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 The semantics of this linkage follow the ELF object file model: the
255 symbol is weak until linked, if not linked, the symbol becomes null
256 instead of being an undefined reference.
257 ``linkonce_odr``, ``weak_odr``
258 Some languages allow differing globals to be merged, such as two
259 functions with different semantics. Other languages, such as
260 ``C++``, ensure that only equivalent globals are ever merged (the
261 "one definition rule" --- "ODR"). Such languages can use the
262 ``linkonce_odr`` and ``weak_odr`` linkage types to indicate that the
263 global will only be merged with equivalent globals. These linkage
264 types are otherwise the same as their non-``odr`` versions.
266 If none of the above identifiers are used, the global is externally
267 visible, meaning that it participates in linkage and can be used to
268 resolve external symbol references.
270 It is illegal for a function *declaration* to have any linkage type
271 other than ``external`` or ``extern_weak``.
278 LLVM :ref:`functions <functionstructure>`, :ref:`calls <i_call>` and
279 :ref:`invokes <i_invoke>` can all have an optional calling convention
280 specified for the call. The calling convention of any pair of dynamic
281 caller/callee must match, or the behavior of the program is undefined.
282 The following calling conventions are supported by LLVM, and more may be
285 "``ccc``" - The C calling convention
286 This calling convention (the default if no other calling convention
287 is specified) matches the target C calling conventions. This calling
288 convention supports varargs function calls and tolerates some
289 mismatch in the declared prototype and implemented declaration of
290 the function (as does normal C).
291 "``fastcc``" - The fast calling convention
292 This calling convention attempts to make calls as fast as possible
293 (e.g. by passing things in registers). This calling convention
294 allows the target to use whatever tricks it wants to produce fast
295 code for the target, without having to conform to an externally
296 specified ABI (Application Binary Interface). `Tail calls can only
297 be optimized when this, the GHC or the HiPE convention is
298 used. <CodeGenerator.html#id80>`_ This calling convention does not
299 support varargs and requires the prototype of all callees to exactly
300 match the prototype of the function definition.
301 "``coldcc``" - The cold calling convention
302 This calling convention attempts to make code in the caller as
303 efficient as possible under the assumption that the call is not
304 commonly executed. As such, these calls often preserve all registers
305 so that the call does not break any live ranges in the caller side.
306 This calling convention does not support varargs and requires the
307 prototype of all callees to exactly match the prototype of the
308 function definition. Furthermore the inliner doesn't consider such function
310 "``cc 10``" - GHC convention
311 This calling convention has been implemented specifically for use by
312 the `Glasgow Haskell Compiler (GHC) <http://www.haskell.org/ghc>`_.
313 It passes everything in registers, going to extremes to achieve this
314 by disabling callee save registers. This calling convention should
315 not be used lightly but only for specific situations such as an
316 alternative to the *register pinning* performance technique often
317 used when implementing functional programming languages. At the
318 moment only X86 supports this convention and it has the following
321 - On *X86-32* only supports up to 4 bit type parameters. No
322 floating point types are supported.
323 - On *X86-64* only supports up to 10 bit type parameters and 6
324 floating point parameters.
326 This calling convention supports `tail call
327 optimization <CodeGenerator.html#id80>`_ but requires both the
328 caller and callee are using it.
329 "``cc 11``" - The HiPE calling convention
330 This calling convention has been implemented specifically for use by
331 the `High-Performance Erlang
332 (HiPE) <http://www.it.uu.se/research/group/hipe/>`_ compiler, *the*
333 native code compiler of the `Ericsson's Open Source Erlang/OTP
334 system <http://www.erlang.org/download.shtml>`_. It uses more
335 registers for argument passing than the ordinary C calling
336 convention and defines no callee-saved registers. The calling
337 convention properly supports `tail call
338 optimization <CodeGenerator.html#id80>`_ but requires that both the
339 caller and the callee use it. It uses a *register pinning*
340 mechanism, similar to GHC's convention, for keeping frequently
341 accessed runtime components pinned to specific hardware registers.
342 At the moment only X86 supports this convention (both 32 and 64
344 "``webkit_jscc``" - WebKit's JavaScript calling convention
345 This calling convention has been implemented for `WebKit FTL JIT
346 <https://trac.webkit.org/wiki/FTLJIT>`_. It passes arguments on the
347 stack right to left (as cdecl does), and returns a value in the
348 platform's customary return register.
349 "``anyregcc``" - Dynamic calling convention for code patching
350 This is a special convention that supports patching an arbitrary code
351 sequence in place of a call site. This convention forces the call
352 arguments into registers but allows them to be dynamically
353 allocated. This can currently only be used with calls to
354 llvm.experimental.patchpoint because only this intrinsic records
355 the location of its arguments in a side table. See :doc:`StackMaps`.
356 "``preserve_mostcc``" - The `PreserveMost` calling convention
357 This calling convention attempts to make the code in the caller as
358 unintrusive as possible. This convention behaves identically to the `C`
359 calling convention on how arguments and return values are passed, but it
360 uses a different set of caller/callee-saved registers. This alleviates the
361 burden of saving and recovering a large register set before and after the
362 call in the caller. If the arguments are passed in callee-saved registers,
363 then they will be preserved by the callee across the call. This doesn't
364 apply for values returned in callee-saved registers.
366 - On X86-64 the callee preserves all general purpose registers, except for
367 R11. R11 can be used as a scratch register. Floating-point registers
368 (XMMs/YMMs) are not preserved and need to be saved by the caller.
370 The idea behind this convention is to support calls to runtime functions
371 that have a hot path and a cold path. The hot path is usually a small piece
372 of code that doesn't use many registers. The cold path might need to call out to
373 another function and therefore only needs to preserve the caller-saved
374 registers, which haven't already been saved by the caller. The
375 `PreserveMost` calling convention is very similar to the `cold` calling
376 convention in terms of caller/callee-saved registers, but they are used for
377 different types of function calls. `coldcc` is for function calls that are
378 rarely executed, whereas `preserve_mostcc` function calls are intended to be
379 on the hot path and definitely executed a lot. Furthermore `preserve_mostcc`
380 doesn't prevent the inliner from inlining the function call.
382 This calling convention will be used by a future version of the ObjectiveC
383 runtime and should therefore still be considered experimental at this time.
384 Although this convention was created to optimize certain runtime calls to
385 the ObjectiveC runtime, it is not limited to this runtime and might be used
386 by other runtimes in the future too. The current implementation only
387 supports X86-64, but the intention is to support more architectures in the
389 "``preserve_allcc``" - The `PreserveAll` calling convention
390 This calling convention attempts to make the code in the caller even less
391 intrusive than the `PreserveMost` calling convention. This calling
392 convention also behaves identical to the `C` calling convention on how
393 arguments and return values are passed, but it uses a different set of
394 caller/callee-saved registers. This removes the burden of saving and
395 recovering a large register set before and after the call in the caller. If
396 the arguments are passed in callee-saved registers, then they will be
397 preserved by the callee across the call. This doesn't apply for values
398 returned in callee-saved registers.
400 - On X86-64 the callee preserves all general purpose registers, except for
401 R11. R11 can be used as a scratch register. Furthermore it also preserves
402 all floating-point registers (XMMs/YMMs).
404 The idea behind this convention is to support calls to runtime functions
405 that don't need to call out to any other functions.
407 This calling convention, like the `PreserveMost` calling convention, will be
408 used by a future version of the ObjectiveC runtime and should be considered
409 experimental at this time.
410 "``cxx_fast_tlscc``" - The `CXX_FAST_TLS` calling convention for access functions
411 Clang generates an access function to access C++-style TLS. The access
412 function generally has an entry block, an exit block and an initialization
413 block that is run at the first time. The entry and exit blocks can access
414 a few TLS IR variables, each access will be lowered to a platform-specific
417 This calling convention aims to minimize overhead in the caller by
418 preserving as many registers as possible (all the registers that are
419 perserved on the fast path, composed of the entry and exit blocks).
421 This calling convention behaves identical to the `C` calling convention on
422 how arguments and return values are passed, but it uses a different set of
423 caller/callee-saved registers.
425 Given that each platform has its own lowering sequence, hence its own set
426 of preserved registers, we can't use the existing `PreserveMost`.
428 - On X86-64 the callee preserves all general purpose registers, except for
430 "``cc <n>``" - Numbered convention
431 Any calling convention may be specified by number, allowing
432 target-specific calling conventions to be used. Target specific
433 calling conventions start at 64.
435 More calling conventions can be added/defined on an as-needed basis, to
436 support Pascal conventions or any other well-known target-independent
439 .. _visibilitystyles:
444 All Global Variables and Functions have one of the following visibility
447 "``default``" - Default style
448 On targets that use the ELF object file format, default visibility
449 means that the declaration is visible to other modules and, in
450 shared libraries, means that the declared entity may be overridden.
451 On Darwin, default visibility means that the declaration is visible
452 to other modules. Default visibility corresponds to "external
453 linkage" in the language.
454 "``hidden``" - Hidden style
455 Two declarations of an object with hidden visibility refer to the
456 same object if they are in the same shared object. Usually, hidden
457 visibility indicates that the symbol will not be placed into the
458 dynamic symbol table, so no other module (executable or shared
459 library) can reference it directly.
460 "``protected``" - Protected style
461 On ELF, protected visibility indicates that the symbol will be
462 placed in the dynamic symbol table, but that references within the
463 defining module will bind to the local symbol. That is, the symbol
464 cannot be overridden by another module.
466 A symbol with ``internal`` or ``private`` linkage must have ``default``
474 All Global Variables, Functions and Aliases can have one of the following
478 "``dllimport``" causes the compiler to reference a function or variable via
479 a global pointer to a pointer that is set up by the DLL exporting the
480 symbol. On Microsoft Windows targets, the pointer name is formed by
481 combining ``__imp_`` and the function or variable name.
483 "``dllexport``" causes the compiler to provide a global pointer to a pointer
484 in a DLL, so that it can be referenced with the ``dllimport`` attribute. On
485 Microsoft Windows targets, the pointer name is formed by combining
486 ``__imp_`` and the function or variable name. Since this storage class
487 exists for defining a dll interface, the compiler, assembler and linker know
488 it is externally referenced and must refrain from deleting the symbol.
492 Thread Local Storage Models
493 ---------------------------
495 A variable may be defined as ``thread_local``, which means that it will
496 not be shared by threads (each thread will have a separated copy of the
497 variable). Not all targets support thread-local variables. Optionally, a
498 TLS model may be specified:
501 For variables that are only used within the current shared library.
503 For variables in modules that will not be loaded dynamically.
505 For variables defined in the executable and only used within it.
507 If no explicit model is given, the "general dynamic" model is used.
509 The models correspond to the ELF TLS models; see `ELF Handling For
510 Thread-Local Storage <http://people.redhat.com/drepper/tls.pdf>`_ for
511 more information on under which circumstances the different models may
512 be used. The target may choose a different TLS model if the specified
513 model is not supported, or if a better choice of model can be made.
515 A model can also be specified in an alias, but then it only governs how
516 the alias is accessed. It will not have any effect in the aliasee.
518 For platforms without linker support of ELF TLS model, the -femulated-tls
519 flag can be used to generate GCC compatible emulated TLS code.
526 LLVM IR allows you to specify both "identified" and "literal" :ref:`structure
527 types <t_struct>`. Literal types are uniqued structurally, but identified types
528 are never uniqued. An :ref:`opaque structural type <t_opaque>` can also be used
529 to forward declare a type that is not yet available.
531 An example of an identified structure specification is:
535 %mytype = type { %mytype*, i32 }
537 Prior to the LLVM 3.0 release, identified types were structurally uniqued. Only
538 literal types are uniqued in recent versions of LLVM.
545 Global variables define regions of memory allocated at compilation time
548 Global variable definitions must be initialized.
550 Global variables in other translation units can also be declared, in which
551 case they don't have an initializer.
