foo() {}`` in C++).
4017 .. code-block:: text
4019 !0 = !BasicType(name: "int", size: 32, align: 32, DW_ATE_signed)
4020 !1 = !BasicType(name: "char", size: 8, align: 8, DW_ATE_signed_char)
4021 !2 = !DISubroutineType(types: !{null, !0, !1}) ; void (int, char)
4028 ``DIDerivedType`` nodes represent types derived from other types, such as
4031 .. code-block:: text
4033 !0 = !DIBasicType(name: "unsigned char", size: 8, align: 8,
4034 encoding: DW_ATE_unsigned_char)
4035 !1 = !DIDerivedType(tag: DW_TAG_pointer_type, baseType: !0, size: 32,
4038 The following ``tag:`` values are valid:
4040 .. code-block:: text
4043 DW_TAG_pointer_type = 15
4044 DW_TAG_reference_type = 16
4046 DW_TAG_inheritance = 28
4047 DW_TAG_ptr_to_member_type = 31
4048 DW_TAG_const_type = 38
4050 DW_TAG_volatile_type = 53
4051 DW_TAG_restrict_type = 55
4053 .. _DIDerivedTypeMember:
4055 ``DW_TAG_member`` is used to define a member of a :ref:`composite type
4056 <DICompositeType>`. The type of the member is the ``baseType:``. The
4057 ``offset:`` is the member's bit offset. If the composite type has an ODR
4058 ``identifier:`` and does not set ``flags: DIFwdDecl``, then the member is
4059 uniqued based only on its ``name:`` and ``scope:``.
4061 ``DW_TAG_inheritance`` and ``DW_TAG_friend`` are used in the ``elements:``
4062 field of :ref:`composite types <DICompositeType>` to describe parents and
4065 ``DW_TAG_typedef`` is used to provide a name for the ``baseType:``.
4067 ``DW_TAG_pointer_type``, ``DW_TAG_reference_type``, ``DW_TAG_const_type``,
4068 ``DW_TAG_volatile_type`` and ``DW_TAG_restrict_type`` are used to qualify the
4071 Note that the ``void *`` type is expressed as a type derived from NULL.
4073 .. _DICompositeType:
4078 ``DICompositeType`` nodes represent types composed of other types, like
4079 structures and unions. ``elements:`` points to a tuple of the composed types.
4081 If the source language supports ODR, the ``identifier:`` field gives the unique
4082 identifier used for type merging between modules. When specified,
4083 :ref:`subprogram declarations <DISubprogramDeclaration>` and :ref:`member
4084 derived types <DIDerivedTypeMember>` that reference the ODR-type in their
4085 ``scope:`` change uniquing rules.
4087 For a given ``identifier:``, there should only be a single composite type that
4088 does not have ``flags: DIFlagFwdDecl`` set. LLVM tools that link modules
4089 together will unique such definitions at parse time via the ``identifier:``
4090 field, even if the nodes are ``distinct``.
4092 .. code-block:: text
4094 !0 = !DIEnumerator(name: "SixKind", value: 7)
4095 !1 = !DIEnumerator(name: "SevenKind", value: 7)
4096 !2 = !DIEnumerator(name: "NegEightKind", value: -8)
4097 !3 = !DICompositeType(tag: DW_TAG_enumeration_type, name: "Enum", file: !12,
4098 line: 2, size: 32, align: 32, identifier: "_M4Enum",
4099 elements: !{!0, !1, !2})
4101 The following ``tag:`` values are valid:
4103 .. code-block:: text
4105 DW_TAG_array_type = 1
4106 DW_TAG_class_type = 2
4107 DW_TAG_enumeration_type = 4
4108 DW_TAG_structure_type = 19
4109 DW_TAG_union_type = 23
4111 For ``DW_TAG_array_type``, the ``elements:`` should be :ref:`subrange
4112 descriptors <DISubrange>`, each representing the range of subscripts at that
4113 level of indexing. The ``DIFlagVector`` flag to ``flags:`` indicates that an
4114 array type is a native packed vector.
4116 For ``DW_TAG_enumeration_type``, the ``elements:`` should be :ref:`enumerator
4117 descriptors <DIEnumerator>`, each representing the definition of an enumeration
4118 value for the set. All enumeration type descriptors are collected in the
4119 ``enums:`` field of the :ref:`compile unit <DICompileUnit>`.
4121 For ``DW_TAG_structure_type``, ``DW_TAG_class_type``, and
4122 ``DW_TAG_union_type``, the ``elements:`` should be :ref:`derived types
4123 <DIDerivedType>` with ``tag: DW_TAG_member``, ``tag: DW_TAG_inheritance``, or
4124 ``tag: DW_TAG_friend``; or :ref:`subprograms <DISubprogram>` with
4125 ``isDefinition: false``.
4132 ``DISubrange`` nodes are the elements for ``DW_TAG_array_type`` variants of
4133 :ref:`DICompositeType`. ``count: -1`` indicates an empty array.
4135 .. code-block:: llvm
4137 !0 = !DISubrange(count: 5, lowerBound: 0) ; array counting from 0
4138 !1 = !DISubrange(count: 5, lowerBound: 1) ; array counting from 1
4139 !2 = !DISubrange(count: -1) ; empty array.
4146 ``DIEnumerator`` nodes are the elements for ``DW_TAG_enumeration_type``
4147 variants of :ref:`DICompositeType`.
4149 .. code-block:: llvm
4151 !0 = !DIEnumerator(name: "SixKind", value: 7)
4152 !1 = !DIEnumerator(name: "SevenKind", value: 7)
4153 !2 = !DIEnumerator(name: "NegEightKind", value: -8)
4155 DITemplateTypeParameter
4156 """""""""""""""""""""""
4158 ``DITemplateTypeParameter`` nodes represent type parameters to generic source
4159 language constructs. They are used (optionally) in :ref:`DICompositeType` and
4160 :ref:`DISubprogram` ``templateParams:`` fields.
4162 .. code-block:: llvm
4164 !0 = !DITemplateTypeParameter(name: "Ty", type: !1)
4166 DITemplateValueParameter
4167 """"""""""""""""""""""""
4169 ``DITemplateValueParameter`` nodes represent value parameters to generic source
4170 language constructs. ``tag:`` defaults to ``DW_TAG_template_value_parameter``,
4171 but if specified can also be set to ``DW_TAG_GNU_template_template_param`` or
4172 ``DW_TAG_GNU_template_param_pack``. They are used (optionally) in
4173 :ref:`DICompositeType` and :ref:`DISubprogram` ``templateParams:`` fields.
4175 .. code-block:: llvm
4177 !0 = !DITemplateValueParameter(name: "Ty", type: !1, value: i32 7)
4182 ``DINamespace`` nodes represent namespaces in the source language.
4184 .. code-block:: llvm
4186 !0 = !DINamespace(name: "myawesomeproject", scope: !1, file: !2, line: 7)
4191 ``DIGlobalVariable`` nodes represent global variables in the source language.
4193 .. code-block:: llvm
4195 !0 = !DIGlobalVariable(name: "foo", linkageName: "foo", scope: !1,
4196 file: !2, line: 7, type: !3, isLocal: true,
4197 isDefinition: false, variable: i32* @foo,
4200 All global variables should be referenced by the `globals:` field of a
4201 :ref:`compile unit <DICompileUnit>`.
4208 ``DISubprogram`` nodes represent functions from the source language. A
4209 ``DISubprogram`` may be attached to a function definition using ``!dbg``
4210 metadata. The ``variables:`` field points at :ref:`variables <DILocalVariable>`
4211 that must be retained, even if their IR counterparts are optimized out of
4212 the IR. The ``type:`` field must point at an :ref:`DISubroutineType`.
4214 .. _DISubprogramDeclaration:
4216 When ``isDefinition: false``, subprograms describe a declaration in the type
4217 tree as opposed to a definition of a function. If the scope is a composite
4218 type with an ODR ``identifier:`` and that does not set ``flags: DIFwdDecl``,
4219 then the subprogram declaration is uniqued based only on its ``linkageName:``
4222 .. code-block:: text
4224 define void @_Z3foov() !dbg !0 {
4228 !0 = distinct !DISubprogram(name: "foo", linkageName: "_Zfoov", scope: !1,
4229 file: !2, line: 7, type: !3, isLocal: true,
4230 isDefinition: true, scopeLine: 8,
4232 virtuality: DW_VIRTUALITY_pure_virtual,
4233 virtualIndex: 10, flags: DIFlagPrototyped,
4234 isOptimized: true, templateParams: !5,
4235 declaration: !6, variables: !7)
4242 ``DILexicalBlock`` nodes describe nested blocks within a :ref:`subprogram
4243 <DISubprogram>`. The line number and column numbers are used to distinguish
4244 two lexical blocks at same depth. They are valid targets for ``scope:``
4247 .. code-block:: text
4249 !0 = distinct !DILexicalBlock(scope: !1, file: !2, line: 7, column: 35)
4251 Usually lexical blocks are ``distinct`` to prevent node merging based on
4254 .. _DILexicalBlockFile:
4259 ``DILexicalBlockFile`` nodes are used to discriminate between sections of a
4260 :ref:`lexical block <DILexicalBlock>`. The ``file:`` field can be changed to
4261 indicate textual inclusion, or the ``discriminator:`` field can be used to
4262 discriminate between control flow within a single block in the source language.
4264 .. code-block:: llvm
4266 !0 = !DILexicalBlock(scope: !3, file: !4, line: 7, column: 35)
4267 !1 = !DILexicalBlockFile(scope: !0, file: !4, discriminator: 0)
4268 !2 = !DILexicalBlockFile(scope: !0, file: !4, discriminator: 1)
4275 ``DILocation`` nodes represent source debug locations. The ``scope:`` field is
4276 mandatory, and points at an :ref:`DILexicalBlockFile`, an
4277 :ref:`DILexicalBlock`, or an :ref:`DISubprogram`.
4279 .. code-block:: llvm
4281 !0 = !DILocation(line: 2900, column: 42, scope: !1, inlinedAt: !2)
4283 .. _DILocalVariable:
4288 ``DILocalVariable`` nodes represent local variables in the source language. If
4289 the ``arg:`` field is set to non-zero, then this variable is a subprogram
4290 parameter, and it will be included in the ``variables:`` field of its
4291 :ref:`DISubprogram`.
4293 .. code-block:: text
4295 !0 = !DILocalVariable(name: "this", arg: 1, scope: !3, file: !2, line: 7,
4296 type: !3, flags: DIFlagArtificial)
4297 !1 = !DILocalVariable(name: "x", arg: 2, scope: !4, file: !2, line: 7,
4299 !2 = !DILocalVariable(name: "y", scope: !5, file: !2, line: 7, type: !3)
4304 ``DIExpression`` nodes represent DWARF expression sequences. They are used in
4305 :ref:`debug intrinsics<dbg_intrinsics>` (such as ``llvm.dbg.declare``) to
4306 describe how the referenced LLVM variable relates to the source language
4309 The current supported vocabulary is limited:
4311 - ``DW_OP_deref`` dereferences the working expression.
4312 - ``DW_OP_plus, 93`` adds ``93`` to the working expression.
4313 - ``DW_OP_bit_piece, 16, 8`` specifies the offset and size (``16`` and ``8``
4314 here, respectively) of the variable piece from the working expression.
4316 .. code-block:: text
4318 !0 = !DIExpression(DW_OP_deref)
4319 !1 = !DIExpression(DW_OP_plus, 3)
4320 !2 = !DIExpression(DW_OP_bit_piece, 3, 7)
4321 !3 = !DIExpression(DW_OP_deref, DW_OP_plus, 3, DW_OP_bit_piece, 3, 7)
4326 ``DIObjCProperty`` nodes represent Objective-C property nodes.
4328 .. code-block:: llvm
4330 !3 = !DIObjCProperty(name: "foo", file: !1, line: 7, setter: "setFoo",
4331 getter: "getFoo", attributes: 7, type: !2)
4336 ``DIImportedEntity`` nodes represent entities (such as modules) imported into a
4339 .. code-block:: text
4341 !2 = !DIImportedEntity(tag: DW_TAG_imported_module, name: "foo", scope: !0,
4342 entity: !1, line: 7)
4347 ``DIMacro`` nodes represent definition or undefinition of a macro identifiers.
4348 The ``name:`` field is the macro identifier, followed by macro parameters when
4349 defining a function-like macro, and the ``value`` field is the token-string
4350 used to expand the macro identifier.
4352 .. code-block:: text
4354 !2 = !DIMacro(macinfo: DW_MACINFO_define, line: 7, name: "foo(x)",
4356 !3 = !DIMacro(macinfo: DW_MACINFO_undef, line: 30, name: "foo")
4361 ``DIMacroFile`` nodes represent inclusion of source files.
4362 The ``nodes:`` field is a list of ``DIMacro`` and ``DIMacroFile`` nodes that
4363 appear in the included source file.
4365 .. code-block:: text
4367 !2 = !DIMacroFile(macinfo: DW_MACINFO_start_file, line: 7, file: !2,
4373 In LLVM IR, memory does not have types, so LLVM's own type system is not
4374 suitable for doing TBAA. Instead, metadata is added to the IR to
4375 describe a type system of a higher level language. This can be used to
4376 implement typical C/C++ TBAA, but it can also be used to implement
4377 custom alias analysis behavior for other languages.
4379 The current metadata format is very simple. TBAA metadata nodes have up
4380 to three fields, e.g.:
4382 .. code-block:: llvm
4384 !0 = !{ !"an example type tree" }
4385 !1 = !{ !"int", !0 }
4386 !2 = !{ !"float", !0 }
4387 !3 = !{ !"const float", !2, i64 1 }
4389 The first field is an identity field. It can be any value, usually a
4390 metadata string, which uniquely identifies the type. The most important
4391 name in the tree is the name of the root node. Two trees with different
4392 root node names are entirely disjoint, even if they have leaves with
4395 The second field identifies the type's parent node in the tree, or is
4396 null or omitted for a root node. A type is considered to alias all of
4397 its descendants and all of its ancestors in the tree. Also, a type is
4398 considered to alias all types in other trees, so that bitcode produced
4399 from multiple front-ends is handled conservatively.
4401 If the third field is present, it's an integer which if equal to 1
4402 indicates that the type is "constant" (meaning
4403 ``pointsToConstantMemory`` should return true; see `other useful
4404 AliasAnalysis methods <AliasAnalysis.html#OtherItfs>`_).
4406 '``tbaa.struct``' Metadata
4407 ^^^^^^^^^^^^^^^^^^^^^^^^^^
4409 The :ref:`llvm.memcpy <int_memcpy>` is often used to implement
4410 aggregate assignment operations in C and similar languages, however it
4411 is defined to copy a contiguous region of memory, which is more than
4412 strictly necessary for aggregate types which contain holes due to
4413 padding. Also, it doesn't contain any TBAA information about the fields
4416 ``!tbaa.struct`` metadata can describe which memory subregions in a
4417 memcpy are padding and what the TBAA tags of the struct are.
4419 The current metadata format is very simple. ``!tbaa.struct`` metadata
4420 nodes are a list of operands which are in conceptual groups of three.
4421 For each group of three, the first operand gives the byte offset of a
4422 field in bytes, the second gives its size in bytes, and the third gives
4425 .. code-block:: llvm
4427 !4 = !{ i64 0, i64 4, !1, i64 8, i64 4, !2 }
4429 This describes a struct with two fields. The first is at offset 0 bytes
4430 with size 4 bytes, and has tbaa tag !1. The second is at offset 8 bytes
4431 and has size 4 bytes and has tbaa tag !2.
4433 Note that the fields need not be contiguous. In this example, there is a
4434 4 byte gap between the two fields. This gap represents padding which
4435 does not carry useful data and need not be preserved.
4437 '``noalias``' and '``alias.scope``' Metadata
4438 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4440 ``noalias`` and ``alias.scope`` metadata provide the ability to specify generic
4441 noalias memory-access sets. This means that some collection of memory access
4442 instructions (loads, stores, memory-accessing calls, etc.) that carry
4443 ``noalias`` metadata can specifically be specified not to alias with some other
4444 collection of memory access instructions that carry ``alias.scope`` metadata.
4445 Each type of metadata specifies a list of scopes where each scope has an id and
4448 When evaluating an aliasing query, if for some domain, the set
4449 of scopes with that domain in one instruction's ``alias.scope`` list is a
4450 subset of (or equal to) the set of scopes for that domain in another
4451 instruction's ``noalias`` list, then the two memory accesses are assumed not to
4454 Because scopes in one domain don't affect scopes in other domains, separate
4455 domains can be used to compose multiple independent noalias sets. This is
4456 used for example during inlining. As the noalias function parameters are
4457 turned into noalias scope metadata, a new domain is used every time the
4458 function is inlined.
4460 The metadata identifying each domain is itself a list containing one or two
4461 entries. The first entry is the name of the domain. Note that if the name is a
4462 string then it can be combined across functions and translation units. A
4463 self-reference can be used to create globally unique domain names. A
4464 descriptive string may optionally be provided as a second list entry.
4466 The metadata identifying each scope is also itself a list containing two or
4467 three entries. The first entry is the name of the scope. Note that if the name
4468 is a string then it can be combined across functions and translation units. A
4469 self-reference can be used to create globally unique scope names. A metadata
4470 reference to the scope's domain is the second entry. A descriptive string may
4471 optionally be provided as a third list entry.
4475 .. code-block:: llvm
4477 ; Two scope domains:
4481 ; Some scopes in these domains:
4487 !5 = !{!4} ; A list containing only scope !4
4491 ; These two instructions don't alias:
4492 %0 = load float, float* %c, align 4, !alias.scope !5
4493 store float %0, float* %arrayidx.i, align 4, !noalias !5
4495 ; These two instructions also don't alias (for domain !1, the set of scopes
4496 ; in the !alias.scope equals that in the !noalias list):
4497 %2 = load float, float* %c, align 4, !alias.scope !5
4498 store float %2, float* %arrayidx.i2, align 4, !noalias !6
4500 ; These two instructions may alias (for domain !0, the set of scopes in
4501 ; the !noalias list is not a superset of, or equal to, the scopes in the
4502 ; !alias.scope list):
4503 %2 = load float, float* %c, align 4, !alias.scope !6
4504 store float %0, float* %arrayidx.i, align 4, !noalias !7
4506 '``fpmath``' Metadata
4507 ^^^^^^^^^^^^^^^^^^^^^
4509 ``fpmath`` metadata may be attached to any instruction of floating point
4510 type. It can be used to express the maximum acceptable error in the
4511 result of that instruction, in ULPs, thus potentially allowing the
4512 compiler to use a more efficient but less accurate method of computing
4513 it. ULP is defined as follows:
4515 If ``x`` is a real number that lies between two finite consecutive
4516 floating-point numbers ``a`` and ``b``, without being equal to one
4517 of them, then ``ulp(x) = |b - a|``, otherwise ``ulp(x)`` is the
4518 distance between the two non-equal finite floating-point numbers
4519 nearest ``x``. Moreover, ``ulp(NaN)`` is ``NaN``.
4521 The metadata node shall consist of a single positive float type number
4522 representing the maximum relative error, for example:
4524 .. code-block:: llvm
4526 !0 = !{ float 2.5 } ; maximum acceptable inaccuracy is 2.5 ULPs
4530 '``range``' Metadata
4531 ^^^^^^^^^^^^^^^^^^^^
4533 ``range`` metadata may be attached only to ``load``, ``call`` and ``invoke`` of
4534 integer types. It expresses the possible ranges the loaded value or the value
4535 returned by the called function at this call site is in. The ranges are
4536 represented with a flattened list of integers. The loaded value or the value
4537 returned is known to be in the union of the ranges defined by each consecutive
4538 pair. Each pair has the following properties:
4540 - The type must match the type loaded by the instruction.
4541 - The pair ``a,b`` represents the range ``[a,b)``.
4542 - Both ``a`` and ``b`` are constants.
4543 - The range is allowed to wrap.
4544 - The range should not represent the full or empty set. That is,
4547 In addition, the pairs must be in signed order of the lower bound and
4548 they must be non-contiguous.
4552 .. code-block:: llvm
4554 %a = load i8, i8* %x, align 1, !range !0 ; Can only be 0 or 1
4555 %b = load i8, i8* %y, align 1, !range !1 ; Can only be 255 (-1), 0 or 1
4556 %c = call i8 @foo(), !range !2 ; Can only be 0, 1, 3, 4 or 5
4557 %d = invoke i8 @bar() to label %cont
4558 unwind label %lpad, !range !3 ; Can only be -2, -1, 3, 4 or 5
4560 !0 = !{ i8 0, i8 2 }
4561 !1 = !{ i8 255, i8 2 }
4562 !2 = !{ i8 0, i8 2, i8 3, i8 6 }
4563 !3 = !{ i8 -2, i8 0, i8 3, i8 6 }
4565 '``unpredictable``' Metadata
4566 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4568 ``unpredictable`` metadata may be attached to any branch or switch
4569 instruction. It can be used to express the unpredictability of control
4570 flow. Similar to the llvm.expect intrinsic, it may be used to alter
4571 optimizations related to compare and branch instructions. The metadata
4572 is treated as a boolean value; if it exists, it signals that the branch
4573 or switch that it is attached to is completely unpredictable.
4578 It is sometimes useful to attach information to loop constructs. Currently,
4579 loop metadata is implemented as metadata attached to the branch instruction
4580 in the loop latch block. This type of metadata refer to a metadata node that is
4581 guaranteed to be separate for each loop. The loop identifier metadata is
4582 specified with the name ``llvm.loop``.
4584 The loop identifier metadata is implemented using a metadata that refers to
4585 itself to avoid merging it with any other identifier metadata, e.g.,
4586 during module linkage or function inlining. That is, each loop should refer
4587 to their own identification metadata even if they reside in separate functions.
4588 The following example contains loop identifier metadata for two separate loop
4591 .. code-block:: llvm
4596 The loop identifier metadata can be used to specify additional
4597 per-loop metadata. Any operands after the first operand can be treated
4598 as user-defined metadata. For example the ``llvm.loop.unroll.count``
4599 suggests an unroll factor to the loop unroller:
4601 .. code-block:: llvm
4603 br i1 %exitcond, label %._crit_edge, label %.lr.ph, !llvm.loop !0
4606 !1 = !{!"llvm.loop.unroll.count", i32 4}
4608 '``llvm.loop.vectorize``' and '``llvm.loop.interleave``'
4609 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4611 Metadata prefixed with ``llvm.loop.vectorize`` or ``llvm.loop.interleave`` are
4612 used to control per-loop vectorization and interleaving parameters such as
4613 vectorization width and interleave count. These metadata should be used in
4614 conjunction with ``llvm.loop`` loop identification metadata. The
4615 ``llvm.loop.vectorize`` and ``llvm.loop.interleave`` metadata are only
4616 optimization hints and the optimizer will only interleave and vectorize loops if
4617 it believes it is safe to do so. The ``llvm.mem.parallel_loop_access`` metadata
4618 which contains information about loop-carried memory dependencies can be helpful
4619 in determining the safety of these transformations.
4621 '``llvm.loop.interleave.count``' Metadata
4622 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4624 This metadata suggests an interleave count to the loop interleaver.
4625 The first operand is the string ``llvm.loop.interleave.count`` and the
4626 second operand is an integer specifying the interleave count. For
4629 .. code-block:: llvm
4631 !0 = !{!"llvm.loop.interleave.count", i32 4}
4633 Note that setting ``llvm.loop.interleave.count`` to 1 disables interleaving
4634 multiple iterations of the loop. If ``llvm.loop.interleave.count`` is set to 0
4635 then the interleave count will be determined automatically.
4637 '``llvm.loop.vectorize.enable``' Metadata
4638 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4640 This metadata selectively enables or disables vectorization for the loop. The
4641 first operand is the string ``llvm.loop.vectorize.enable`` and the second operand
4642 is a bit. If the bit operand value is 1 vectorization is enabled. A value of
4643 0 disables vectorization:
4645 .. code-block:: llvm
4647 !0 = !{!"llvm.loop.vectorize.enable", i1 0}
4648 !1 = !{!"llvm.loop.vectorize.enable", i1 1}
4650 '``llvm.loop.vectorize.width``' Metadata
4651 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4653 This metadata sets the target width of the vectorizer. The first
4654 operand is the string ``llvm.loop.vectorize.width`` and the second
4655 operand is an integer specifying the width. For example:
4657 .. code-block:: llvm
4659 !0 = !{!"llvm.loop.vectorize.width", i32 4}
4661 Note that setting ``llvm.loop.vectorize.width`` to 1 disables
4662 vectorization of the loop. If ``llvm.loop.vectorize.width`` is set to
4663 0 or if the loop does not have this metadata the width will be
4664 determined automatically.
4666 '``llvm.loop.unroll``'
4667 ^^^^^^^^^^^^^^^^^^^^^^
4669 Metadata prefixed with ``llvm.loop.unroll`` are loop unrolling
4670 optimization hints such as the unroll factor. ``llvm.loop.unroll``
4671 metadata should be used in conjunction with ``llvm.loop`` loop
4672 identification metadata. The ``llvm.loop.unroll`` metadata are only
4673 optimization hints and the unrolling will only be performed if the
4674 optimizer believes it is safe to do so.
4676 '``llvm.loop.unroll.count``' Metadata
4677 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4679 This metadata suggests an unroll factor to the loop unroller. The
4680 first operand is the string ``llvm.loop.unroll.count`` and the second
4681 operand is a positive integer specifying the unroll factor. For
4684 .. code-block:: llvm
4686 !0 = !{!"llvm.loop.unroll.count", i32 4}
4688 If the trip count of the loop is less than the unroll count the loop
4689 will be partially unrolled.
4691 '``llvm.loop.unroll.disable``' Metadata
4692 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4694 This metadata disables loop unrolling. The metadata has a single operand
4695 which is the string ``llvm.loop.unroll.disable``. For example:
4697 .. code-block:: llvm
4699 !0 = !{!"llvm.loop.unroll.disable"}
4701 '``llvm.loop.unroll.runtime.disable``' Metadata
4702 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4704 This metadata disables runtime loop unrolling. The metadata has a single
4705 operand which is the string ``llvm.loop.unroll.runtime.disable``. For example:
4707 .. code-block:: llvm
4709 !0 = !{!"llvm.loop.unroll.runtime.disable"}
4711 '``llvm.loop.unroll.enable``' Metadata
4712 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4714 This metadata suggests that the loop should be fully unrolled if the trip count
4715 is known at compile time and partially unrolled if the trip count is not known
4716 at compile time. The metadata has a single operand which is the string
4717 ``llvm.loop.unroll.enable``. For example:
4719 .. code-block:: llvm
4721 !0 = !{!"llvm.loop.unroll.enable"}
4723 '``llvm.loop.unroll.full``' Metadata
4724 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4726 This metadata suggests that the loop should be unrolled fully. The
4727 metadata has a single operand which is the string ``llvm.loop.unroll.full``.
4730 .. code-block:: llvm
4732 !0 = !{!"llvm.loop.unroll.full"}
4734 '``llvm.loop.licm_versioning.disable``' Metadata
4735 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4737 This metadata indicates that the loop should not be versioned for the purpose
4738 of enabling loop-invariant code motion (LICM). The metadata has a single operand
4739 which is the string ``llvm.loop.licm_versioning.disable``. For example:
4741 .. code-block:: llvm
4743 !0 = !{!"llvm.loop.licm_versioning.disable"}
4745 '``llvm.loop.distribute.enable``' Metadata
4746 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4748 Loop distribution allows splitting a loop into multiple loops. Currently,
4749 this is only performed if the entire loop cannot be vectorized due to unsafe
4750 memory dependencies. The transformation will atempt to isolate the unsafe
4751 dependencies into their own loop.
4753 This metadata can be used to selectively enable or disable distribution of the
4754 loop. The first operand is the string ``llvm.loop.distribute.enable`` and the
4755 second operand is a bit. If the bit operand value is 1 distribution is
4756 enabled. A value of 0 disables distribution:
4758 .. code-block:: llvm
4760 !0 = !{!"llvm.loop.distribute.enable", i1 0}
4761 !1 = !{!"llvm.loop.distribute.enable", i1 1}
4763 This metadata should be used in conjunction with ``llvm.loop`` loop
4764 identification metadata.
4769 Metadata types used to annotate memory accesses with information helpful
4770 for optimizations are prefixed with ``llvm.mem``.
4772 '``llvm.mem.parallel_loop_access``' Metadata
4773 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4775 The ``llvm.mem.parallel_loop_access`` metadata refers to a loop identifier,
4776 or metadata containing a list of loop identifiers for nested loops.
4777 The metadata is attached to memory accessing instructions and denotes that
4778 no loop carried memory dependence exist between it and other instructions denoted
4779 with the same loop identifier. The metadata on memory reads also implies that
4780 if conversion (i.e. speculative execution within a loop iteration) is safe.
4782 Precisely, given two instructions ``m1`` and ``m2`` that both have the
4783 ``llvm.mem.parallel_loop_access`` metadata, with ``L1`` and ``L2`` being the
4784 set of loops associated with that metadata, respectively, then there is no loop
4785 carried dependence between ``m1`` and ``m2`` for loops in both ``L1`` and
4788 As a special case, if all memory accessing instructions in a loop have
4789 ``llvm.mem.parallel_loop_access`` metadata that refers to that loop, then the
4790 loop has no loop carried memory dependences and is considered to be a parallel
4793 Note that if not all memory access instructions have such metadata referring to
4794 the loop, then the loop is considered not being trivially parallel. Additional
4795 memory dependence analysis is required to make that determination. As a fail
4796 safe mechanism, this causes loops that were originally parallel to be considered
4797 sequential (if optimization passes that are unaware of the parallel semantics
4798 insert new memory instructions into the loop body).
4800 Example of a loop that is considered parallel due to its correct use of
4801 both ``llvm.loop`` and ``llvm.mem.parallel_loop_access``
4802 metadata types that refer to the same loop identifier metadata.
4804 .. code-block:: llvm
4808 %val0 = load i32, i32* %arrayidx, !llvm.mem.parallel_loop_access !0
4810 store i32 %val0, i32* %arrayidx1, !llvm.mem.parallel_loop_access !0
4812 br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !0
4818 It is also possible to have nested parallel loops. In that case the
4819 memory accesses refer to a list of loop identifier metadata nodes instead of
4820 the loop identifier metadata node directly:
4822 .. code-block:: llvm
4826 %val1 = load i32, i32* %arrayidx3, !llvm.mem.parallel_loop_access !2
4828 br label %inner.for.body
4832 %val0 = load i32, i32* %arrayidx1, !llvm.mem.parallel_loop_access !0
4834 store i32 %val0, i32* %arrayidx2, !llvm.mem.parallel_loop_access !0
4836 br i1 %exitcond, label %inner.for.end, label %inner.for.body, !llvm.loop !1
4840 store i32 %val1, i32* %arrayidx4, !llvm.mem.parallel_loop_access !2
4842 br i1 %exitcond, label %outer.for.end, label %outer.for.body, !llvm.loop !2
4844 outer.for.end: ; preds = %for.body
4846 !0 = !{!1, !2} ; a list of loop identifiers
4847 !1 = !{!1} ; an identifier for the inner loop
4848 !2 = !{!2} ; an identifier for the outer loop
4850 '``invariant.group``' Metadata
4851 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4853 The ``invariant.group`` metadata may be attached to ``load``/``store`` instructions.