553 Either global variable definitions or declarations may have an explicit section
554 to be placed in and may have an optional explicit alignment specified.
556 A variable may be defined as a global ``constant``, which indicates that
557 the contents of the variable will **never** be modified (enabling better
558 optimization, allowing the global data to be placed in the read-only
559 section of an executable, etc). Note that variables that need runtime
560 initialization cannot be marked ``constant`` as there is a store to the
563 LLVM explicitly allows *declarations* of global variables to be marked
564 constant, even if the final definition of the global is not. This
565 capability can be used to enable slightly better optimization of the
566 program, but requires the language definition to guarantee that
567 optimizations based on the 'constantness' are valid for the translation
568 units that do not include the definition.
570 As SSA values, global variables define pointer values that are in scope
571 (i.e. they dominate) all basic blocks in the program. Global variables
572 always define a pointer to their "content" type because they describe a
573 region of memory, and all memory objects in LLVM are accessed through
576 Global variables can be marked with ``unnamed_addr`` which indicates
577 that the address is not significant, only the content. Constants marked
578 like this can be merged with other constants if they have the same
579 initializer. Note that a constant with significant address *can* be
580 merged with a ``unnamed_addr`` constant, the result being a constant
581 whose address is significant.
583 A global variable may be declared to reside in a target-specific
584 numbered address space. For targets that support them, address spaces
585 may affect how optimizations are performed and/or what target
586 instructions are used to access the variable. The default address space
587 is zero. The address space qualifier must precede any other attributes.
589 LLVM allows an explicit section to be specified for globals. If the
590 target supports it, it will emit globals to the section specified.
591 Additionally, the global can placed in a comdat if the target has the necessary
594 By default, global initializers are optimized by assuming that global
595 variables defined within the module are not modified from their
596 initial values before the start of the global initializer. This is
597 true even for variables potentially accessible from outside the
598 module, including those with external linkage or appearing in
599 ``@llvm.used`` or dllexported variables. This assumption may be suppressed
600 by marking the variable with ``externally_initialized``.
602 An explicit alignment may be specified for a global, which must be a
603 power of 2. If not present, or if the alignment is set to zero, the
604 alignment of the global is set by the target to whatever it feels
605 convenient. If an explicit alignment is specified, the global is forced
606 to have exactly that alignment. Targets and optimizers are not allowed
607 to over-align the global if the global has an assigned section. In this
608 case, the extra alignment could be observable: for example, code could
609 assume that the globals are densely packed in their section and try to
610 iterate over them as an array, alignment padding would break this
611 iteration. The maximum alignment is ``1 << 29``.
613 Globals can also have a :ref:`DLL storage class <dllstorageclass>`.
615 Variables and aliases can have a
616 :ref:`Thread Local Storage Model <tls_model>`.
620 [@<GlobalVarName> =] [Linkage] [Visibility] [DLLStorageClass] [ThreadLocal]
621 [unnamed_addr] [AddrSpace] [ExternallyInitialized]
622 <global | constant> <Type> [<InitializerConstant>]
623 [, section "name"] [, comdat [($name)]]
624 [, align <Alignment>]
626 For example, the following defines a global in a numbered address space
627 with an initializer, section, and alignment:
631 @G = addrspace(5) constant float 1.0, section "foo", align 4
633 The following example just declares a global variable
637 @G = external global i32
639 The following example defines a thread-local global with the
640 ``initialexec`` TLS model:
644 @G = thread_local(initialexec) global i32 0, align 4
646 .. _functionstructure:
651 LLVM function definitions consist of the "``define``" keyword, an
652 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
653 style <visibility>`, an optional :ref:`DLL storage class <dllstorageclass>`,
654 an optional :ref:`calling convention <callingconv>`,
655 an optional ``unnamed_addr`` attribute, a return type, an optional
656 :ref:`parameter attribute <paramattrs>` for the return type, a function
657 name, a (possibly empty) argument list (each with optional :ref:`parameter
658 attributes <paramattrs>`), optional :ref:`function attributes <fnattrs>`,
659 an optional section, an optional alignment,
660 an optional :ref:`comdat <langref_comdats>`,
661 an optional :ref:`garbage collector name <gc>`, an optional :ref:`prefix <prefixdata>`,
662 an optional :ref:`prologue <prologuedata>`,
663 an optional :ref:`personality <personalityfn>`,
664 an optional list of attached :ref:`metadata <metadata>`,
665 an opening curly brace, a list of basic blocks, and a closing curly brace.
667 LLVM function declarations consist of the "``declare``" keyword, an
668 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
669 style <visibility>`, an optional :ref:`DLL storage class <dllstorageclass>`,
670 an optional :ref:`calling convention <callingconv>`,
671 an optional ``unnamed_addr`` attribute, a return type, an optional
672 :ref:`parameter attribute <paramattrs>` for the return type, a function
673 name, a possibly empty list of arguments, an optional alignment, an optional
674 :ref:`garbage collector name <gc>`, an optional :ref:`prefix <prefixdata>`,
675 and an optional :ref:`prologue <prologuedata>`.
677 A function definition contains a list of basic blocks, forming the CFG (Control
678 Flow Graph) for the function. Each basic block may optionally start with a label
679 (giving the basic block a symbol table entry), contains a list of instructions,
680 and ends with a :ref:`terminator <terminators>` instruction (such as a branch or
681 function return). If an explicit label is not provided, a block is assigned an
682 implicit numbered label, using the next value from the same counter as used for
683 unnamed temporaries (:ref:`see above<identifiers>`). For example, if a function
684 entry block does not have an explicit label, it will be assigned label "%0",
685 then the first unnamed temporary in that block will be "%1", etc.
687 The first basic block in a function is special in two ways: it is
688 immediately executed on entrance to the function, and it is not allowed
689 to have predecessor basic blocks (i.e. there can not be any branches to
690 the entry block of a function). Because the block can have no
691 predecessors, it also cannot have any :ref:`PHI nodes <i_phi>`.
693 LLVM allows an explicit section to be specified for functions. If the
694 target supports it, it will emit functions to the section specified.
695 Additionally, the function can be placed in a COMDAT.
697 An explicit alignment may be specified for a function. If not present,
698 or if the alignment is set to zero, the alignment of the function is set
699 by the target to whatever it feels convenient. If an explicit alignment
700 is specified, the function is forced to have at least that much
701 alignment. All alignments must be a power of 2.
703 If the ``unnamed_addr`` attribute is given, the address is known to not
704 be significant and two identical functions can be merged.
708 define [linkage] [visibility] [DLLStorageClass]
710 <ResultType> @<FunctionName> ([argument list])
711 [unnamed_addr] [fn Attrs] [section "name"] [comdat [($name)]]
712 [align N] [gc] [prefix Constant] [prologue Constant]
713 [personality Constant] (!name !N)* { ... }
715 The argument list is a comma separated sequence of arguments where each
716 argument is of the following form:
720 <type> [parameter Attrs] [name]
728 Aliases, unlike function or variables, don't create any new data. They
729 are just a new symbol and metadata for an existing position.
731 Aliases have a name and an aliasee that is either a global value or a
734 Aliases may have an optional :ref:`linkage type <linkage>`, an optional
735 :ref:`visibility style <visibility>`, an optional :ref:`DLL storage class
736 <dllstorageclass>` and an optional :ref:`tls model <tls_model>`.
740 @<Name> = [Linkage] [Visibility] [DLLStorageClass] [ThreadLocal] [unnamed_addr] alias <AliaseeTy>, <AliaseeTy>* @<Aliasee>
742 The linkage must be one of ``private``, ``internal``, ``linkonce``, ``weak``,
743 ``linkonce_odr``, ``weak_odr``, ``external``. Note that some system linkers
744 might not correctly handle dropping a weak symbol that is aliased.
746 Aliases that are not ``unnamed_addr`` are guaranteed to have the same address as
747 the aliasee expression. ``unnamed_addr`` ones are only guaranteed to point
750 Since aliases are only a second name, some restrictions apply, of which
751 some can only be checked when producing an object file:
753 * The expression defining the aliasee must be computable at assembly
754 time. Since it is just a name, no relocations can be used.
756 * No alias in the expression can be weak as the possibility of the
757 intermediate alias being overridden cannot be represented in an
760 * No global value in the expression can be a declaration, since that
761 would require a relocation, which is not possible.
768 Comdat IR provides access to COFF and ELF object file COMDAT functionality.
770 Comdats have a name which represents the COMDAT key. All global objects that
771 specify this key will only end up in the final object file if the linker chooses
772 that key over some other key. Aliases are placed in the same COMDAT that their
773 aliasee computes to, if any.
775 Comdats have a selection kind to provide input on how the linker should
776 choose between keys in two different object files.
780 $<Name> = comdat SelectionKind
782 The selection kind must be one of the following:
785 The linker may choose any COMDAT key, the choice is arbitrary.
787 The linker may choose any COMDAT key but the sections must contain the
790 The linker will choose the section containing the largest COMDAT key.
792 The linker requires that only section with this COMDAT key exist.
794 The linker may choose any COMDAT key but the sections must contain the
797 Note that the Mach-O platform doesn't support COMDATs and ELF only supports
798 ``any`` as a selection kind.
800 Here is an example of a COMDAT group where a function will only be selected if
801 the COMDAT key's section is the largest:
805 $foo = comdat largest
806 @foo = global i32 2, comdat($foo)
808 define void @bar() comdat($foo) {
812 As a syntactic sugar the ``$name`` can be omitted if the name is the same as
818 @foo = global i32 2, comdat
821 In a COFF object file, this will create a COMDAT section with selection kind
822 ``IMAGE_COMDAT_SELECT_LARGEST`` containing the contents of the ``@foo`` symbol
823 and another COMDAT section with selection kind
824 ``IMAGE_COMDAT_SELECT_ASSOCIATIVE`` which is associated with the first COMDAT
825 section and contains the contents of the ``@bar`` symbol.
827 There are some restrictions on the properties of the global object.
828 It, or an alias to it, must have the same name as the COMDAT group when
830 The contents and size of this object may be used during link-time to determine
831 which COMDAT groups get selected depending on the selection kind.
832 Because the name of the object must match the name of the COMDAT group, the
833 linkage of the global object must not be local; local symbols can get renamed
834 if a collision occurs in the symbol table.
836 The combined use of COMDATS and section attributes may yield surprising results.
843 @g1 = global i32 42, section "sec", comdat($foo)
844 @g2 = global i32 42, section "sec", comdat($bar)
846 From the object file perspective, this requires the creation of two sections
847 with the same name. This is necessary because both globals belong to different
848 COMDAT groups and COMDATs, at the object file level, are represented by
851 Note that certain IR constructs like global variables and functions may
852 create COMDATs in the object file in addition to any which are specified using
853 COMDAT IR. This arises when the code generator is configured to emit globals
854 in individual sections (e.g. when `-data-sections` or `-function-sections`
855 is supplied to `llc`).
857 .. _namedmetadatastructure:
862 Named metadata is a collection of metadata. :ref:`Metadata
863 nodes <metadata>` (but not metadata strings) are the only valid
864 operands for a named metadata.
866 #. Named metadata are represented as a string of characters with the
867 metadata prefix. The rules for metadata names are the same as for
868 identifiers, but quoted names are not allowed. ``"\xx"`` type escapes
869 are still valid, which allows any character to be part of a name.
873 ; Some unnamed metadata nodes, which are referenced by the named metadata.
878 !name = !{!0, !1, !2}
885 The return type and each parameter of a function type may have a set of
886 *parameter attributes* associated with them. Parameter attributes are
887 used to communicate additional information about the result or
888 parameters of a function. Parameter attributes are considered to be part
889 of the function, not of the function type, so functions with different
890 parameter attributes can have the same function type.
892 Parameter attributes are simple keywords that follow the type specified.
893 If multiple parameter attributes are needed, they are space separated.
898 declare i32 @printf(i8* noalias nocapture, ...)
899 declare i32 @atoi(i8 zeroext)
900 declare signext i8 @returns_signed_char()
902 Note that any attributes for the function result (``nounwind``,
903 ``readonly``) come immediately after the argument list.
905 Currently, only the following parameter attributes are defined:
908 This indicates to the code generator that the parameter or return
909 value should be zero-extended to the extent required by the target's
910 ABI (which is usually 32-bits, but is 8-bits for a i1 on x86-64) by
911 the caller (for a parameter) or the callee (for a return value).
913 This indicates to the code generator that the parameter or return
914 value should be sign-extended to the extent required by the target's
915 ABI (which is usually 32-bits) by the caller (for a parameter) or
916 the callee (for a return value).
918 This indicates that this parameter or return value should be treated
919 in a special target-dependent fashion while emitting code for
920 a function call or return (usually, by putting it in a register as
921 opposed to memory, though some targets use it to distinguish between
922 two different kinds of registers). Use of this attribute is
925 This indicates that the pointer parameter should really be passed by
926 value to the function. The attribute implies that a hidden copy of
927 the pointee is made between the caller and the callee, so the callee
928 is unable to modify the value in the caller. This attribute is only
929 valid on LLVM pointer arguments. It is generally used to pass
930 structs and arrays by value, but is also valid on pointers to
931 scalars. The copy is considered to belong to the caller not the
932 callee (for example, ``readonly`` functions should not write to
933 ``byval`` parameters). This is not a valid attribute for return
936 The byval attribute also supports specifying an alignment with the
937 align attribute. It indicates the alignment of the stack slot to
938 form and the known alignment of the pointer specified to the call
939 site. If the alignment is not specified, then the code generator
940 makes a target-specific assumption.
946 The ``inalloca`` argument attribute allows the caller to take the
947 address of outgoing stack arguments. An ``inalloca`` argument must
948 be a pointer to stack memory produced by an ``alloca`` instruction.
949 The alloca, or argument allocation, must also be tagged with the
950 inalloca keyword. Only the last argument may have the ``inalloca``
951 attribute, and that argument is guaranteed to be passed in memory.
953 An argument allocation may be used by a call at most once because
954 the call may deallocate it. The ``inalloca`` attribute cannot be
955 used in conjunction with other attributes that affect argument
956 storage, like ``inreg``, ``nest``, ``sret``, or ``byval``. The
957 ``inalloca`` attribute also disables LLVM's implicit lowering of
958 large aggregate return values, which means that frontend authors
959 must lower them with ``sret`` pointers.
961 When the call site is reached, the argument allocation must have
962 been the most recent stack allocation that is still live, or the
963 results are undefined. It is possible to allocate additional stack
964 space after an argument allocation and before its call site, but it
965 must be cleared off with :ref:`llvm.stackrestore
968 See :doc:`InAlloca` for more information on how to use this
972 This indicates that the pointer parameter specifies the address of a
973 structure that is the return value of the function in the source
974 program. This pointer must be guaranteed by the caller to be valid:
975 loads and stores to the structure may be assumed by the callee
976 not to trap and to be properly aligned. This may only be applied to
977 the first parameter. This is not a valid attribute for return
981 This indicates that the pointer value may be assumed by the optimizer to
982 have the specified alignment.
984 Note that this attribute has additional semantics when combined with the
990 This indicates that objects accessed via pointer values
991 :ref:`based <pointeraliasing>` on the argument or return value are not also
992 accessed, during the execution of the function, via pointer values not
993 *based* on the argument or return value. The attribute on a return value
994 also has additional semantics described below. The caller shares the
995 responsibility with the callee for ensuring that these requirements are met.
996 For further details, please see the discussion of the NoAlias response in
997 :ref:`alias analysis <Must, May, or No>`.
999 Note that this definition of ``noalias`` is intentionally similar
1000 to the definition of ``restrict`` in C99 for function arguments.
1002 For function return values, C99's ``restrict`` is not meaningful,
1003 while LLVM's ``noalias`` is. Furthermore, the semantics of the ``noalias``
1004 attribute on return values are stronger than the semantics of the attribute
1005 when used on function arguments. On function return values, the ``noalias``
1006 attribute indicates that the function acts like a system memory allocation
1007 function, returning a pointer to allocated storage disjoint from the
1008 storage for any other object accessible to the caller.
1011 This indicates that the callee does not make any copies of the
1012 pointer that outlive the callee itself. This is not a valid
1013 attribute for return values.
1018 This indicates that the pointer parameter can be excised using the
1019 :ref:`trampoline intrinsics <int_trampoline>`. This is not a valid
1020 attribute for return values and can only be applied to one parameter.
1023 This indicates that the function always returns the argument as its return
1024 value. This is an optimization hint to the code generator when generating
1025 the caller, allowing tail call optimization and omission of register saves
1026 and restores in some cases; it is not checked or enforced when generating
1027 the callee. The parameter and the function return type must be valid
1028 operands for the :ref:`bitcast instruction <i_bitcast>`. This is not a
1029 valid attribute for return values and can only be applied to one parameter.
1032 This indicates that the parameter or return pointer is not null. This
1033 attribute may only be applied to pointer typed parameters. This is not
1034 checked or enforced by LLVM, the caller must ensure that the pointer
1035 passed in is non-null, or the callee must ensure that the returned pointer
1038 ``dereferenceable(<n>)``
1039 This indicates that the parameter or return pointer is dereferenceable. This
1040 attribute may only be applied to pointer typed parameters. A pointer that
1041 is dereferenceable can be loaded from speculatively without a risk of
1042 trapping. The number of bytes known to be dereferenceable must be provided
1043 in parentheses. It is legal for the number of bytes to be less than the
1044 size of the pointee type. The ``nonnull`` attribute does not imply
1045 dereferenceability (consider a pointer to one element past the end of an
1046 array), however ``dereferenceable(<n>)`` does imply ``nonnull`` in
1047 ``addrspace(0)`` (which is the default address space).
1049 ``dereferenceable_or_null(<n>)``
1050 This indicates that the parameter or return value isn't both
1051 non-null and non-dereferenceable (up to ``<n>`` bytes) at the same
1052 time. All non-null pointers tagged with
1053 ``dereferenceable_or_null(<n>)`` are ``dereferenceable(<n>)``.
1054 For address space 0 ``dereferenceable_or_null(<n>)`` implies that
1055 a pointer is exactly one of ``dereferenceable(<n>)`` or ``null``,
1056 and in other address spaces ``dereferenceable_or_null(<n>)``
1057 implies that a pointer is at least one of ``dereferenceable(<n>)``
1058 or ``null`` (i.e. it may be both ``null`` and
1059 ``dereferenceable(<n>)``). This attribute may only be applied to
1060 pointer typed parameters.
1064 Garbage Collector Strategy Names
1065 --------------------------------
1067 Each function may specify a garbage collector strategy name, which is simply a
1070 .. code-block:: llvm
1072 define void @f() gc "name" { ... }
1074 The supported values of *name* includes those :ref:`built in to LLVM
1075 <builtin-gc-strategies>` and any provided by loaded plugins. Specifying a GC
1076 strategy will cause the compiler to alter its output in order to support the
1077 named garbage collection algorithm. Note that LLVM itself does not contain a
1078 garbage collector, this functionality is restricted to generating machine code
1079 which can interoperate with a collector provided externally.
1086 Prefix data is data associated with a function which the code
1087 generator will emit immediately before the function's entrypoint.
1088 The purpose of this feature is to allow frontends to associate
1089 language-specific runtime metadata with specific functions and make it
1090 available through the function pointer while still allowing the
1091 function pointer to be called.
1093 To access the data for a given function, a program may bitcast the
1094 function pointer to a pointer to the constant's type and dereference
1095 index -1. This implies that the IR symbol points just past the end of
1096 the prefix data. For instance, take the example of a function annotated
1097 with a single ``i32``,
1099 .. code-block:: llvm
1101 define void @f() prefix i32 123 { ... }
1103 The prefix data can be referenced as,
1105 .. code-block:: llvm
1107 %0 = bitcast void* () @f to i32*
1108 %a = getelementptr inbounds i32, i32* %0, i32 -1
1109 %b = load i32, i32* %a
1111 Prefix data is laid out as if it were an initializer for a global variable
1112 of the prefix data's type. The function will be placed such that the
1113 beginning of the prefix data is aligned. This means that if the size
1114 of the prefix data is not a multiple of the alignment size, the
1115 function's entrypoint will not be aligned. If alignment of the
1116 function's entrypoint is desired, padding must be added to the prefix
1119 A function may have prefix data but no body. This has similar semantics
1120 to the ``available_externally`` linkage in that the data may be used by the
1121 optimizers but will not be emitted in the object file.
1128 The ``prologue`` attribute allows arbitrary code (encoded as bytes) to
1129 be inserted prior to the function body. This can be used for enabling
1130 function hot-patching and instrumentation.
1132 To maintain the semantics of ordinary function calls, the prologue data must
1133 have a particular format. Specifically, it must begin with a sequence of
1134 bytes which decode to a sequence of machine instructions, valid for the
1135 module's target, which transfer control to the point immediately succeeding
1136 the prologue data, without performing any other visible action. This allows
1137 the inliner and other passes to reason about the semantics of the function
1138 definition without needing to reason about the prologue data. Obviously this
1139 makes the format of the prologue data highly target dependent.
1141 A trivial example of valid prologue data for the x86 architecture is ``i8 144``,
1142 which encodes the ``nop`` instruction:
1144 .. code-block:: llvm
1146 define void @f() prologue i8 144 { ... }
1148 Generally prologue data can be formed by encoding a relative branch instruction
1149 which skips the metadata, as in this example of valid prologue data for the
1150 x86_64 architecture, where the first two bytes encode ``jmp .+10``:
1152 .. code-block:: llvm
1154 %0 = type <{ i8, i8, i8* }>
1156 define void @f() prologue %0 <{ i8 235, i8 8, i8* @md}> { ... }
1158 A function may have prologue data but no body. This has similar semantics
1159 to the ``available_externally`` linkage in that the data may be used by the
1160 optimizers but will not be emitted in the object file.
1164 Personality Function
1165 --------------------
1167 The ``personality`` attribute permits functions to specify what function
1168 to use for exception handling.
1175 Attribute groups are groups of attributes that are referenced by objects within
1176 the IR. They are important for keeping ``.ll`` files readable, because a lot of
1177 functions will use the same set of attributes. In the degenerative case of a
1178 ``.ll`` file that corresponds to a single ``.c`` file, the single attribute
1179 group will capture the important command line flags used to build that file.
1181 An attribute group is a module-level object. To use an attribute group, an
1182 object references the attribute group's ID (e.g. ``#37``). An object may refer
1183 to more than one attribute group. In that situation, the attributes from the
1184 different groups are merged.
1186 Here is an example of attribute groups for a function that should always be
1187 inlined, has a stack alignment of 4, and which shouldn't use SSE instructions:
1189 .. code-block:: llvm
1191 ; Target-independent attributes:
1192 attributes #0 = { alwaysinline alignstack=4 }
1194 ; Target-dependent attributes:
1195 attributes #1 = { "no-sse" }
1197 ; Function @f has attributes: alwaysinline, alignstack=4, and "no-sse".
1198 define void @f() #0 #1 { ... }
1205 Function attributes are set to communicate additional information about
1206 a function. Function attributes are considered to be part of the
1207 function, not of the function type, so functions with different function
1208 attributes can have the same function type.
1210 Function attributes are simple keywords that follow the type specified.
1211 If multiple attributes are needed, they are space separated. For
1214 .. code-block:: llvm
1216 define void @f() noinline { ... }
1217 define void @f() alwaysinline { ... }
1218 define void @f() alwaysinline optsize { ... }
1219 define void @f() optsize { ... }
1222 This attribute indicates that, when emitting the prologue and
1223 epilogue, the backend should forcibly align the stack pointer.
1224 Specify the desired alignment, which must be a power of two, in
1227 This attribute indicates that the inliner should attempt to inline
1228 this function into callers whenever possible, ignoring any active
1229 inlining size threshold for this caller.