4854 The existence of the ``invariant.group`` metadata on the instruction tells
4855 the optimizer that every ``load`` and ``store`` to the same pointer operand
4856 within the same invariant group can be assumed to load or store the same
4857 value (but see the ``llvm.invariant.group.barrier`` intrinsic which affects
4858 when two pointers are considered the same).
4862 .. code-block:: llvm
4864 @unknownPtr = external global i8
4867 store i8 42, i8* %ptr, !invariant.group !0
4868 call void @foo(i8* %ptr)
4870 %a = load i8, i8* %ptr, !invariant.group !0 ; Can assume that value under %ptr didn't change
4871 call void @foo(i8* %ptr)
4872 %b = load i8, i8* %ptr, !invariant.group !1 ; Can't assume anything, because group changed
4874 %newPtr = call i8* @getPointer(i8* %ptr)
4875 %c = load i8, i8* %newPtr, !invariant.group !0 ; Can't assume anything, because we only have information about %ptr
4877 %unknownValue = load i8, i8* @unknownPtr
4878 store i8 %unknownValue, i8* %ptr, !invariant.group !0 ; Can assume that %unknownValue == 42
4880 call void @foo(i8* %ptr)
4881 %newPtr2 = call i8* @llvm.invariant.group.barrier(i8* %ptr)
4882 %d = load i8, i8* %newPtr2, !invariant.group !0 ; Can't step through invariant.group.barrier to get value of %ptr
4885 declare void @foo(i8*)
4886 declare i8* @getPointer(i8*)
4887 declare i8* @llvm.invariant.group.barrier(i8*)
4889 !0 = !{!"magic ptr"}
4890 !1 = !{!"other ptr"}
4894 Module Flags Metadata
4895 =====================
4897 Information about the module as a whole is difficult to convey to LLVM's
4898 subsystems. The LLVM IR isn't sufficient to transmit this information.
4899 The ``llvm.module.flags`` named metadata exists in order to facilitate
4900 this. These flags are in the form of key / value pairs --- much like a
4901 dictionary --- making it easy for any subsystem who cares about a flag to
4904 The ``llvm.module.flags`` metadata contains a list of metadata triplets.
4905 Each triplet has the following form:
4907 - The first element is a *behavior* flag, which specifies the behavior
4908 when two (or more) modules are merged together, and it encounters two
4909 (or more) metadata with the same ID. The supported behaviors are
4911 - The second element is a metadata string that is a unique ID for the
4912 metadata. Each module may only have one flag entry for each unique ID (not
4913 including entries with the **Require** behavior).
4914 - The third element is the value of the flag.
4916 When two (or more) modules are merged together, the resulting
4917 ``llvm.module.flags`` metadata is the union of the modules' flags. That is, for
4918 each unique metadata ID string, there will be exactly one entry in the merged
4919 modules ``llvm.module.flags`` metadata table, and the value for that entry will
4920 be determined by the merge behavior flag, as described below. The only exception
4921 is that entries with the *Require* behavior are always preserved.
4923 The following behaviors are supported:
4934 Emits an error if two values disagree, otherwise the resulting value
4935 is that of the operands.
4939 Emits a warning if two values disagree. The result value will be the
4940 operand for the flag from the first module being linked.
4944 Adds a requirement that another module flag be present and have a
4945 specified value after linking is performed. The value must be a
4946 metadata pair, where the first element of the pair is the ID of the
4947 module flag to be restricted, and the second element of the pair is
4948 the value the module flag should be restricted to. This behavior can
4949 be used to restrict the allowable results (via triggering of an
4950 error) of linking IDs with the **Override** behavior.
4954 Uses the specified value, regardless of the behavior or value of the
4955 other module. If both modules specify **Override**, but the values
4956 differ, an error will be emitted.
4960 Appends the two values, which are required to be metadata nodes.
4964 Appends the two values, which are required to be metadata
4965 nodes. However, duplicate entries in the second list are dropped
4966 during the append operation.
4968 It is an error for a particular unique flag ID to have multiple behaviors,
4969 except in the case of **Require** (which adds restrictions on another metadata
4970 value) or **Override**.
4972 An example of module flags:
4974 .. code-block:: llvm
4976 !0 = !{ i32 1, !"foo", i32 1 }
4977 !1 = !{ i32 4, !"bar", i32 37 }
4978 !2 = !{ i32 2, !"qux", i32 42 }
4979 !3 = !{ i32 3, !"qux",
4984 !llvm.module.flags = !{ !0, !1, !2, !3 }
4986 - Metadata ``!0`` has the ID ``!"foo"`` and the value '1'. The behavior
4987 if two or more ``!"foo"`` flags are seen is to emit an error if their
4988 values are not equal.
4990 - Metadata ``!1`` has the ID ``!"bar"`` and the value '37'. The
4991 behavior if two or more ``!"bar"`` flags are seen is to use the value
4994 - Metadata ``!2`` has the ID ``!"qux"`` and the value '42'. The
4995 behavior if two or more ``!"qux"`` flags are seen is to emit a
4996 warning if their values are not equal.
4998 - Metadata ``!3`` has the ID ``!"qux"`` and the value:
5004 The behavior is to emit an error if the ``llvm.module.flags`` does not
5005 contain a flag with the ID ``!"foo"`` that has the value '1' after linking is
5008 Objective-C Garbage Collection Module Flags Metadata
5009 ----------------------------------------------------
5011 On the Mach-O platform, Objective-C stores metadata about garbage
5012 collection in a special section called "image info". The metadata
5013 consists of a version number and a bitmask specifying what types of
5014 garbage collection are supported (if any) by the file. If two or more
5015 modules are linked together their garbage collection metadata needs to
5016 be merged rather than appended together.
5018 The Objective-C garbage collection module flags metadata consists of the
5019 following key-value pairs:
5028 * - ``Objective-C Version``
5029 - **[Required]** --- The Objective-C ABI version. Valid values are 1 and 2.
5031 * - ``Objective-C Image Info Version``
5032 - **[Required]** --- The version of the image info section. Currently
5035 * - ``Objective-C Image Info Section``
5036 - **[Required]** --- The section to place the metadata. Valid values are
5037 ``"__OBJC, __image_info, regular"`` for Objective-C ABI version 1, and
5038 ``"__DATA,__objc_imageinfo, regular, no_dead_strip"`` for
5039 Objective-C ABI version 2.
5041 * - ``Objective-C Garbage Collection``
5042 - **[Required]** --- Specifies whether garbage collection is supported or
5043 not. Valid values are 0, for no garbage collection, and 2, for garbage
5044 collection supported.
5046 * - ``Objective-C GC Only``
5047 - **[Optional]** --- Specifies that only garbage collection is supported.
5048 If present, its value must be 6. This flag requires that the
5049 ``Objective-C Garbage Collection`` flag have the value 2.
5051 Some important flag interactions:
5053 - If a module with ``Objective-C Garbage Collection`` set to 0 is
5054 merged with a module with ``Objective-C Garbage Collection`` set to
5055 2, then the resulting module has the
5056 ``Objective-C Garbage Collection`` flag set to 0.
5057 - A module with ``Objective-C Garbage Collection`` set to 0 cannot be
5058 merged with a module with ``Objective-C GC Only`` set to 6.
5060 Automatic Linker Flags Module Flags Metadata
5061 --------------------------------------------
5063 Some targets support embedding flags to the linker inside individual object
5064 files. Typically this is used in conjunction with language extensions which
5065 allow source files to explicitly declare the libraries they depend on, and have
5066 these automatically be transmitted to the linker via object files.
5068 These flags are encoded in the IR using metadata in the module flags section,
5069 using the ``Linker Options`` key. The merge behavior for this flag is required
5070 to be ``AppendUnique``, and the value for the key is expected to be a metadata
5071 node which should be a list of other metadata nodes, each of which should be a
5072 list of metadata strings defining linker options.
5074 For example, the following metadata section specifies two separate sets of
5075 linker options, presumably to link against ``libz`` and the ``Cocoa``
5078 !0 = !{ i32 6, !"Linker Options",
5081 !{ !"-framework", !"Cocoa" } } }
5082 !llvm.module.flags = !{ !0 }
5084 The metadata encoding as lists of lists of options, as opposed to a collapsed
5085 list of options, is chosen so that the IR encoding can use multiple option
5086 strings to specify e.g., a single library, while still having that specifier be
5087 preserved as an atomic element that can be recognized by a target specific
5088 assembly writer or object file emitter.
5090 Each individual option is required to be either a valid option for the target's
5091 linker, or an option that is reserved by the target specific assembly writer or
5092 object file emitter. No other aspect of these options is defined by the IR.
5094 C type width Module Flags Metadata
5095 ----------------------------------
5097 The ARM backend emits a section into each generated object file describing the
5098 options that it was compiled with (in a compiler-independent way) to prevent
5099 linking incompatible objects, and to allow automatic library selection. Some
5100 of these options are not visible at the IR level, namely wchar_t width and enum
5103 To pass this information to the backend, these options are encoded in module
5104 flags metadata, using the following key-value pairs:
5114 - * 0 --- sizeof(wchar_t) == 4
5115 * 1 --- sizeof(wchar_t) == 2
5118 - * 0 --- Enums are at least as large as an ``int``.
5119 * 1 --- Enums are stored in the smallest integer type which can
5120 represent all of its values.
5122 For example, the following metadata section specifies that the module was
5123 compiled with a ``wchar_t`` width of 4 bytes, and the underlying type of an
5124 enum is the smallest type which can represent all of its values::
5126 !llvm.module.flags = !{!0, !1}
5127 !0 = !{i32 1, !"short_wchar", i32 1}
5128 !1 = !{i32 1, !"short_enum", i32 0}
5130 .. _intrinsicglobalvariables:
5132 Intrinsic Global Variables
5133 ==========================
5135 LLVM has a number of "magic" global variables that contain data that
5136 affect code generation or other IR semantics. These are documented here.
5137 All globals of this sort should have a section specified as
5138 "``llvm.metadata``". This section and all globals that start with
5139 "``llvm.``" are reserved for use by LLVM.
5143 The '``llvm.used``' Global Variable
5144 -----------------------------------
5146 The ``@llvm.used`` global is an array which has
5147 :ref:`appending linkage <linkage_appending>`. This array contains a list of
5148 pointers to named global variables, functions and aliases which may optionally
5149 have a pointer cast formed of bitcast or getelementptr. For example, a legal
5152 .. code-block:: llvm
5157 @llvm.used = appending global [2 x i8*] [
5159 i8* bitcast (i32* @Y to i8*)
5160 ], section "llvm.metadata"
5162 If a symbol appears in the ``@llvm.used`` list, then the compiler, assembler,
5163 and linker are required to treat the symbol as if there is a reference to the
5164 symbol that it cannot see (which is why they have to be named). For example, if
5165 a variable has internal linkage and no references other than that from the
5166 ``@llvm.used`` list, it cannot be deleted. This is commonly used to represent
5167 references from inline asms and other things the compiler cannot "see", and
5168 corresponds to "``attribute((used))``" in GNU C.
5170 On some targets, the code generator must emit a directive to the
5171 assembler or object file to prevent the assembler and linker from
5172 molesting the symbol.
5174 .. _gv_llvmcompilerused:
5176 The '``llvm.compiler.used``' Global Variable
5177 --------------------------------------------
5179 The ``@llvm.compiler.used`` directive is the same as the ``@llvm.used``
5180 directive, except that it only prevents the compiler from touching the
5181 symbol. On targets that support it, this allows an intelligent linker to
5182 optimize references to the symbol without being impeded as it would be
5185 This is a rare construct that should only be used in rare circumstances,
5186 and should not be exposed to source languages.
5188 .. _gv_llvmglobalctors:
5190 The '``llvm.global_ctors``' Global Variable
5191 -------------------------------------------
5193 .. code-block:: llvm
5195 %0 = type { i32, void ()*, i8* }
5196 @llvm.global_ctors = appending global [1 x %0] [%0 { i32 65535, void ()* @ctor, i8* @data }]
5198 The ``@llvm.global_ctors`` array contains a list of constructor
5199 functions, priorities, and an optional associated global or function.
5200 The functions referenced by this array will be called in ascending order
5201 of priority (i.e. lowest first) when the module is loaded. The order of
5202 functions with the same priority is not defined.
5204 If the third field is present, non-null, and points to a global variable
5205 or function, the initializer function will only run if the associated
5206 data from the current module is not discarded.
5208 .. _llvmglobaldtors:
5210 The '``llvm.global_dtors``' Global Variable
5211 -------------------------------------------
5213 .. code-block:: llvm
5215 %0 = type { i32, void ()*, i8* }
5216 @llvm.global_dtors = appending global [1 x %0] [%0 { i32 65535, void ()* @dtor, i8* @data }]
5218 The ``@llvm.global_dtors`` array contains a list of destructor
5219 functions, priorities, and an optional associated global or function.
5220 The functions referenced by this array will be called in descending
5221 order of priority (i.e. highest first) when the module is unloaded. The
5222 order of functions with the same priority is not defined.
5224 If the third field is present, non-null, and points to a global variable
5225 or function, the destructor function will only run if the associated
5226 data from the current module is not discarded.
5228 Instruction Reference
5229 =====================
5231 The LLVM instruction set consists of several different classifications
5232 of instructions: :ref:`terminator instructions <terminators>`, :ref:`binary
5233 instructions <binaryops>`, :ref:`bitwise binary
5234 instructions <bitwiseops>`, :ref:`memory instructions <memoryops>`, and
5235 :ref:`other instructions <otherops>`.
5239 Terminator Instructions
5240 -----------------------
5242 As mentioned :ref:`previously <functionstructure>`, every basic block in a
5243 program ends with a "Terminator" instruction, which indicates which
5244 block should be executed after the current block is finished. These
5245 terminator instructions typically yield a '``void``' value: they produce
5246 control flow, not values (the one exception being the
5247 ':ref:`invoke <i_invoke>`' instruction).
5249 The terminator instructions are: ':ref:`ret <i_ret>`',
5250 ':ref:`br <i_br>`', ':ref:`switch <i_switch>`',
5251 ':ref:`indirectbr <i_indirectbr>`', ':ref:`invoke <i_invoke>`',
5252 ':ref:`resume <i_resume>`', ':ref:`catchswitch <i_catchswitch>`',
5253 ':ref:`catchret <i_catchret>`',
5254 ':ref:`cleanupret <i_cleanupret>`',
5255 and ':ref:`unreachable <i_unreachable>`'.
5259 '``ret``' Instruction
5260 ^^^^^^^^^^^^^^^^^^^^^
5267 ret <type> <value> ; Return a value from a non-void function
5268 ret void ; Return from void function
5273 The '``ret``' instruction is used to return control flow (and optionally
5274 a value) from a function back to the caller.
5276 There are two forms of the '``ret``' instruction: one that returns a
5277 value and then causes control flow, and one that just causes control
5283 The '``ret``' instruction optionally accepts a single argument, the
5284 return value. The type of the return value must be a ':ref:`first
5285 class <t_firstclass>`' type.
5287 A function is not :ref:`well formed <wellformed>` if it it has a non-void
5288 return type and contains a '``ret``' instruction with no return value or
5289 a return value with a type that does not match its type, or if it has a
5290 void return type and contains a '``ret``' instruction with a return
5296 When the '``ret``' instruction is executed, control flow returns back to
5297 the calling function's context. If the caller is a
5298 ":ref:`call <i_call>`" instruction, execution continues at the
5299 instruction after the call. If the caller was an
5300 ":ref:`invoke <i_invoke>`" instruction, execution continues at the
5301 beginning of the "normal" destination block. If the instruction returns
5302 a value, that value shall set the call or invoke instruction's return
5308 .. code-block:: llvm
5310 ret i32 5 ; Return an integer value of 5
5311 ret void ; Return from a void function
5312 ret { i32, i8 } { i32 4, i8 2 } ; Return a struct of values 4 and 2
5316 '``br``' Instruction
5317 ^^^^^^^^^^^^^^^^^^^^
5324 br i1 <cond>, label <iftrue>, label <iffalse>
5325 br label <dest> ; Unconditional branch
5330 The '``br``' instruction is used to cause control flow to transfer to a
5331 different basic block in the current function. There are two forms of
5332 this instruction, corresponding to a conditional branch and an
5333 unconditional branch.
5338 The conditional branch form of the '``br``' instruction takes a single
5339 '``i1``' value and two '``label``' values. The unconditional form of the
5340 '``br``' instruction takes a single '``label``' value as a target.
5345 Upon execution of a conditional '``br``' instruction, the '``i1``'
5346 argument is evaluated. If the value is ``true``, control flows to the
5347 '``iftrue``' ``label`` argument. If "cond" is ``false``, control flows
5348 to the '``iffalse``' ``label`` argument.
5353 .. code-block:: llvm
5356 %cond = icmp eq i32 %a, %b
5357 br i1 %cond, label %IfEqual, label %IfUnequal
5365 '``switch``' Instruction
5366 ^^^^^^^^^^^^^^^^^^^^^^^^
5373 switch <intty> <value>, label <defaultdest> [ <intty> <val>, label <dest> ... ]
5378 The '``switch``' instruction is used to transfer control flow to one of
5379 several different places. It is a generalization of the '``br``'
5380 instruction, allowing a branch to occur to one of many possible
5386 The '``switch``' instruction uses three parameters: an integer
5387 comparison value '``value``', a default '``label``' destination, and an
5388 array of pairs of comparison value constants and '``label``'s. The table
5389 is not allowed to contain duplicate constant entries.
5394 The ``switch`` instruction specifies a table of values and destinations.
5395 When the '``switch``' instruction is executed, this table is searched
5396 for the given value. If the value is found, control flow is transferred
5397 to the corresponding destination; otherwise, control flow is transferred
5398 to the default destination.
5403 Depending on properties of the target machine and the particular
5404 ``switch`` instruction, this instruction may be code generated in
5405 different ways. For example, it could be generated as a series of
5406 chained conditional branches or with a lookup table.
5411 .. code-block:: llvm
5413 ; Emulate a conditional br instruction
5414 %Val = zext i1 %value to i32
5415 switch i32 %Val, label %truedest [ i32 0, label %falsedest ]
5417 ; Emulate an unconditional br instruction
5418 switch i32 0, label %dest [ ]
5420 ; Implement a jump table:
5421 switch i32 %val, label %otherwise [ i32 0, label %onzero
5423 i32 2, label %ontwo ]
5427 '``indirectbr``' Instruction
5428 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5435 indirectbr <somety>* <address>, [ label <dest1>, label <dest2>, ... ]
5440 The '``indirectbr``' instruction implements an indirect branch to a
5441 label within the current function, whose address is specified by
5442 "``address``". Address must be derived from a
5443 :ref:`blockaddress <blockaddress>` constant.
5448 The '``address``' argument is the address of the label to jump to. The
5449 rest of the arguments indicate the full set of possible destinations
5450 that the address may point to. Blocks are allowed to occur multiple
5451 times in the destination list, though this isn't particularly useful.
5453 This destination list is required so that dataflow analysis has an
5454 accurate understanding of the CFG.
5459 Control transfers to the block specified in the address argument. All
5460 possible destination blocks must be listed in the label list, otherwise
5461 this instruction has undefined behavior. This implies that jumps to
5462 labels defined in other functions have undefined behavior as well.
5467 This is typically implemented with a jump through a register.
5472 .. code-block:: llvm
5474 indirectbr i8* %Addr, [ label %bb1, label %bb2, label %bb3 ]
5478 '``invoke``' Instruction
5479 ^^^^^^^^^^^^^^^^^^^^^^^^
5486 <result> = invoke [cconv] [ret attrs] <ty>|<fnty> <fnptrval>(<function args>) [fn attrs]
5487 [operand bundles] to label <normal label> unwind label <exception label>
5492 The '``invoke``' instruction causes control to transfer to a specified
5493 function, with the possibility of control flow transfer to either the
5494 '``normal``' label or the '``exception``' label. If the callee function
5495 returns with the "``ret``" instruction, control flow will return to the
5496 "normal" label. If the callee (or any indirect callees) returns via the
5497 ":ref:`resume <i_resume>`" instruction or other exception handling
5498 mechanism, control is interrupted and continued at the dynamically
5499 nearest "exception" label.
5501 The '``exception``' label is a `landing
5502 pad <ExceptionHandling.html#overview>`_ for the exception. As such,
5503 '``exception``' label is required to have the
5504 ":ref:`landingpad <i_landingpad>`" instruction, which contains the
5505 information about the behavior of the program after unwinding happens,
5506 as its first non-PHI instruction. The restrictions on the
5507 "``landingpad``" instruction's tightly couples it to the "``invoke``"
5508 instruction, so that the important information contained within the
5509 "``landingpad``" instruction can't be lost through normal code motion.
5514 This instruction requires several arguments:
5516 #. The optional "cconv" marker indicates which :ref:`calling
5517 convention <callingconv>` the call should use. If none is
5518 specified, the call defaults to using C calling conventions.
5519 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
5520 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
5522 #. '``ty``': the type of the call instruction itself which is also the
5523 type of the return value. Functions that return no value are marked
5525 #. '``fnty``': shall be the signature of the function being invoked. The
5526 argument types must match the types implied by this signature. This
5527 type can be omitted if the function is not varargs.
5528 #. '``fnptrval``': An LLVM value containing a pointer to a function to
5529 be invoked. In most cases, this is a direct function invocation, but
5530 indirect ``invoke``'s are just as possible, calling an arbitrary pointer
5532 #. '``function args``': argument list whose types match the function
5533 signature argument types and parameter attributes. All arguments must
5534 be of :ref:`first class <t_firstclass>` type. If the function signature
5535 indicates the function accepts a variable number of arguments, the
5536 extra arguments can be specified.
5537 #. '``normal label``': the label reached when the called function
5538 executes a '``ret``' instruction.
5539 #. '``exception label``': the label reached when a callee returns via
5540 the :ref:`resume <i_resume>` instruction or other exception handling
5542 #. The optional :ref:`function attributes <fnattrs>` list. Only
5543 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
5544 attributes are valid here.
5545 #. The optional :ref:`operand bundles <opbundles>` list.
5550 This instruction is designed to operate as a standard '``call``'
5551 instruction in most regards. The primary difference is that it
5552 establishes an association with a label, which is used by the runtime
5553 library to unwind the stack.
5555 This instruction is used in languages with destructors to ensure that
5556 proper cleanup is performed in the case of either a ``longjmp`` or a
5557 thrown exception. Additionally, this is important for implementation of
5558 '``catch``' clauses in high-level languages that support them.
5560 For the purposes of the SSA form, the definition of the value returned
5561 by the '``invoke``' instruction is deemed to occur on the edge from the
5562 current block to the "normal" label. If the callee unwinds then no
5563 return value is available.
5568 .. code-block:: llvm
5570 %retval = invoke i32 @Test(i32 15) to label %Continue
5571 unwind label %TestCleanup ; i32:retval set
5572 %retval = invoke coldcc i32 %Testfnptr(i32 15) to label %Continue
5573 unwind label %TestCleanup ; i32:retval set
5577 '``resume``' Instruction
5578 ^^^^^^^^^^^^^^^^^^^^^^^^
5585 resume <type> <value>
5590 The '``resume``' instruction is a terminator instruction that has no
5596 The '``resume``' instruction requires one argument, which must have the
5597 same type as the result of any '``landingpad``' instruction in the same
5603 The '``resume``' instruction resumes propagation of an existing
5604 (in-flight) exception whose unwinding was interrupted with a
5605 :ref:`landingpad <i_landingpad>` instruction.
5610 .. code-block:: llvm
5612 resume { i8*, i32 } %exn
5616 '``catchswitch``' Instruction
5617 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5624 <resultval> = catchswitch within <parent> [ label <handler1>, label <handler2>, ... ] unwind to caller
5625 <resultval> = catchswitch within <parent> [ label <handler1>, label <handler2>, ... ] unwind label <default>
5630 The '``catchswitch``' instruction is used by `LLVM's exception handling system
5631 <ExceptionHandling.html#overview>`_ to describe the set of possible catch handlers
5632 that may be executed by the :ref:`EH personality routine <personalityfn>`.
5637 The ``parent`` argument is the token of the funclet that contains the
5638 ``catchswitch`` instruction. If the ``catchswitch`` is not inside a funclet,
5639 this operand may be the token ``none``.
5641 The ``default`` argument is the label of another basic block beginning with
5642 either a ``cleanuppad`` or ``catchswitch`` instruction. This unwind destination
5643 must be a legal target with respect to the ``parent`` links, as described in
5644 the `exception handling documentation\ <ExceptionHandling.html#wineh-constraints>`_.
5646 The ``handlers`` are a nonempty list of successor blocks that each begin with a
5647 :ref:`catchpad <i_catchpad>` instruction.
5652 Executing this instruction transfers control to one of the successors in
5653 ``handlers``, if appropriate, or continues to unwind via the unwind label if
5656 The ``catchswitch`` is both a terminator and a "pad" instruction, meaning that
5657 it must be both the first non-phi instruction and last instruction in the basic
5658 block. Therefore, it must be the only non-phi instruction in the block.
5663 .. code-block:: text
5666 %cs1 = catchswitch within none [label %handler0, label %handler1] unwind to caller
5668 %cs2 = catchswitch within %parenthandler [label %handler0] unwind label %cleanup
5672 '``catchret``' Instruction
5673 ^^^^^^^^^^^^^^^^^^^^^^^^^^
5680 catchret from <token> to label <normal>
5685 The '``catchret``' instruction is a terminator instruction that has a
5692 The first argument to a '``catchret``' indicates which ``catchpad`` it
5693 exits. It must be a :ref:`catchpad <i_catchpad>`.
5694 The second argument to a '``catchret``' specifies where control will
5700 The '``catchret``' instruction ends an existing (in-flight) exception whose
5701 unwinding was interrupted with a :ref:`catchpad <i_catchpad>` instruction. The
5702 :ref:`personality function <personalityfn>` gets a chance to execute arbitrary
5703 code to, for example, destroy the active exception. Control then transfers to
5706 The ``token`` argument must be a token produced by a ``catchpad`` instruction.
5707 If the specified ``catchpad`` is not the most-recently-entered not-yet-exited
5708 funclet pad (as described in the `EH documentation\ <ExceptionHandling.html#wineh-constraints>`_),
5709 the ``catchret``'s behavior is undefined.
5714 .. code-block:: text
5716 catchret from %catch label %continue
5720 '``cleanupret``' Instruction
5721 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5728 cleanupret from <value> unwind label <continue>
5729 cleanupret from <value> unwind to caller
5734 The '``cleanupret``' instruction is a terminator instruction that has
5735 an optional successor.
5741 The '``cleanupret``' instruction requires one argument, which indicates
5742 which ``cleanuppad`` it exits, and must be a :ref:`cleanuppad <i_cleanuppad>`.
5743 If the specified ``cleanuppad`` is not the most-recently-entered not-yet-exited
5744 funclet pad (as described in the `EH documentation\ <ExceptionHandling.html#wineh-constraints>`_),
5745 the ``cleanupret``'s behavior is undefined.
5747 The '``cleanupret``' instruction also has an optional successor, ``continue``,
5748 which must be the label of another basic block beginning with either a
5749 ``cleanuppad`` or ``catchswitch`` instruction. This unwind destination must
5750 be a legal target with respect to the ``parent`` links, as described in the
5751 `exception handling documentation\ <ExceptionHandling.html#wineh-constraints>`_.
5756 The '``cleanupret``' instruction indicates to the
5757 :ref:`personality function <personalityfn>` that one
5758 :ref:`cleanuppad <i_cleanuppad>` it transferred control to has ended.
5759 It transfers control to ``continue`` or unwinds out of the function.
5764 .. code-block:: text
5766 cleanupret from %cleanup unwind to caller
5767 cleanupret from %cleanup unwind label %continue
5771 '``unreachable``' Instruction
5772 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5784 The '``unreachable``' instruction has no defined semantics. This
5785 instruction is used to inform the optimizer that a particular portion of
5786 the code is not reachable. This can be used to indicate that the code
5787 after a no-return function cannot be reached, and other facts.
5792 The '``unreachable``' instruction has no defined semantics.
5799 Binary operators are used to do most of the computation in a program.
5800 They require two operands of the same type, execute an operation on
5801 them, and produce a single value. The operands might represent multiple
5802 data, as is the case with the :ref:`vector <t_vector>` data type. The
5803 result value has the same type as its operands.
5805 There are several different binary operators:
5809 '``add``' Instruction
5810 ^^^^^^^^^^^^^^^^^^^^^
5817 <result> = add <ty> <op1>, <op2> ; yields ty:result
5818 <result> = add nuw <ty> <op1>, <op2> ; yields ty:result
5819 <result> = add nsw <ty> <op1>, <op2> ; yields ty:result
5820 <result> = add nuw nsw <ty> <op1>, <op2> ; yields ty:result
5825 The '``add``' instruction returns the sum of its two operands.
5830 The two arguments to the '``add``' instruction must be
5831 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
5832 arguments must have identical types.
5837 The value produced is the integer sum of the two operands.
5839 If the sum has unsigned overflow, the result returned is the
5840 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
5843 Because LLVM integers use a two's complement representation, this
5844 instruction is appropriate for both signed and unsigned integers.
5846 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
5847 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
5848 result value of the ``add`` is a :ref:`poison value <poisonvalues>` if
5849 unsigned and/or signed overflow, respectively, occurs.
5854 .. code-block:: text
5856 <result> = add i32 4, %var ; yields i32:result = 4 + %var
5860 '``fadd``' Instruction
5861 ^^^^^^^^^^^^^^^^^^^^^^
5868 <result> = fadd [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
5873 The '``fadd``' instruction returns the sum of its two operands.
5878 The two arguments to the '``fadd``' instruction must be :ref:`floating
5879 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
5880 Both arguments must have identical types.
5885 The value produced is the floating point sum of the two operands. This
5886 instruction can also take any number of :ref:`fast-math flags <fastmath>`,
5887 which are optimization hints to enable otherwise unsafe floating point
5893 .. code-block:: text
5895 <result> = fadd float 4.0, %var ; yields float:result = 4.0 + %var
5897 '``sub``' Instruction
5898 ^^^^^^^^^^^^^^^^^^^^^
5905 <result> = sub <ty> <op1>, <op2> ; yields ty:result
5906 <result> = sub nuw <ty> <op1>, <op2> ; yields ty:result
5907 <result> = sub nsw <ty> <op1>, <op2> ; yields ty:result
5908 <result> = sub nuw nsw <ty> <op1>, <op2> ; yields ty:result
5913 The '``sub``' instruction returns the difference of its two operands.
5915 Note that the '``sub``' instruction is used to represent the '``neg``'
5916 instruction present in most other intermediate representations.