1231 This indicates that the callee function at a call site should be
1232 recognized as a built-in function, even though the function's declaration
1233 uses the ``nobuiltin`` attribute. This is only valid at call sites for
1234 direct calls to functions that are declared with the ``nobuiltin``
1237 This attribute indicates that this function is rarely called. When
1238 computing edge weights, basic blocks post-dominated by a cold
1239 function call are also considered to be cold; and, thus, given low
1242 This attribute indicates that the callee is dependent on a convergent
1243 thread execution pattern under certain parallel execution models.
1244 Transformations that are execution model agnostic may not make the execution
1245 of a convergent operation control dependent on any additional values.
1247 This attribute indicates that the source code contained a hint that
1248 inlining this function is desirable (such as the "inline" keyword in
1249 C/C++). It is just a hint; it imposes no requirements on the
1252 This attribute indicates that the function should be added to a
1253 jump-instruction table at code-generation time, and that all address-taken
1254 references to this function should be replaced with a reference to the
1255 appropriate jump-instruction-table function pointer. Note that this creates
1256 a new pointer for the original function, which means that code that depends
1257 on function-pointer identity can break. So, any function annotated with
1258 ``jumptable`` must also be ``unnamed_addr``.
1260 This attribute suggests that optimization passes and code generator
1261 passes make choices that keep the code size of this function as small
1262 as possible and perform optimizations that may sacrifice runtime
1263 performance in order to minimize the size of the generated code.
1265 This attribute disables prologue / epilogue emission for the
1266 function. This can have very system-specific consequences.
1268 This indicates that the callee function at a call site is not recognized as
1269 a built-in function. LLVM will retain the original call and not replace it
1270 with equivalent code based on the semantics of the built-in function, unless
1271 the call site uses the ``builtin`` attribute. This is valid at call sites
1272 and on function declarations and definitions.
1274 This attribute indicates that calls to the function cannot be
1275 duplicated. A call to a ``noduplicate`` function may be moved
1276 within its parent function, but may not be duplicated within
1277 its parent function.
1279 A function containing a ``noduplicate`` call may still
1280 be an inlining candidate, provided that the call is not
1281 duplicated by inlining. That implies that the function has
1282 internal linkage and only has one call site, so the original
1283 call is dead after inlining.
1285 This attributes disables implicit floating point instructions.
1287 This attribute indicates that the inliner should never inline this
1288 function in any situation. This attribute may not be used together
1289 with the ``alwaysinline`` attribute.
1291 This attribute suppresses lazy symbol binding for the function. This
1292 may make calls to the function faster, at the cost of extra program
1293 startup time if the function is not called during program startup.
1295 This attribute indicates that the code generator should not use a
1296 red zone, even if the target-specific ABI normally permits it.
1298 This function attribute indicates that the function never returns
1299 normally. This produces undefined behavior at runtime if the
1300 function ever does dynamically return.
1302 This function attribute indicates that the function does not call itself
1303 either directly or indirectly down any possible call path. This produces
1304 undefined behavior at runtime if the function ever does recurse.
1306 This function attribute indicates that the function never raises an
1307 exception. If the function does raise an exception, its runtime
1308 behavior is undefined. However, functions marked nounwind may still
1309 trap or generate asynchronous exceptions. Exception handling schemes
1310 that are recognized by LLVM to handle asynchronous exceptions, such
1311 as SEH, will still provide their implementation defined semantics.
1313 This function attribute indicates that most optimization passes will skip
1314 this function, with the exception of interprocedural optimization passes.
1315 Code generation defaults to the "fast" instruction selector.
1316 This attribute cannot be used together with the ``alwaysinline``
1317 attribute; this attribute is also incompatible
1318 with the ``minsize`` attribute and the ``optsize`` attribute.
1320 This attribute requires the ``noinline`` attribute to be specified on
1321 the function as well, so the function is never inlined into any caller.
1322 Only functions with the ``alwaysinline`` attribute are valid
1323 candidates for inlining into the body of this function.
1325 This attribute suggests that optimization passes and code generator
1326 passes make choices that keep the code size of this function low,
1327 and otherwise do optimizations specifically to reduce code size as
1328 long as they do not significantly impact runtime performance.
1330 On a function, this attribute indicates that the function computes its
1331 result (or decides to unwind an exception) based strictly on its arguments,
1332 without dereferencing any pointer arguments or otherwise accessing
1333 any mutable state (e.g. memory, control registers, etc) visible to
1334 caller functions. It does not write through any pointer arguments
1335 (including ``byval`` arguments) and never changes any state visible
1336 to callers. This means that it cannot unwind exceptions by calling
1337 the ``C++`` exception throwing methods.
1339 On an argument, this attribute indicates that the function does not
1340 dereference that pointer argument, even though it may read or write the
1341 memory that the pointer points to if accessed through other pointers.
1343 On a function, this attribute indicates that the function does not write
1344 through any pointer arguments (including ``byval`` arguments) or otherwise
1345 modify any state (e.g. memory, control registers, etc) visible to
1346 caller functions. It may dereference pointer arguments and read
1347 state that may be set in the caller. A readonly function always
1348 returns the same value (or unwinds an exception identically) when
1349 called with the same set of arguments and global state. It cannot
1350 unwind an exception by calling the ``C++`` exception throwing
1353 On an argument, this attribute indicates that the function does not write
1354 through this pointer argument, even though it may write to the memory that
1355 the pointer points to.
1357 This attribute indicates that the only memory accesses inside function are
1358 loads and stores from objects pointed to by its pointer-typed arguments,
1359 with arbitrary offsets. Or in other words, all memory operations in the
1360 function can refer to memory only using pointers based on its function
1362 Note that ``argmemonly`` can be used together with ``readonly`` attribute
1363 in order to specify that function reads only from its arguments.
1365 This attribute indicates that this function can return twice. The C
1366 ``setjmp`` is an example of such a function. The compiler disables
1367 some optimizations (like tail calls) in the caller of these
1370 This attribute indicates that
1371 `SafeStack <http://clang.llvm.org/docs/SafeStack.html>`_
1372 protection is enabled for this function.
1374 If a function that has a ``safestack`` attribute is inlined into a
1375 function that doesn't have a ``safestack`` attribute or which has an
1376 ``ssp``, ``sspstrong`` or ``sspreq`` attribute, then the resulting
1377 function will have a ``safestack`` attribute.
1378 ``sanitize_address``
1379 This attribute indicates that AddressSanitizer checks
1380 (dynamic address safety analysis) are enabled for this function.
1382 This attribute indicates that MemorySanitizer checks (dynamic detection
1383 of accesses to uninitialized memory) are enabled for this function.
1385 This attribute indicates that ThreadSanitizer checks
1386 (dynamic thread safety analysis) are enabled for this function.
1388 This attribute indicates that the function should emit a stack
1389 smashing protector. It is in the form of a "canary" --- a random value
1390 placed on the stack before the local variables that's checked upon
1391 return from the function to see if it has been overwritten. A
1392 heuristic is used to determine if a function needs stack protectors
1393 or not. The heuristic used will enable protectors for functions with:
1395 - Character arrays larger than ``ssp-buffer-size`` (default 8).
1396 - Aggregates containing character arrays larger than ``ssp-buffer-size``.
1397 - Calls to alloca() with variable sizes or constant sizes greater than
1398 ``ssp-buffer-size``.
1400 Variables that are identified as requiring a protector will be arranged
1401 on the stack such that they are adjacent to the stack protector guard.
1403 If a function that has an ``ssp`` attribute is inlined into a
1404 function that doesn't have an ``ssp`` attribute, then the resulting
1405 function will have an ``ssp`` attribute.
1407 This attribute indicates that the function should *always* emit a
1408 stack smashing protector. This overrides the ``ssp`` function
1411 Variables that are identified as requiring a protector will be arranged
1412 on the stack such that they are adjacent to the stack protector guard.
1413 The specific layout rules are:
1415 #. Large arrays and structures containing large arrays
1416 (``>= ssp-buffer-size``) are closest to the stack protector.
1417 #. Small arrays and structures containing small arrays
1418 (``< ssp-buffer-size``) are 2nd closest to the protector.
1419 #. Variables that have had their address taken are 3rd closest to the
1422 If a function that has an ``sspreq`` attribute is inlined into a
1423 function that doesn't have an ``sspreq`` attribute or which has an
1424 ``ssp`` or ``sspstrong`` attribute, then the resulting function will have
1425 an ``sspreq`` attribute.
1427 This attribute indicates that the function should emit a stack smashing
1428 protector. This attribute causes a strong heuristic to be used when
1429 determining if a function needs stack protectors. The strong heuristic
1430 will enable protectors for functions with:
1432 - Arrays of any size and type
1433 - Aggregates containing an array of any size and type.
1434 - Calls to alloca().
1435 - Local variables that have had their address taken.
1437 Variables that are identified as requiring a protector will be arranged
1438 on the stack such that they are adjacent to the stack protector guard.
1439 The specific layout rules are:
1441 #. Large arrays and structures containing large arrays
1442 (``>= ssp-buffer-size``) are closest to the stack protector.
1443 #. Small arrays and structures containing small arrays
1444 (``< ssp-buffer-size``) are 2nd closest to the protector.
1445 #. Variables that have had their address taken are 3rd closest to the
1448 This overrides the ``ssp`` function attribute.
1450 If a function that has an ``sspstrong`` attribute is inlined into a
1451 function that doesn't have an ``sspstrong`` attribute, then the
1452 resulting function will have an ``sspstrong`` attribute.
1454 This attribute indicates that the function will delegate to some other
1455 function with a tail call. The prototype of a thunk should not be used for
1456 optimization purposes. The caller is expected to cast the thunk prototype to
1457 match the thunk target prototype.
1459 This attribute indicates that the ABI being targeted requires that
1460 an unwind table entry be produced for this function even if we can
1461 show that no exceptions passes by it. This is normally the case for
1462 the ELF x86-64 abi, but it can be disabled for some compilation
1471 Note: operand bundles are a work in progress, and they should be
1472 considered experimental at this time.
1474 Operand bundles are tagged sets of SSA values that can be associated
1475 with certain LLVM instructions (currently only ``call`` s and
1476 ``invoke`` s). In a way they are like metadata, but dropping them is
1477 incorrect and will change program semantics.
1481 operand bundle set ::= '[' operand bundle (, operand bundle )* ']'
1482 operand bundle ::= tag '(' [ bundle operand ] (, bundle operand )* ')'
1483 bundle operand ::= SSA value
1484 tag ::= string constant
1486 Operand bundles are **not** part of a function's signature, and a
1487 given function may be called from multiple places with different kinds
1488 of operand bundles. This reflects the fact that the operand bundles
1489 are conceptually a part of the ``call`` (or ``invoke``), not the
1490 callee being dispatched to.
1492 Operand bundles are a generic mechanism intended to support
1493 runtime-introspection-like functionality for managed languages. While
1494 the exact semantics of an operand bundle depend on the bundle tag,
1495 there are certain limitations to how much the presence of an operand
1496 bundle can influence the semantics of a program. These restrictions
1497 are described as the semantics of an "unknown" operand bundle. As
1498 long as the behavior of an operand bundle is describable within these
1499 restrictions, LLVM does not need to have special knowledge of the
1500 operand bundle to not miscompile programs containing it.
1502 - The bundle operands for an unknown operand bundle escape in unknown
1503 ways before control is transferred to the callee or invokee.
1504 - Calls and invokes with operand bundles have unknown read / write
1505 effect on the heap on entry and exit (even if the call target is
1506 ``readnone`` or ``readonly``), unless they're overriden with
1507 callsite specific attributes.
1508 - An operand bundle at a call site cannot change the implementation
1509 of the called function. Inter-procedural optimizations work as
1510 usual as long as they take into account the first two properties.