5921 The two arguments to the '``sub``' instruction must be
5922 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
5923 arguments must have identical types.
5928 The value produced is the integer difference of the two operands.
5930 If the difference has unsigned overflow, the result returned is the
5931 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
5934 Because LLVM integers use a two's complement representation, this
5935 instruction is appropriate for both signed and unsigned integers.
5937 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
5938 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
5939 result value of the ``sub`` is a :ref:`poison value <poisonvalues>` if
5940 unsigned and/or signed overflow, respectively, occurs.
5945 .. code-block:: text
5947 <result> = sub i32 4, %var ; yields i32:result = 4 - %var
5948 <result> = sub i32 0, %val ; yields i32:result = -%var
5952 '``fsub``' Instruction
5953 ^^^^^^^^^^^^^^^^^^^^^^
5960 <result> = fsub [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
5965 The '``fsub``' instruction returns the difference of its two operands.
5967 Note that the '``fsub``' instruction is used to represent the '``fneg``'
5968 instruction present in most other intermediate representations.
5973 The two arguments to the '``fsub``' instruction must be :ref:`floating
5974 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
5975 Both arguments must have identical types.
5980 The value produced is the floating point difference of the two operands.
5981 This instruction can also take any number of :ref:`fast-math
5982 flags <fastmath>`, which are optimization hints to enable otherwise
5983 unsafe floating point optimizations:
5988 .. code-block:: text
5990 <result> = fsub float 4.0, %var ; yields float:result = 4.0 - %var
5991 <result> = fsub float -0.0, %val ; yields float:result = -%var
5993 '``mul``' Instruction
5994 ^^^^^^^^^^^^^^^^^^^^^
6001 <result> = mul <ty> <op1>, <op2> ; yields ty:result
6002 <result> = mul nuw <ty> <op1>, <op2> ; yields ty:result
6003 <result> = mul nsw <ty> <op1>, <op2> ; yields ty:result
6004 <result> = mul nuw nsw <ty> <op1>, <op2> ; yields ty:result
6009 The '``mul``' instruction returns the product of its two operands.
6014 The two arguments to the '``mul``' instruction must be
6015 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6016 arguments must have identical types.
6021 The value produced is the integer product of the two operands.
6023 If the result of the multiplication has unsigned overflow, the result
6024 returned is the mathematical result modulo 2\ :sup:`n`\ , where n is the
6025 bit width of the result.
6027 Because LLVM integers use a two's complement representation, and the
6028 result is the same width as the operands, this instruction returns the
6029 correct result for both signed and unsigned integers. If a full product
6030 (e.g. ``i32`` * ``i32`` -> ``i64``) is needed, the operands should be
6031 sign-extended or zero-extended as appropriate to the width of the full
6034 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
6035 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
6036 result value of the ``mul`` is a :ref:`poison value <poisonvalues>` if
6037 unsigned and/or signed overflow, respectively, occurs.
6042 .. code-block:: text
6044 <result> = mul i32 4, %var ; yields i32:result = 4 * %var
6048 '``fmul``' Instruction
6049 ^^^^^^^^^^^^^^^^^^^^^^
6056 <result> = fmul [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
6061 The '``fmul``' instruction returns the product of its two operands.
6066 The two arguments to the '``fmul``' instruction must be :ref:`floating
6067 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
6068 Both arguments must have identical types.
6073 The value produced is the floating point product of the two operands.
6074 This instruction can also take any number of :ref:`fast-math
6075 flags <fastmath>`, which are optimization hints to enable otherwise
6076 unsafe floating point optimizations:
6081 .. code-block:: text
6083 <result> = fmul float 4.0, %var ; yields float:result = 4.0 * %var
6085 '``udiv``' Instruction
6086 ^^^^^^^^^^^^^^^^^^^^^^
6093 <result> = udiv <ty> <op1>, <op2> ; yields ty:result
6094 <result> = udiv exact <ty> <op1>, <op2> ; yields ty:result
6099 The '``udiv``' instruction returns the quotient of its two operands.
6104 The two arguments to the '``udiv``' instruction must be
6105 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6106 arguments must have identical types.
6111 The value produced is the unsigned integer quotient of the two operands.
6113 Note that unsigned integer division and signed integer division are
6114 distinct operations; for signed integer division, use '``sdiv``'.
6116 Division by zero leads to undefined behavior.
6118 If the ``exact`` keyword is present, the result value of the ``udiv`` is
6119 a :ref:`poison value <poisonvalues>` if %op1 is not a multiple of %op2 (as
6120 such, "((a udiv exact b) mul b) == a").
6125 .. code-block:: text
6127 <result> = udiv i32 4, %var ; yields i32:result = 4 / %var
6129 '``sdiv``' Instruction
6130 ^^^^^^^^^^^^^^^^^^^^^^
6137 <result> = sdiv <ty> <op1>, <op2> ; yields ty:result
6138 <result> = sdiv exact <ty> <op1>, <op2> ; yields ty:result
6143 The '``sdiv``' instruction returns the quotient of its two operands.
6148 The two arguments to the '``sdiv``' instruction must be
6149 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6150 arguments must have identical types.
6155 The value produced is the signed integer quotient of the two operands
6156 rounded towards zero.
6158 Note that signed integer division and unsigned integer division are
6159 distinct operations; for unsigned integer division, use '``udiv``'.
6161 Division by zero leads to undefined behavior. Overflow also leads to
6162 undefined behavior; this is a rare case, but can occur, for example, by
6163 doing a 32-bit division of -2147483648 by -1.
6165 If the ``exact`` keyword is present, the result value of the ``sdiv`` is
6166 a :ref:`poison value <poisonvalues>` if the result would be rounded.
6171 .. code-block:: text
6173 <result> = sdiv i32 4, %var ; yields i32:result = 4 / %var
6177 '``fdiv``' Instruction
6178 ^^^^^^^^^^^^^^^^^^^^^^
6185 <result> = fdiv [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
6190 The '``fdiv``' instruction returns the quotient of its two operands.
6195 The two arguments to the '``fdiv``' instruction must be :ref:`floating
6196 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
6197 Both arguments must have identical types.
6202 The value produced is the floating point quotient of the two operands.
6203 This instruction can also take any number of :ref:`fast-math
6204 flags <fastmath>`, which are optimization hints to enable otherwise
6205 unsafe floating point optimizations:
6210 .. code-block:: text
6212 <result> = fdiv float 4.0, %var ; yields float:result = 4.0 / %var
6214 '``urem``' Instruction
6215 ^^^^^^^^^^^^^^^^^^^^^^
6222 <result> = urem <ty> <op1>, <op2> ; yields ty:result
6227 The '``urem``' instruction returns the remainder from the unsigned
6228 division of its two arguments.
6233 The two arguments to the '``urem``' instruction must be
6234 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6235 arguments must have identical types.
6240 This instruction returns the unsigned integer *remainder* of a division.
6241 This instruction always performs an unsigned division to get the
6244 Note that unsigned integer remainder and signed integer remainder are
6245 distinct operations; for signed integer remainder, use '``srem``'.
6247 Taking the remainder of a division by zero leads to undefined behavior.
6252 .. code-block:: text
6254 <result> = urem i32 4, %var ; yields i32:result = 4 % %var
6256 '``srem``' Instruction
6257 ^^^^^^^^^^^^^^^^^^^^^^
6264 <result> = srem <ty> <op1>, <op2> ; yields ty:result
6269 The '``srem``' instruction returns the remainder from the signed
6270 division of its two operands. This instruction can also take
6271 :ref:`vector <t_vector>` versions of the values in which case the elements
6277 The two arguments to the '``srem``' instruction must be
6278 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6279 arguments must have identical types.
6284 This instruction returns the *remainder* of a division (where the result
6285 is either zero or has the same sign as the dividend, ``op1``), not the
6286 *modulo* operator (where the result is either zero or has the same sign
6287 as the divisor, ``op2``) of a value. For more information about the
6288 difference, see `The Math
6289 Forum <http://mathforum.org/dr.math/problems/anne.4.28.99.html>`_. For a
6290 table of how this is implemented in various languages, please see
6292 operation <http://en.wikipedia.org/wiki/Modulo_operation>`_.
6294 Note that signed integer remainder and unsigned integer remainder are
6295 distinct operations; for unsigned integer remainder, use '``urem``'.
6297 Taking the remainder of a division by zero leads to undefined behavior.
6298 Overflow also leads to undefined behavior; this is a rare case, but can
6299 occur, for example, by taking the remainder of a 32-bit division of
6300 -2147483648 by -1. (The remainder doesn't actually overflow, but this
6301 rule lets srem be implemented using instructions that return both the
6302 result of the division and the remainder.)
6307 .. code-block:: text
6309 <result> = srem i32 4, %var ; yields i32:result = 4 % %var
6313 '``frem``' Instruction
6314 ^^^^^^^^^^^^^^^^^^^^^^
6321 <result> = frem [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
6326 The '``frem``' instruction returns the remainder from the division of
6332 The two arguments to the '``frem``' instruction must be :ref:`floating
6333 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
6334 Both arguments must have identical types.
6339 This instruction returns the *remainder* of a division. The remainder
6340 has the same sign as the dividend. This instruction can also take any
6341 number of :ref:`fast-math flags <fastmath>`, which are optimization hints
6342 to enable otherwise unsafe floating point optimizations:
6347 .. code-block:: text
6349 <result> = frem float 4.0, %var ; yields float:result = 4.0 % %var
6353 Bitwise Binary Operations
6354 -------------------------
6356 Bitwise binary operators are used to do various forms of bit-twiddling
6357 in a program. They are generally very efficient instructions and can
6358 commonly be strength reduced from other instructions. They require two
6359 operands of the same type, execute an operation on them, and produce a
6360 single value. The resulting value is the same type as its operands.
6362 '``shl``' Instruction
6363 ^^^^^^^^^^^^^^^^^^^^^
6370 <result> = shl <ty> <op1>, <op2> ; yields ty:result
6371 <result> = shl nuw <ty> <op1>, <op2> ; yields ty:result
6372 <result> = shl nsw <ty> <op1>, <op2> ; yields ty:result
6373 <result> = shl nuw nsw <ty> <op1>, <op2> ; yields ty:result
6378 The '``shl``' instruction returns the first operand shifted to the left
6379 a specified number of bits.
6384 Both arguments to the '``shl``' instruction must be the same
6385 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
6386 '``op2``' is treated as an unsigned value.
6391 The value produced is ``op1`` \* 2\ :sup:`op2` mod 2\ :sup:`n`,
6392 where ``n`` is the width of the result. If ``op2`` is (statically or
6393 dynamically) equal to or larger than the number of bits in
6394 ``op1``, the result is undefined. If the arguments are vectors, each
6395 vector element of ``op1`` is shifted by the corresponding shift amount
6398 If the ``nuw`` keyword is present, then the shift produces a :ref:`poison
6399 value <poisonvalues>` if it shifts out any non-zero bits. If the
6400 ``nsw`` keyword is present, then the shift produces a :ref:`poison
6401 value <poisonvalues>` if it shifts out any bits that disagree with the
6402 resultant sign bit. As such, NUW/NSW have the same semantics as they
6403 would if the shift were expressed as a mul instruction with the same
6404 nsw/nuw bits in (mul %op1, (shl 1, %op2)).
6409 .. code-block:: text
6411 <result> = shl i32 4, %var ; yields i32: 4 << %var
6412 <result> = shl i32 4, 2 ; yields i32: 16
6413 <result> = shl i32 1, 10 ; yields i32: 1024
6414 <result> = shl i32 1, 32 ; undefined
6415 <result> = shl <2 x i32> < i32 1, i32 1>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 2, i32 4>
6417 '``lshr``' Instruction
6418 ^^^^^^^^^^^^^^^^^^^^^^
6425 <result> = lshr <ty> <op1>, <op2> ; yields ty:result
6426 <result> = lshr exact <ty> <op1>, <op2> ; yields ty:result
6431 The '``lshr``' instruction (logical shift right) returns the first
6432 operand shifted to the right a specified number of bits with zero fill.
6437 Both arguments to the '``lshr``' instruction must be the same
6438 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
6439 '``op2``' is treated as an unsigned value.
6444 This instruction always performs a logical shift right operation. The
6445 most significant bits of the result will be filled with zero bits after
6446 the shift. If ``op2`` is (statically or dynamically) equal to or larger
6447 than the number of bits in ``op1``, the result is undefined. If the
6448 arguments are vectors, each vector element of ``op1`` is shifted by the
6449 corresponding shift amount in ``op2``.
6451 If the ``exact`` keyword is present, the result value of the ``lshr`` is
6452 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
6458 .. code-block:: text
6460 <result> = lshr i32 4, 1 ; yields i32:result = 2
6461 <result> = lshr i32 4, 2 ; yields i32:result = 1
6462 <result> = lshr i8 4, 3 ; yields i8:result = 0
6463 <result> = lshr i8 -2, 1 ; yields i8:result = 0x7F
6464 <result> = lshr i32 1, 32 ; undefined
6465 <result> = lshr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 0x7FFFFFFF, i32 1>
6467 '``ashr``' Instruction
6468 ^^^^^^^^^^^^^^^^^^^^^^
6475 <result> = ashr <ty> <op1>, <op2> ; yields ty:result
6476 <result> = ashr exact <ty> <op1>, <op2> ; yields ty:result
6481 The '``ashr``' instruction (arithmetic shift right) returns the first
6482 operand shifted to the right a specified number of bits with sign
6488 Both arguments to the '``ashr``' instruction must be the same
6489 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
6490 '``op2``' is treated as an unsigned value.
6495 This instruction always performs an arithmetic shift right operation,
6496 The most significant bits of the result will be filled with the sign bit
6497 of ``op1``. If ``op2`` is (statically or dynamically) equal to or larger
6498 than the number of bits in ``op1``, the result is undefined. If the
6499 arguments are vectors, each vector element of ``op1`` is shifted by the
6500 corresponding shift amount in ``op2``.
6502 If the ``exact`` keyword is present, the result value of the ``ashr`` is
6503 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
6509 .. code-block:: text
6511 <result> = ashr i32 4, 1 ; yields i32:result = 2
6512 <result> = ashr i32 4, 2 ; yields i32:result = 1
6513 <result> = ashr i8 4, 3 ; yields i8:result = 0
6514 <result> = ashr i8 -2, 1 ; yields i8:result = -1
6515 <result> = ashr i32 1, 32 ; undefined
6516 <result> = ashr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 3> ; yields: result=<2 x i32> < i32 -1, i32 0>
6518 '``and``' Instruction
6519 ^^^^^^^^^^^^^^^^^^^^^
6526 <result> = and <ty> <op1>, <op2> ; yields ty:result
6531 The '``and``' instruction returns the bitwise logical and of its two
6537 The two arguments to the '``and``' instruction must be
6538 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6539 arguments must have identical types.
6544 The truth table used for the '``and``' instruction is:
6561 .. code-block:: text
6563 <result> = and i32 4, %var ; yields i32:result = 4 & %var
6564 <result> = and i32 15, 40 ; yields i32:result = 8
6565 <result> = and i32 4, 8 ; yields i32:result = 0
6567 '``or``' Instruction
6568 ^^^^^^^^^^^^^^^^^^^^
6575 <result> = or <ty> <op1>, <op2> ; yields ty:result
6580 The '``or``' instruction returns the bitwise logical inclusive or of its
6586 The two arguments to the '``or``' instruction must be
6587 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6588 arguments must have identical types.
6593 The truth table used for the '``or``' instruction is:
6612 <result> = or i32 4, %var ; yields i32:result = 4 | %var
6613 <result> = or i32 15, 40 ; yields i32:result = 47
6614 <result> = or i32 4, 8 ; yields i32:result = 12
6616 '``xor``' Instruction
6617 ^^^^^^^^^^^^^^^^^^^^^
6624 <result> = xor <ty> <op1>, <op2> ; yields ty:result
6629 The '``xor``' instruction returns the bitwise logical exclusive or of
6630 its two operands. The ``xor`` is used to implement the "one's
6631 complement" operation, which is the "~" operator in C.
6636 The two arguments to the '``xor``' instruction must be
6637 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6638 arguments must have identical types.
6643 The truth table used for the '``xor``' instruction is:
6660 .. code-block:: text
6662 <result> = xor i32 4, %var ; yields i32:result = 4 ^ %var
6663 <result> = xor i32 15, 40 ; yields i32:result = 39
6664 <result> = xor i32 4, 8 ; yields i32:result = 12
6665 <result> = xor i32 %V, -1 ; yields i32:result = ~%V
6670 LLVM supports several instructions to represent vector operations in a
6671 target-independent manner. These instructions cover the element-access
6672 and vector-specific operations needed to process vectors effectively.
6673 While LLVM does directly support these vector operations, many
6674 sophisticated algorithms will want to use target-specific intrinsics to
6675 take full advantage of a specific target.
6677 .. _i_extractelement:
6679 '``extractelement``' Instruction
6680 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6687 <result> = extractelement <n x <ty>> <val>, <ty2> <idx> ; yields <ty>
6692 The '``extractelement``' instruction extracts a single scalar element
6693 from a vector at a specified index.
6698 The first operand of an '``extractelement``' instruction is a value of
6699 :ref:`vector <t_vector>` type. The second operand is an index indicating
6700 the position from which to extract the element. The index may be a
6701 variable of any integer type.
6706 The result is a scalar of the same type as the element type of ``val``.
6707 Its value is the value at position ``idx`` of ``val``. If ``idx``
6708 exceeds the length of ``val``, the results are undefined.
6713 .. code-block:: text
6715 <result> = extractelement <4 x i32> %vec, i32 0 ; yields i32
6717 .. _i_insertelement:
6719 '``insertelement``' Instruction
6720 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6727 <result> = insertelement <n x <ty>> <val>, <ty> <elt>, <ty2> <idx> ; yields <n x <ty>>
6732 The '``insertelement``' instruction inserts a scalar element into a
6733 vector at a specified index.
6738 The first operand of an '``insertelement``' instruction is a value of
6739 :ref:`vector <t_vector>` type. The second operand is a scalar value whose
6740 type must equal the element type of the first operand. The third operand
6741 is an index indicating the position at which to insert the value. The
6742 index may be a variable of any integer type.
6747 The result is a vector of the same type as ``val``. Its element values
6748 are those of ``val`` except at position ``idx``, where it gets the value
6749 ``elt``. If ``idx`` exceeds the length of ``val``, the results are
6755 .. code-block:: text
6757 <result> = insertelement <4 x i32> %vec, i32 1, i32 0 ; yields <4 x i32>
6759 .. _i_shufflevector:
6761 '``shufflevector``' Instruction
6762 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6769 <result> = shufflevector <n x <ty>> <v1>, <n x <ty>> <v2>, <m x i32> <mask> ; yields <m x <ty>>
6774 The '``shufflevector``' instruction constructs a permutation of elements
6775 from two input vectors, returning a vector with the same element type as
6776 the input and length that is the same as the shuffle mask.
6781 The first two operands of a '``shufflevector``' instruction are vectors
6782 with the same type. The third argument is a shuffle mask whose element
6783 type is always 'i32'. The result of the instruction is a vector whose
6784 length is the same as the shuffle mask and whose element type is the
6785 same as the element type of the first two operands.
6787 The shuffle mask operand is required to be a constant vector with either
6788 constant integer or undef values.
6793 The elements of the two input vectors are numbered from left to right
6794 across both of the vectors. The shuffle mask operand specifies, for each
6795 element of the result vector, which element of the two input vectors the
6796 result element gets. The element selector may be undef (meaning "don't
6797 care") and the second operand may be undef if performing a shuffle from
6803 .. code-block:: text
6805 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
6806 <4 x i32> <i32 0, i32 4, i32 1, i32 5> ; yields <4 x i32>
6807 <result> = shufflevector <4 x i32> %v1, <4 x i32> undef,
6808 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32> - Identity shuffle.
6809 <result> = shufflevector <8 x i32> %v1, <8 x i32> undef,
6810 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32>
6811 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
6812 <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7 > ; yields <8 x i32>
6814 Aggregate Operations
6815 --------------------
6817 LLVM supports several instructions for working with
6818 :ref:`aggregate <t_aggregate>` values.
6822 '``extractvalue``' Instruction
6823 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6830 <result> = extractvalue <aggregate type> <val>, <idx>{, <idx>}*
6835 The '``extractvalue``' instruction extracts the value of a member field
6836 from an :ref:`aggregate <t_aggregate>` value.
6841 The first operand of an '``extractvalue``' instruction is a value of
6842 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The other operands are
6843 constant indices to specify which value to extract in a similar manner
6844 as indices in a '``getelementptr``' instruction.
6846 The major differences to ``getelementptr`` indexing are:
6848 - Since the value being indexed is not a pointer, the first index is
6849 omitted and assumed to be zero.
6850 - At least one index must be specified.
6851 - Not only struct indices but also array indices must be in bounds.
6856 The result is the value at the position in the aggregate specified by
6862 .. code-block:: text
6864 <result> = extractvalue {i32, float} %agg, 0 ; yields i32
6868 '``insertvalue``' Instruction
6869 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6876 <result> = insertvalue <aggregate type> <val>, <ty> <elt>, <idx>{, <idx>}* ; yields <aggregate type>
6881 The '``insertvalue``' instruction inserts a value into a member field in
6882 an :ref:`aggregate <t_aggregate>` value.
6887 The first operand of an '``insertvalue``' instruction is a value of
6888 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The second operand is
6889 a first-class value to insert. The following operands are constant
6890 indices indicating the position at which to insert the value in a
6891 similar manner as indices in a '``extractvalue``' instruction. The value
6892 to insert must have the same type as the value identified by the
6898 The result is an aggregate of the same type as ``val``. Its value is
6899 that of ``val`` except that the value at the position specified by the
6900 indices is that of ``elt``.
6905 .. code-block:: llvm
6907 %agg1 = insertvalue {i32, float} undef, i32 1, 0 ; yields {i32 1, float undef}
6908 %agg2 = insertvalue {i32, float} %agg1, float %val, 1 ; yields {i32 1, float %val}
6909 %agg3 = insertvalue {i32, {float}} undef, float %val, 1, 0 ; yields {i32 undef, {float %val}}
6913 Memory Access and Addressing Operations
6914 ---------------------------------------
6916 A key design point of an SSA-based representation is how it represents
6917 memory. In LLVM, no memory locations are in SSA form, which makes things
6918 very simple. This section describes how to read, write, and allocate
6923 '``alloca``' Instruction
6924 ^^^^^^^^^^^^^^^^^^^^^^^^
6931 <result> = alloca [inalloca] <type> [, <ty> <NumElements>] [, align <alignment>] ; yields type*:result
6936 The '``alloca``' instruction allocates memory on the stack frame of the
6937 currently executing function, to be automatically released when this
6938 function returns to its caller. The object is always allocated in the
6939 generic address space (address space zero).
6944 The '``alloca``' instruction allocates ``sizeof(<type>)*NumElements``
6945 bytes of memory on the runtime stack, returning a pointer of the
6946 appropriate type to the program. If "NumElements" is specified, it is
6947 the number of elements allocated, otherwise "NumElements" is defaulted
6948 to be one. If a constant alignment is specified, the value result of the
6949 allocation is guaranteed to be aligned to at least that boundary. The
6950 alignment may not be greater than ``1 << 29``. If not specified, or if
6951 zero, the target can choose to align the allocation on any convenient
6952 boundary compatible with the type.
6954 '``type``' may be any sized type.
6959 Memory is allocated; a pointer is returned. The operation is undefined
6960 if there is insufficient stack space for the allocation. '``alloca``'d
6961 memory is automatically released when the function returns. The
6962 '``alloca``' instruction is commonly used to represent automatic
6963 variables that must have an address available. When the function returns
6964 (either with the ``ret`` or ``resume`` instructions), the memory is
6965 reclaimed. Allocating zero bytes is legal, but the result is undefined.
6966 The order in which memory is allocated (ie., which way the stack grows)
6972 .. code-block:: llvm
6974 %ptr = alloca i32 ; yields i32*:ptr
6975 %ptr = alloca i32, i32 4 ; yields i32*:ptr
6976 %ptr = alloca i32, i32 4, align 1024 ; yields i32*:ptr
6977 %ptr = alloca i32, align 1024 ; yields i32*:ptr
6981 '``load``' Instruction
6982 ^^^^^^^^^^^^^^^^^^^^^^
6989 <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>]
6990 <result> = load atomic [volatile] <ty>, <ty>* <pointer> [singlethread] <ordering>, align <alignment> [, !invariant.group !<index>]
6991 !<index> = !{ i32 1 }
6992 !<deref_bytes_node> = !{i64 <dereferenceable_bytes>}
6993 !<align_node> = !{ i64 <value_alignment> }
6998 The '``load``' instruction is used to read from memory.
7003 The argument to the ``load`` instruction specifies the memory address from which
7004 to load. The type specified must be a :ref:`first class <t_firstclass>` type of
7005 known size (i.e. not containing an :ref:`opaque structural type <t_opaque>`). If
7006 the ``load`` is marked as ``volatile``, then the optimizer is not allowed to
7007 modify the number or order of execution of this ``load`` with other
7008 :ref:`volatile operations <volatile>`.
7010 If the ``load`` is marked as ``atomic``, it takes an extra :ref:`ordering
7011 <ordering>` and optional ``singlethread`` argument. The ``release`` and
7012 ``acq_rel`` orderings are not valid on ``load`` instructions. Atomic loads
7013 produce :ref:`defined <memmodel>` results when they may see multiple atomic
7014 stores. The type of the pointee must be an integer, pointer, or floating-point
7015 type whose bit width is a power of two greater than or equal to eight and less
7016 than or equal to a target-specific size limit. ``align`` must be explicitly
7017 specified on atomic loads, and the load has undefined behavior if the alignment
7018 is not set to a value which is at least the size in bytes of the
7019 pointee. ``!nontemporal`` does not have any defined semantics for atomic loads.
7021 The optional constant ``align`` argument specifies the alignment of the
7022 operation (that is, the alignment of the memory address). A value of 0
7023 or an omitted ``align`` argument means that the operation has the ABI
7024 alignment for the target. It is the responsibility of the code emitter
7025 to ensure that the alignment information is correct. Overestimating the
7026 alignment results in undefined behavior. Underestimating the alignment
7027 may produce less efficient code. An alignment of 1 is always safe. The
7028 maximum possible alignment is ``1 << 29``. An alignment value higher
7029 than the size of the loaded type implies memory up to the alignment
7030 value bytes can be safely loaded without trapping in the default
7031 address space. Access of the high bytes can interfere with debugging
7032 tools, so should not be accessed if the function has the
7033 ``sanitize_thread`` or ``sanitize_address`` attributes.
7035 The optional ``!nontemporal`` metadata must reference a single
7036 metadata name ``<index>`` corresponding to a metadata node with one
7037 ``i32`` entry of value 1. The existence of the ``!nontemporal``
7038 metadata on the instruction tells the optimizer and code generator
7039 that this load is not expected to be reused in the cache. The code
7040 generator may select special instructions to save cache bandwidth, such
7041 as the ``MOVNT`` instruction on x86.
7043 The optional ``!invariant.load`` metadata must reference a single
7044 metadata name ``<index>`` corresponding to a metadata node with no
7045 entries. The existence of the ``!invariant.load`` metadata on the
7046 instruction tells the optimizer and code generator that the address
7047 operand to this load points to memory which can be assumed unchanged.
7048 Being invariant does not imply that a location is dereferenceable,
7049 but it does imply that once the location is known dereferenceable
7050 its value is henceforth unchanging.
7052 The optional ``!invariant.group`` metadata must reference a single metadata name
7053 ``<index>`` corresponding to a metadata node. See ``invariant.group`` metadata.
7055 The optional ``!nonnull`` metadata must reference a single
7056 metadata name ``<index>`` corresponding to a metadata node with no
7057 entries. The existence of the ``!nonnull`` metadata on the
7058 instruction tells the optimizer that the value loaded is known to
7059 never be null. This is analogous to the ``nonnull`` attribute
7060 on parameters and return values. This metadata can only be applied
7061 to loads of a pointer type.
7063 The optional ``!dereferenceable`` metadata must reference a single metadata
7064 name ``<deref_bytes_node>`` corresponding to a metadata node with one ``i64``
7065 entry. The existence of the ``!dereferenceable`` metadata on the instruction
7066 tells the optimizer that the value loaded is known to be dereferenceable.
7067 The number of bytes known to be dereferenceable is specified by the integer
7068 value in the metadata node. This is analogous to the ''dereferenceable''
7069 attribute on parameters and return values. This metadata can only be applied
7070 to loads of a pointer type.
7072 The optional ``!dereferenceable_or_null`` metadata must reference a single
7073 metadata name ``<deref_bytes_node>`` corresponding to a metadata node with one
7074 ``i64`` entry. The existence of the ``!dereferenceable_or_null`` metadata on the
7075 instruction tells the optimizer that the value loaded is known to be either
7076 dereferenceable or null.
7077 The number of bytes known to be dereferenceable is specified by the integer
7078 value in the metadata node. This is analogous to the ''dereferenceable_or_null''
7079 attribute on parameters and return values. This metadata can only be applied
7080 to loads of a pointer type.
7082 The optional ``!align`` metadata must reference a single metadata name
7083 ``<align_node>`` corresponding to a metadata node with one ``i64`` entry.
7084 The existence of the ``!align`` metadata on the instruction tells the
7085 optimizer that the value loaded is known to be aligned to a boundary specified
7086 by the integer value in the metadata node. The alignment must be a power of 2.
7087 This is analogous to the ''align'' attribute on parameters and return values.
7088 This metadata can only be applied to loads of a pointer type.
7093 The location of memory pointed to is loaded. If the value being loaded
7094 is of scalar type then the number of bytes read does not exceed the
7095 minimum number of bytes needed to hold all bits of the type. For
7096 example, loading an ``i24`` reads at most three bytes. When loading a
7097 value of a type like ``i20`` with a size that is not an integral number
7098 of bytes, the result is undefined if the value was not originally
7099 written using a store of the same type.
7104 .. code-block:: llvm
7106 %ptr = alloca i32 ; yields i32*:ptr
7107 store i32 3, i32* %ptr ; yields void
7108 %val = load i32, i32* %ptr ; yields i32:val = i32 3
7112 '``store``' Instruction
7113 ^^^^^^^^^^^^^^^^^^^^^^^
7120 store [volatile] <ty> <value>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>][, !invariant.group !<index>] ; yields void
7121 store atomic [volatile] <ty> <value>, <ty>* <pointer> [singlethread] <ordering>, align <alignment> [, !invariant.group !<index>] ; yields void
7126 The '``store``' instruction is used to write to memory.
7131 There are two arguments to the ``store`` instruction: a value to store and an
7132 address at which to store it. The type of the ``<pointer>`` operand must be a
7133 pointer to the :ref:`first class <t_firstclass>` type of the ``<value>``
7134 operand. If the ``store`` is marked as ``volatile``, then the optimizer is not
7135 allowed to modify the number or order of execution of this ``store`` with other
7136 :ref:`volatile operations <volatile>`. Only values of :ref:`first class
7137 <t_firstclass>` types of known size (i.e. not containing an :ref:`opaque
7138 structural type <t_opaque>`) can be stored.
7140 If the ``store`` is marked as ``atomic``, it takes an extra :ref:`ordering
7141 <ordering>` and optional ``singlethread`` argument. The ``acquire`` and
7142 ``acq_rel`` orderings aren't valid on ``store`` instructions. Atomic loads
7143 produce :ref:`defined <memmodel>` results when they may see multiple atomic
7144 stores. The type of the pointee must be an integer, pointer, or floating-point
7145 type whose bit width is a power of two greater than or equal to eight and less
7146 than or equal to a target-specific size limit. ``align`` must be explicitly
7147 specified on atomic stores, and the store has undefined behavior if the
7148 alignment is not set to a value which is at least the size in bytes of the
7149 pointee. ``!nontemporal`` does not have any defined semantics for atomic stores.
7151 The optional constant ``align`` argument specifies the alignment of the
7152 operation (that is, the alignment of the memory address). A value of 0
7153 or an omitted ``align`` argument means that the operation has the ABI
7154 alignment for the target. It is the responsibility of the code emitter
7155 to ensure that the alignment information is correct. Overestimating the
7156 alignment results in undefined behavior. Underestimating the
7157 alignment may produce less efficient code. An alignment of 1 is always
7158 safe. The maximum possible alignment is ``1 << 29``. An alignment
7159 value higher than the size of the stored type implies memory up to the
7160 alignment value bytes can be stored to without trapping in the default
7161 address space. Storing to the higher bytes however may result in data
7162 races if another thread can access the same address. Introducing a
7163 data race is not allowed. Storing to the extra bytes is not allowed
7164 even in situations where a data race is known to not exist if the
7165 function has the ``sanitize_address`` attribute.
7167 The optional ``!nontemporal`` metadata must reference a single metadata
7168 name ``<index>`` corresponding to a metadata node with one ``i32`` entry of
7169 value 1. The existence of the ``!nontemporal`` metadata on the instruction
7170 tells the optimizer and code generator that this load is not expected to
7171 be reused in the cache. The code generator may select special
7172 instructions to save cache bandwidth, such as the ``MOVNT`` instruction on
7175 The optional ``!invariant.group`` metadata must reference a
7176 single metadata name ``<index>``. See ``invariant.group`` metadata.
7181 The contents of memory are updated to contain ``<value>`` at the
7182 location specified by the ``<pointer>`` operand. If ``<value>`` is
7183 of scalar type then the number of bytes written does not exceed the
7184 minimum number of bytes needed to hold all bits of the type. For
7185 example, storing an ``i24`` writes at most three bytes. When writing a
7186 value of a type like ``i20`` with a size that is not an integral number
7187 of bytes, it is unspecified what happens to the extra bits that do not
7188 belong to the type, but they will typically be overwritten.
7193 .. code-block:: llvm
7195 %ptr = alloca i32 ; yields i32*:ptr
7196 store i32 3, i32* %ptr ; yields void
7197 %val = load i32, i32* %ptr ; yields i32:val = i32 3
7201 '``fence``' Instruction
7202 ^^^^^^^^^^^^^^^^^^^^^^^
7209 fence [singlethread] <ordering> ; yields void
7214 The '``fence``' instruction is used to introduce happens-before edges
7220 '``fence``' instructions take an :ref:`ordering <ordering>` argument which
7221 defines what *synchronizes-with* edges they add. They can only be given
7222 ``acquire``, ``release``, ``acq_rel``, and ``seq_cst`` orderings.
7227 A fence A which has (at least) ``release`` ordering semantics
7228 *synchronizes with* a fence B with (at least) ``acquire`` ordering
7229 semantics if and only if there exist atomic operations X and Y, both
7230 operating on some atomic object M, such that A is sequenced before X, X
7231 modifies M (either directly or through some side effect of a sequence
7232 headed by X), Y is sequenced before B, and Y observes M. This provides a
7233 *happens-before* dependency between A and B. Rather than an explicit
7234 ``fence``, one (but not both) of the atomic operations X or Y might
7235 provide a ``release`` or ``acquire`` (resp.) ordering constraint and
7236 still *synchronize-with* the explicit ``fence`` and establish the
7237 *happens-before* edge.
7239 A ``fence`` which has ``seq_cst`` ordering, in addition to having both
7240 ``acquire`` and ``release`` semantics specified above, participates in
7241 the global program order of other ``seq_cst`` operations and/or fences.
7243 The optional ":ref:`singlethread <singlethread>`" argument specifies
7244 that the fence only synchronizes with other fences in the same thread.
7245 (This is useful for interacting with signal handlers.)
7250 .. code-block:: llvm
7252 fence acquire ; yields void
7253 fence singlethread seq_cst ; yields void
7257 '``cmpxchg``' Instruction
7258 ^^^^^^^^^^^^^^^^^^^^^^^^^
7265 cmpxchg [weak] [volatile] <ty>* <pointer>, <ty> <cmp>, <ty> <new> [singlethread] <success ordering> <failure ordering> ; yields { ty, i1 }
7270 The '``cmpxchg``' instruction is used to atomically modify memory. It
7271 loads a value in memory and compares it to a given value. If they are
7272 equal, it tries to store a new value into the memory.
7277 There are three arguments to the '``cmpxchg``' instruction: an address
7278 to operate on, a value to compare to the value currently be at that
7279 address, and a new value to place at that address if the compared values
7280 are equal. The type of '<cmp>' must be an integer or pointer type whose
7281 bit width is a power of two greater than or equal to eight and less
7282 than or equal to a target-specific size limit. '<cmp>' and '<new>' must
7283 have the same type, and the type of '<pointer>' must be a pointer to
7284 that type. If the ``cmpxchg`` is marked as ``volatile``, then the
7285 optimizer is not allowed to modify the number or order of execution of
7286 this ``cmpxchg`` with other :ref:`volatile operations <volatile>`.
7288 The success and failure :ref:`ordering <ordering>` arguments specify how this
7289 ``cmpxchg`` synchronizes with other atomic operations. Both ordering parameters
7290 must be at least ``monotonic``, the ordering constraint on failure must be no
7291 stronger than that on success, and the failure ordering cannot be either
7292 ``release`` or ``acq_rel``.
7294 The optional "``singlethread``" argument declares that the ``cmpxchg``
7295 is only atomic with respect to code (usually signal handlers) running in
7296 the same thread as the ``cmpxchg``. Otherwise the cmpxchg is atomic with
7297 respect to all other code in the system.
7299 The pointer passed into cmpxchg must have alignment greater than or
7300 equal to the size in memory of the operand.
7305 The contents of memory at the location specified by the '``<pointer>``' operand
7306 is read and compared to '``<cmp>``'; if the read value is the equal, the
7307 '``<new>``' is written. The original value at the location is returned, together
7308 with a flag indicating success (true) or failure (false).
7310 If the cmpxchg operation is marked as ``weak`` then a spurious failure is
7311 permitted: the operation may not write ``<new>`` even if the comparison
7314 If the cmpxchg operation is strong (the default), the i1 value is 1 if and only
7315 if the value loaded equals ``cmp``.
7317 A successful ``cmpxchg`` is a read-modify-write instruction for the purpose of
7318 identifying release sequences. A failed ``cmpxchg`` is equivalent to an atomic
7319 load with an ordering parameter determined the second ordering parameter.
7324 .. code-block:: llvm
7327 %orig = load atomic i32, i32* %ptr unordered, align 4 ; yields i32
7331 %cmp = phi i32 [ %orig, %entry ], [%value_loaded, %loop]
7332 %squared = mul i32 %cmp, %cmp
7333 %val_success = cmpxchg i32* %ptr, i32 %cmp, i32 %squared acq_rel monotonic ; yields { i32, i1 }
7334 %value_loaded = extractvalue { i32, i1 } %val_success, 0
7335 %success = extractvalue { i32, i1 } %val_success, 1
7336 br i1 %success, label %done, label %loop
7343 '``atomicrmw``' Instruction
7344 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7351 atomicrmw [volatile] <operation> <ty>* <pointer>, <ty> <value> [singlethread] <ordering> ; yields ty
7356 The '``atomicrmw``' instruction is used to atomically modify memory.
7361 There are three arguments to the '``atomicrmw``' instruction: an
7362 operation to apply, an address whose value to modify, an argument to the
7363 operation. The operation must be one of the following keywords:
7377 The type of '<value>' must be an integer type whose bit width is a power
7378 of two greater than or equal to eight and less than or equal to a
7379 target-specific size limit. The type of the '``<pointer>``' operand must
7380 be a pointer to that type. If the ``atomicrmw`` is marked as
7381 ``volatile``, then the optimizer is not allowed to modify the number or
7382 order of execution of this ``atomicrmw`` with other :ref:`volatile
7383 operations <volatile>`.
7388 The contents of memory at the location specified by the '``<pointer>``'
7389 operand are atomically read, modified, and written back. The original
7390 value at the location is returned. The modification is specified by the
7393 - xchg: ``*ptr = val``
7394 - add: ``*ptr = *ptr + val``
7395 - sub: ``*ptr = *ptr - val``
7396 - and: ``*ptr = *ptr & val``
7397 - nand: ``*ptr = ~(*ptr & val)``
7398 - or: ``*ptr = *ptr | val``
7399 - xor: ``*ptr = *ptr ^ val``
7400 - max: ``*ptr = *ptr > val ? *ptr : val`` (using a signed comparison)
7401 - min: ``*ptr = *ptr < val ? *ptr : val`` (using a signed comparison)
7402 - umax: ``*ptr = *ptr > val ? *ptr : val`` (using an unsigned
7404 - umin: ``*ptr = *ptr < val ? *ptr : val`` (using an unsigned
7410 .. code-block:: llvm
7412 %old = atomicrmw add i32* %ptr, i32 1 acquire ; yields i32
7414 .. _i_getelementptr:
7416 '``getelementptr``' Instruction
7417 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7424 <result> = getelementptr <ty>, <ty>* <ptrval>{, <ty> <idx>}*
7425 <result> = getelementptr inbounds <ty>, <ty>* <ptrval>{, <ty> <idx>}*
7426 <result> = getelementptr <ty>, <ptr vector> <ptrval>, <vector index type> <idx>
7431 The '``getelementptr``' instruction is used to get the address of a
7432 subelement of an :ref:`aggregate <t_aggregate>` data structure. It performs
7433 address calculation only and does not access memory. The instruction can also
7434 be used to calculate a vector of such addresses.
7439 The first argument is always a type used as the basis for the calculations.
7440 The second argument is always a pointer or a vector of pointers, and is the
7441 base address to start from. The remaining arguments are indices
7442 that indicate which of the elements of the aggregate object are indexed.
7443 The interpretation of each index is dependent on the type being indexed
7444 into. The first index always indexes the pointer value given as the
7445 first argument, the second index indexes a value of the type pointed to
7446 (not necessarily the value directly pointed to, since the first index
7447 can be non-zero), etc. The first type indexed into must be a pointer
7448 value, subsequent types can be arrays, vectors, and structs. Note that
7449 subsequent types being indexed into can never be pointers, since that
7450 would require loading the pointer before continuing calculation.
7452 The type of each index argument depends on the type it is indexing into.
7453 When indexing into a (optionally packed) structure, only ``i32`` integer
7454 **constants** are allowed (when using a vector of indices they must all
7455 be the **same** ``i32`` integer constant). When indexing into an array,
7456 pointer or vector, integers of any width are allowed, and they are not
7457 required to be constant. These integers are treated as signed values
7460 For example, let's consider a C code fragment and how it gets compiled
7476 int *foo(struct ST *s) {
7477 return &s[1].Z.B[5][13];
7480 The LLVM code generated by Clang is:
7482 .. code-block:: llvm
7484 %struct.RT = type { i8, [10 x [20 x i32]], i8 }
7485 %struct.ST = type { i32, double, %struct.RT }
7487 define i32* @foo(%struct.ST* %s) nounwind uwtable readnone optsize ssp {
7489 %arrayidx = getelementptr inbounds %struct.ST, %struct.ST* %s, i64 1, i32 2, i32 1, i64 5, i64 13
7496 In the example above, the first index is indexing into the
7497 '``%struct.ST*``' type, which is a pointer, yielding a '``%struct.ST``'
7498 = '``{ i32, double, %struct.RT }``' type, a structure. The second index
7499 indexes into the third element of the structure, yielding a
7500 '``%struct.RT``' = '``{ i8 , [10 x [20 x i32]], i8 }``' type, another
7501 structure. The third index indexes into the second element of the
7502 structure, yielding a '``[10 x [20 x i32]]``' type, an array. The two
7503 dimensions of the array are subscripted into, yielding an '``i32``'
7504 type. The '``getelementptr``' instruction returns a pointer to this
7505 element, thus computing a value of '``i32*``' type.
7507 Note that it is perfectly legal to index partially through a structure,
7508 returning a pointer to an inner element. Because of this, the LLVM code
7509 for the given testcase is equivalent to:
7511 .. code-block:: llvm
7513 define i32* @foo(%struct.ST* %s) {
7514 %t1 = getelementptr %struct.ST, %struct.ST* %s, i32 1 ; yields %struct.ST*:%t1
7515 %t2 = getelementptr %struct.ST, %struct.ST* %t1, i32 0, i32 2 ; yields %struct.RT*:%t2
7516 %t3 = getelementptr %struct.RT, %struct.RT* %t2, i32 0, i32 1 ; yields [10 x [20 x i32]]*:%t3
7517 %t4 = getelementptr [10 x [20 x i32]], [10 x [20 x i32]]* %t3, i32 0, i32 5 ; yields [20 x i32]*:%t4
7518 %t5 = getelementptr [20 x i32], [20 x i32]* %t4, i32 0, i32 13 ; yields i32*:%t5
7522 If the ``inbounds`` keyword is present, the result value of the
7523 ``getelementptr`` is a :ref:`poison value <poisonvalues>` if the base
7524 pointer is not an *in bounds* address of an allocated object, or if any
7525 of the addresses that would be formed by successive addition of the
7526 offsets implied by the indices to the base address with infinitely
7527 precise signed arithmetic are not an *in bounds* address of that
7528 allocated object. The *in bounds* addresses for an allocated object are
7529 all the addresses that point into the object, plus the address one byte
7530 past the end. In cases where the base is a vector of pointers the
7531 ``inbounds`` keyword applies to each of the computations element-wise.
7533 If the ``inbounds`` keyword is not present, the offsets are added to the
7534 base address with silently-wrapping two's complement arithmetic. If the
7535 offsets have a different width from the pointer, they are sign-extended
7536 or truncated to the width of the pointer. The result value of the
7537 ``getelementptr`` may be outside the object pointed to by the base
7538 pointer. The result value may not necessarily be used to access memory
7539 though, even if it happens to point into allocated storage. See the
7540 :ref:`Pointer Aliasing Rules <pointeraliasing>` section for more
7543 The getelementptr instruction is often confusing. For some more insight
7544 into how it works, see :doc:`the getelementptr FAQ <GetElementPtr>`.
7549 .. code-block:: llvm
7551 ; yields [12 x i8]*:aptr
7552 %aptr = getelementptr {i32, [12 x i8]}, {i32, [12 x i8]}* %saptr, i64 0, i32 1
7554 %vptr = getelementptr {i32, <2 x i8>}, {i32, <2 x i8>}* %svptr, i64 0, i32 1, i32 1
7556 %eptr = getelementptr [12 x i8], [12 x i8]* %aptr, i64 0, i32 1
7558 %iptr = getelementptr [10 x i32], [10 x i32]* @arr, i16 0, i16 0
7563 The ``getelementptr`` returns a vector of pointers, instead of a single address,
7564 when one or more of its arguments is a vector. In such cases, all vector
7565 arguments should have the same number of elements, and every scalar argument
7566 will be effectively broadcast into a vector during address calculation.
7568 .. code-block:: llvm
7570 ; All arguments are vectors:
7571 ; A[i] = ptrs[i] + offsets[i]*sizeof(i8)
7572 %A = getelementptr i8, <4 x i8*> %ptrs, <4 x i64> %offsets
7574 ; Add the same scalar offset to each pointer of a vector:
7575 ; A[i] = ptrs[i] + offset*sizeof(i8)
7576 %A = getelementptr i8, <4 x i8*> %ptrs, i64 %offset
7578 ; Add distinct offsets to the same pointer:
7579 ; A[i] = ptr + offsets[i]*sizeof(i8)
7580 %A = getelementptr i8, i8* %ptr, <4 x i64> %offsets
7582 ; In all cases described above the type of the result is <4 x i8*>
7584 The two following instructions are equivalent:
7586 .. code-block:: llvm
7588 getelementptr %struct.ST, <4 x %struct.ST*> %s, <4 x i64> %ind1,
7589 <4 x i32> <i32 2, i32 2, i32 2, i32 2>,
7590 <4 x i32> <i32 1, i32 1, i32 1, i32 1>,
7592 <4 x i64> <i64 13, i64 13, i64 13, i64 13>
7594 getelementptr %struct.ST, <4 x %struct.ST*> %s, <4 x i64> %ind1,
7595 i32 2, i32 1, <4 x i32> %ind4, i64 13
7597 Let's look at the C code, where the vector version of ``getelementptr``
7602 // Let's assume that we vectorize the following loop:
7603 double *A, B; int *C;
7604 for (int i = 0; i < size; ++i) {
7608 .. code-block:: llvm
7610 ; get pointers for 8 elements from array B
7611 %ptrs = getelementptr double, double* %B, <8 x i32> %C
7612 ; load 8 elements from array B into A
7613 %A = call <8 x double> @llvm.masked.gather.v8f64(<8 x double*> %ptrs,
7614 i32 8, <8 x i1> %mask, <8 x double> %passthru)
7616 Conversion Operations
7617 ---------------------
7619 The instructions in this category are the conversion instructions
7620 (casting) which all take a single operand and a type. They perform
7621 various bit conversions on the operand.
7623 '``trunc .. to``' Instruction
7624 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7631 <result> = trunc <ty> <value> to <ty2> ; yields ty2
7636 The '``trunc``' instruction truncates its operand to the type ``ty2``.
7641 The '``trunc``' instruction takes a value to trunc, and a type to trunc
7642 it to. Both types must be of :ref:`integer <t_integer>` types, or vectors
7643 of the same number of integers. The bit size of the ``value`` must be
7644 larger than the bit size of the destination type, ``ty2``. Equal sized
7645 types are not allowed.
7650 The '``trunc``' instruction truncates the high order bits in ``value``
7651 and converts the remaining bits to ``ty2``. Since the source size must
7652 be larger than the destination size, ``trunc`` cannot be a *no-op cast*.
7653 It will always truncate bits.
7658 .. code-block:: llvm
7660 %X = trunc i32 257 to i8 ; yields i8:1
7661 %Y = trunc i32 123 to i1 ; yields i1:true
7662 %Z = trunc i32 122 to i1 ; yields i1:false
7663 %W = trunc <2 x i16> <i16 8, i16 7> to <2 x i8> ; yields <i8 8, i8 7>
7665 '``zext .. to``' Instruction
7666 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7673 <result> = zext <ty> <value> to <ty2> ; yields ty2
7678 The '``zext``' instruction zero extends its operand to type ``ty2``.
7683 The '``zext``' instruction takes a value to cast, and a type to cast it
7684 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
7685 the same number of integers. The bit size of the ``value`` must be
7686 smaller than the bit size of the destination type, ``ty2``.
7691 The ``zext`` fills the high order bits of the ``value`` with zero bits
7692 until it reaches the size of the destination type, ``ty2``.
7694 When zero extending from i1, the result will always be either 0 or 1.
7699 .. code-block:: llvm
7701 %X = zext i32 257 to i64 ; yields i64:257
7702 %Y = zext i1 true to i32 ; yields i32:1
7703 %Z = zext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
7705 '``sext .. to``' Instruction
7706 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7713 <result> = sext <ty> <value> to <ty2> ; yields ty2
7718 The '``sext``' sign extends ``value`` to the type ``ty2``.
7723 The '``sext``' instruction takes a value to cast, and a type to cast it
7724 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
7725 the same number of integers. The bit size of the ``value`` must be
7726 smaller than the bit size of the destination type, ``ty2``.
7731 The '``sext``' instruction performs a sign extension by copying the sign
7732 bit (highest order bit) of the ``value`` until it reaches the bit size
7733 of the type ``ty2``.
7735 When sign extending from i1, the extension always results in -1 or 0.
7740 .. code-block:: llvm
7742 %X = sext i8 -1 to i16 ; yields i16 :65535
7743 %Y = sext i1 true to i32 ; yields i32:-1
7744 %Z = sext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
7746 '``fptrunc .. to``' Instruction
7747 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7754 <result> = fptrunc <ty> <value> to <ty2> ; yields ty2
7759 The '``fptrunc``' instruction truncates ``value`` to type ``ty2``.
7764 The '``fptrunc``' instruction takes a :ref:`floating point <t_floating>`
7765 value to cast and a :ref:`floating point <t_floating>` type to cast it to.
7766 The size of ``value`` must be larger than the size of ``ty2``. This
7767 implies that ``fptrunc`` cannot be used to make a *no-op cast*.
7772 The '``fptrunc``' instruction casts a ``value`` from a larger
7773 :ref:`floating point <t_floating>` type to a smaller :ref:`floating
7774 point <t_floating>` type. If the value cannot fit (i.e. overflows) within the
7775 destination type, ``ty2``, then the results are undefined. If the cast produces
7776 an inexact result, how rounding is performed (e.g. truncation, also known as
7777 round to zero) is undefined.
7782 .. code-block:: llvm
7784 %X = fptrunc double 123.0 to float ; yields float:123.0
7785 %Y = fptrunc double 1.0E+300 to float ; yields undefined
7787 '``fpext .. to``' Instruction
7788 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7795 <result> = fpext <ty> <value> to <ty2> ; yields ty2
7800 The '``fpext``' extends a floating point ``value`` to a larger floating
7806 The '``fpext``' instruction takes a :ref:`floating point <t_floating>`
7807 ``value`` to cast, and a :ref:`floating point <t_floating>` type to cast it
7808 to. The source type must be smaller than the destination type.
7813 The '``fpext``' instruction extends the ``value`` from a smaller
7814 :ref:`floating point <t_floating>` type to a larger :ref:`floating
7815 point <t_floating>` type. The ``fpext`` cannot be used to make a
7816 *no-op cast* because it always changes bits. Use ``bitcast`` to make a
7817 *no-op cast* for a floating point cast.
7822 .. code-block:: llvm
7824 %X = fpext float 3.125 to double ; yields double:3.125000e+00
7825 %Y = fpext double %X to fp128 ; yields fp128:0xL00000000000000004000900000000000
7827 '``fptoui .. to``' Instruction
7828 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7835 <result> = fptoui <ty> <value> to <ty2> ; yields ty2
7840 The '``fptoui``' converts a floating point ``value`` to its unsigned
7841 integer equivalent of type ``ty2``.
7846 The '``fptoui``' instruction takes a value to cast, which must be a
7847 scalar or vector :ref:`floating point <t_floating>` value, and a type to
7848 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
7849 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
7850 type with the same number of elements as ``ty``
7855 The '``fptoui``' instruction converts its :ref:`floating
7856 point <t_floating>` operand into the nearest (rounding towards zero)
7857 unsigned integer value. If the value cannot fit in ``ty2``, the results
7863 .. code-block:: llvm
7865 %X = fptoui double 123.0 to i32 ; yields i32:123
7866 %Y = fptoui float 1.0E+300 to i1 ; yields undefined:1
7867 %Z = fptoui float 1.04E+17 to i8 ; yields undefined:1
7869 '``fptosi .. to``' Instruction
7870 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7877 <result> = fptosi <ty> <value> to <ty2> ; yields ty2
7882 The '``fptosi``' instruction converts :ref:`floating point <t_floating>`
7883 ``value`` to type ``ty2``.
7888 The '``fptosi``' instruction takes a value to cast, which must be a
7889 scalar or vector :ref:`floating point <t_floating>` value, and a type to
7890 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
7891 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
7892 type with the same number of elements as ``ty``
7897 The '``fptosi``' instruction converts its :ref:`floating
7898 point <t_floating>` operand into the nearest (rounding towards zero)
7899 signed integer value. If the value cannot fit in ``ty2``, the results
7905 .. code-block:: llvm
7907 %X = fptosi double -123.0 to i32 ; yields i32:-123
7908 %Y = fptosi float 1.0E-247 to i1 ; yields undefined:1
7909 %Z = fptosi float 1.04E+17 to i8 ; yields undefined:1
7911 '``uitofp .. to``' Instruction
7912 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7919 <result> = uitofp <ty> <value> to <ty2> ; yields ty2
7924 The '``uitofp``' instruction regards ``value`` as an unsigned integer
7925 and converts that value to the ``ty2`` type.
7930 The '``uitofp``' instruction takes a value to cast, which must be a
7931 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
7932 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
7933 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
7934 type with the same number of elements as ``ty``
7939 The '``uitofp``' instruction interprets its operand as an unsigned
7940 integer quantity and converts it to the corresponding floating point
7941 value. If the value cannot fit in the floating point value, the results
7947 .. code-block:: llvm
7949 %X = uitofp i32 257 to float ; yields float:257.0
7950 %Y = uitofp i8 -1 to double ; yields double:255.0
7952 '``sitofp .. to``' Instruction
7953 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7960 <result> = sitofp <ty> <value> to <ty2> ; yields ty2
7965 The '``sitofp``' instruction regards ``value`` as a signed integer and
7966 converts that value to the ``ty2`` type.
7971 The '``sitofp``' instruction takes a value to cast, which must be a
7972 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
7973 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
7974 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
7975 type with the same number of elements as ``ty``
7980 The '``sitofp``' instruction interprets its operand as a signed integer
7981 quantity and converts it to the corresponding floating point value. If
7982 the value cannot fit in the floating point value, the results are
7988 .. code-block:: llvm
7990 %X = sitofp i32 257 to float ; yields float:257.0
7991 %Y = sitofp i8 -1 to double ; yields double:-1.0
7995 '``ptrtoint .. to``' Instruction
7996 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8003 <result> = ptrtoint <ty> <value> to <ty2> ; yields ty2
8008 The '``ptrtoint``' instruction converts the pointer or a vector of
8009 pointers ``value`` to the integer (or vector of integers) type ``ty2``.
8014 The '``ptrtoint``' instruction takes a ``value`` to cast, which must be
8015 a value of type :ref:`pointer <t_pointer>` or a vector of pointers, and a
8016 type to cast it to ``ty2``, which must be an :ref:`integer <t_integer>` or
8017 a vector of integers type.
8022 The '``ptrtoint``' instruction converts ``value`` to integer type
8023 ``ty2`` by interpreting the pointer value as an integer and either
8024 truncating or zero extending that value to the size of the integer type.
8025 If ``value`` is smaller than ``ty2`` then a zero extension is done. If
8026 ``value`` is larger than ``ty2`` then a truncation is done. If they are
8027 the same size, then nothing is done (*no-op cast*) other than a type
8033 .. code-block:: llvm
8035 %X = ptrtoint i32* %P to i8 ; yields truncation on 32-bit architecture
8036 %Y = ptrtoint i32* %P to i64 ; yields zero extension on 32-bit architecture
8037 %Z = ptrtoint <4 x i32*> %P to <4 x i64>; yields vector zero extension for a vector of addresses on 32-bit architecture
8041 '``inttoptr .. to``' Instruction
8042 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8049 <result> = inttoptr <ty> <value> to <ty2> ; yields ty2
8054 The '``inttoptr``' instruction converts an integer ``value`` to a
8055 pointer type, ``ty2``.
8060 The '``inttoptr``' instruction takes an :ref:`integer <t_integer>` value to
8061 cast, and a type to cast it to, which must be a :ref:`pointer <t_pointer>`
8067 The '``inttoptr``' instruction converts ``value`` to type ``ty2`` by
8068 applying either a zero extension or a truncation depending on the size
8069 of the integer ``value``. If ``value`` is larger than the size of a
8070 pointer then a truncation is done. If ``value`` is smaller than the size
8071 of a pointer then a zero extension is done. If they are the same size,
8072 nothing is done (*no-op cast*).
8077 .. code-block:: llvm
8079 %X = inttoptr i32 255 to i32* ; yields zero extension on 64-bit architecture
8080 %Y = inttoptr i32 255 to i32* ; yields no-op on 32-bit architecture
8081 %Z = inttoptr i64 0 to i32* ; yields truncation on 32-bit architecture
8082 %Z = inttoptr <4 x i32> %G to <4 x i8*>; yields truncation of vector G to four pointers
8086 '``bitcast .. to``' Instruction
8087 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8094 <result> = bitcast <ty> <value> to <ty2> ; yields ty2
8099 The '``bitcast``' instruction converts ``value`` to type ``ty2`` without
8105 The '``bitcast``' instruction takes a value to cast, which must be a
8106 non-aggregate first class value, and a type to cast it to, which must
8107 also be a non-aggregate :ref:`first class <t_firstclass>` type. The
8108 bit sizes of ``value`` and the destination type, ``ty2``, must be
8109 identical. If the source type is a pointer, the destination type must
8110 also be a pointer of the same size. This instruction supports bitwise
8111 conversion of vectors to integers and to vectors of other types (as
8112 long as they have the same size).
8117 The '``bitcast``' instruction converts ``value`` to type ``ty2``. It
8118 is always a *no-op cast* because no bits change with this
8119 conversion. The conversion is done as if the ``value`` had been stored
8120 to memory and read back as type ``ty2``. Pointer (or vector of
8121 pointers) types may only be converted to other pointer (or vector of
8122 pointers) types with the same address space through this instruction.
8123 To convert pointers to other types, use the :ref:`inttoptr <i_inttoptr>`
8124 or :ref:`ptrtoint <i_ptrtoint>` instructions first.
8129 .. code-block:: text
8131 %X = bitcast i8 255 to i8 ; yields i8 :-1
8132 %Y = bitcast i32* %x to sint* ; yields sint*:%x
8133 %Z = bitcast <2 x int> %V to i64; ; yields i64: %V
8134 %Z = bitcast <2 x i32*> %V to <2 x i64*> ; yields <2 x i64*>
8136 .. _i_addrspacecast:
8138 '``addrspacecast .. to``' Instruction
8139 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8146 <result> = addrspacecast <pty> <ptrval> to <pty2> ; yields pty2
8151 The '``addrspacecast``' instruction converts ``ptrval`` from ``pty`` in
8152 address space ``n`` to type ``pty2`` in address space ``m``.