1512 More specific types of operand bundles are described below.
1514 Deoptimization Operand Bundles
1515 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1517 Deoptimization operand bundles are characterized by the ``"deopt"``
1518 operand bundle tag. These operand bundles represent an alternate
1519 "safe" continuation for the call site they're attached to, and can be
1520 used by a suitable runtime to deoptimize the compiled frame at the
1521 specified call site. There can be at most one ``"deopt"`` operand
1522 bundle attached to a call site. Exact details of deoptimization is
1523 out of scope for the language reference, but it usually involves
1524 rewriting a compiled frame into a set of interpreted frames.
1526 From the compiler's perspective, deoptimization operand bundles make
1527 the call sites they're attached to at least ``readonly``. They read
1528 through all of their pointer typed operands (even if they're not
1529 otherwise escaped) and the entire visible heap. Deoptimization
1530 operand bundles do not capture their operands except during
1531 deoptimization, in which case control will not be returned to the
1534 The inliner knows how to inline through calls that have deoptimization
1535 operand bundles. Just like inlining through a normal call site
1536 involves composing the normal and exceptional continuations, inlining
1537 through a call site with a deoptimization operand bundle needs to
1538 appropriately compose the "safe" deoptimization continuation. The
1539 inliner does this by prepending the parent's deoptimization
1540 continuation to every deoptimization continuation in the inlined body.
1541 E.g. inlining ``@f`` into ``@g`` in the following example
1543 .. code-block:: llvm
1546 call void @x() ;; no deopt state
1547 call void @y() [ "deopt"(i32 10) ]
1548 call void @y() [ "deopt"(i32 10), "unknown"(i8* null) ]
1553 call void @f() [ "deopt"(i32 20) ]
1559 .. code-block:: llvm
1562 call void @x() ;; still no deopt state
1563 call void @y() [ "deopt"(i32 20, i32 10) ]
1564 call void @y() [ "deopt"(i32 20, i32 10), "unknown"(i8* null) ]
1568 It is the frontend's responsibility to structure or encode the
1569 deoptimization state in a way that syntactically prepending the
1570 caller's deoptimization state to the callee's deoptimization state is
1571 semantically equivalent to composing the caller's deoptimization
1572 continuation after the callee's deoptimization continuation.
1576 Module-Level Inline Assembly
1577 ----------------------------
1579 Modules may contain "module-level inline asm" blocks, which corresponds
1580 to the GCC "file scope inline asm" blocks. These blocks are internally
1581 concatenated by LLVM and treated as a single unit, but may be separated
1582 in the ``.ll`` file if desired. The syntax is very simple:
1584 .. code-block:: llvm
1586 module asm "inline asm code goes here"
1587 module asm "more can go here"
1589 The strings can contain any character by escaping non-printable
1590 characters. The escape sequence used is simply "\\xx" where "xx" is the
1591 two digit hex code for the number.
1593 Note that the assembly string *must* be parseable by LLVM's integrated assembler
1594 (unless it is disabled), even when emitting a ``.s`` file.
1596 .. _langref_datalayout:
1601 A module may specify a target specific data layout string that specifies
1602 how data is to be laid out in memory. The syntax for the data layout is
1605 .. code-block:: llvm
1607 target datalayout = "layout specification"
1609 The *layout specification* consists of a list of specifications
1610 separated by the minus sign character ('-'). Each specification starts
1611 with a letter and may include other information after the letter to
1612 define some aspect of the data layout. The specifications accepted are
1616 Specifies that the target lays out data in big-endian form. That is,
1617 the bits with the most significance have the lowest address
1620 Specifies that the target lays out data in little-endian form. That
1621 is, the bits with the least significance have the lowest address
1624 Specifies the natural alignment of the stack in bits. Alignment
1625 promotion of stack variables is limited to the natural stack
1626 alignment to avoid dynamic stack realignment. The stack alignment
1627 must be a multiple of 8-bits. If omitted, the natural stack
1628 alignment defaults to "unspecified", which does not prevent any
1629 alignment promotions.
1630 ``p[n]:<size>:<abi>:<pref>``
1631 This specifies the *size* of a pointer and its ``<abi>`` and
1632 ``<pref>``\erred alignments for address space ``n``. All sizes are in
1633 bits. The address space, ``n``, is optional, and if not specified,
1634 denotes the default address space 0. The value of ``n`` must be
1635 in the range [1,2^23).
1636 ``i<size>:<abi>:<pref>``
1637 This specifies the alignment for an integer type of a given bit
1638 ``<size>``. The value of ``<size>`` must be in the range [1,2^23).
1639 ``v<size>:<abi>:<pref>``
1640 This specifies the alignment for a vector type of a given bit
1642 ``f<size>:<abi>:<pref>``
1643 This specifies the alignment for a floating point type of a given bit
1644 ``<size>``. Only values of ``<size>`` that are supported by the target
1645 will work. 32 (float) and 64 (double) are supported on all targets; 80
1646 or 128 (different flavors of long double) are also supported on some
1649 This specifies the alignment for an object of aggregate type.
1651 If present, specifies that llvm names are mangled in the output. The
1654 * ``e``: ELF mangling: Private symbols get a ``.L`` prefix.
1655 * ``m``: Mips mangling: Private symbols get a ``$`` prefix.
1656 * ``o``: Mach-O mangling: Private symbols get ``L`` prefix. Other
1657 symbols get a ``_`` prefix.
1658 * ``w``: Windows COFF prefix: Similar to Mach-O, but stdcall and fastcall
1659 functions also get a suffix based on the frame size.
1660 * ``x``: Windows x86 COFF prefix: Similar to Windows COFF, but use a ``_``
1661 prefix for ``__cdecl`` functions.
1662 ``n<size1>:<size2>:<size3>...``
1663 This specifies a set of native integer widths for the target CPU in
1664 bits. For example, it might contain ``n32`` for 32-bit PowerPC,
1665 ``n32:64`` for PowerPC 64, or ``n8:16:32:64`` for X86-64. Elements of
1666 this set are considered to support most general arithmetic operations
1669 On every specification that takes a ``<abi>:<pref>``, specifying the
1670 ``<pref>`` alignment is optional. If omitted, the preceding ``:``
1671 should be omitted too and ``<pref>`` will be equal to ``<abi>``.
1673 When constructing the data layout for a given target, LLVM starts with a
1674 default set of specifications which are then (possibly) overridden by
1675 the specifications in the ``datalayout`` keyword. The default
1676 specifications are given in this list:
1678 - ``E`` - big endian
1679 - ``p:64:64:64`` - 64-bit pointers with 64-bit alignment.
1680 - ``p[n]:64:64:64`` - Other address spaces are assumed to be the
1681 same as the default address space.
1682 - ``S0`` - natural stack alignment is unspecified
1683 - ``i1:8:8`` - i1 is 8-bit (byte) aligned
1684 - ``i8:8:8`` - i8 is 8-bit (byte) aligned
1685 - ``i16:16:16`` - i16 is 16-bit aligned
1686 - ``i32:32:32`` - i32 is 32-bit aligned
1687 - ``i64:32:64`` - i64 has ABI alignment of 32-bits but preferred
1688 alignment of 64-bits
1689 - ``f16:16:16`` - half is 16-bit aligned
1690 - ``f32:32:32`` - float is 32-bit aligned
1691 - ``f64:64:64`` - double is 64-bit aligned
1692 - ``f128:128:128`` - quad is 128-bit aligned
1693 - ``v64:64:64`` - 64-bit vector is 64-bit aligned
1694 - ``v128:128:128`` - 128-bit vector is 128-bit aligned
1695 - ``a:0:64`` - aggregates are 64-bit aligned
1697 When LLVM is determining the alignment for a given type, it uses the
1700 #. If the type sought is an exact match for one of the specifications,
1701 that specification is used.
1702 #. If no match is found, and the type sought is an integer type, then
1703 the smallest integer type that is larger than the bitwidth of the
1704 sought type is used. If none of the specifications are larger than
1705 the bitwidth then the largest integer type is used. For example,
1706 given the default specifications above, the i7 type will use the
1707 alignment of i8 (next largest) while both i65 and i256 will use the
1708 alignment of i64 (largest specified).
1709 #. If no match is found, and the type sought is a vector type, then the
1710 largest vector type that is smaller than the sought vector type will
1711 be used as a fall back. This happens because <128 x double> can be
1712 implemented in terms of 64 <2 x double>, for example.
1714 The function of the data layout string may not be what you expect.
1715 Notably, this is not a specification from the frontend of what alignment
1716 the code generator should use.
1718 Instead, if specified, the target data layout is required to match what
1719 the ultimate *code generator* expects. This string is used by the
1720 mid-level optimizers to improve code, and this only works if it matches
1721 what the ultimate code generator uses. There is no way to generate IR
1722 that does not embed this target-specific detail into the IR. If you
1723 don't specify the string, the default specifications will be used to
1724 generate a Data Layout and the optimization phases will operate
1725 accordingly and introduce target specificity into the IR with respect to
1726 these default specifications.
1733 A module may specify a target triple string that describes the target
1734 host. The syntax for the target triple is simply:
1736 .. code-block:: llvm
1738 target triple = "x86_64-apple-macosx10.7.0"
1740 The *target triple* string consists of a series of identifiers delimited
1741 by the minus sign character ('-'). The canonical forms are:
1745 ARCHITECTURE-VENDOR-OPERATING_SYSTEM
1746 ARCHITECTURE-VENDOR-OPERATING_SYSTEM-ENVIRONMENT
1748 This information is passed along to the backend so that it generates
1749 code for the proper architecture. It's possible to override this on the
1750 command line with the ``-mtriple`` command line option.
1752 .. _pointeraliasing:
1754 Pointer Aliasing Rules
1755 ----------------------
1757 Any memory access must be done through a pointer value associated with
1758 an address range of the memory access, otherwise the behavior is
1759 undefined. Pointer values are associated with address ranges according
1760 to the following rules:
1762 - A pointer value is associated with the addresses associated with any
1763 value it is *based* on.
1764 - An address of a global variable is associated with the address range
1765 of the variable's storage.
1766 - The result value of an allocation instruction is associated with the
1767 address range of the allocated storage.
1768 - A null pointer in the default address-space is associated with no
1770 - An integer constant other than zero or a pointer value returned from
1771 a function not defined within LLVM may be associated with address
1772 ranges allocated through mechanisms other than those provided by
1773 LLVM. Such ranges shall not overlap with any ranges of addresses
1774 allocated by mechanisms provided by LLVM.
1776 A pointer value is *based* on another pointer value according to the
1779 - A pointer value formed from a ``getelementptr`` operation is *based*
1780 on the first value operand of the ``getelementptr``.
1781 - The result value of a ``bitcast`` is *based* on the operand of the
1783 - A pointer value formed by an ``inttoptr`` is *based* on all pointer
1784 values that contribute (directly or indirectly) to the computation of
1785 the pointer's value.
1786 - The "*based* on" relationship is transitive.
1788 Note that this definition of *"based"* is intentionally similar to the
1789 definition of *"based"* in C99, though it is slightly weaker.
1791 LLVM IR does not associate types with memory. The result type of a
1792 ``load`` merely indicates the size and alignment of the memory from
1793 which to load, as well as the interpretation of the value. The first
1794 operand type of a ``store`` similarly only indicates the size and
1795 alignment of the store.
1797 Consequently, type-based alias analysis, aka TBAA, aka
1798 ``-fstrict-aliasing``, is not applicable to general unadorned LLVM IR.
1799 :ref:`Metadata <metadata>` may be used to encode additional information
1800 which specialized optimization passes may use to implement type-based
1805 Volatile Memory Accesses
1806 ------------------------
1808 Certain memory accesses, such as :ref:`load <i_load>`'s,
1809 :ref:`store <i_store>`'s, and :ref:`llvm.memcpy <int_memcpy>`'s may be
1810 marked ``volatile``. The optimizers must not change the number of
1811 volatile operations or change their order of execution relative to other
1812 volatile operations. The optimizers *may* change the order of volatile
1813 operations relative to non-volatile operations. This is not Java's
1814 "volatile" and has no cross-thread synchronization behavior.
1816 IR-level volatile loads and stores cannot safely be optimized into
1817 llvm.memcpy or llvm.memmove intrinsics even when those intrinsics are
1818 flagged volatile. Likewise, the backend should never split or merge
1819 target-legal volatile load/store instructions.
1821 .. admonition:: Rationale
1823 Platforms may rely on volatile loads and stores of natively supported
1824 data width to be executed as single instruction. For example, in C
1825 this holds for an l-value of volatile primitive type with native
1826 hardware support, but not necessarily for aggregate types. The
1827 frontend upholds these expectations, which are intentionally
1828 unspecified in the IR. The rules above ensure that IR transformations
1829 do not violate the frontend's contract with the language.
1833 Memory Model for Concurrent Operations
1834 --------------------------------------
1836 The LLVM IR does not define any way to start parallel threads of
1837 execution or to register signal handlers. Nonetheless, there are
1838 platform-specific ways to create them, and we define LLVM IR's behavior
1839 in their presence. This model is inspired by the C++0x memory model.
1841 For a more informal introduction to this model, see the :doc:`Atomics`.
1843 We define a *happens-before* partial order as the least partial order
1846 - Is a superset of single-thread program order, and
1847 - When a *synchronizes-with* ``b``, includes an edge from ``a`` to
1848 ``b``. *Synchronizes-with* pairs are introduced by platform-specific
1849 techniques, like pthread locks, thread creation, thread joining,
1850 etc., and by atomic instructions. (See also :ref:`Atomic Memory Ordering
1851 Constraints <ordering>`).
1853 Note that program order does not introduce *happens-before* edges
1854 between a thread and signals executing inside that thread.
1856 Every (defined) read operation (load instructions, memcpy, atomic
1857 loads/read-modify-writes, etc.) R reads a series of bytes written by
1858 (defined) write operations (store instructions, atomic
1859 stores/read-modify-writes, memcpy, etc.). For the purposes of this
1860 section, initialized globals are considered to have a write of the
1861 initializer which is atomic and happens before any other read or write
1862 of the memory in question. For each byte of a read R, R\ :sub:`byte`
1863 may see any write to the same byte, except:
1865 - If write\ :sub:`1` happens before write\ :sub:`2`, and
1866 write\ :sub:`2` happens before R\ :sub:`byte`, then
1867 R\ :sub:`byte` does not see write\ :sub:`1`.
1868 - If R\ :sub:`byte` happens before write\ :sub:`3`, then
1869 R\ :sub:`byte` does not see write\ :sub:`3`.
1871 Given that definition, R\ :sub:`byte` is defined as follows:
1873 - If R is volatile, the result is target-dependent. (Volatile is
1874 supposed to give guarantees which can support ``sig_atomic_t`` in
1875 C/C++, and may be used for accesses to addresses that do not behave
1876 like normal memory. It does not generally provide cross-thread
1878 - Otherwise, if there is no write to the same byte that happens before
1879 R\ :sub:`byte`, R\ :sub:`byte` returns ``undef`` for that byte.
1880 - Otherwise, if R\ :sub:`byte` may see exactly one write,
1881 R\ :sub:`byte` returns the value written by that write.
1882 - Otherwise, if R is atomic, and all the writes R\ :sub:`byte` may
1883 see are atomic, it chooses one of the values written. See the :ref:`Atomic
1884 Memory Ordering Constraints <ordering>` section for additional
1885 constraints on how the choice is made.
1886 - Otherwise R\ :sub:`byte` returns ``undef``.
1888 R returns the value composed of the series of bytes it read. This
1889 implies that some bytes within the value may be ``undef`` **without**
1890 the entire value being ``undef``. Note that this only defines the
1891 semantics of the operation; it doesn't mean that targets will emit more
1892 than one instruction to read the series of bytes.
1894 Note that in cases where none of the atomic intrinsics are used, this
1895 model places only one restriction on IR transformations on top of what
1896 is required for single-threaded execution: introducing a store to a byte
1897 which might not otherwise be stored is not allowed in general.
1898 (Specifically, in the case where another thread might write to and read
1899 from an address, introducing a store can change a load that may see
1900 exactly one write into a load that may see multiple writes.)
1904 Atomic Memory Ordering Constraints
1905 ----------------------------------
1907 Atomic instructions (:ref:`cmpxchg <i_cmpxchg>`,
1908 :ref:`atomicrmw <i_atomicrmw>`, :ref:`fence <i_fence>`,
1909 :ref:`atomic load <i_load>`, and :ref:`atomic store <i_store>`) take
1910 ordering parameters that determine which other atomic instructions on
1911 the same address they *synchronize with*. These semantics are borrowed
1912 from Java and C++0x, but are somewhat more colloquial. If these
1913 descriptions aren't precise enough, check those specs (see spec
1914 references in the :doc:`atomics guide <Atomics>`).
1915 :ref:`fence <i_fence>` instructions treat these orderings somewhat
1916 differently since they don't take an address. See that instruction's
1917 documentation for details.
1919 For a simpler introduction to the ordering constraints, see the
1923 The set of values that can be read is governed by the happens-before
1924 partial order. A value cannot be read unless some operation wrote
1925 it. This is intended to provide a guarantee strong enough to model
1926 Java's non-volatile shared variables. This ordering cannot be
1927 specified for read-modify-write operations; it is not strong enough
1928 to make them atomic in any interesting way.
1930 In addition to the guarantees of ``unordered``, there is a single
1931 total order for modifications by ``monotonic`` operations on each
1932 address. All modification orders must be compatible with the
1933 happens-before order. There is no guarantee that the modification
1934 orders can be combined to a global total order for the whole program
1935 (and this often will not be possible). The read in an atomic
1936 read-modify-write operation (:ref:`cmpxchg <i_cmpxchg>` and
1937 :ref:`atomicrmw <i_atomicrmw>`) reads the value in the modification
1938 order immediately before the value it writes. If one atomic read
1939 happens before another atomic read of the same address, the later
1940 read must see the same value or a later value in the address's
1941 modification order. This disallows reordering of ``monotonic`` (or
1942 stronger) operations on the same address. If an address is written
1943 ``monotonic``-ally by one thread, and other threads ``monotonic``-ally
1944 read that address repeatedly, the other threads must eventually see
1945 the write. This corresponds to the C++0x/C1x
1946 ``memory_order_relaxed``.
1948 In addition to the guarantees of ``monotonic``, a
1949 *synchronizes-with* edge may be formed with a ``release`` operation.
1950 This is intended to model C++'s ``memory_order_acquire``.
1952 In addition to the guarantees of ``monotonic``, if this operation
1953 writes a value which is subsequently read by an ``acquire``
1954 operation, it *synchronizes-with* that operation. (This isn't a
1955 complete description; see the C++0x definition of a release
1956 sequence.) This corresponds to the C++0x/C1x
1957 ``memory_order_release``.
1958 ``acq_rel`` (acquire+release)
1959 Acts as both an ``acquire`` and ``release`` operation on its
1960 address. This corresponds to the C++0x/C1x ``memory_order_acq_rel``.
1961 ``seq_cst`` (sequentially consistent)
1962 In addition to the guarantees of ``acq_rel`` (``acquire`` for an
1963 operation that only reads, ``release`` for an operation that only
1964 writes), there is a global total order on all
1965 sequentially-consistent operations on all addresses, which is
1966 consistent with the *happens-before* partial order and with the
1967 modification orders of all the affected addresses. Each
1968 sequentially-consistent read sees the last preceding write to the
1969 same address in this global order. This corresponds to the C++0x/C1x
1970 ``memory_order_seq_cst`` and Java volatile.
1974 If an atomic operation is marked ``singlethread``, it only *synchronizes
1975 with* or participates in modification and seq\_cst total orderings with
1976 other operations running in the same thread (for example, in signal
1984 LLVM IR floating-point binary ops (:ref:`fadd <i_fadd>`,
1985 :ref:`fsub <i_fsub>`, :ref:`fmul <i_fmul>`, :ref:`fdiv <i_fdiv>`,
1986 :ref:`frem <i_frem>`, :ref:`fcmp <i_fcmp>`) have the following flags that can
1987 be set to enable otherwise unsafe floating point operations
1990 No NaNs - Allow optimizations to assume the arguments and result are not
1991 NaN. Such optimizations are required to retain defined behavior over
1992 NaNs, but the value of the result is undefined.
1995 No Infs - Allow optimizations to assume the arguments and result are not
1996 +/-Inf. Such optimizations are required to retain defined behavior over
1997 +/-Inf, but the value of the result is undefined.
2000 No Signed Zeros - Allow optimizations to treat the sign of a zero
2001 argument or result as insignificant.
2004 Allow Reciprocal - Allow optimizations to use the reciprocal of an
2005 argument rather than perform division.
2008 Fast - Allow algebraically equivalent transformations that may
2009 dramatically change results in floating point (e.g. reassociate). This
2010 flag implies all the others.
2014 Use-list Order Directives
2015 -------------------------
2017 Use-list directives encode the in-memory order of each use-list, allowing the
2018 order to be recreated. ``<order-indexes>`` is a comma-separated list of
2019 indexes that are assigned to the referenced value's uses. The referenced
2020 value's use-list is immediately sorted by these indexes.
2022 Use-list directives may appear at function scope or global scope. They are not
2023 instructions, and have no effect on the semantics of the IR. When they're at
2024 function scope, they must appear after the terminator of the final basic block.
2026 If basic blocks have their address taken via ``blockaddress()`` expressions,
2027 ``uselistorder_bb`` can be used to reorder their use-lists from outside their
2034 uselistorder <ty> <value>, { <order-indexes> }
2035 uselistorder_bb @function, %block { <order-indexes> }
2041 define void @foo(i32 %arg1, i32 %arg2) {
2043 ; ... instructions ...
2045 ; ... instructions ...
2047 ; At function scope.
2048 uselistorder i32 %arg1, { 1, 0, 2 }
2049 uselistorder label %bb, { 1, 0 }
2053 uselistorder i32* @global, { 1, 2, 0 }
2054 uselistorder i32 7, { 1, 0 }
2055 uselistorder i32 (i32) @bar, { 1, 0 }
2056 uselistorder_bb @foo, %bb, { 5, 1, 3, 2, 0, 4 }
2063 The LLVM type system is one of the most important features of the
2064 intermediate representation. Being typed enables a number of
2065 optimizations to be performed on the intermediate representation
2066 directly, without having to do extra analyses on the side before the
2067 transformation. A strong type system makes it easier to read the
2068 generated code and enables novel analyses and transformations that are
2069 not feasible to perform on normal three address code representations.
2079 The void type does not represent any value and has no size.
2097 The function type can be thought of as a function signature. It consists of a
2098 return type and a list of formal parameter types. The return type of a function
2099 type is a void type or first class type --- except for :ref:`label <t_label>`
2100 and :ref:`metadata <t_metadata>` types.
2106 <returntype> (<parameter list>)
2108 ...where '``<parameter list>``' is a comma-separated list of type
2109 specifiers. Optionally, the parameter list may include a type ``...``, which
2110 indicates that the function takes a variable number of arguments. Variable
2111 argument functions can access their arguments with the :ref:`variable argument
2112 handling intrinsic <int_varargs>` functions. '``<returntype>``' is any type
2113 except :ref:`label <t_label>` and :ref:`metadata <t_metadata>`.
2117 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2118 | ``i32 (i32)`` | function taking an ``i32``, returning an ``i32`` |
2119 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2120 | ``float (i16, i32 *) *`` | :ref:`Pointer <t_pointer>` to a function that takes an ``i16`` and a :ref:`pointer <t_pointer>` to ``i32``, returning ``float``. |
2121 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2122 | ``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. |
2123 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2124 | ``{i32, i32} (i32)`` | A function taking an ``i32``, returning a :ref:`structure <t_struct>` containing two ``i32`` values |
2125 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2132 The :ref:`first class <t_firstclass>` types are perhaps the most important.