8157 The '``addrspacecast``' instruction takes a pointer or vector of pointer value
8158 to cast and a pointer type to cast it to, which must have a different
8164 The '``addrspacecast``' instruction converts the pointer value
8165 ``ptrval`` to type ``pty2``. It can be a *no-op cast* or a complex
8166 value modification, depending on the target and the address space
8167 pair. Pointer conversions within the same address space must be
8168 performed with the ``bitcast`` instruction. Note that if the address space
8169 conversion is legal then both result and operand refer to the same memory
8175 .. code-block:: llvm
8177 %X = addrspacecast i32* %x to i32 addrspace(1)* ; yields i32 addrspace(1)*:%x
8178 %Y = addrspacecast i32 addrspace(1)* %y to i64 addrspace(2)* ; yields i64 addrspace(2)*:%y
8179 %Z = addrspacecast <4 x i32*> %z to <4 x float addrspace(3)*> ; yields <4 x float addrspace(3)*>:%z
8186 The instructions in this category are the "miscellaneous" instructions,
8187 which defy better classification.
8191 '``icmp``' Instruction
8192 ^^^^^^^^^^^^^^^^^^^^^^
8199 <result> = icmp <cond> <ty> <op1>, <op2> ; yields i1 or <N x i1>:result
8204 The '``icmp``' instruction returns a boolean value or a vector of
8205 boolean values based on comparison of its two integer, integer vector,
8206 pointer, or pointer vector operands.
8211 The '``icmp``' instruction takes three operands. The first operand is
8212 the condition code indicating the kind of comparison to perform. It is
8213 not a value, just a keyword. The possible condition codes are:
8216 #. ``ne``: not equal
8217 #. ``ugt``: unsigned greater than
8218 #. ``uge``: unsigned greater or equal
8219 #. ``ult``: unsigned less than
8220 #. ``ule``: unsigned less or equal
8221 #. ``sgt``: signed greater than
8222 #. ``sge``: signed greater or equal
8223 #. ``slt``: signed less than
8224 #. ``sle``: signed less or equal
8226 The remaining two arguments must be :ref:`integer <t_integer>` or
8227 :ref:`pointer <t_pointer>` or integer :ref:`vector <t_vector>` typed. They
8228 must also be identical types.
8233 The '``icmp``' compares ``op1`` and ``op2`` according to the condition
8234 code given as ``cond``. The comparison performed always yields either an
8235 :ref:`i1 <t_integer>` or vector of ``i1`` result, as follows:
8237 #. ``eq``: yields ``true`` if the operands are equal, ``false``
8238 otherwise. No sign interpretation is necessary or performed.
8239 #. ``ne``: yields ``true`` if the operands are unequal, ``false``
8240 otherwise. No sign interpretation is necessary or performed.
8241 #. ``ugt``: interprets the operands as unsigned values and yields
8242 ``true`` if ``op1`` is greater than ``op2``.
8243 #. ``uge``: interprets the operands as unsigned values and yields
8244 ``true`` if ``op1`` is greater than or equal to ``op2``.
8245 #. ``ult``: interprets the operands as unsigned values and yields
8246 ``true`` if ``op1`` is less than ``op2``.
8247 #. ``ule``: interprets the operands as unsigned values and yields
8248 ``true`` if ``op1`` is less than or equal to ``op2``.
8249 #. ``sgt``: interprets the operands as signed values and yields ``true``
8250 if ``op1`` is greater than ``op2``.
8251 #. ``sge``: interprets the operands as signed values and yields ``true``
8252 if ``op1`` is greater than or equal to ``op2``.
8253 #. ``slt``: interprets the operands as signed values and yields ``true``
8254 if ``op1`` is less than ``op2``.
8255 #. ``sle``: interprets the operands as signed values and yields ``true``
8256 if ``op1`` is less than or equal to ``op2``.
8258 If the operands are :ref:`pointer <t_pointer>` typed, the pointer values
8259 are compared as if they were integers.
8261 If the operands are integer vectors, then they are compared element by
8262 element. The result is an ``i1`` vector with the same number of elements
8263 as the values being compared. Otherwise, the result is an ``i1``.
8268 .. code-block:: text
8270 <result> = icmp eq i32 4, 5 ; yields: result=false
8271 <result> = icmp ne float* %X, %X ; yields: result=false
8272 <result> = icmp ult i16 4, 5 ; yields: result=true
8273 <result> = icmp sgt i16 4, 5 ; yields: result=false
8274 <result> = icmp ule i16 -4, 5 ; yields: result=false
8275 <result> = icmp sge i16 4, 5 ; yields: result=false
8279 '``fcmp``' Instruction
8280 ^^^^^^^^^^^^^^^^^^^^^^
8287 <result> = fcmp [fast-math flags]* <cond> <ty> <op1>, <op2> ; yields i1 or <N x i1>:result
8292 The '``fcmp``' instruction returns a boolean value or vector of boolean
8293 values based on comparison of its operands.
8295 If the operands are floating point scalars, then the result type is a
8296 boolean (:ref:`i1 <t_integer>`).
8298 If the operands are floating point vectors, then the result type is a
8299 vector of boolean with the same number of elements as the operands being
8305 The '``fcmp``' instruction takes three operands. The first operand is
8306 the condition code indicating the kind of comparison to perform. It is
8307 not a value, just a keyword. The possible condition codes are:
8309 #. ``false``: no comparison, always returns false
8310 #. ``oeq``: ordered and equal
8311 #. ``ogt``: ordered and greater than
8312 #. ``oge``: ordered and greater than or equal
8313 #. ``olt``: ordered and less than
8314 #. ``ole``: ordered and less than or equal
8315 #. ``one``: ordered and not equal
8316 #. ``ord``: ordered (no nans)
8317 #. ``ueq``: unordered or equal
8318 #. ``ugt``: unordered or greater than
8319 #. ``uge``: unordered or greater than or equal
8320 #. ``ult``: unordered or less than
8321 #. ``ule``: unordered or less than or equal
8322 #. ``une``: unordered or not equal
8323 #. ``uno``: unordered (either nans)
8324 #. ``true``: no comparison, always returns true
8326 *Ordered* means that neither operand is a QNAN while *unordered* means
8327 that either operand may be a QNAN.
8329 Each of ``val1`` and ``val2`` arguments must be either a :ref:`floating
8330 point <t_floating>` type or a :ref:`vector <t_vector>` of floating point
8331 type. They must have identical types.
8336 The '``fcmp``' instruction compares ``op1`` and ``op2`` according to the
8337 condition code given as ``cond``. If the operands are vectors, then the
8338 vectors are compared element by element. Each comparison performed
8339 always yields an :ref:`i1 <t_integer>` result, as follows:
8341 #. ``false``: always yields ``false``, regardless of operands.
8342 #. ``oeq``: yields ``true`` if both operands are not a QNAN and ``op1``
8343 is equal to ``op2``.
8344 #. ``ogt``: yields ``true`` if both operands are not a QNAN and ``op1``
8345 is greater than ``op2``.
8346 #. ``oge``: yields ``true`` if both operands are not a QNAN and ``op1``
8347 is greater than or equal to ``op2``.
8348 #. ``olt``: yields ``true`` if both operands are not a QNAN and ``op1``
8349 is less than ``op2``.
8350 #. ``ole``: yields ``true`` if both operands are not a QNAN and ``op1``
8351 is less than or equal to ``op2``.
8352 #. ``one``: yields ``true`` if both operands are not a QNAN and ``op1``
8353 is not equal to ``op2``.
8354 #. ``ord``: yields ``true`` if both operands are not a QNAN.
8355 #. ``ueq``: yields ``true`` if either operand is a QNAN or ``op1`` is
8357 #. ``ugt``: yields ``true`` if either operand is a QNAN or ``op1`` is
8358 greater than ``op2``.
8359 #. ``uge``: yields ``true`` if either operand is a QNAN or ``op1`` is
8360 greater than or equal to ``op2``.
8361 #. ``ult``: yields ``true`` if either operand is a QNAN or ``op1`` is
8363 #. ``ule``: yields ``true`` if either operand is a QNAN or ``op1`` is
8364 less than or equal to ``op2``.
8365 #. ``une``: yields ``true`` if either operand is a QNAN or ``op1`` is
8366 not equal to ``op2``.
8367 #. ``uno``: yields ``true`` if either operand is a QNAN.
8368 #. ``true``: always yields ``true``, regardless of operands.
8370 The ``fcmp`` instruction can also optionally take any number of
8371 :ref:`fast-math flags <fastmath>`, which are optimization hints to enable
8372 otherwise unsafe floating point optimizations.
8374 Any set of fast-math flags are legal on an ``fcmp`` instruction, but the
8375 only flags that have any effect on its semantics are those that allow
8376 assumptions to be made about the values of input arguments; namely
8377 ``nnan``, ``ninf``, and ``nsz``. See :ref:`fastmath` for more information.
8382 .. code-block:: text
8384 <result> = fcmp oeq float 4.0, 5.0 ; yields: result=false
8385 <result> = fcmp one float 4.0, 5.0 ; yields: result=true
8386 <result> = fcmp olt float 4.0, 5.0 ; yields: result=true
8387 <result> = fcmp ueq double 1.0, 2.0 ; yields: result=false
8391 '``phi``' Instruction
8392 ^^^^^^^^^^^^^^^^^^^^^
8399 <result> = phi <ty> [ <val0>, <label0>], ...
8404 The '``phi``' instruction is used to implement the φ node in the SSA
8405 graph representing the function.
8410 The type of the incoming values is specified with the first type field.
8411 After this, the '``phi``' instruction takes a list of pairs as
8412 arguments, with one pair for each predecessor basic block of the current
8413 block. Only values of :ref:`first class <t_firstclass>` type may be used as
8414 the value arguments to the PHI node. Only labels may be used as the
8417 There must be no non-phi instructions between the start of a basic block
8418 and the PHI instructions: i.e. PHI instructions must be first in a basic
8421 For the purposes of the SSA form, the use of each incoming value is
8422 deemed to occur on the edge from the corresponding predecessor block to
8423 the current block (but after any definition of an '``invoke``'
8424 instruction's return value on the same edge).
8429 At runtime, the '``phi``' instruction logically takes on the value
8430 specified by the pair corresponding to the predecessor basic block that
8431 executed just prior to the current block.
8436 .. code-block:: llvm
8438 Loop: ; Infinite loop that counts from 0 on up...
8439 %indvar = phi i32 [ 0, %LoopHeader ], [ %nextindvar, %Loop ]
8440 %nextindvar = add i32 %indvar, 1
8445 '``select``' Instruction
8446 ^^^^^^^^^^^^^^^^^^^^^^^^
8453 <result> = select selty <cond>, <ty> <val1>, <ty> <val2> ; yields ty
8455 selty is either i1 or {<N x i1>}
8460 The '``select``' instruction is used to choose one value based on a
8461 condition, without IR-level branching.
8466 The '``select``' instruction requires an 'i1' value or a vector of 'i1'
8467 values indicating the condition, and two values of the same :ref:`first
8468 class <t_firstclass>` type.
8473 If the condition is an i1 and it evaluates to 1, the instruction returns
8474 the first value argument; otherwise, it returns the second value
8477 If the condition is a vector of i1, then the value arguments must be
8478 vectors of the same size, and the selection is done element by element.
8480 If the condition is an i1 and the value arguments are vectors of the
8481 same size, then an entire vector is selected.
8486 .. code-block:: llvm
8488 %X = select i1 true, i8 17, i8 42 ; yields i8:17
8492 '``call``' Instruction
8493 ^^^^^^^^^^^^^^^^^^^^^^
8500 <result> = [tail | musttail | notail ] call [fast-math flags] [cconv] [ret attrs] <ty>|<fnty> <fnptrval>(<function args>) [fn attrs]
8506 The '``call``' instruction represents a simple function call.
8511 This instruction requires several arguments:
8513 #. The optional ``tail`` and ``musttail`` markers indicate that the optimizers
8514 should perform tail call optimization. The ``tail`` marker is a hint that
8515 `can be ignored <CodeGenerator.html#sibcallopt>`_. The ``musttail`` marker
8516 means that the call must be tail call optimized in order for the program to
8517 be correct. The ``musttail`` marker provides these guarantees:
8519 #. The call will not cause unbounded stack growth if it is part of a
8520 recursive cycle in the call graph.
8521 #. Arguments with the :ref:`inalloca <attr_inalloca>` attribute are
8524 Both markers imply that the callee does not access allocas or varargs from
8525 the caller. Calls marked ``musttail`` must obey the following additional
8528 - The call must immediately precede a :ref:`ret <i_ret>` instruction,
8529 or a pointer bitcast followed by a ret instruction.
8530 - The ret instruction must return the (possibly bitcasted) value
8531 produced by the call or void.
8532 - The caller and callee prototypes must match. Pointer types of
8533 parameters or return types may differ in pointee type, but not
8535 - The calling conventions of the caller and callee must match.
8536 - All ABI-impacting function attributes, such as sret, byval, inreg,
8537 returned, and inalloca, must match.
8538 - The callee must be varargs iff the caller is varargs. Bitcasting a
8539 non-varargs function to the appropriate varargs type is legal so
8540 long as the non-varargs prefixes obey the other rules.
8542 Tail call optimization for calls marked ``tail`` is guaranteed to occur if
8543 the following conditions are met:
8545 - Caller and callee both have the calling convention ``fastcc``.
8546 - The call is in tail position (ret immediately follows call and ret
8547 uses value of call or is void).
8548 - Option ``-tailcallopt`` is enabled, or
8549 ``llvm::GuaranteedTailCallOpt`` is ``true``.
8550 - `Platform-specific constraints are
8551 met. <CodeGenerator.html#tailcallopt>`_
8553 #. The optional ``notail`` marker indicates that the optimizers should not add
8554 ``tail`` or ``musttail`` markers to the call. It is used to prevent tail
8555 call optimization from being performed on the call.
8557 #. The optional ``fast-math flags`` marker indicates that the call has one or more
8558 :ref:`fast-math flags <fastmath>`, which are optimization hints to enable
8559 otherwise unsafe floating-point optimizations. Fast-math flags are only valid
8560 for calls that return a floating-point scalar or vector type.
8562 #. The optional "cconv" marker indicates which :ref:`calling
8563 convention <callingconv>` the call should use. If none is
8564 specified, the call defaults to using C calling conventions. The
8565 calling convention of the call must match the calling convention of
8566 the target function, or else the behavior is undefined.
8567 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
8568 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
8570 #. '``ty``': the type of the call instruction itself which is also the
8571 type of the return value. Functions that return no value are marked
8573 #. '``fnty``': shall be the signature of the function being called. The
8574 argument types must match the types implied by this signature. This
8575 type can be omitted if the function is not varargs.
8576 #. '``fnptrval``': An LLVM value containing a pointer to a function to
8577 be called. In most cases, this is a direct function call, but
8578 indirect ``call``'s are just as possible, calling an arbitrary pointer
8580 #. '``function args``': argument list whose types match the function
8581 signature argument types and parameter attributes. All arguments must
8582 be of :ref:`first class <t_firstclass>` type. If the function signature
8583 indicates the function accepts a variable number of arguments, the
8584 extra arguments can be specified.
8585 #. The optional :ref:`function attributes <fnattrs>` list. Only
8586 '``noreturn``', '``nounwind``', '``readonly``' , '``readnone``',
8587 and '``convergent``' attributes are valid here.
8588 #. The optional :ref:`operand bundles <opbundles>` list.
8593 The '``call``' instruction is used to cause control flow to transfer to
8594 a specified function, with its incoming arguments bound to the specified
8595 values. Upon a '``ret``' instruction in the called function, control
8596 flow continues with the instruction after the function call, and the
8597 return value of the function is bound to the result argument.
8602 .. code-block:: llvm
8604 %retval = call i32 @test(i32 %argc)
8605 call i32 (i8*, ...)* @printf(i8* %msg, i32 12, i8 42) ; yields i32
8606 %X = tail call i32 @foo() ; yields i32
8607 %Y = tail call fastcc i32 @foo() ; yields i32
8608 call void %foo(i8 97 signext)
8610 %struct.A = type { i32, i8 }
8611 %r = call %struct.A @foo() ; yields { i32, i8 }
8612 %gr = extractvalue %struct.A %r, 0 ; yields i32
8613 %gr1 = extractvalue %struct.A %r, 1 ; yields i8
8614 %Z = call void @foo() noreturn ; indicates that %foo never returns normally
8615 %ZZ = call zeroext i32 @bar() ; Return value is %zero extended
8617 llvm treats calls to some functions with names and arguments that match
8618 the standard C99 library as being the C99 library functions, and may
8619 perform optimizations or generate code for them under that assumption.
8620 This is something we'd like to change in the future to provide better
8621 support for freestanding environments and non-C-based languages.
8625 '``va_arg``' Instruction
8626 ^^^^^^^^^^^^^^^^^^^^^^^^
8633 <resultval> = va_arg <va_list*> <arglist>, <argty>
8638 The '``va_arg``' instruction is used to access arguments passed through
8639 the "variable argument" area of a function call. It is used to implement
8640 the ``va_arg`` macro in C.
8645 This instruction takes a ``va_list*`` value and the type of the
8646 argument. It returns a value of the specified argument type and
8647 increments the ``va_list`` to point to the next argument. The actual
8648 type of ``va_list`` is target specific.
8653 The '``va_arg``' instruction loads an argument of the specified type
8654 from the specified ``va_list`` and causes the ``va_list`` to point to
8655 the next argument. For more information, see the variable argument
8656 handling :ref:`Intrinsic Functions <int_varargs>`.
8658 It is legal for this instruction to be called in a function which does
8659 not take a variable number of arguments, for example, the ``vfprintf``
8662 ``va_arg`` is an LLVM instruction instead of an :ref:`intrinsic
8663 function <intrinsics>` because it takes a type as an argument.
8668 See the :ref:`variable argument processing <int_varargs>` section.
8670 Note that the code generator does not yet fully support va\_arg on many
8671 targets. Also, it does not currently support va\_arg with aggregate
8672 types on any target.
8676 '``landingpad``' Instruction
8677 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8684 <resultval> = landingpad <resultty> <clause>+
8685 <resultval> = landingpad <resultty> cleanup <clause>*
8687 <clause> := catch <type> <value>
8688 <clause> := filter <array constant type> <array constant>
8693 The '``landingpad``' instruction is used by `LLVM's exception handling
8694 system <ExceptionHandling.html#overview>`_ to specify that a basic block
8695 is a landing pad --- one where the exception lands, and corresponds to the
8696 code found in the ``catch`` portion of a ``try``/``catch`` sequence. It
8697 defines values supplied by the :ref:`personality function <personalityfn>` upon
8698 re-entry to the function. The ``resultval`` has the type ``resultty``.
8704 ``cleanup`` flag indicates that the landing pad block is a cleanup.
8706 A ``clause`` begins with the clause type --- ``catch`` or ``filter`` --- and
8707 contains the global variable representing the "type" that may be caught
8708 or filtered respectively. Unlike the ``catch`` clause, the ``filter``
8709 clause takes an array constant as its argument. Use
8710 "``[0 x i8**] undef``" for a filter which cannot throw. The
8711 '``landingpad``' instruction must contain *at least* one ``clause`` or
8712 the ``cleanup`` flag.
8717 The '``landingpad``' instruction defines the values which are set by the
8718 :ref:`personality function <personalityfn>` upon re-entry to the function, and
8719 therefore the "result type" of the ``landingpad`` instruction. As with
8720 calling conventions, how the personality function results are
8721 represented in LLVM IR is target specific.
8723 The clauses are applied in order from top to bottom. If two
8724 ``landingpad`` instructions are merged together through inlining, the
8725 clauses from the calling function are appended to the list of clauses.
8726 When the call stack is being unwound due to an exception being thrown,
8727 the exception is compared against each ``clause`` in turn. If it doesn't
8728 match any of the clauses, and the ``cleanup`` flag is not set, then
8729 unwinding continues further up the call stack.
8731 The ``landingpad`` instruction has several restrictions:
8733 - A landing pad block is a basic block which is the unwind destination
8734 of an '``invoke``' instruction.
8735 - A landing pad block must have a '``landingpad``' instruction as its
8736 first non-PHI instruction.
8737 - There can be only one '``landingpad``' instruction within the landing
8739 - A basic block that is not a landing pad block may not include a
8740 '``landingpad``' instruction.
8745 .. code-block:: llvm
8747 ;; A landing pad which can catch an integer.
8748 %res = landingpad { i8*, i32 }
8750 ;; A landing pad that is a cleanup.
8751 %res = landingpad { i8*, i32 }
8753 ;; A landing pad which can catch an integer and can only throw a double.
8754 %res = landingpad { i8*, i32 }
8756 filter [1 x i8**] [@_ZTId]
8760 '``catchpad``' Instruction
8761 ^^^^^^^^^^^^^^^^^^^^^^^^^^
8768 <resultval> = catchpad within <catchswitch> [<args>*]
8773 The '``catchpad``' instruction is used by `LLVM's exception handling
8774 system <ExceptionHandling.html#overview>`_ to specify that a basic block
8775 begins a catch handler --- one where a personality routine attempts to transfer
8776 control to catch an exception.
8781 The ``catchswitch`` operand must always be a token produced by a
8782 :ref:`catchswitch <i_catchswitch>` instruction in a predecessor block. This
8783 ensures that each ``catchpad`` has exactly one predecessor block, and it always
8784 terminates in a ``catchswitch``.
8786 The ``args`` correspond to whatever information the personality routine
8787 requires to know if this is an appropriate handler for the exception. Control
8788 will transfer to the ``catchpad`` if this is the first appropriate handler for
8791 The ``resultval`` has the type :ref:`token <t_token>` and is used to match the
8792 ``catchpad`` to corresponding :ref:`catchrets <i_catchret>` and other nested EH
8798 When the call stack is being unwound due to an exception being thrown, the
8799 exception is compared against the ``args``. If it doesn't match, control will
8800 not reach the ``catchpad`` instruction. The representation of ``args`` is
8801 entirely target and personality function-specific.
8803 Like the :ref:`landingpad <i_landingpad>` instruction, the ``catchpad``
8804 instruction must be the first non-phi of its parent basic block.
8806 The meaning of the tokens produced and consumed by ``catchpad`` and other "pad"
8807 instructions is described in the
8808 `Windows exception handling documentation\ <ExceptionHandling.html#wineh>`_.
8810 When a ``catchpad`` has been "entered" but not yet "exited" (as
8811 described in the `EH documentation\ <ExceptionHandling.html#wineh-constraints>`_),
8812 it is undefined behavior to execute a :ref:`call <i_call>` or :ref:`invoke <i_invoke>`
8813 that does not carry an appropriate :ref:`"funclet" bundle <ob_funclet>`.
8818 .. code-block:: text
8821 %cs = catchswitch within none [label %handler0] unwind to caller
8822 ;; A catch block which can catch an integer.
8824 %tok = catchpad within %cs [i8** @_ZTIi]
8828 '``cleanuppad``' Instruction
8829 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8836 <resultval> = cleanuppad within <parent> [<args>*]
8841 The '``cleanuppad``' instruction is used by `LLVM's exception handling
8842 system <ExceptionHandling.html#overview>`_ to specify that a basic block
8843 is a cleanup block --- one where a personality routine attempts to
8844 transfer control to run cleanup actions.
8845 The ``args`` correspond to whatever additional
8846 information the :ref:`personality function <personalityfn>` requires to
8847 execute the cleanup.
8848 The ``resultval`` has the type :ref:`token <t_token>` and is used to
8849 match the ``cleanuppad`` to corresponding :ref:`cleanuprets <i_cleanupret>`.
8850 The ``parent`` argument is the token of the funclet that contains the
8851 ``cleanuppad`` instruction. If the ``cleanuppad`` is not inside a funclet,
8852 this operand may be the token ``none``.
8857 The instruction takes a list of arbitrary values which are interpreted
8858 by the :ref:`personality function <personalityfn>`.
8863 When the call stack is being unwound due to an exception being thrown,
8864 the :ref:`personality function <personalityfn>` transfers control to the
8865 ``cleanuppad`` with the aid of the personality-specific arguments.
8866 As with calling conventions, how the personality function results are
8867 represented in LLVM IR is target specific.
8869 The ``cleanuppad`` instruction has several restrictions:
8871 - A cleanup block is a basic block which is the unwind destination of
8872 an exceptional instruction.
8873 - A cleanup block must have a '``cleanuppad``' instruction as its
8874 first non-PHI instruction.
8875 - There can be only one '``cleanuppad``' instruction within the
8877 - A basic block that is not a cleanup block may not include a
8878 '``cleanuppad``' instruction.
8880 When a ``cleanuppad`` has been "entered" but not yet "exited" (as
8881 described in the `EH documentation\ <ExceptionHandling.html#wineh-constraints>`_),
8882 it is undefined behavior to execute a :ref:`call <i_call>` or :ref:`invoke <i_invoke>`
8883 that does not carry an appropriate :ref:`"funclet" bundle <ob_funclet>`.
8888 .. code-block:: text
8890 %tok = cleanuppad within %cs []
8897 LLVM supports the notion of an "intrinsic function". These functions
8898 have well known names and semantics and are required to follow certain
8899 restrictions. Overall, these intrinsics represent an extension mechanism
8900 for the LLVM language that does not require changing all of the
8901 transformations in LLVM when adding to the language (or the bitcode
8902 reader/writer, the parser, etc...).
8904 Intrinsic function names must all start with an "``llvm.``" prefix. This
8905 prefix is reserved in LLVM for intrinsic names; thus, function names may
8906 not begin with this prefix. Intrinsic functions must always be external
8907 functions: you cannot define the body of intrinsic functions. Intrinsic
8908 functions may only be used in call or invoke instructions: it is illegal
8909 to take the address of an intrinsic function. Additionally, because
8910 intrinsic functions are part of the LLVM language, it is required if any
8911 are added that they be documented here.
8913 Some intrinsic functions can be overloaded, i.e., the intrinsic
8914 represents a family of functions that perform the same operation but on
8915 different data types. Because LLVM can represent over 8 million
8916 different integer types, overloading is used commonly to allow an
8917 intrinsic function to operate on any integer type. One or more of the
8918 argument types or the result type can be overloaded to accept any
8919 integer type. Argument types may also be defined as exactly matching a
8920 previous argument's type or the result type. This allows an intrinsic
8921 function which accepts multiple arguments, but needs all of them to be
8922 of the same type, to only be overloaded with respect to a single
8923 argument or the result.
8925 Overloaded intrinsics will have the names of its overloaded argument
8926 types encoded into its function name, each preceded by a period. Only
8927 those types which are overloaded result in a name suffix. Arguments
8928 whose type is matched against another type do not. For example, the
8929 ``llvm.ctpop`` function can take an integer of any width and returns an
8930 integer of exactly the same integer width. This leads to a family of
8931 functions such as ``i8 @llvm.ctpop.i8(i8 %val)`` and
8932 ``i29 @llvm.ctpop.i29(i29 %val)``. Only one type, the return type, is
8933 overloaded, and only one type suffix is required. Because the argument's
8934 type is matched against the return type, it does not require its own
8937 To learn how to add an intrinsic function, please see the `Extending
8938 LLVM Guide <ExtendingLLVM.html>`_.
8942 Variable Argument Handling Intrinsics
8943 -------------------------------------
8945 Variable argument support is defined in LLVM with the
8946 :ref:`va_arg <i_va_arg>` instruction and these three intrinsic
8947 functions. These functions are related to the similarly named macros
8948 defined in the ``<stdarg.h>`` header file.
8950 All of these functions operate on arguments that use a target-specific
8951 value type "``va_list``". The LLVM assembly language reference manual
8952 does not define what this type is, so all transformations should be
8953 prepared to handle these functions regardless of the type used.
8955 This example shows how the :ref:`va_arg <i_va_arg>` instruction and the
8956 variable argument handling intrinsic functions are used.
8958 .. code-block:: llvm
8960 ; This struct is different for every platform. For most platforms,
8961 ; it is merely an i8*.
8962 %struct.va_list = type { i8* }
8964 ; For Unix x86_64 platforms, va_list is the following struct:
8965 ; %struct.va_list = type { i32, i32, i8*, i8* }
8967 define i32 @test(i32 %X, ...) {
8968 ; Initialize variable argument processing
8969 %ap = alloca %struct.va_list
8970 %ap2 = bitcast %struct.va_list* %ap to i8*
8971 call void @llvm.va_start(i8* %ap2)
8973 ; Read a single integer argument
8974 %tmp = va_arg i8* %ap2, i32
8976 ; Demonstrate usage of llvm.va_copy and llvm.va_end
8978 %aq2 = bitcast i8** %aq to i8*
8979 call void @llvm.va_copy(i8* %aq2, i8* %ap2)
8980 call void @llvm.va_end(i8* %aq2)
8982 ; Stop processing of arguments.
8983 call void @llvm.va_end(i8* %ap2)
8987 declare void @llvm.va_start(i8*)
8988 declare void @llvm.va_copy(i8*, i8*)
8989 declare void @llvm.va_end(i8*)
8993 '``llvm.va_start``' Intrinsic
8994 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9001 declare void @llvm.va_start(i8* <arglist>)
9006 The '``llvm.va_start``' intrinsic initializes ``*<arglist>`` for
9007 subsequent use by ``va_arg``.
9012 The argument is a pointer to a ``va_list`` element to initialize.
9017 The '``llvm.va_start``' intrinsic works just like the ``va_start`` macro
9018 available in C. In a target-dependent way, it initializes the
9019 ``va_list`` element to which the argument points, so that the next call
9020 to ``va_arg`` will produce the first variable argument passed to the
9021 function. Unlike the C ``va_start`` macro, this intrinsic does not need
9022 to know the last argument of the function as the compiler can figure
9025 '``llvm.va_end``' Intrinsic
9026 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9033 declare void @llvm.va_end(i8* <arglist>)
9038 The '``llvm.va_end``' intrinsic destroys ``*<arglist>``, which has been
9039 initialized previously with ``llvm.va_start`` or ``llvm.va_copy``.
9044 The argument is a pointer to a ``va_list`` to destroy.
9049 The '``llvm.va_end``' intrinsic works just like the ``va_end`` macro
9050 available in C. In a target-dependent way, it destroys the ``va_list``
9051 element to which the argument points. Calls to
9052 :ref:`llvm.va_start <int_va_start>` and
9053 :ref:`llvm.va_copy <int_va_copy>` must be matched exactly with calls to
9058 '``llvm.va_copy``' Intrinsic
9059 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9066 declare void @llvm.va_copy(i8* <destarglist>, i8* <srcarglist>)
9071 The '``llvm.va_copy``' intrinsic copies the current argument position
9072 from the source argument list to the destination argument list.
9077 The first argument is a pointer to a ``va_list`` element to initialize.
9078 The second argument is a pointer to a ``va_list`` element to copy from.
9083 The '``llvm.va_copy``' intrinsic works just like the ``va_copy`` macro
9084 available in C. In a target-dependent way, it copies the source
9085 ``va_list`` element into the destination ``va_list`` element. This
9086 intrinsic is necessary because the `` llvm.va_start`` intrinsic may be
9087 arbitrarily complex and require, for example, memory allocation.
9089 Accurate Garbage Collection Intrinsics
9090 --------------------------------------
9092 LLVM's support for `Accurate Garbage Collection <GarbageCollection.html>`_
9093 (GC) requires the frontend to generate code containing appropriate intrinsic
9094 calls and select an appropriate GC strategy which knows how to lower these
9095 intrinsics in a manner which is appropriate for the target collector.
9097 These intrinsics allow identification of :ref:`GC roots on the
9098 stack <int_gcroot>`, as well as garbage collector implementations that
9099 require :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers.
9100 Frontends for type-safe garbage collected languages should generate
9101 these intrinsics to make use of the LLVM garbage collectors. For more
9102 details, see `Garbage Collection with LLVM <GarbageCollection.html>`_.
9104 Experimental Statepoint Intrinsics
9105 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9107 LLVM provides an second experimental set of intrinsics for describing garbage
9108 collection safepoints in compiled code. These intrinsics are an alternative
9109 to the ``llvm.gcroot`` intrinsics, but are compatible with the ones for
9110 :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers. The
9111 differences in approach are covered in the `Garbage Collection with LLVM
9112 <GarbageCollection.html>`_ documentation. The intrinsics themselves are
9113 described in :doc:`Statepoints`.
9117 '``llvm.gcroot``' Intrinsic
9118 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9125 declare void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
9130 The '``llvm.gcroot``' intrinsic declares the existence of a GC root to
9131 the code generator, and allows some metadata to be associated with it.
9136 The first argument specifies the address of a stack object that contains
9137 the root pointer. The second pointer (which must be either a constant or
9138 a global value address) contains the meta-data to be associated with the
9144 At runtime, a call to this intrinsic stores a null pointer into the
9145 "ptrloc" location. At compile-time, the code generator generates
9146 information to allow the runtime to find the pointer at GC safe points.
9147 The '``llvm.gcroot``' intrinsic may only be used in a function which
9148 :ref:`specifies a GC algorithm <gc>`.
9152 '``llvm.gcread``' Intrinsic
9153 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9160 declare i8* @llvm.gcread(i8* %ObjPtr, i8** %Ptr)
9165 The '``llvm.gcread``' intrinsic identifies reads of references from heap
9166 locations, allowing garbage collector implementations that require read
9172 The second argument is the address to read from, which should be an
9173 address allocated from the garbage collector. The first object is a
9174 pointer to the start of the referenced object, if needed by the language
9175 runtime (otherwise null).
9180 The '``llvm.gcread``' intrinsic has the same semantics as a load
9181 instruction, but may be replaced with substantially more complex code by
9182 the garbage collector runtime, as needed. The '``llvm.gcread``'
9183 intrinsic may only be used in a function which :ref:`specifies a GC
9188 '``llvm.gcwrite``' Intrinsic
9189 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9196 declare void @llvm.gcwrite(i8* %P1, i8* %Obj, i8** %P2)
9201 The '``llvm.gcwrite``' intrinsic identifies writes of references to heap
9202 locations, allowing garbage collector implementations that require write
9203 barriers (such as generational or reference counting collectors).
9208 The first argument is the reference to store, the second is the start of
9209 the object to store it to, and the third is the address of the field of
9210 Obj to store to. If the runtime does not require a pointer to the
9211 object, Obj may be null.
9216 The '``llvm.gcwrite``' intrinsic has the same semantics as a store
9217 instruction, but may be replaced with substantially more complex code by
9218 the garbage collector runtime, as needed. The '``llvm.gcwrite``'
9219 intrinsic may only be used in a function which :ref:`specifies a GC
9222 Code Generator Intrinsics
9223 -------------------------
9225 These intrinsics are provided by LLVM to expose special features that
9226 may only be implemented with code generator support.
9228 '``llvm.returnaddress``' Intrinsic
9229 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9236 declare i8 *@llvm.returnaddress(i32 <level>)
9241 The '``llvm.returnaddress``' intrinsic attempts to compute a
9242 target-specific value indicating the return address of the current
9243 function or one of its callers.
9248 The argument to this intrinsic indicates which function to return the
9249 address for. Zero indicates the calling function, one indicates its
9250 caller, etc. The argument is **required** to be a constant integer
9256 The '``llvm.returnaddress``' intrinsic either returns a pointer
9257 indicating the return address of the specified call frame, or zero if it
9258 cannot be identified. The value returned by this intrinsic is likely to
9259 be incorrect or 0 for arguments other than zero, so it should only be
9260 used for debugging purposes.
9262 Note that calling this intrinsic does not prevent function inlining or
9263 other aggressive transformations, so the value returned may not be that
9264 of the obvious source-language caller.
9266 '``llvm.frameaddress``' Intrinsic
9267 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9274 declare i8* @llvm.frameaddress(i32 <level>)
9279 The '``llvm.frameaddress``' intrinsic attempts to return the
9280 target-specific frame pointer value for the specified stack frame.
9285 The argument to this intrinsic indicates which function to return the
9286 frame pointer for. Zero indicates the calling function, one indicates
9287 its caller, etc. The argument is **required** to be a constant integer
9293 The '``llvm.frameaddress``' intrinsic either returns a pointer
9294 indicating the frame address of the specified call frame, or zero if it
9295 cannot be identified. The value returned by this intrinsic is likely to
9296 be incorrect or 0 for arguments other than zero, so it should only be
9297 used for debugging purposes.
9299 Note that calling this intrinsic does not prevent function inlining or
9300 other aggressive transformations, so the value returned may not be that
9301 of the obvious source-language caller.
9303 '``llvm.localescape``' and '``llvm.localrecover``' Intrinsics
9304 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9311 declare void @llvm.localescape(...)
9312 declare i8* @llvm.localrecover(i8* %func, i8* %fp, i32 %idx)
9317 The '``llvm.localescape``' intrinsic escapes offsets of a collection of static
9318 allocas, and the '``llvm.localrecover``' intrinsic applies those offsets to a
9319 live frame pointer to recover the address of the allocation. The offset is
9320 computed during frame layout of the caller of ``llvm.localescape``.
9325 All arguments to '``llvm.localescape``' must be pointers to static allocas or
9326 casts of static allocas. Each function can only call '``llvm.localescape``'
9327 once, and it can only do so from the entry block.
9329 The ``func`` argument to '``llvm.localrecover``' must be a constant
9330 bitcasted pointer to a function defined in the current module. The code
9331 generator cannot determine the frame allocation offset of functions defined in
9334 The ``fp`` argument to '``llvm.localrecover``' must be a frame pointer of a
9335 call frame that is currently live. The return value of '``llvm.localaddress``'
9336 is one way to produce such a value, but various runtimes also expose a suitable
9337 pointer in platform-specific ways.
9339 The ``idx`` argument to '``llvm.localrecover``' indicates which alloca passed to
9340 '``llvm.localescape``' to recover. It is zero-indexed.
9345 These intrinsics allow a group of functions to share access to a set of local
9346 stack allocations of a one parent function. The parent function may call the
9347 '``llvm.localescape``' intrinsic once from the function entry block, and the
9348 child functions can use '``llvm.localrecover``' to access the escaped allocas.
9349 The '``llvm.localescape``' intrinsic blocks inlining, as inlining changes where
9350 the escaped allocas are allocated, which would break attempts to use
9351 '``llvm.localrecover``'.
9353 .. _int_read_register:
9354 .. _int_write_register:
9356 '``llvm.read_register``' and '``llvm.write_register``' Intrinsics
9357 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9364 declare i32 @llvm.read_register.i32(metadata)
9365 declare i64 @llvm.read_register.i64(metadata)
9366 declare void @llvm.write_register.i32(metadata, i32 @value)
9367 declare void @llvm.write_register.i64(metadata, i64 @value)
9373 The '``llvm.read_register``' and '``llvm.write_register``' intrinsics
9374 provides access to the named register. The register must be valid on
9375 the architecture being compiled to. The type needs to be compatible
9376 with the register being read.
9381 The '``llvm.read_register``' intrinsic returns the current value of the
9382 register, where possible. The '``llvm.write_register``' intrinsic sets
9383 the current value of the register, where possible.
9385 This is useful to implement named register global variables that need
9386 to always be mapped to a specific register, as is common practice on
9387 bare-metal programs including OS kernels.
9389 The compiler doesn't check for register availability or use of the used
9390 register in surrounding code, including inline assembly. Because of that,
9391 allocatable registers are not supported.
9393 Warning: So far it only works with the stack pointer on selected
9394 architectures (ARM, AArch64, PowerPC and x86_64). Significant amount of
9395 work is needed to support other registers and even more so, allocatable
9400 '``llvm.stacksave``' Intrinsic
9401 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9408 declare i8* @llvm.stacksave()
9413 The '``llvm.stacksave``' intrinsic is used to remember the current state
9414 of the function stack, for use with
9415 :ref:`llvm.stackrestore <int_stackrestore>`. This is useful for
9416 implementing language features like scoped automatic variable sized
9422 This intrinsic returns a opaque pointer value that can be passed to
9423 :ref:`llvm.stackrestore <int_stackrestore>`. When an
9424 ``llvm.stackrestore`` intrinsic is executed with a value saved from
9425 ``llvm.stacksave``, it effectively restores the state of the stack to
9426 the state it was in when the ``llvm.stacksave`` intrinsic executed. In
9427 practice, this pops any :ref:`alloca <i_alloca>` blocks from the stack that
9428 were allocated after the ``llvm.stacksave`` was executed.
9430 .. _int_stackrestore:
9432 '``llvm.stackrestore``' Intrinsic
9433 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9440 declare void @llvm.stackrestore(i8* %ptr)
9445 The '``llvm.stackrestore``' intrinsic is used to restore the state of
9446 the function stack to the state it was in when the corresponding
9447 :ref:`llvm.stacksave <int_stacksave>` intrinsic executed. This is
9448 useful for implementing language features like scoped automatic variable
9449 sized arrays in C99.
9454 See the description for :ref:`llvm.stacksave <int_stacksave>`.
9456 .. _int_get_dynamic_area_offset:
9458 '``llvm.get.dynamic.area.offset``' Intrinsic
9459 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9466 declare i32 @llvm.get.dynamic.area.offset.i32()
9467 declare i64 @llvm.get.dynamic.area.offset.i64()
9472 The '``llvm.get.dynamic.area.offset.*``' intrinsic family is used to
9473 get the offset from native stack pointer to the address of the most
9474 recent dynamic alloca on the caller's stack. These intrinsics are
9475 intendend for use in combination with
9476 :ref:`llvm.stacksave <int_stacksave>` to get a
9477 pointer to the most recent dynamic alloca. This is useful, for example,
9478 for AddressSanitizer's stack unpoisoning routines.
9483 These intrinsics return a non-negative integer value that can be used to
9484 get the address of the most recent dynamic alloca, allocated by :ref:`alloca <i_alloca>`
9485 on the caller's stack. In particular, for targets where stack grows downwards,
9486 adding this offset to the native stack pointer would get the address of the most
9487 recent dynamic alloca. For targets where stack grows upwards, the situation is a bit more
9488 complicated, because substracting this value from stack pointer would get the address
9489 one past the end of the most recent dynamic alloca.
9491 Although for most targets `llvm.get.dynamic.area.offset <int_get_dynamic_area_offset>`
9492 returns just a zero, for others, such as PowerPC and PowerPC64, it returns a
9493 compile-time-known constant value.
9495 The return value type of :ref:`llvm.get.dynamic.area.offset <int_get_dynamic_area_offset>`
9496 must match the target's generic address space's (address space 0) pointer type.
9498 '``llvm.prefetch``' Intrinsic
9499 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9506 declare void @llvm.prefetch(i8* <address>, i32 <rw>, i32 <locality>, i32 <cache type>)
9511 The '``llvm.prefetch``' intrinsic is a hint to the code generator to
9512 insert a prefetch instruction if supported; otherwise, it is a noop.
9513 Prefetches have no effect on the behavior of the program but can change
9514 its performance characteristics.
9519 ``address`` is the address to be prefetched, ``rw`` is the specifier
9520 determining if the fetch should be for a read (0) or write (1), and
9521 ``locality`` is a temporal locality specifier ranging from (0) - no
9522 locality, to (3) - extremely local keep in cache. The ``cache type``
9523 specifies whether the prefetch is performed on the data (1) or
9524 instruction (0) cache. The ``rw``, ``locality`` and ``cache type``
9525 arguments must be constant integers.
9530 This intrinsic does not modify the behavior of the program. In
9531 particular, prefetches cannot trap and do not produce a value. On
9532 targets that support this intrinsic, the prefetch can provide hints to
9533 the processor cache for better performance.
9535 '``llvm.pcmarker``' Intrinsic
9536 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9543 declare void @llvm.pcmarker(i32 <id>)
9548 The '``llvm.pcmarker``' intrinsic is a method to export a Program
9549 Counter (PC) in a region of code to simulators and other tools. The
9550 method is target specific, but it is expected that the marker will use
9551 exported symbols to transmit the PC of the marker. The marker makes no
9552 guarantees that it will remain with any specific instruction after
9553 optimizations. It is possible that the presence of a marker will inhibit
9554 optimizations. The intended use is to be inserted after optimizations to
9555 allow correlations of simulation runs.
9560 ``id`` is a numerical id identifying the marker.
9565 This intrinsic does not modify the behavior of the program. Backends
9566 that do not support this intrinsic may ignore it.
9568 '``llvm.readcyclecounter``' Intrinsic
9569 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9576 declare i64 @llvm.readcyclecounter()
9581 The '``llvm.readcyclecounter``' intrinsic provides access to the cycle
9582 counter register (or similar low latency, high accuracy clocks) on those
9583 targets that support it. On X86, it should map to RDTSC. On Alpha, it
9584 should map to RPCC. As the backing counters overflow quickly (on the
9585 order of 9 seconds on alpha), this should only be used for small
9591 When directly supported, reading the cycle counter should not modify any
9592 memory. Implementations are allowed to either return a application
9593 specific value or a system wide value. On backends without support, this
9594 is lowered to a constant 0.
9596 Note that runtime support may be conditional on the privilege-level code is
9597 running at and the host platform.
9599 '``llvm.clear_cache``' Intrinsic
9600 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9607 declare void @llvm.clear_cache(i8*, i8*)
9612 The '``llvm.clear_cache``' intrinsic ensures visibility of modifications
9613 in the specified range to the execution unit of the processor. On
9614 targets with non-unified instruction and data cache, the implementation
9615 flushes the instruction cache.
9620 On platforms with coherent instruction and data caches (e.g. x86), this
9621 intrinsic is a nop. On platforms with non-coherent instruction and data
9622 cache (e.g. ARM, MIPS), the intrinsic is lowered either to appropriate
9623 instructions or a system call, if cache flushing requires special
9626 The default behavior is to emit a call to ``__clear_cache`` from the run
9629 This instrinsic does *not* empty the instruction pipeline. Modifications
9630 of the current function are outside the scope of the intrinsic.
9632 '``llvm.instrprof_increment``' Intrinsic
9633 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9640 declare void @llvm.instrprof_increment(i8* <name>, i64 <hash>,
9641 i32 <num-counters>, i32 <index>)
9646 The '``llvm.instrprof_increment``' intrinsic can be emitted by a
9647 frontend for use with instrumentation based profiling. These will be
9648 lowered by the ``-instrprof`` pass to generate execution counts of a
9654 The first argument is a pointer to a global variable containing the
9655 name of the entity being instrumented. This should generally be the
9656 (mangled) function name for a set of counters.
9658 The second argument is a hash value that can be used by the consumer
9659 of the profile data to detect changes to the instrumented source, and
9660 the third is the number of counters associated with ``name``. It is an
9661 error if ``hash`` or ``num-counters`` differ between two instances of
9662 ``instrprof_increment`` that refer to the same name.
9664 The last argument refers to which of the counters for ``name`` should
9665 be incremented. It should be a value between 0 and ``num-counters``.
9670 This intrinsic represents an increment of a profiling counter. It will
9671 cause the ``-instrprof`` pass to generate the appropriate data
9672 structures and the code to increment the appropriate value, in a
9673 format that can be written out by a compiler runtime and consumed via
9674 the ``llvm-profdata`` tool.
9676 '``llvm.instrprof_value_profile``' Intrinsic
9677 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9684 declare void @llvm.instrprof_value_profile(i8* <name>, i64 <hash>,
9685 i64 <value>, i32 <value_kind>,
9691 The '``llvm.instrprof_value_profile``' intrinsic can be emitted by a
9692 frontend for use with instrumentation based profiling. This will be
9693 lowered by the ``-instrprof`` pass to find out the target values,
9694 instrumented expressions take in a program at runtime.
9699 The first argument is a pointer to a global variable containing the
9700 name of the entity being instrumented. ``name`` should generally be the
9701 (mangled) function name for a set of counters.
9703 The second argument is a hash value that can be used by the consumer
9704 of the profile data to detect changes to the instrumented source. It
9705 is an error if ``hash`` differs between two instances of
9706 ``llvm.instrprof_*`` that refer to the same name.
9708 The third argument is the value of the expression being profiled. The profiled
9709 expression's value should be representable as an unsigned 64-bit value. The
9710 fourth argument represents the kind of value profiling that is being done. The
9711 supported value profiling kinds are enumerated through the
9712 ``InstrProfValueKind`` type declared in the
9713 ``<include/llvm/ProfileData/InstrProf.h>`` header file. The last argument is the
9714 index of the instrumented expression within ``name``. It should be >= 0.
9719 This intrinsic represents the point where a call to a runtime routine
9720 should be inserted for value profiling of target expressions. ``-instrprof``
9721 pass will generate the appropriate data structures and replace the
9722 ``llvm.instrprof_value_profile`` intrinsic with the call to the profile
9723 runtime library with proper arguments.
9725 '``llvm.thread.pointer``' Intrinsic
9726 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9733 declare i8* @llvm.thread.pointer()
9738 The '``llvm.thread.pointer``' intrinsic returns the value of the thread
9744 The '``llvm.thread.pointer``' intrinsic returns a pointer to the TLS area
9745 for the current thread. The exact semantics of this value are target
9746 specific: it may point to the start of TLS area, to the end, or somewhere
9747 in the middle. Depending on the target, this intrinsic may read a register,
9748 call a helper function, read from an alternate memory space, or perform
9749 other operations necessary to locate the TLS area. Not all targets support
9752 Standard C Library Intrinsics
9753 -----------------------------
9755 LLVM provides intrinsics for a few important standard C library
9756 functions. These intrinsics allow source-language front-ends to pass
9757 information about the alignment of the pointer arguments to the code
9758 generator, providing opportunity for more efficient code generation.
9762 '``llvm.memcpy``' Intrinsic
9763 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9768 This is an overloaded intrinsic. You can use ``llvm.memcpy`` on any
9769 integer bit width and for different address spaces. Not all targets
9770 support all bit widths however.
9774 declare void @llvm.memcpy.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
9775 i32 <len>, i32 <align>, i1 <isvolatile>)
9776 declare void @llvm.memcpy.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
9777 i64 <len>, i32 <align>, i1 <isvolatile>)
9782 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
9783 source location to the destination location.
9785 Note that, unlike the standard libc function, the ``llvm.memcpy.*``
9786 intrinsics do not return a value, takes extra alignment/isvolatile
9787 arguments and the pointers can be in specified address spaces.
9792 The first argument is a pointer to the destination, the second is a
9793 pointer to the source. The third argument is an integer argument
9794 specifying the number of bytes to copy, the fourth argument is the
9795 alignment of the source and destination locations, and the fifth is a
9796 boolean indicating a volatile access.
9798 If the call to this intrinsic has an alignment value that is not 0 or 1,
9799 then the caller guarantees that both the source and destination pointers
9800 are aligned to that boundary.
9802 If the ``isvolatile`` parameter is ``true``, the ``llvm.memcpy`` call is
9803 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
9804 very cleanly specified and it is unwise to depend on it.
9809 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
9810 source location to the destination location, which are not allowed to
9811 overlap. It copies "len" bytes of memory over. If the argument is known
9812 to be aligned to some boundary, this can be specified as the fourth
9813 argument, otherwise it should be set to 0 or 1 (both meaning no alignment).
9815 '``llvm.memmove``' Intrinsic
9816 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9821 This is an overloaded intrinsic. You can use llvm.memmove on any integer
9822 bit width and for different address space. Not all targets support all
9827 declare void @llvm.memmove.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
9828 i32 <len>, i32 <align>, i1 <isvolatile>)
9829 declare void @llvm.memmove.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
9830 i64 <len>, i32 <align>, i1 <isvolatile>)
9835 The '``llvm.memmove.*``' intrinsics move a block of memory from the
9836 source location to the destination location. It is similar to the
9837 '``llvm.memcpy``' intrinsic but allows the two memory locations to
9840 Note that, unlike the standard libc function, the ``llvm.memmove.*``
9841 intrinsics do not return a value, takes extra alignment/isvolatile
9842 arguments and the pointers can be in specified address spaces.
9847 The first argument is a pointer to the destination, the second is a
9848 pointer to the source. The third argument is an integer argument
9849 specifying the number of bytes to copy, the fourth argument is the
9850 alignment of the source and destination locations, and the fifth is a
9851 boolean indicating a volatile access.
9853 If the call to this intrinsic has an alignment value that is not 0 or 1,
9854 then the caller guarantees that the source and destination pointers are
9855 aligned to that boundary.
9857 If the ``isvolatile`` parameter is ``true``, the ``llvm.memmove`` call
9858 is a :ref:`volatile operation <volatile>`. The detailed access behavior is
9859 not very cleanly specified and it is unwise to depend on it.
9864 The '``llvm.memmove.*``' intrinsics copy a block of memory from the
9865 source location to the destination location, which may overlap. It
9866 copies "len" bytes of memory over. If the argument is known to be
9867 aligned to some boundary, this can be specified as the fourth argument,
9868 otherwise it should be set to 0 or 1 (both meaning no alignment).
9870 '``llvm.memset.*``' Intrinsics
9871 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9876 This is an overloaded intrinsic. You can use llvm.memset on any integer
9877 bit width and for different address spaces. However, not all targets
9878 support all bit widths.
9882 declare void @llvm.memset.p0i8.i32(i8* <dest>, i8 <val>,
9883 i32 <len>, i32 <align>, i1 <isvolatile>)
9884 declare void @llvm.memset.p0i8.i64(i8* <dest>, i8 <val>,
9885 i64 <len>, i32 <align>, i1 <isvolatile>)
9890 The '``llvm.memset.*``' intrinsics fill a block of memory with a
9891 particular byte value.
9893 Note that, unlike the standard libc function, the ``llvm.memset``
9894 intrinsic does not return a value and takes extra alignment/volatile
9895 arguments. Also, the destination can be in an arbitrary address space.
9900 The first argument is a pointer to the destination to fill, the second
9901 is the byte value with which to fill it, the third argument is an
9902 integer argument specifying the number of bytes to fill, and the fourth
9903 argument is the known alignment of the destination location.
9905 If the call to this intrinsic has an alignment value that is not 0 or 1,
9906 then the caller guarantees that the destination pointer is aligned to
9909 If the ``isvolatile`` parameter is ``true``, the ``llvm.memset`` call is
9910 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
9911 very cleanly specified and it is unwise to depend on it.
9916 The '``llvm.memset.*``' intrinsics fill "len" bytes of memory starting
9917 at the destination location. If the argument is known to be aligned to
9918 some boundary, this can be specified as the fourth argument, otherwise
9919 it should be set to 0 or 1 (both meaning no alignment).
9921 '``llvm.sqrt.*``' Intrinsic
9922 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9927 This is an overloaded intrinsic. You can use ``llvm.sqrt`` on any
9928 floating point or vector of floating point type. Not all targets support
9933 declare float @llvm.sqrt.f32(float %Val)
9934 declare double @llvm.sqrt.f64(double %Val)
9935 declare x86_fp80 @llvm.sqrt.f80(x86_fp80 %Val)
9936 declare fp128 @llvm.sqrt.f128(fp128 %Val)
9937 declare ppc_fp128 @llvm.sqrt.ppcf128(ppc_fp128 %Val)
9942 The '``llvm.sqrt``' intrinsics return the sqrt of the specified operand,
9943 returning the same value as the libm '``sqrt``' functions would. Unlike
9944 ``sqrt`` in libm, however, ``llvm.sqrt`` has undefined behavior for
9945 negative numbers other than -0.0 (which allows for better optimization,
9946 because there is no need to worry about errno being set).
9947 ``llvm.sqrt(-0.0)`` is defined to return -0.0 like IEEE sqrt.
9952 The argument and return value are floating point numbers of the same
9958 This function returns the sqrt of the specified operand if it is a
9959 nonnegative floating point number.
9961 '``llvm.powi.*``' Intrinsic
9962 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9967 This is an overloaded intrinsic. You can use ``llvm.powi`` on any
9968 floating point or vector of floating point type. Not all targets support
9973 declare float @llvm.powi.f32(float %Val, i32 %power)
9974 declare double @llvm.powi.f64(double %Val, i32 %power)
9975 declare x86_fp80 @llvm.powi.f80(x86_fp80 %Val, i32 %power)
9976 declare fp128 @llvm.powi.f128(fp128 %Val, i32 %power)
9977 declare ppc_fp128 @llvm.powi.ppcf128(ppc_fp128 %Val, i32 %power)
9982 The '``llvm.powi.*``' intrinsics return the first operand raised to the
9983 specified (positive or negative) power. The order of evaluation of
9984 multiplications is not defined. When a vector of floating point type is
9985 used, the second argument remains a scalar integer value.
9990 The second argument is an integer power, and the first is a value to
9991 raise to that power.
9996 This function returns the first value raised to the second power with an
9997 unspecified sequence of rounding operations.
9999 '``llvm.sin.*``' Intrinsic
10000 ^^^^^^^^^^^^^^^^^^^^^^^^^^
10005 This is an overloaded intrinsic. You can use ``llvm.sin`` on any
10006 floating point or vector of floating point type. Not all targets support
10011 declare float @llvm.sin.f32(float %Val)
10012 declare double @llvm.sin.f64(double %Val)
10013 declare x86_fp80 @llvm.sin.f80(x86_fp80 %Val)
10014 declare fp128 @llvm.sin.f128(fp128 %Val)
10015 declare ppc_fp128 @llvm.sin.ppcf128(ppc_fp128 %Val)
10020 The '``llvm.sin.*``' intrinsics return the sine of the operand.
10025 The argument and return value are floating point numbers of the same
10031 This function returns the sine of the specified operand, returning the
10032 same values as the libm ``sin`` functions would, and handles error
10033 conditions in the same way.
10035 '``llvm.cos.*``' Intrinsic
10036 ^^^^^^^^^^^^^^^^^^^^^^^^^^
10041 This is an overloaded intrinsic. You can use ``llvm.cos`` on any
10042 floating point or vector of floating point type. Not all targets support
10047 declare float @llvm.cos.f32(float %Val)
10048 declare double @llvm.cos.f64(double %Val)
10049 declare x86_fp80 @llvm.cos.f80(x86_fp80 %Val)
10050 declare fp128 @llvm.cos.f128(fp128 %Val)
10051 declare ppc_fp128 @llvm.cos.ppcf128(ppc_fp128 %Val)
10056 The '``llvm.cos.*``' intrinsics return the cosine of the operand.
10061 The argument and return value are floating point numbers of the same
10067 This function returns the cosine of the specified operand, returning the
10068 same values as the libm ``cos`` functions would, and handles error
10069 conditions in the same way.
10071 '``llvm.pow.*``' Intrinsic
10072 ^^^^^^^^^^^^^^^^^^^^^^^^^^
10077 This is an overloaded intrinsic. You can use ``llvm.pow`` on any
10078 floating point or vector of floating point type. Not all targets support
10083 declare float @llvm.pow.f32(float %Val, float %Power)
10084 declare double @llvm.pow.f64(double %Val, double %Power)
10085 declare x86_fp80 @llvm.pow.f80(x86_fp80 %Val, x86_fp80 %Power)
10086 declare fp128 @llvm.pow.f128(fp128 %Val, fp128 %Power)
10087 declare ppc_fp128 @llvm.pow.ppcf128(ppc_fp128 %Val, ppc_fp128 Power)
10092 The '``llvm.pow.*``' intrinsics return the first operand raised to the
10093 specified (positive or negative) power.
10098 The second argument is a floating point power, and the first is a value
10099 to raise to that power.
10104 This function returns the first value raised to the second power,
10105 returning the same values as the libm ``pow`` functions would, and
10106 handles error conditions in the same way.
10108 '``llvm.exp.*``' Intrinsic
10109 ^^^^^^^^^^^^^^^^^^^^^^^^^^
10114 This is an overloaded intrinsic. You can use ``llvm.exp`` on any
10115 floating point or vector of floating point type. Not all targets support
10120 declare float @llvm.exp.f32(float %Val)
10121 declare double @llvm.exp.f64(double %Val)
10122 declare x86_fp80 @llvm.exp.f80(x86_fp80 %Val)
10123 declare fp128 @llvm.exp.f128(fp128 %Val)
10124 declare ppc_fp128 @llvm.exp.ppcf128(ppc_fp128 %Val)
10129 The '``llvm.exp.*``' intrinsics perform the exp function.
10134 The argument and return value are floating point numbers of the same
10140 This function returns the same values as the libm ``exp`` functions
10141 would, and handles error conditions in the same way.
10143 '``llvm.exp2.*``' Intrinsic
10144 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10149 This is an overloaded intrinsic. You can use ``llvm.exp2`` on any
10150 floating point or vector of floating point type. Not all targets support
10155 declare float @llvm.exp2.f32(float %Val)
10156 declare double @llvm.exp2.f64(double %Val)
10157 declare x86_fp80 @llvm.exp2.f80(x86_fp80 %Val)
10158 declare fp128 @llvm.exp2.f128(fp128 %Val)
10159 declare ppc_fp128 @llvm.exp2.ppcf128(ppc_fp128 %Val)
10164 The '``llvm.exp2.*``' intrinsics perform the exp2 function.
10169 The argument and return value are floating point numbers of the same
10175 This function returns the same values as the libm ``exp2`` functions
10176 would, and handles error conditions in the same way.
10178 '``llvm.log.*``' Intrinsic
10179 ^^^^^^^^^^^^^^^^^^^^^^^^^^
10184 This is an overloaded intrinsic. You can use ``llvm.log`` on any
10185 floating point or vector of floating point type. Not all targets support
10190 declare float @llvm.log.f32(float %Val)
10191 declare double @llvm.log.f64(double %Val)
10192 declare x86_fp80 @llvm.log.f80(x86_fp80 %Val)
10193 declare fp128 @llvm.log.f128(fp128 %Val)
10194 declare ppc_fp128 @llvm.log.ppcf128(ppc_fp128 %Val)
10199 The '``llvm.log.*``' intrinsics perform the log function.
10204 The argument and return value are floating point numbers of the same
10210 This function returns the same values as the libm ``log`` functions
10211 would, and handles error conditions in the same way.