2133 Values of these types are the only ones which can be produced by
2141 These are the types that are valid in registers from CodeGen's perspective.
2150 The integer type is a very simple type that simply specifies an
2151 arbitrary bit width for the integer type desired. Any bit width from 1
2152 bit to 2\ :sup:`23`\ -1 (about 8 million) can be specified.
2160 The number of bits the integer will occupy is specified by the ``N``
2166 +----------------+------------------------------------------------+
2167 | ``i1`` | a single-bit integer. |
2168 +----------------+------------------------------------------------+
2169 | ``i32`` | a 32-bit integer. |
2170 +----------------+------------------------------------------------+
2171 | ``i1942652`` | a really big integer of over 1 million bits. |
2172 +----------------+------------------------------------------------+
2176 Floating Point Types
2177 """"""""""""""""""""
2186 - 16-bit floating point value
2189 - 32-bit floating point value
2192 - 64-bit floating point value
2195 - 128-bit floating point value (112-bit mantissa)
2198 - 80-bit floating point value (X87)
2201 - 128-bit floating point value (two 64-bits)
2208 The x86_mmx type represents a value held in an MMX register on an x86
2209 machine. The operations allowed on it are quite limited: parameters and
2210 return values, load and store, and bitcast. User-specified MMX
2211 instructions are represented as intrinsic or asm calls with arguments
2212 and/or results of this type. There are no arrays, vectors or constants
2229 The pointer type is used to specify memory locations. Pointers are
2230 commonly used to reference objects in memory.
2232 Pointer types may have an optional address space attribute defining the
2233 numbered address space where the pointed-to object resides. The default
2234 address space is number zero. The semantics of non-zero address spaces
2235 are target-specific.
2237 Note that LLVM does not permit pointers to void (``void*``) nor does it
2238 permit pointers to labels (``label*``). Use ``i8*`` instead.
2248 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2249 | ``[4 x i32]*`` | A :ref:`pointer <t_pointer>` to :ref:`array <t_array>` of four ``i32`` values. |
2250 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2251 | ``i32 (i32*) *`` | A :ref:`pointer <t_pointer>` to a :ref:`function <t_function>` that takes an ``i32*``, returning an ``i32``. |
2252 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2253 | ``i32 addrspace(5)*`` | A :ref:`pointer <t_pointer>` to an ``i32`` value that resides in address space #5. |
2254 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2263 A vector type is a simple derived type that represents a vector of
2264 elements. Vector types are used when multiple primitive data are
2265 operated in parallel using a single instruction (SIMD). A vector type
2266 requires a size (number of elements) and an underlying primitive data
2267 type. Vector types are considered :ref:`first class <t_firstclass>`.
2273 < <# elements> x <elementtype> >
2275 The number of elements is a constant integer value larger than 0;
2276 elementtype may be any integer, floating point or pointer type. Vectors
2277 of size zero are not allowed.
2281 +-------------------+--------------------------------------------------+
2282 | ``<4 x i32>`` | Vector of 4 32-bit integer values. |
2283 +-------------------+--------------------------------------------------+
2284 | ``<8 x float>`` | Vector of 8 32-bit floating-point values. |
2285 +-------------------+--------------------------------------------------+
2286 | ``<2 x i64>`` | Vector of 2 64-bit integer values. |
2287 +-------------------+--------------------------------------------------+
2288 | ``<4 x i64*>`` | Vector of 4 pointers to 64-bit integer values. |
2289 +-------------------+--------------------------------------------------+
2298 The label type represents code labels.
2313 The token type is used when a value is associated with an instruction
2314 but all uses of the value must not attempt to introspect or obscure it.
2315 As such, it is not appropriate to have a :ref:`phi <i_phi>` or
2316 :ref:`select <i_select>` of type token.
2333 The metadata type represents embedded metadata. No derived types may be
2334 created from metadata except for :ref:`function <t_function>` arguments.
2347 Aggregate Types are a subset of derived types that can contain multiple
2348 member types. :ref:`Arrays <t_array>` and :ref:`structs <t_struct>` are
2349 aggregate types. :ref:`Vectors <t_vector>` are not considered to be
2359 The array type is a very simple derived type that arranges elements
2360 sequentially in memory. The array type requires a size (number of
2361 elements) and an underlying data type.
2367 [<# elements> x <elementtype>]
2369 The number of elements is a constant integer value; ``elementtype`` may
2370 be any type with a size.
2374 +------------------+--------------------------------------+
2375 | ``[40 x i32]`` | Array of 40 32-bit integer values. |
2376 +------------------+--------------------------------------+
2377 | ``[41 x i32]`` | Array of 41 32-bit integer values. |
2378 +------------------+--------------------------------------+
2379 | ``[4 x i8]`` | Array of 4 8-bit integer values. |
2380 +------------------+--------------------------------------+
2382 Here are some examples of multidimensional arrays:
2384 +-----------------------------+----------------------------------------------------------+
2385 | ``[3 x [4 x i32]]`` | 3x4 array of 32-bit integer values. |
2386 +-----------------------------+----------------------------------------------------------+
2387 | ``[12 x [10 x float]]`` | 12x10 array of single precision floating point values. |
2388 +-----------------------------+----------------------------------------------------------+
2389 | ``[2 x [3 x [4 x i16]]]`` | 2x3x4 array of 16-bit integer values. |
2390 +-----------------------------+----------------------------------------------------------+
2392 There is no restriction on indexing beyond the end of the array implied
2393 by a static type (though there are restrictions on indexing beyond the
2394 bounds of an allocated object in some cases). This means that
2395 single-dimension 'variable sized array' addressing can be implemented in
2396 LLVM with a zero length array type. An implementation of 'pascal style
2397 arrays' in LLVM could use the type "``{ i32, [0 x float]}``", for
2407 The structure type is used to represent a collection of data members
2408 together in memory. The elements of a structure may be any type that has
2411 Structures in memory are accessed using '``load``' and '``store``' by
2412 getting a pointer to a field with the '``getelementptr``' instruction.
2413 Structures in registers are accessed using the '``extractvalue``' and
2414 '``insertvalue``' instructions.
2416 Structures may optionally be "packed" structures, which indicate that
2417 the alignment of the struct is one byte, and that there is no padding
2418 between the elements. In non-packed structs, padding between field types
2419 is inserted as defined by the DataLayout string in the module, which is
2420 required to match what the underlying code generator expects.
2422 Structures can either be "literal" or "identified". A literal structure
2423 is defined inline with other types (e.g. ``{i32, i32}*``) whereas
2424 identified types are always defined at the top level with a name.
2425 Literal types are uniqued by their contents and can never be recursive
2426 or opaque since there is no way to write one. Identified types can be
2427 recursive, can be opaqued, and are never uniqued.
2433 %T1 = type { <type list> } ; Identified normal struct type
2434 %T2 = type <{ <type list> }> ; Identified packed struct type
2438 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2439 | ``{ i32, i32, i32 }`` | A triple of three ``i32`` values |
2440 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2441 | ``{ 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``. |
2442 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2443 | ``<{ i8, i32 }>`` | A packed struct known to be 5 bytes in size. |
2444 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2448 Opaque Structure Types
2449 """"""""""""""""""""""
2453 Opaque structure types are used to represent named structure types that
2454 do not have a body specified. This corresponds (for example) to the C
2455 notion of a forward declared structure.
2466 +--------------+-------------------+
2467 | ``opaque`` | An opaque type. |
2468 +--------------+-------------------+
2475 LLVM has several different basic types of constants. This section
2476 describes them all and their syntax.
2481 **Boolean constants**
2482 The two strings '``true``' and '``false``' are both valid constants
2484 **Integer constants**
2485 Standard integers (such as '4') are constants of the
2486 :ref:`integer <t_integer>` type. Negative numbers may be used with
2488 **Floating point constants**
2489 Floating point constants use standard decimal notation (e.g.
2490 123.421), exponential notation (e.g. 1.23421e+2), or a more precise
2491 hexadecimal notation (see below). The assembler requires the exact
2492 decimal value of a floating-point constant. For example, the
2493 assembler accepts 1.25 but rejects 1.3 because 1.3 is a repeating
2494 decimal in binary. Floating point constants must have a :ref:`floating
2495 point <t_floating>` type.
2496 **Null pointer constants**
2497 The identifier '``null``' is recognized as a null pointer constant
2498 and must be of :ref:`pointer type <t_pointer>`.
2500 The identifier '``none``' is recognized as an empty token constant
2501 and must be of :ref:`token type <t_token>`.
2503 The one non-intuitive notation for constants is the hexadecimal form of
2504 floating point constants. For example, the form
2505 '``double 0x432ff973cafa8000``' is equivalent to (but harder to read
2506 than) '``double 4.5e+15``'. The only time hexadecimal floating point
2507 constants are required (and the only time that they are generated by the
2508 disassembler) is when a floating point constant must be emitted but it
2509 cannot be represented as a decimal floating point number in a reasonable
2510 number of digits. For example, NaN's, infinities, and other special
2511 values are represented in their IEEE hexadecimal format so that assembly
2512 and disassembly do not cause any bits to change in the constants.
2514 When using the hexadecimal form, constants of types half, float, and
2515 double are represented using the 16-digit form shown above (which
2516 matches the IEEE754 representation for double); half and float values
2517 must, however, be exactly representable as IEEE 754 half and single
2518 precision, respectively. Hexadecimal format is always used for long
2519 double, and there are three forms of long double. The 80-bit format used
2520 by x86 is represented as ``0xK`` followed by 20 hexadecimal digits. The
2521 128-bit format used by PowerPC (two adjacent doubles) is represented by
2522 ``0xM`` followed by 32 hexadecimal digits. The IEEE 128-bit format is
2523 represented by ``0xL`` followed by 32 hexadecimal digits. Long doubles
2524 will only work if they match the long double format on your target.
2525 The IEEE 16-bit format (half precision) is represented by ``0xH``
2526 followed by 4 hexadecimal digits. All hexadecimal formats are big-endian
2527 (sign bit at the left).
2529 There are no constants of type x86_mmx.
2531 .. _complexconstants:
2536 Complex constants are a (potentially recursive) combination of simple
2537 constants and smaller complex constants.
2539 **Structure constants**
2540 Structure constants are represented with notation similar to
2541 structure type definitions (a comma separated list of elements,
2542 surrounded by braces (``{}``)). For example:
2543 "``{ i32 4, float 17.0, i32* @G }``", where "``@G``" is declared as
2544 "``@G = external global i32``". Structure constants must have
2545 :ref:`structure type <t_struct>`, and the number and types of elements
2546 must match those specified by the type.
2548 Array constants are represented with notation similar to array type
2549 definitions (a comma separated list of elements, surrounded by
2550 square brackets (``[]``)). For example:
2551 "``[ i32 42, i32 11, i32 74 ]``". Array constants must have
2552 :ref:`array type <t_array>`, and the number and types of elements must
2553 match those specified by the type. As a special case, character array
2554 constants may also be represented as a double-quoted string using the ``c``
2555 prefix. For example: "``c"Hello World\0A\00"``".
2556 **Vector constants**
2557 Vector constants are represented with notation similar to vector
2558 type definitions (a comma separated list of elements, surrounded by
2559 less-than/greater-than's (``<>``)). For example:
2560 "``< i32 42, i32 11, i32 74, i32 100 >``". Vector constants
2561 must have :ref:`vector type <t_vector>`, and the number and types of
2562 elements must match those specified by the type.
2563 **Zero initialization**
2564 The string '``zeroinitializer``' can be used to zero initialize a
2565 value to zero of *any* type, including scalar and
2566 :ref:`aggregate <t_aggregate>` types. This is often used to avoid
2567 having to print large zero initializers (e.g. for large arrays) and
2568 is always exactly equivalent to using explicit zero initializers.
2570 A metadata node is a constant tuple without types. For example:
2571 "``!{!0, !{!2, !0}, !"test"}``". Metadata can reference constant values,
2572 for example: "``!{!0, i32 0, i8* @global, i64 (i64)* @function, !"str"}``".