10213 '``llvm.log10.*``' Intrinsic
10214 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10219 This is an overloaded intrinsic. You can use ``llvm.log10`` on any
10220 floating point or vector of floating point type. Not all targets support
10225 declare float @llvm.log10.f32(float %Val)
10226 declare double @llvm.log10.f64(double %Val)
10227 declare x86_fp80 @llvm.log10.f80(x86_fp80 %Val)
10228 declare fp128 @llvm.log10.f128(fp128 %Val)
10229 declare ppc_fp128 @llvm.log10.ppcf128(ppc_fp128 %Val)
10234 The '``llvm.log10.*``' intrinsics perform the log10 function.
10239 The argument and return value are floating point numbers of the same
10245 This function returns the same values as the libm ``log10`` functions
10246 would, and handles error conditions in the same way.
10248 '``llvm.log2.*``' Intrinsic
10249 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10254 This is an overloaded intrinsic. You can use ``llvm.log2`` on any
10255 floating point or vector of floating point type. Not all targets support
10260 declare float @llvm.log2.f32(float %Val)
10261 declare double @llvm.log2.f64(double %Val)
10262 declare x86_fp80 @llvm.log2.f80(x86_fp80 %Val)
10263 declare fp128 @llvm.log2.f128(fp128 %Val)
10264 declare ppc_fp128 @llvm.log2.ppcf128(ppc_fp128 %Val)
10269 The '``llvm.log2.*``' intrinsics perform the log2 function.
10274 The argument and return value are floating point numbers of the same
10280 This function returns the same values as the libm ``log2`` functions
10281 would, and handles error conditions in the same way.
10283 '``llvm.fma.*``' Intrinsic
10284 ^^^^^^^^^^^^^^^^^^^^^^^^^^
10289 This is an overloaded intrinsic. You can use ``llvm.fma`` on any
10290 floating point or vector of floating point type. Not all targets support
10295 declare float @llvm.fma.f32(float %a, float %b, float %c)
10296 declare double @llvm.fma.f64(double %a, double %b, double %c)
10297 declare x86_fp80 @llvm.fma.f80(x86_fp80 %a, x86_fp80 %b, x86_fp80 %c)
10298 declare fp128 @llvm.fma.f128(fp128 %a, fp128 %b, fp128 %c)
10299 declare ppc_fp128 @llvm.fma.ppcf128(ppc_fp128 %a, ppc_fp128 %b, ppc_fp128 %c)
10304 The '``llvm.fma.*``' intrinsics perform the fused multiply-add
10310 The argument and return value are floating point numbers of the same
10316 This function returns the same values as the libm ``fma`` functions
10317 would, and does not set errno.
10319 '``llvm.fabs.*``' Intrinsic
10320 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10325 This is an overloaded intrinsic. You can use ``llvm.fabs`` on any
10326 floating point or vector of floating point type. Not all targets support
10331 declare float @llvm.fabs.f32(float %Val)
10332 declare double @llvm.fabs.f64(double %Val)
10333 declare x86_fp80 @llvm.fabs.f80(x86_fp80 %Val)
10334 declare fp128 @llvm.fabs.f128(fp128 %Val)
10335 declare ppc_fp128 @llvm.fabs.ppcf128(ppc_fp128 %Val)
10340 The '``llvm.fabs.*``' intrinsics return the absolute value of the
10346 The argument and return value are floating point numbers of the same
10352 This function returns the same values as the libm ``fabs`` functions
10353 would, and handles error conditions in the same way.
10355 '``llvm.minnum.*``' Intrinsic
10356 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10361 This is an overloaded intrinsic. You can use ``llvm.minnum`` on any
10362 floating point or vector of floating point type. Not all targets support
10367 declare float @llvm.minnum.f32(float %Val0, float %Val1)
10368 declare double @llvm.minnum.f64(double %Val0, double %Val1)
10369 declare x86_fp80 @llvm.minnum.f80(x86_fp80 %Val0, x86_fp80 %Val1)
10370 declare fp128 @llvm.minnum.f128(fp128 %Val0, fp128 %Val1)
10371 declare ppc_fp128 @llvm.minnum.ppcf128(ppc_fp128 %Val0, ppc_fp128 %Val1)
10376 The '``llvm.minnum.*``' intrinsics return the minimum of the two
10383 The arguments and return value are floating point numbers of the same
10389 Follows the IEEE-754 semantics for minNum, which also match for libm's
10392 If either operand is a NaN, returns the other non-NaN operand. Returns
10393 NaN only if both operands are NaN. If the operands compare equal,
10394 returns a value that compares equal to both operands. This means that
10395 fmin(+/-0.0, +/-0.0) could return either -0.0 or 0.0.
10397 '``llvm.maxnum.*``' Intrinsic
10398 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10403 This is an overloaded intrinsic. You can use ``llvm.maxnum`` on any
10404 floating point or vector of floating point type. Not all targets support
10409 declare float @llvm.maxnum.f32(float %Val0, float %Val1l)
10410 declare double @llvm.maxnum.f64(double %Val0, double %Val1)
10411 declare x86_fp80 @llvm.maxnum.f80(x86_fp80 %Val0, x86_fp80 %Val1)
10412 declare fp128 @llvm.maxnum.f128(fp128 %Val0, fp128 %Val1)
10413 declare ppc_fp128 @llvm.maxnum.ppcf128(ppc_fp128 %Val0, ppc_fp128 %Val1)
10418 The '``llvm.maxnum.*``' intrinsics return the maximum of the two
10425 The arguments and return value are floating point numbers of the same
10430 Follows the IEEE-754 semantics for maxNum, which also match for libm's
10433 If either operand is a NaN, returns the other non-NaN operand. Returns
10434 NaN only if both operands are NaN. If the operands compare equal,
10435 returns a value that compares equal to both operands. This means that
10436 fmax(+/-0.0, +/-0.0) could return either -0.0 or 0.0.
10438 '``llvm.copysign.*``' Intrinsic
10439 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10444 This is an overloaded intrinsic. You can use ``llvm.copysign`` on any
10445 floating point or vector of floating point type. Not all targets support
10450 declare float @llvm.copysign.f32(float %Mag, float %Sgn)
10451 declare double @llvm.copysign.f64(double %Mag, double %Sgn)
10452 declare x86_fp80 @llvm.copysign.f80(x86_fp80 %Mag, x86_fp80 %Sgn)
10453 declare fp128 @llvm.copysign.f128(fp128 %Mag, fp128 %Sgn)
10454 declare ppc_fp128 @llvm.copysign.ppcf128(ppc_fp128 %Mag, ppc_fp128 %Sgn)
10459 The '``llvm.copysign.*``' intrinsics return a value with the magnitude of the
10460 first operand and the sign of the second operand.
10465 The arguments and return value are floating point numbers of the same
10471 This function returns the same values as the libm ``copysign``
10472 functions would, and handles error conditions in the same way.
10474 '``llvm.floor.*``' Intrinsic
10475 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10480 This is an overloaded intrinsic. You can use ``llvm.floor`` on any
10481 floating point or vector of floating point type. Not all targets support
10486 declare float @llvm.floor.f32(float %Val)
10487 declare double @llvm.floor.f64(double %Val)
10488 declare x86_fp80 @llvm.floor.f80(x86_fp80 %Val)
10489 declare fp128 @llvm.floor.f128(fp128 %Val)
10490 declare ppc_fp128 @llvm.floor.ppcf128(ppc_fp128 %Val)
10495 The '``llvm.floor.*``' intrinsics return the floor of the operand.
10500 The argument and return value are floating point numbers of the same
10506 This function returns the same values as the libm ``floor`` functions
10507 would, and handles error conditions in the same way.
10509 '``llvm.ceil.*``' Intrinsic
10510 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10515 This is an overloaded intrinsic. You can use ``llvm.ceil`` on any
10516 floating point or vector of floating point type. Not all targets support
10521 declare float @llvm.ceil.f32(float %Val)
10522 declare double @llvm.ceil.f64(double %Val)
10523 declare x86_fp80 @llvm.ceil.f80(x86_fp80 %Val)
10524 declare fp128 @llvm.ceil.f128(fp128 %Val)
10525 declare ppc_fp128 @llvm.ceil.ppcf128(ppc_fp128 %Val)
10530 The '``llvm.ceil.*``' intrinsics return the ceiling of the operand.
10535 The argument and return value are floating point numbers of the same
10541 This function returns the same values as the libm ``ceil`` functions
10542 would, and handles error conditions in the same way.
10544 '``llvm.trunc.*``' Intrinsic
10545 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10550 This is an overloaded intrinsic. You can use ``llvm.trunc`` on any
10551 floating point or vector of floating point type. Not all targets support
10556 declare float @llvm.trunc.f32(float %Val)
10557 declare double @llvm.trunc.f64(double %Val)
10558 declare x86_fp80 @llvm.trunc.f80(x86_fp80 %Val)
10559 declare fp128 @llvm.trunc.f128(fp128 %Val)
10560 declare ppc_fp128 @llvm.trunc.ppcf128(ppc_fp128 %Val)
10565 The '``llvm.trunc.*``' intrinsics returns the operand rounded to the
10566 nearest integer not larger in magnitude than the operand.
10571 The argument and return value are floating point numbers of the same
10577 This function returns the same values as the libm ``trunc`` functions
10578 would, and handles error conditions in the same way.
10580 '``llvm.rint.*``' Intrinsic
10581 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10586 This is an overloaded intrinsic. You can use ``llvm.rint`` on any
10587 floating point or vector of floating point type. Not all targets support
10592 declare float @llvm.rint.f32(float %Val)
10593 declare double @llvm.rint.f64(double %Val)
10594 declare x86_fp80 @llvm.rint.f80(x86_fp80 %Val)
10595 declare fp128 @llvm.rint.f128(fp128 %Val)
10596 declare ppc_fp128 @llvm.rint.ppcf128(ppc_fp128 %Val)
10601 The '``llvm.rint.*``' intrinsics returns the operand rounded to the
10602 nearest integer. It may raise an inexact floating-point exception if the
10603 operand isn't an integer.
10608 The argument and return value are floating point numbers of the same
10614 This function returns the same values as the libm ``rint`` functions
10615 would, and handles error conditions in the same way.
10617 '``llvm.nearbyint.*``' Intrinsic
10618 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10623 This is an overloaded intrinsic. You can use ``llvm.nearbyint`` on any
10624 floating point or vector of floating point type. Not all targets support
10629 declare float @llvm.nearbyint.f32(float %Val)
10630 declare double @llvm.nearbyint.f64(double %Val)
10631 declare x86_fp80 @llvm.nearbyint.f80(x86_fp80 %Val)
10632 declare fp128 @llvm.nearbyint.f128(fp128 %Val)
10633 declare ppc_fp128 @llvm.nearbyint.ppcf128(ppc_fp128 %Val)
10638 The '``llvm.nearbyint.*``' intrinsics returns the operand rounded to the
10644 The argument and return value are floating point numbers of the same
10650 This function returns the same values as the libm ``nearbyint``
10651 functions would, and handles error conditions in the same way.
10653 '``llvm.round.*``' Intrinsic
10654 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10659 This is an overloaded intrinsic. You can use ``llvm.round`` on any
10660 floating point or vector of floating point type. Not all targets support
10665 declare float @llvm.round.f32(float %Val)
10666 declare double @llvm.round.f64(double %Val)
10667 declare x86_fp80 @llvm.round.f80(x86_fp80 %Val)
10668 declare fp128 @llvm.round.f128(fp128 %Val)
10669 declare ppc_fp128 @llvm.round.ppcf128(ppc_fp128 %Val)
10674 The '``llvm.round.*``' intrinsics returns the operand rounded to the
10680 The argument and return value are floating point numbers of the same
10686 This function returns the same values as the libm ``round``
10687 functions would, and handles error conditions in the same way.
10689 Bit Manipulation Intrinsics
10690 ---------------------------
10692 LLVM provides intrinsics for a few important bit manipulation
10693 operations. These allow efficient code generation for some algorithms.
10695 '``llvm.bitreverse.*``' Intrinsics
10696 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10701 This is an overloaded intrinsic function. You can use bitreverse on any
10706 declare i16 @llvm.bitreverse.i16(i16 <id>)
10707 declare i32 @llvm.bitreverse.i32(i32 <id>)
10708 declare i64 @llvm.bitreverse.i64(i64 <id>)
10713 The '``llvm.bitreverse``' family of intrinsics is used to reverse the
10714 bitpattern of an integer value; for example ``0b10110110`` becomes
10720 The ``llvm.bitreverse.iN`` intrinsic returns an i16 value that has bit
10721 ``M`` in the input moved to bit ``N-M`` in the output.
10723 '``llvm.bswap.*``' Intrinsics
10724 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10729 This is an overloaded intrinsic function. You can use bswap on any
10730 integer type that is an even number of bytes (i.e. BitWidth % 16 == 0).
10734 declare i16 @llvm.bswap.i16(i16 <id>)
10735 declare i32 @llvm.bswap.i32(i32 <id>)
10736 declare i64 @llvm.bswap.i64(i64 <id>)
10741 The '``llvm.bswap``' family of intrinsics is used to byte swap integer
10742 values with an even number of bytes (positive multiple of 16 bits).
10743 These are useful for performing operations on data that is not in the
10744 target's native byte order.
10749 The ``llvm.bswap.i16`` intrinsic returns an i16 value that has the high
10750 and low byte of the input i16 swapped. Similarly, the ``llvm.bswap.i32``
10751 intrinsic returns an i32 value that has the four bytes of the input i32
10752 swapped, so that if the input bytes are numbered 0, 1, 2, 3 then the
10753 returned i32 will have its bytes in 3, 2, 1, 0 order. The
10754 ``llvm.bswap.i48``, ``llvm.bswap.i64`` and other intrinsics extend this
10755 concept to additional even-byte lengths (6 bytes, 8 bytes and more,
10758 '``llvm.ctpop.*``' Intrinsic
10759 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10764 This is an overloaded intrinsic. You can use llvm.ctpop on any integer
10765 bit width, or on any vector with integer elements. Not all targets
10766 support all bit widths or vector types, however.
10770 declare i8 @llvm.ctpop.i8(i8 <src>)
10771 declare i16 @llvm.ctpop.i16(i16 <src>)
10772 declare i32 @llvm.ctpop.i32(i32 <src>)
10773 declare i64 @llvm.ctpop.i64(i64 <src>)
10774 declare i256 @llvm.ctpop.i256(i256 <src>)
10775 declare <2 x i32> @llvm.ctpop.v2i32(<2 x i32> <src>)
10780 The '``llvm.ctpop``' family of intrinsics counts the number of bits set
10786 The only argument is the value to be counted. The argument may be of any
10787 integer type, or a vector with integer elements. The return type must
10788 match the argument type.
10793 The '``llvm.ctpop``' intrinsic counts the 1's in a variable, or within
10794 each element of a vector.
10796 '``llvm.ctlz.*``' Intrinsic
10797 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10802 This is an overloaded intrinsic. You can use ``llvm.ctlz`` on any
10803 integer bit width, or any vector whose elements are integers. Not all
10804 targets support all bit widths or vector types, however.
10808 declare i8 @llvm.ctlz.i8 (i8 <src>, i1 <is_zero_undef>)
10809 declare i16 @llvm.ctlz.i16 (i16 <src>, i1 <is_zero_undef>)
10810 declare i32 @llvm.ctlz.i32 (i32 <src>, i1 <is_zero_undef>)
10811 declare i64 @llvm.ctlz.i64 (i64 <src>, i1 <is_zero_undef>)
10812 declare i256 @llvm.ctlz.i256(i256 <src>, i1 <is_zero_undef>)
10813 declare <2 x i32> @llvm.ctlz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
10818 The '``llvm.ctlz``' family of intrinsic functions counts the number of
10819 leading zeros in a variable.
10824 The first argument is the value to be counted. This argument may be of
10825 any integer type, or a vector with integer element type. The return
10826 type must match the first argument type.
10828 The second argument must be a constant and is a flag to indicate whether
10829 the intrinsic should ensure that a zero as the first argument produces a
10830 defined result. Historically some architectures did not provide a
10831 defined result for zero values as efficiently, and many algorithms are
10832 now predicated on avoiding zero-value inputs.
10837 The '``llvm.ctlz``' intrinsic counts the leading (most significant)
10838 zeros in a variable, or within each element of the vector. If
10839 ``src == 0`` then the result is the size in bits of the type of ``src``
10840 if ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
10841 ``llvm.ctlz(i32 2) = 30``.
10843 '``llvm.cttz.*``' Intrinsic
10844 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10849 This is an overloaded intrinsic. You can use ``llvm.cttz`` on any
10850 integer bit width, or any vector of integer elements. Not all targets
10851 support all bit widths or vector types, however.
10855 declare i8 @llvm.cttz.i8 (i8 <src>, i1 <is_zero_undef>)
10856 declare i16 @llvm.cttz.i16 (i16 <src>, i1 <is_zero_undef>)
10857 declare i32 @llvm.cttz.i32 (i32 <src>, i1 <is_zero_undef>)
10858 declare i64 @llvm.cttz.i64 (i64 <src>, i1 <is_zero_undef>)
10859 declare i256 @llvm.cttz.i256(i256 <src>, i1 <is_zero_undef>)
10860 declare <2 x i32> @llvm.cttz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
10865 The '``llvm.cttz``' family of intrinsic functions counts the number of
10871 The first argument is the value to be counted. This argument may be of
10872 any integer type, or a vector with integer element type. The return
10873 type must match the first argument type.
10875 The second argument must be a constant and is a flag to indicate whether
10876 the intrinsic should ensure that a zero as the first argument produces a
10877 defined result. Historically some architectures did not provide a
10878 defined result for zero values as efficiently, and many algorithms are
10879 now predicated on avoiding zero-value inputs.
10884 The '``llvm.cttz``' intrinsic counts the trailing (least significant)
10885 zeros in a variable, or within each element of a vector. If ``src == 0``
10886 then the result is the size in bits of the type of ``src`` if
10887 ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
10888 ``llvm.cttz(2) = 1``.
10892 Arithmetic with Overflow Intrinsics
10893 -----------------------------------
10895 LLVM provides intrinsics for fast arithmetic overflow checking.
10897 Each of these intrinsics returns a two-element struct. The first
10898 element of this struct contains the result of the corresponding
10899 arithmetic operation modulo 2\ :sup:`n`\ , where n is the bit width of
10900 the result. Therefore, for example, the first element of the struct
10901 returned by ``llvm.sadd.with.overflow.i32`` is always the same as the
10902 result of a 32-bit ``add`` instruction with the same operands, where
10903 the ``add`` is *not* modified by an ``nsw`` or ``nuw`` flag.
10905 The second element of the result is an ``i1`` that is 1 if the
10906 arithmetic operation overflowed and 0 otherwise. An operation
10907 overflows if, for any values of its operands ``A`` and ``B`` and for
10908 any ``N`` larger than the operands' width, ``ext(A op B) to iN`` is
10909 not equal to ``(ext(A) to iN) op (ext(B) to iN)`` where ``ext`` is
10910 ``sext`` for signed overflow and ``zext`` for unsigned overflow, and
10911 ``op`` is the underlying arithmetic operation.
10913 The behavior of these intrinsics is well-defined for all argument
10916 '``llvm.sadd.with.overflow.*``' Intrinsics
10917 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10922 This is an overloaded intrinsic. You can use ``llvm.sadd.with.overflow``
10923 on any integer bit width.
10927 declare {i16, i1} @llvm.sadd.with.overflow.i16(i16 %a, i16 %b)
10928 declare {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
10929 declare {i64, i1} @llvm.sadd.with.overflow.i64(i64 %a, i64 %b)
10934 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
10935 a signed addition of the two arguments, and indicate whether an overflow
10936 occurred during the signed summation.
10941 The arguments (%a and %b) and the first element of the result structure
10942 may be of integer types of any bit width, but they must have the same
10943 bit width. The second element of the result structure must be of type
10944 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
10950 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
10951 a signed addition of the two variables. They return a structure --- the
10952 first element of which is the signed summation, and the second element
10953 of which is a bit specifying if the signed summation resulted in an
10959 .. code-block:: llvm
10961 %res = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
10962 %sum = extractvalue {i32, i1} %res, 0
10963 %obit = extractvalue {i32, i1} %res, 1
10964 br i1 %obit, label %overflow, label %normal
10966 '``llvm.uadd.with.overflow.*``' Intrinsics
10967 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10972 This is an overloaded intrinsic. You can use ``llvm.uadd.with.overflow``
10973 on any integer bit width.
10977 declare {i16, i1} @llvm.uadd.with.overflow.i16(i16 %a, i16 %b)
10978 declare {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
10979 declare {i64, i1} @llvm.uadd.with.overflow.i64(i64 %a, i64 %b)
10984 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
10985 an unsigned addition of the two arguments, and indicate whether a carry
10986 occurred during the unsigned summation.
10991 The arguments (%a and %b) and the first element of the result structure
10992 may be of integer types of any bit width, but they must have the same
10993 bit width. The second element of the result structure must be of type
10994 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
11000 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
11001 an unsigned addition of the two arguments. They return a structure --- the
11002 first element of which is the sum, and the second element of which is a
11003 bit specifying if the unsigned summation resulted in a carry.
11008 .. code-block:: llvm
11010 %res = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
11011 %sum = extractvalue {i32, i1} %res, 0
11012 %obit = extractvalue {i32, i1} %res, 1
11013 br i1 %obit, label %carry, label %normal
11015 '``llvm.ssub.with.overflow.*``' Intrinsics
11016 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11021 This is an overloaded intrinsic. You can use ``llvm.ssub.with.overflow``
11022 on any integer bit width.
11026 declare {i16, i1} @llvm.ssub.with.overflow.i16(i16 %a, i16 %b)
11027 declare {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
11028 declare {i64, i1} @llvm.ssub.with.overflow.i64(i64 %a, i64 %b)
11033 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
11034 a signed subtraction of the two arguments, and indicate whether an
11035 overflow occurred during the signed subtraction.
11040 The arguments (%a and %b) and the first element of the result structure
11041 may be of integer types of any bit width, but they must have the same
11042 bit width. The second element of the result structure must be of type
11043 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
11049 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
11050 a signed subtraction of the two arguments. They return a structure --- the
11051 first element of which is the subtraction, and the second element of
11052 which is a bit specifying if the signed subtraction resulted in an
11058 .. code-block:: llvm
11060 %res = call {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
11061 %sum = extractvalue {i32, i1} %res, 0
11062 %obit = extractvalue {i32, i1} %res, 1
11063 br i1 %obit, label %overflow, label %normal
11065 '``llvm.usub.with.overflow.*``' Intrinsics
11066 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11071 This is an overloaded intrinsic. You can use ``llvm.usub.with.overflow``
11072 on any integer bit width.
11076 declare {i16, i1} @llvm.usub.with.overflow.i16(i16 %a, i16 %b)
11077 declare {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
11078 declare {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
11083 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
11084 an unsigned subtraction of the two arguments, and indicate whether an
11085 overflow occurred during the unsigned subtraction.
11090 The arguments (%a and %b) and the first element of the result structure
11091 may be of integer types of any bit width, but they must have the same
11092 bit width. The second element of the result structure must be of type
11093 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
11099 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
11100 an unsigned subtraction of the two arguments. They return a structure ---
11101 the first element of which is the subtraction, and the second element of
11102 which is a bit specifying if the unsigned subtraction resulted in an
11108 .. code-block:: llvm
11110 %res = call {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
11111 %sum = extractvalue {i32, i1} %res, 0
11112 %obit = extractvalue {i32, i1} %res, 1
11113 br i1 %obit, label %overflow, label %normal
11115 '``llvm.smul.with.overflow.*``' Intrinsics
11116 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11121 This is an overloaded intrinsic. You can use ``llvm.smul.with.overflow``
11122 on any integer bit width.
11126 declare {i16, i1} @llvm.smul.with.overflow.i16(i16 %a, i16 %b)
11127 declare {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
11128 declare {i64, i1} @llvm.smul.with.overflow.i64(i64 %a, i64 %b)
11133 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
11134 a signed multiplication of the two arguments, and indicate whether an
11135 overflow occurred during the signed multiplication.
11140 The arguments (%a and %b) and the first element of the result structure
11141 may be of integer types of any bit width, but they must have the same
11142 bit width. The second element of the result structure must be of type
11143 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
11149 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
11150 a signed multiplication of the two arguments. They return a structure ---
11151 the first element of which is the multiplication, and the second element
11152 of which is a bit specifying if the signed multiplication resulted in an
11158 .. code-block:: llvm
11160 %res = call {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
11161 %sum = extractvalue {i32, i1} %res, 0
11162 %obit = extractvalue {i32, i1} %res, 1
11163 br i1 %obit, label %overflow, label %normal
11165 '``llvm.umul.with.overflow.*``' Intrinsics
11166 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11171 This is an overloaded intrinsic. You can use ``llvm.umul.with.overflow``
11172 on any integer bit width.
11176 declare {i16, i1} @llvm.umul.with.overflow.i16(i16 %a, i16 %b)
11177 declare {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
11178 declare {i64, i1} @llvm.umul.with.overflow.i64(i64 %a, i64 %b)
11183 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
11184 a unsigned multiplication of the two arguments, and indicate whether an
11185 overflow occurred during the unsigned multiplication.
11190 The arguments (%a and %b) and the first element of the result structure
11191 may be of integer types of any bit width, but they must have the same
11192 bit width. The second element of the result structure must be of type
11193 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
11199 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
11200 an unsigned multiplication of the two arguments. They return a structure ---
11201 the first element of which is the multiplication, and the second
11202 element of which is a bit specifying if the unsigned multiplication
11203 resulted in an overflow.
11208 .. code-block:: llvm
11210 %res = call {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
11211 %sum = extractvalue {i32, i1} %res, 0
11212 %obit = extractvalue {i32, i1} %res, 1
11213 br i1 %obit, label %overflow, label %normal
11215 Specialised Arithmetic Intrinsics
11216 ---------------------------------
11218 '``llvm.canonicalize.*``' Intrinsic
11219 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11226 declare float @llvm.canonicalize.f32(float %a)
11227 declare double @llvm.canonicalize.f64(double %b)
11232 The '``llvm.canonicalize.*``' intrinsic returns the platform specific canonical
11233 encoding of a floating point number. This canonicalization is useful for
11234 implementing certain numeric primitives such as frexp. The canonical encoding is
11235 defined by IEEE-754-2008 to be:
11239 2.1.8 canonical encoding: The preferred encoding of a floating-point
11240 representation in a format. Applied to declets, significands of finite
11241 numbers, infinities, and NaNs, especially in decimal formats.
11243 This operation can also be considered equivalent to the IEEE-754-2008
11244 conversion of a floating-point value to the same format. NaNs are handled
11245 according to section 6.2.
11247 Examples of non-canonical encodings:
11249 - x87 pseudo denormals, pseudo NaNs, pseudo Infinity, Unnormals. These are
11250 converted to a canonical representation per hardware-specific protocol.
11251 - Many normal decimal floating point numbers have non-canonical alternative
11253 - Some machines, like GPUs or ARMv7 NEON, do not support subnormal values.
11254 These are treated as non-canonical encodings of zero and will be flushed to
11255 a zero of the same sign by this operation.
11257 Note that per IEEE-754-2008 6.2, systems that support signaling NaNs with
11258 default exception handling must signal an invalid exception, and produce a
11261 This function should always be implementable as multiplication by 1.0, provided
11262 that the compiler does not constant fold the operation. Likewise, division by
11263 1.0 and ``llvm.minnum(x, x)`` are possible implementations. Addition with
11264 -0.0 is also sufficient provided that the rounding mode is not -Infinity.
11266 ``@llvm.canonicalize`` must preserve the equality relation. That is:
11268 - ``(@llvm.canonicalize(x) == x)`` is equivalent to ``(x == x)``
11269 - ``(@llvm.canonicalize(x) == @llvm.canonicalize(y))`` is equivalent to
11272 Additionally, the sign of zero must be conserved:
11273 ``@llvm.canonicalize(-0.0) = -0.0`` and ``@llvm.canonicalize(+0.0) = +0.0``
11275 The payload bits of a NaN must be conserved, with two exceptions.
11276 First, environments which use only a single canonical representation of NaN
11277 must perform said canonicalization. Second, SNaNs must be quieted per the
11280 The canonicalization operation may be optimized away if:
11282 - The input is known to be canonical. For example, it was produced by a
11283 floating-point operation that is required by the standard to be canonical.
11284 - The result is consumed only by (or fused with) other floating-point
11285 operations. That is, the bits of the floating point value are not examined.
11287 '``llvm.fmuladd.*``' Intrinsic
11288 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11295 declare float @llvm.fmuladd.f32(float %a, float %b, float %c)
11296 declare double @llvm.fmuladd.f64(double %a, double %b, double %c)
11301 The '``llvm.fmuladd.*``' intrinsic functions represent multiply-add
11302 expressions that can be fused if the code generator determines that (a) the
11303 target instruction set has support for a fused operation, and (b) that the
11304 fused operation is more efficient than the equivalent, separate pair of mul
11305 and add instructions.
11310 The '``llvm.fmuladd.*``' intrinsics each take three arguments: two
11311 multiplicands, a and b, and an addend c.
11320 %0 = call float @llvm.fmuladd.f32(%a, %b, %c)
11322 is equivalent to the expression a \* b + c, except that rounding will
11323 not be performed between the multiplication and addition steps if the
11324 code generator fuses the operations. Fusion is not guaranteed, even if
11325 the target platform supports it. If a fused multiply-add is required the
11326 corresponding llvm.fma.\* intrinsic function should be used
11327 instead. This never sets errno, just as '``llvm.fma.*``'.
11332 .. code-block:: llvm
11334 %r2 = call float @llvm.fmuladd.f32(float %a, float %b, float %c) ; yields float:r2 = (a * b) + c
11336 Half Precision Floating Point Intrinsics
11337 ----------------------------------------
11339 For most target platforms, half precision floating point is a
11340 storage-only format. This means that it is a dense encoding (in memory)
11341 but does not support computation in the format.
11343 This means that code must first load the half-precision floating point
11344 value as an i16, then convert it to float with
11345 :ref:`llvm.convert.from.fp16 <int_convert_from_fp16>`. Computation can
11346 then be performed on the float value (including extending to double
11347 etc). To store the value back to memory, it is first converted to float
11348 if needed, then converted to i16 with
11349 :ref:`llvm.convert.to.fp16 <int_convert_to_fp16>`, then storing as an
11352 .. _int_convert_to_fp16:
11354 '``llvm.convert.to.fp16``' Intrinsic
11355 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11362 declare i16 @llvm.convert.to.fp16.f32(float %a)
11363 declare i16 @llvm.convert.to.fp16.f64(double %a)
11368 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion from a
11369 conventional floating point type to half precision floating point format.
11374 The intrinsic function contains single argument - the value to be
11380 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion from a
11381 conventional floating point format to half precision floating point format. The
11382 return value is an ``i16`` which contains the converted number.
11387 .. code-block:: llvm
11389 %res = call i16 @llvm.convert.to.fp16.f32(float %a)
11390 store i16 %res, i16* @x, align 2
11392 .. _int_convert_from_fp16:
11394 '``llvm.convert.from.fp16``' Intrinsic
11395 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11402 declare float @llvm.convert.from.fp16.f32(i16 %a)
11403 declare double @llvm.convert.from.fp16.f64(i16 %a)
11408 The '``llvm.convert.from.fp16``' intrinsic function performs a
11409 conversion from half precision floating point format to single precision
11410 floating point format.
11415 The intrinsic function contains single argument - the value to be
11421 The '``llvm.convert.from.fp16``' intrinsic function performs a
11422 conversion from half single precision floating point format to single
11423 precision floating point format. The input half-float value is
11424 represented by an ``i16`` value.
11429 .. code-block:: llvm
11431 %a = load i16, i16* @x, align 2
11432 %res = call float @llvm.convert.from.fp16(i16 %a)
11434 .. _dbg_intrinsics:
11436 Debugger Intrinsics
11437 -------------------
11439 The LLVM debugger intrinsics (which all start with ``llvm.dbg.``
11440 prefix), are described in the `LLVM Source Level
11441 Debugging <SourceLevelDebugging.html#format_common_intrinsics>`_
11444 Exception Handling Intrinsics
11445 -----------------------------
11447 The LLVM exception handling intrinsics (which all start with
11448 ``llvm.eh.`` prefix), are described in the `LLVM Exception
11449 Handling <ExceptionHandling.html#format_common_intrinsics>`_ document.
11451 .. _int_trampoline:
11453 Trampoline Intrinsics
11454 ---------------------
11456 These intrinsics make it possible to excise one parameter, marked with
11457 the :ref:`nest <nest>` attribute, from a function. The result is a
11458 callable function pointer lacking the nest parameter - the caller does
11459 not need to provide a value for it. Instead, the value to use is stored
11460 in advance in a "trampoline", a block of memory usually allocated on the
11461 stack, which also contains code to splice the nest value into the
11462 argument list. This is used to implement the GCC nested function address
11465 For example, if the function is ``i32 f(i8* nest %c, i32 %x, i32 %y)``
11466 then the resulting function pointer has signature ``i32 (i32, i32)*``.
11467 It can be created as follows:
11469 .. code-block:: llvm
11471 %tramp = alloca [10 x i8], align 4 ; size and alignment only correct for X86
11472 %tramp1 = getelementptr [10 x i8], [10 x i8]* %tramp, i32 0, i32 0
11473 call i8* @llvm.init.trampoline(i8* %tramp1, i8* bitcast (i32 (i8*, i32, i32)* @f to i8*), i8* %nval)
11474 %p = call i8* @llvm.adjust.trampoline(i8* %tramp1)
11475 %fp = bitcast i8* %p to i32 (i32, i32)*
11477 The call ``%val = call i32 %fp(i32 %x, i32 %y)`` is then equivalent to
11478 ``%val = call i32 %f(i8* %nval, i32 %x, i32 %y)``.
11482 '``llvm.init.trampoline``' Intrinsic
11483 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11490 declare void @llvm.init.trampoline(i8* <tramp>, i8* <func>, i8* <nval>)
11495 This fills the memory pointed to by ``tramp`` with executable code,
11496 turning it into a trampoline.
11501 The ``llvm.init.trampoline`` intrinsic takes three arguments, all
11502 pointers. The ``tramp`` argument must point to a sufficiently large and
11503 sufficiently aligned block of memory; this memory is written to by the
11504 intrinsic. Note that the size and the alignment are target-specific -
11505 LLVM currently provides no portable way of determining them, so a
11506 front-end that generates this intrinsic needs to have some
11507 target-specific knowledge. The ``func`` argument must hold a function
11508 bitcast to an ``i8*``.
11513 The block of memory pointed to by ``tramp`` is filled with target
11514 dependent code, turning it into a function. Then ``tramp`` needs to be
11515 passed to :ref:`llvm.adjust.trampoline <int_at>` to get a pointer which can
11516 be :ref:`bitcast (to a new function) and called <int_trampoline>`. The new
11517 function's signature is the same as that of ``func`` with any arguments
11518 marked with the ``nest`` attribute removed. At most one such ``nest``
11519 argument is allowed, and it must be of pointer type. Calling the new
11520 function is equivalent to calling ``func`` with the same argument list,
11521 but with ``nval`` used for the missing ``nest`` argument. If, after
11522 calling ``llvm.init.trampoline``, the memory pointed to by ``tramp`` is
11523 modified, then the effect of any later call to the returned function
11524 pointer is undefined.
11528 '``llvm.adjust.trampoline``' Intrinsic
11529 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11536 declare i8* @llvm.adjust.trampoline(i8* <tramp>)
11541 This performs any required machine-specific adjustment to the address of
11542 a trampoline (passed as ``tramp``).
11547 ``tramp`` must point to a block of memory which already has trampoline
11548 code filled in by a previous call to
11549 :ref:`llvm.init.trampoline <int_it>`.
11554 On some architectures the address of the code to be executed needs to be
11555 different than the address where the trampoline is actually stored. This
11556 intrinsic returns the executable address corresponding to ``tramp``
11557 after performing the required machine specific adjustments. The pointer
11558 returned can then be :ref:`bitcast and executed <int_trampoline>`.
11560 .. _int_mload_mstore:
11562 Masked Vector Load and Store Intrinsics
11563 ---------------------------------------
11565 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.
11569 '``llvm.masked.load.*``' Intrinsics
11570 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11574 This is an overloaded intrinsic. The loaded data is a vector of any integer, floating point or pointer data type.
11578 declare <16 x float> @llvm.masked.load.v16f32.p0v16f32 (<16 x float>* <ptr>, i32 <alignment>, <16 x i1> <mask>, <16 x float> <passthru>)
11579 declare <2 x double> @llvm.masked.load.v2f64.p0v2f64 (<2 x double>* <ptr>, i32 <alignment>, <2 x i1> <mask>, <2 x double> <passthru>)
11580 ;; The data is a vector of pointers to double
11581 declare <8 x double*> @llvm.masked.load.v8p0f64.p0v8p0f64 (<8 x double*>* <ptr>, i32 <alignment>, <8 x i1> <mask>, <8 x double*> <passthru>)
11582 ;; The data is a vector of function pointers
11583 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>)
11588 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.
11594 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.
11600 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.
11601 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.
11606 %res = call <16 x float> @llvm.masked.load.v16f32.p0v16f32 (<16 x float>* %ptr, i32 4, <16 x i1>%mask, <16 x float> %passthru)
11608 ;; The result of the two following instructions is identical aside from potential memory access exception
11609 %loadlal = load <16 x float>, <16 x float>* %ptr, align 4
11610 %res = select <16 x i1> %mask, <16 x float> %loadlal, <16 x float> %passthru
11614 '``llvm.masked.store.*``' Intrinsics
11615 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11619 This is an overloaded intrinsic. The data stored in memory is a vector of any integer, floating point or pointer data type.
11623 declare void @llvm.masked.store.v8i32.p0v8i32 (<8 x i32> <value>, <8 x i32>* <ptr>, i32 <alignment>, <8 x i1> <mask>)
11624 declare void @llvm.masked.store.v16f32.p0v16f32 (<16 x float> <value>, <16 x float>* <ptr>, i32 <alignment>, <16 x i1> <mask>)
11625 ;; The data is a vector of pointers to double
11626 declare void @llvm.masked.store.v8p0f64.p0v8p0f64 (<8 x double*> <value>, <8 x double*>* <ptr>, i32 <alignment>, <8 x i1> <mask>)
11627 ;; The data is a vector of function pointers
11628 declare void @llvm.masked.store.v4p0f_i32f.p0v4p0f_i32f (<4 x i32 ()*> <value>, <4 x i32 ()*>* <ptr>, i32 <alignment>, <4 x i1> <mask>)
11633 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.
11638 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.
11644 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.
11645 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.
11649 call void @llvm.masked.store.v16f32.p0v16f32(<16 x float> %value, <16 x float>* %ptr, i32 4, <16 x i1> %mask)
11651 ;; The result of the following instructions is identical aside from potential data races and memory access exceptions
11652 %oldval = load <16 x float>, <16 x float>* %ptr, align 4
11653 %res = select <16 x i1> %mask, <16 x float> %value, <16 x float> %oldval
11654 store <16 x float> %res, <16 x float>* %ptr, align 4
11657 Masked Vector Gather and Scatter Intrinsics
11658 -------------------------------------------
11660 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.
11664 '``llvm.masked.gather.*``' Intrinsics
11665 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11669 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.
11673 declare <16 x float> @llvm.masked.gather.v16f32 (<16 x float*> <ptrs>, i32 <alignment>, <16 x i1> <mask>, <16 x float> <passthru>)
11674 declare <2 x double> @llvm.masked.gather.v2f64 (<2 x double*> <ptrs>, i32 <alignment>, <2 x i1> <mask>, <2 x double> <passthru>)
11675 declare <8 x float*> @llvm.masked.gather.v8p0f32 (<8 x float**> <ptrs>, i32 <alignment>, <8 x i1> <mask>, <8 x float*> <passthru>)
11680 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.
11686 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.
11692 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.
11693 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.
11698 %res = call <4 x double> @llvm.masked.gather.v4f64 (<4 x double*> %ptrs, i32 8, <4 x i1>%mask, <4 x double> <true, true, true, true>)
11700 ;; The gather with all-true mask is equivalent to the following instruction sequence
11701 %ptr0 = extractelement <4 x double*> %ptrs, i32 0
11702 %ptr1 = extractelement <4 x double*> %ptrs, i32 1
11703 %ptr2 = extractelement <4 x double*> %ptrs, i32 2
11704 %ptr3 = extractelement <4 x double*> %ptrs, i32 3
11706 %val0 = load double, double* %ptr0, align 8
11707 %val1 = load double, double* %ptr1, align 8
11708 %val2 = load double, double* %ptr2, align 8
11709 %val3 = load double, double* %ptr3, align 8
11711 %vec0 = insertelement <4 x double>undef, %val0, 0
11712 %vec01 = insertelement <4 x double>%vec0, %val1, 1
11713 %vec012 = insertelement <4 x double>%vec01, %val2, 2
11714 %vec0123 = insertelement <4 x double>%vec012, %val3, 3
11718 '``llvm.masked.scatter.*``' Intrinsics
11719 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11723 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.
11727 declare void @llvm.masked.scatter.v8i32 (<8 x i32> <value>, <8 x i32*> <ptrs>, i32 <alignment>, <8 x i1> <mask>)
11728 declare void @llvm.masked.scatter.v16f32 (<16 x float> <value>, <16 x float*> <ptrs>, i32 <alignment>, <16 x i1> <mask>)
11729 declare void @llvm.masked.scatter.v4p0f64 (<4 x double*> <value>, <4 x double**> <ptrs>, i32 <alignment>, <4 x i1> <mask>)
11734 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.
11739 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.
11745 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.
11749 ;; This instruction unconditionally stores data vector in multiple addresses
11750 call @llvm.masked.scatter.v8i32 (<8 x i32> %value, <8 x i32*> %ptrs, i32 4, <8 x i1> <true, true, .. true>)
11752 ;; It is equivalent to a list of scalar stores
11753 %val0 = extractelement <8 x i32> %value, i32 0
11754 %val1 = extractelement <8 x i32> %value, i32 1
11756 %val7 = extractelement <8 x i32> %value, i32 7
11757 %ptr0 = extractelement <8 x i32*> %ptrs, i32 0
11758 %ptr1 = extractelement <8 x i32*> %ptrs, i32 1
11760 %ptr7 = extractelement <8 x i32*> %ptrs, i32 7
11761 ;; Note: the order of the following stores is important when they overlap:
11762 store i32 %val0, i32* %ptr0, align 4
11763 store i32 %val1, i32* %ptr1, align 4
11765 store i32 %val7, i32* %ptr7, align 4
11771 This class of intrinsics provides information about the lifetime of
11772 memory objects and ranges where variables are immutable.
11776 '``llvm.lifetime.start``' Intrinsic
11777 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11784 declare void @llvm.lifetime.start(i64 <size>, i8* nocapture <ptr>)
11789 The '``llvm.lifetime.start``' intrinsic specifies the start of a memory
11795 The first argument is a constant integer representing the size of the
11796 object, or -1 if it is variable sized. The second argument is a pointer
11802 This intrinsic indicates that before this point in the code, the value
11803 of the memory pointed to by ``ptr`` is dead. This means that it is known
11804 to never be used and has an undefined value. A load from the pointer
11805 that precedes this intrinsic can be replaced with ``'undef'``.
11809 '``llvm.lifetime.end``' Intrinsic
11810 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11817 declare void @llvm.lifetime.end(i64 <size>, i8* nocapture <ptr>)
11822 The '``llvm.lifetime.end``' intrinsic specifies the end of a memory
11828 The first argument is a constant integer representing the size of the
11829 object, or -1 if it is variable sized. The second argument is a pointer
11835 This intrinsic indicates that after this point in the code, the value of
11836 the memory pointed to by ``ptr`` is dead. This means that it is known to
11837 never be used and has an undefined value. Any stores into the memory
11838 object following this intrinsic may be removed as dead.
11840 '``llvm.invariant.start``' Intrinsic
11841 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11848 declare {}* @llvm.invariant.start(i64 <size>, i8* nocapture <ptr>)
11853 The '``llvm.invariant.start``' intrinsic specifies that the contents of
11854 a memory object will not change.
11859 The first argument is a constant integer representing the size of the
11860 object, or -1 if it is variable sized. The second argument is a pointer
11866 This intrinsic indicates that until an ``llvm.invariant.end`` that uses
11867 the return value, the referenced memory location is constant and
11870 '``llvm.invariant.end``' Intrinsic
11871 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11878 declare void @llvm.invariant.end({}* <start>, i64 <size>, i8* nocapture <ptr>)
11883 The '``llvm.invariant.end``' intrinsic specifies that the contents of a
11884 memory object are mutable.
11889 The first argument is the matching ``llvm.invariant.start`` intrinsic.
11890 The second argument is a constant integer representing the size of the
11891 object, or -1 if it is variable sized and the third argument is a
11892 pointer to the object.
11897 This intrinsic indicates that the memory is mutable again.
11899 '``llvm.invariant.group.barrier``' Intrinsic
11900 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11907 declare i8* @llvm.invariant.group.barrier(i8* <ptr>)
11912 The '``llvm.invariant.group.barrier``' intrinsic can be used when an invariant
11913 established by invariant.group metadata no longer holds, to obtain a new pointer
11914 value that does not carry the invariant information.
11920 The ``llvm.invariant.group.barrier`` takes only one argument, which is
11921 the pointer to the memory for which the ``invariant.group`` no longer holds.
11926 Returns another pointer that aliases its argument but which is considered different
11927 for the purposes of ``load``/``store`` ``invariant.group`` metadata.
11932 This class of intrinsics is designed to be generic and has no specific
11935 '``llvm.var.annotation``' Intrinsic
11936 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11943 declare void @llvm.var.annotation(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
11948 The '``llvm.var.annotation``' intrinsic.
11953 The first argument is a pointer to a value, the second is a pointer to a
11954 global string, the third is a pointer to a global string which is the
11955 source file name, and the last argument is the line number.
11960 This intrinsic allows annotation of local variables with arbitrary
11961 strings. This can be useful for special purpose optimizations that want
11962 to look for these annotations. These have no other defined use; they are
11963 ignored by code generation and optimization.
11965 '``llvm.ptr.annotation.*``' Intrinsic
11966 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11971 This is an overloaded intrinsic. You can use '``llvm.ptr.annotation``' on a
11972 pointer to an integer of any width. *NOTE* you must specify an address space for
11973 the pointer. The identifier for the default address space is the integer
11978 declare i8* @llvm.ptr.annotation.p<address space>i8(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
11979 declare i16* @llvm.ptr.annotation.p<address space>i16(i16* <val>, i8* <str>, i8* <str>, i32 <int>)
11980 declare i32* @llvm.ptr.annotation.p<address space>i32(i32* <val>, i8* <str>, i8* <str>, i32 <int>)
11981 declare i64* @llvm.ptr.annotation.p<address space>i64(i64* <val>, i8* <str>, i8* <str>, i32 <int>)
11982 declare i256* @llvm.ptr.annotation.p<address space>i256(i256* <val>, i8* <str>, i8* <str>, i32 <int>)
11987 The '``llvm.ptr.annotation``' intrinsic.
11992 The first argument is a pointer to an integer value of arbitrary bitwidth
11993 (result of some expression), the second is a pointer to a global string, the
11994 third is a pointer to a global string which is the source file name, and the
11995 last argument is the line number. It returns the value of the first argument.
12000 This intrinsic allows annotation of a pointer to an integer with arbitrary
12001 strings. This can be useful for special purpose optimizations that want to look
12002 for these annotations. These have no other defined use; they are ignored by code
12003 generation and optimization.
12005 '``llvm.annotation.*``' Intrinsic
12006 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12011 This is an overloaded intrinsic. You can use '``llvm.annotation``' on
12012 any integer bit width.
12016 declare i8 @llvm.annotation.i8(i8 <val>, i8* <str>, i8* <str>, i32 <int>)
12017 declare i16 @llvm.annotation.i16(i16 <val>, i8* <str>, i8* <str>, i32 <int>)
12018 declare i32 @llvm.annotation.i32(i32 <val>, i8* <str>, i8* <str>, i32 <int>)
12019 declare i64 @llvm.annotation.i64(i64 <val>, i8* <str>, i8* <str>, i32 <int>)
12020 declare i256 @llvm.annotation.i256(i256 <val>, i8* <str>, i8* <str>, i32 <int>)
12025 The '``llvm.annotation``' intrinsic.
12030 The first argument is an integer value (result of some expression), the
12031 second is a pointer to a global string, the third is a pointer to a
12032 global string which is the source file name, and the last argument is
12033 the line number. It returns the value of the first argument.
12038 This intrinsic allows annotations to be put on arbitrary expressions
12039 with arbitrary strings. This can be useful for special purpose
12040 optimizations that want to look for these annotations. These have no
12041 other defined use; they are ignored by code generation and optimization.
12043 '``llvm.trap``' Intrinsic
12044 ^^^^^^^^^^^^^^^^^^^^^^^^^
12051 declare void @llvm.trap() noreturn nounwind
12056 The '``llvm.trap``' intrinsic.
12066 This intrinsic is lowered to the target dependent trap instruction. If
12067 the target does not have a trap instruction, this intrinsic will be
12068 lowered to a call of the ``abort()`` function.
12070 '``llvm.debugtrap``' Intrinsic
12071 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12078 declare void @llvm.debugtrap() nounwind
12083 The '``llvm.debugtrap``' intrinsic.
12093 This intrinsic is lowered to code which is intended to cause an
12094 execution trap with the intention of requesting the attention of a
12097 '``llvm.stackprotector``' Intrinsic
12098 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12105 declare void @llvm.stackprotector(i8* <guard>, i8** <slot>)
12110 The ``llvm.stackprotector`` intrinsic takes the ``guard`` and stores it
12111 onto the stack at ``slot``. The stack slot is adjusted to ensure that it
12112 is placed on the stack before local variables.
12117 The ``llvm.stackprotector`` intrinsic requires two pointer arguments.
12118 The first argument is the value loaded from the stack guard
12119 ``@__stack_chk_guard``. The second variable is an ``alloca`` that has
12120 enough space to hold the value of the guard.
12125 This intrinsic causes the prologue/epilogue inserter to force the position of
12126 the ``AllocaInst`` stack slot to be before local variables on the stack. This is
12127 to ensure that if a local variable on the stack is overwritten, it will destroy
12128 the value of the guard. When the function exits, the guard on the stack is
12129 checked against the original guard by ``llvm.stackprotectorcheck``. If they are
12130 different, then ``llvm.stackprotectorcheck`` causes the program to abort by
12131 calling the ``__stack_chk_fail()`` function.
12133 '``llvm.stackguard``' Intrinsic
12134 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12141 declare i8* @llvm.stackguard()
12146 The ``llvm.stackguard`` intrinsic returns the system stack guard value.
12148 It should not be generated by frontends, since it is only for internal usage.
12149 The reason why we create this intrinsic is that we still support IR form Stack
12150 Protector in FastISel.
12160 On some platforms, the value returned by this intrinsic remains unchanged
12161 between loads in the same thread. On other platforms, it returns the same
12162 global variable value, if any, e.g. ``@__stack_chk_guard``.
12164 Currently some platforms have IR-level customized stack guard loading (e.g.
12165 X86 Linux) that is not handled by ``llvm.stackguard()``, while they should be
12168 '``llvm.objectsize``' Intrinsic
12169 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12176 declare i32 @llvm.objectsize.i32(i8* <object>, i1 <min>)
12177 declare i64 @llvm.objectsize.i64(i8* <object>, i1 <min>)
12182 The ``llvm.objectsize`` intrinsic is designed to provide information to
12183 the optimizers to determine at compile time whether a) an operation
12184 (like memcpy) will overflow a buffer that corresponds to an object, or
12185 b) that a runtime check for overflow isn't necessary. An object in this
12186 context means an allocation of a specific class, structure, array, or
12192 The ``llvm.objectsize`` intrinsic takes two arguments. The first
12193 argument is a pointer to or into the ``object``. The second argument is
12194 a boolean and determines whether ``llvm.objectsize`` returns 0 (if true)
12195 or -1 (if false) when the object size is unknown. The second argument
12196 only accepts constants.
12201 The ``llvm.objectsize`` intrinsic is lowered to a constant representing
12202 the size of the object concerned. If the size cannot be determined at
12203 compile time, ``llvm.objectsize`` returns ``i32/i64 -1 or 0`` (depending
12204 on the ``min`` argument).
12206 '``llvm.expect``' Intrinsic
12207 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
12212 This is an overloaded intrinsic. You can use ``llvm.expect`` on any
12217 declare i1 @llvm.expect.i1(i1 <val>, i1 <expected_val>)
12218 declare i32 @llvm.expect.i32(i32 <val>, i32 <expected_val>)
12219 declare i64 @llvm.expect.i64(i64 <val>, i64 <expected_val>)
12224 The ``llvm.expect`` intrinsic provides information about expected (the
12225 most probable) value of ``val``, which can be used by optimizers.
12230 The ``llvm.expect`` intrinsic takes two arguments. The first argument is
12231 a value. The second argument is an expected value, this needs to be a
12232 constant value, variables are not allowed.
12237 This intrinsic is lowered to the ``val``.
12241 '``llvm.assume``' Intrinsic
12242 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12249 declare void @llvm.assume(i1 %cond)
12254 The ``llvm.assume`` allows the optimizer to assume that the provided
12255 condition is true. This information can then be used in simplifying other parts
12261 The condition which the optimizer may assume is always true.
12266 The intrinsic allows the optimizer to assume that the provided condition is
12267 always true whenever the control flow reaches the intrinsic call. No code is
12268 generated for this intrinsic, and instructions that contribute only to the
12269 provided condition are not used for code generation. If the condition is
12270 violated during execution, the behavior is undefined.
12272 Note that the optimizer might limit the transformations performed on values
12273 used by the ``llvm.assume`` intrinsic in order to preserve the instructions
12274 only used to form the intrinsic's input argument. This might prove undesirable
12275 if the extra information provided by the ``llvm.assume`` intrinsic does not cause
12276 sufficient overall improvement in code quality. For this reason,
12277 ``llvm.assume`` should not be used to document basic mathematical invariants
12278 that the optimizer can otherwise deduce or facts that are of little use to the
12283 '``llvm.type.test``' Intrinsic
12284 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12291 declare i1 @llvm.type.test(i8* %ptr, metadata %type) nounwind readnone
12297 The first argument is a pointer to be tested. The second argument is a
12298 metadata object representing a :doc:`type identifier <TypeMetadata>`.
12303 The ``llvm.type.test`` intrinsic tests whether the given pointer is associated
12304 with the given type identifier.
12306 '``llvm.type.checked.load``' Intrinsic
12307 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12314 declare {i8*, i1} @llvm.type.checked.load(i8* %ptr, i32 %offset, metadata %type) argmemonly nounwind readonly
12320 The first argument is a pointer from which to load a function pointer. The
12321 second argument is the byte offset from which to load the function pointer. The
12322 third argument is a metadata object representing a :doc:`type identifier
12328 The ``llvm.type.checked.load`` intrinsic safely loads a function pointer from a
12329 virtual table pointer using type metadata. This intrinsic is used to implement
12330 control flow integrity in conjunction with virtual call optimization. The
12331 virtual call optimization pass will optimize away ``llvm.type.checked.load``
12332 intrinsics associated with devirtualized calls, thereby removing the type
12333 check in cases where it is not needed to enforce the control flow integrity
12336 If the given pointer is associated with a type metadata identifier, this
12337 function returns true as the second element of its return value. (Note that
12338 the function may also return true if the given pointer is not associated
12339 with a type metadata identifier.) If the function's return value's second
12340 element is true, the following rules apply to the first element:
12342 - If the given pointer is associated with the given type metadata identifier,
12343 it is the function pointer loaded from the given byte offset from the given
12346 - If the given pointer is not associated with the given type metadata
12347 identifier, it is one of the following (the choice of which is unspecified):
12349 1. The function pointer that would have been loaded from an arbitrarily chosen
12350 (through an unspecified mechanism) pointer associated with the type
12353 2. If the function has a non-void return type, a pointer to a function that
12354 returns an unspecified value without causing side effects.
12356 If the function's return value's second element is false, the value of the
12357 first element is undefined.
12360 '``llvm.donothing``' Intrinsic
12361 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12368 declare void @llvm.donothing() nounwind readnone
12373 The ``llvm.donothing`` intrinsic doesn't perform any operation. It's one of only
12374 three intrinsics (besides ``llvm.experimental.patchpoint`` and
12375 ``llvm.experimental.gc.statepoint``) that can be called with an invoke
12386 This intrinsic does nothing, and it's removed by optimizers and ignored
12389 '``llvm.experimental.deoptimize``' Intrinsic
12390 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12397 declare type @llvm.experimental.deoptimize(...) [ "deopt"(...) ]
12402 This intrinsic, together with :ref:`deoptimization operand bundles
12403 <deopt_opbundles>`, allow frontends to express transfer of control and
12404 frame-local state from the currently executing (typically more specialized,
12405 hence faster) version of a function into another (typically more generic, hence
12408 In languages with a fully integrated managed runtime like Java and JavaScript
12409 this intrinsic can be used to implement "uncommon trap" or "side exit" like
12410 functionality. In unmanaged languages like C and C++, this intrinsic can be
12411 used to represent the slow paths of specialized functions.
12417 The intrinsic takes an arbitrary number of arguments, whose meaning is
12418 decided by the :ref:`lowering strategy<deoptimize_lowering>`.
12423 The ``@llvm.experimental.deoptimize`` intrinsic executes an attached
12424 deoptimization continuation (denoted using a :ref:`deoptimization
12425 operand bundle <deopt_opbundles>`) and returns the value returned by
12426 the deoptimization continuation. Defining the semantic properties of
12427 the continuation itself is out of scope of the language reference --
12428 as far as LLVM is concerned, the deoptimization continuation can
12429 invoke arbitrary side effects, including reading from and writing to
12432 Deoptimization continuations expressed using ``"deopt"`` operand bundles always
12433 continue execution to the end of the physical frame containing them, so all
12434 calls to ``@llvm.experimental.deoptimize`` must be in "tail position":
12436 - ``@llvm.experimental.deoptimize`` cannot be invoked.
12437 - The call must immediately precede a :ref:`ret <i_ret>` instruction.
12438 - The ``ret`` instruction must return the value produced by the
12439 ``@llvm.experimental.deoptimize`` call if there is one, or void.
12441 Note that the above restrictions imply that the return type for a call to
12442 ``@llvm.experimental.deoptimize`` will match the return type of its immediate
12445 The inliner composes the ``"deopt"`` continuations of the caller into the
12446 ``"deopt"`` continuations present in the inlinee, and also updates calls to this
12447 intrinsic to return directly from the frame of the function it inlined into.
12449 All declarations of ``@llvm.experimental.deoptimize`` must share the
12450 same calling convention.
12452 .. _deoptimize_lowering:
12457 Calls to ``@llvm.experimental.deoptimize`` are lowered to calls to the
12458 symbol ``__llvm_deoptimize`` (it is the frontend's responsibility to
12459 ensure that this symbol is defined). The call arguments to
12460 ``@llvm.experimental.deoptimize`` are lowered as if they were formal
12461 arguments of the specified types, and not as varargs.
12464 '``llvm.experimental.guard``' Intrinsic
12465 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12472 declare void @llvm.experimental.guard(i1, ...) [ "deopt"(...) ]
12477 This intrinsic, together with :ref:`deoptimization operand bundles
12478 <deopt_opbundles>`, allows frontends to express guards or checks on
12479 optimistic assumptions made during compilation. The semantics of
12480 ``@llvm.experimental.guard`` is defined in terms of
12481 ``@llvm.experimental.deoptimize`` -- its body is defined to be
12484 .. code-block:: text
12486 define void @llvm.experimental.guard(i1 %pred, <args...>) {
12487 %realPred = and i1 %pred, undef
12488 br i1 %realPred, label %continue, label %leave [, !make.implicit !{}]
12491 call void @llvm.experimental.deoptimize(<args...>) [ "deopt"() ]
12499 with the optional ``[, !make.implicit !{}]`` present if and only if it
12500 is present on the call site. For more details on ``!make.implicit``,
12501 see :doc:`FaultMaps`.
12503 In words, ``@llvm.experimental.guard`` executes the attached
12504 ``"deopt"`` continuation if (but **not** only if) its first argument
12505 is ``false``. Since the optimizer is allowed to replace the ``undef``
12506 with an arbitrary value, it can optimize guard to fail "spuriously",
12507 i.e. without the original condition being false (hence the "not only
12508 if"); and this allows for "check widening" type optimizations.
12510 ``@llvm.experimental.guard`` cannot be invoked.
12513 '``llvm.load.relative``' Intrinsic
12514 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12521 declare i8* @llvm.load.relative.iN(i8* %ptr, iN %offset) argmemonly nounwind readonly
12526 This intrinsic loads a 32-bit value from the address ``%ptr + %offset``,
12527 adds ``%ptr`` to that value and returns it. The constant folder specifically
12528 recognizes the form of this intrinsic and the constant initializers it may
12529 load from; if a loaded constant initializer is known to have the form
12530 ``i32 trunc(x - %ptr)``, the intrinsic call is folded to ``x``.
12532 LLVM provides that the calculation of such a constant initializer will
12533 not overflow at link time under the medium code model if ``x`` is an
12534 ``unnamed_addr`` function. However, it does not provide this guarantee for
12535 a constant initializer folded into a function body. This intrinsic can be
12536 used to avoid the possibility of overflows when loading from such a constant.
12538 Stack Map Intrinsics
12539 --------------------
12541 LLVM provides experimental intrinsics to support runtime patching
12542 mechanisms commonly desired in dynamic language JITs. These intrinsics
12543 are described in :doc:`StackMaps`.