2573 Unlike other typed constants that are meant to be interpreted as part of
2574 the instruction stream, metadata is a place to attach additional
2575 information such as debug info.
2577 Global Variable and Function Addresses
2578 --------------------------------------
2580 The addresses of :ref:`global variables <globalvars>` and
2581 :ref:`functions <functionstructure>` are always implicitly valid
2582 (link-time) constants. These constants are explicitly referenced when
2583 the :ref:`identifier for the global <identifiers>` is used and always have
2584 :ref:`pointer <t_pointer>` type. For example, the following is a legal LLVM
2587 .. code-block:: llvm
2591 @Z = global [2 x i32*] [ i32* @X, i32* @Y ]
2598 The string '``undef``' can be used anywhere a constant is expected, and
2599 indicates that the user of the value may receive an unspecified
2600 bit-pattern. Undefined values may be of any type (other than '``label``'
2601 or '``void``') and be used anywhere a constant is permitted.
2603 Undefined values are useful because they indicate to the compiler that
2604 the program is well defined no matter what value is used. This gives the
2605 compiler more freedom to optimize. Here are some examples of
2606 (potentially surprising) transformations that are valid (in pseudo IR):
2608 .. code-block:: llvm
2618 This is safe because all of the output bits are affected by the undef
2619 bits. Any output bit can have a zero or one depending on the input bits.
2621 .. code-block:: llvm
2632 These logical operations have bits that are not always affected by the
2633 input. For example, if ``%X`` has a zero bit, then the output of the
2634 '``and``' operation will always be a zero for that bit, no matter what
2635 the corresponding bit from the '``undef``' is. As such, it is unsafe to
2636 optimize or assume that the result of the '``and``' is '``undef``'.
2637 However, it is safe to assume that all bits of the '``undef``' could be
2638 0, and optimize the '``and``' to 0. Likewise, it is safe to assume that
2639 all the bits of the '``undef``' operand to the '``or``' could be set,
2640 allowing the '``or``' to be folded to -1.
2642 .. code-block:: llvm
2644 %A = select undef, %X, %Y
2645 %B = select undef, 42, %Y
2646 %C = select %X, %Y, undef
2656 This set of examples shows that undefined '``select``' (and conditional
2657 branch) conditions can go *either way*, but they have to come from one
2658 of the two operands. In the ``%A`` example, if ``%X`` and ``%Y`` were
2659 both known to have a clear low bit, then ``%A`` would have to have a
2660 cleared low bit. However, in the ``%C`` example, the optimizer is
2661 allowed to assume that the '``undef``' operand could be the same as
2662 ``%Y``, allowing the whole '``select``' to be eliminated.
2664 .. code-block:: llvm
2666 %A = xor undef, undef
2683 This example points out that two '``undef``' operands are not
2684 necessarily the same. This can be surprising to people (and also matches
2685 C semantics) where they assume that "``X^X``" is always zero, even if
2686 ``X`` is undefined. This isn't true for a number of reasons, but the
2687 short answer is that an '``undef``' "variable" can arbitrarily change
2688 its value over its "live range". This is true because the variable
2689 doesn't actually *have a live range*. Instead, the value is logically
2690 read from arbitrary registers that happen to be around when needed, so
2691 the value is not necessarily consistent over time. In fact, ``%A`` and
2692 ``%C`` need to have the same semantics or the core LLVM "replace all
2693 uses with" concept would not hold.
2695 .. code-block:: llvm
2703 These examples show the crucial difference between an *undefined value*
2704 and *undefined behavior*. An undefined value (like '``undef``') is
2705 allowed to have an arbitrary bit-pattern. This means that the ``%A``
2706 operation can be constant folded to '``undef``', because the '``undef``'
2707 could be an SNaN, and ``fdiv`` is not (currently) defined on SNaN's.
2708 However, in the second example, we can make a more aggressive
2709 assumption: because the ``undef`` is allowed to be an arbitrary value,
2710 we are allowed to assume that it could be zero. Since a divide by zero
2711 has *undefined behavior*, we are allowed to assume that the operation
2712 does not execute at all. This allows us to delete the divide and all
2713 code after it. Because the undefined operation "can't happen", the
2714 optimizer can assume that it occurs in dead code.
2716 .. code-block:: llvm
2718 a: store undef -> %X
2719 b: store %X -> undef
2724 These examples reiterate the ``fdiv`` example: a store *of* an undefined
2725 value can be assumed to not have any effect; we can assume that the
2726 value is overwritten with bits that happen to match what was already
2727 there. However, a store *to* an undefined location could clobber
2728 arbitrary memory, therefore, it has undefined behavior.
2735 Poison values are similar to :ref:`undef values <undefvalues>`, however
2736 they also represent the fact that an instruction or constant expression
2737 that cannot evoke side effects has nevertheless detected a condition
2738 that results in undefined behavior.
2740 There is currently no way of representing a poison value in the IR; they
2741 only exist when produced by operations such as :ref:`add <i_add>` with
2744 Poison value behavior is defined in terms of value *dependence*:
2746 - Values other than :ref:`phi <i_phi>` nodes depend on their operands.
2747 - :ref:`Phi <i_phi>` nodes depend on the operand corresponding to
2748 their dynamic predecessor basic block.
2749 - Function arguments depend on the corresponding actual argument values
2750 in the dynamic callers of their functions.
2751 - :ref:`Call <i_call>` instructions depend on the :ref:`ret <i_ret>`
2752 instructions that dynamically transfer control back to them.
2753 - :ref:`Invoke <i_invoke>` instructions depend on the
2754 :ref:`ret <i_ret>`, :ref:`resume <i_resume>`, or exception-throwing
2755 call instructions that dynamically transfer control back to them.
2756 - Non-volatile loads and stores depend on the most recent stores to all
2757 of the referenced memory addresses, following the order in the IR
2758 (including loads and stores implied by intrinsics such as
2759 :ref:`@llvm.memcpy <int_memcpy>`.)
2760 - An instruction with externally visible side effects depends on the
2761 most recent preceding instruction with externally visible side
2762 effects, following the order in the IR. (This includes :ref:`volatile
2763 operations <volatile>`.)
2764 - An instruction *control-depends* on a :ref:`terminator
2765 instruction <terminators>` if the terminator instruction has
2766 multiple successors and the instruction is always executed when
2767 control transfers to one of the successors, and may not be executed
2768 when control is transferred to another.
2769 - Additionally, an instruction also *control-depends* on a terminator
2770 instruction if the set of instructions it otherwise depends on would
2771 be different if the terminator had transferred control to a different
2773 - Dependence is transitive.
2775 Poison values have the same behavior as :ref:`undef values <undefvalues>`,
2776 with the additional effect that any instruction that has a *dependence*
2777 on a poison value has undefined behavior.
2779 Here are some examples:
2781 .. code-block:: llvm
2784 %poison = sub nuw i32 0, 1 ; Results in a poison value.
2785 %still_poison = and i32 %poison, 0 ; 0, but also poison.
2786 %poison_yet_again = getelementptr i32, i32* @h, i32 %still_poison
2787 store i32 0, i32* %poison_yet_again ; memory at @h[0] is poisoned
2789 store i32 %poison, i32* @g ; Poison value stored to memory.
2790 %poison2 = load i32, i32* @g ; Poison value loaded back from memory.
2792 store volatile i32 %poison, i32* @g ; External observation; undefined behavior.
2794 %narrowaddr = bitcast i32* @g to i16*
2795 %wideaddr = bitcast i32* @g to i64*
2796 %poison3 = load i16, i16* %narrowaddr ; Returns a poison value.
2797 %poison4 = load i64, i64* %wideaddr ; Returns a poison value.
2799 %cmp = icmp slt i32 %poison, 0 ; Returns a poison value.
2800 br i1 %cmp, label %true, label %end ; Branch to either destination.
2803 store volatile i32 0, i32* @g ; This is control-dependent on %cmp, so
2804 ; it has undefined behavior.
2808 %p = phi i32 [ 0, %entry ], [ 1, %true ]
2809 ; Both edges into this PHI are
2810 ; control-dependent on %cmp, so this
2811 ; always results in a poison value.
2813 store volatile i32 0, i32* @g ; This would depend on the store in %true
2814 ; if %cmp is true, or the store in %entry
2815 ; otherwise, so this is undefined behavior.
2817 br i1 %cmp, label %second_true, label %second_end
2818 ; The same branch again, but this time the
2819 ; true block doesn't have side effects.
2826 store volatile i32 0, i32* @g ; This time, the instruction always depends
2827 ; on the store in %end. Also, it is
2828 ; control-equivalent to %end, so this is
2829 ; well-defined (ignoring earlier undefined
2830 ; behavior in this example).
2834 Addresses of Basic Blocks
2835 -------------------------
2837 ``blockaddress(@function, %block)``
2839 The '``blockaddress``' constant computes the address of the specified
2840 basic block in the specified function, and always has an ``i8*`` type.
2841 Taking the address of the entry block is illegal.
2843 This value only has defined behavior when used as an operand to the
2844 ':ref:`indirectbr <i_indirectbr>`' instruction, or for comparisons
2845 against null. Pointer equality tests between labels addresses results in
2846 undefined behavior --- though, again, comparison against null is ok, and
2847 no label is equal to the null pointer. This may be passed around as an
2848 opaque pointer sized value as long as the bits are not inspected. This
2849 allows ``ptrtoint`` and arithmetic to be performed on these values so
2850 long as the original value is reconstituted before the ``indirectbr``
2853 Finally, some targets may provide defined semantics when using the value
2854 as the operand to an inline assembly, but that is target specific.
2858 Constant Expressions
2859 --------------------
2861 Constant expressions are used to allow expressions involving other
2862 constants to be used as constants. Constant expressions may be of any
2863 :ref:`first class <t_firstclass>` type and may involve any LLVM operation
2864 that does not have side effects (e.g. load and call are not supported).
2865 The following is the syntax for constant expressions:
2867 ``trunc (CST to TYPE)``
2868 Truncate a constant to another type. The bit size of CST must be
2869 larger than the bit size of TYPE. Both types must be integers.
2870 ``zext (CST to TYPE)``
2871 Zero extend a constant to another type. The bit size of CST must be
2872 smaller than the bit size of TYPE. Both types must be integers.
2873 ``sext (CST to TYPE)``
2874 Sign extend a constant to another type. The bit size of CST must be
2875 smaller than the bit size of TYPE. Both types must be integers.
2876 ``fptrunc (CST to TYPE)``
2877 Truncate a floating point constant to another floating point type.
2878 The size of CST must be larger than the size of TYPE. Both types
2879 must be floating point.
2880 ``fpext (CST to TYPE)``
2881 Floating point extend a constant to another type. The size of CST
2882 must be smaller or equal to the size of TYPE. Both types must be
2884 ``fptoui (CST to TYPE)``
2885 Convert a floating point constant to the corresponding unsigned
2886 integer constant. TYPE must be a scalar or vector integer type. CST
2887 must be of scalar or vector floating point type. Both CST and TYPE
2888 must be scalars, or vectors of the same number of elements. If the
2889 value won't fit in the integer type, the results are undefined.
2890 ``fptosi (CST to TYPE)``
2891 Convert a floating point constant to the corresponding signed
2892 integer constant. TYPE must be a scalar or vector integer type. CST
2893 must be of scalar or vector floating point type. Both CST and TYPE
2894 must be scalars, or vectors of the same number of elements. If the
2895 value won't fit in the integer type, the results are undefined.
2896 ``uitofp (CST to TYPE)``
2897 Convert an unsigned integer constant to the corresponding floating
2898 point constant. TYPE must be a scalar or vector floating point type.
2899 CST must be of scalar or vector integer type. Both CST and TYPE must
2900 be scalars, or vectors of the same number of elements. If the value
2901 won't fit in the floating point type, the results are undefined.
2902 ``sitofp (CST to TYPE)``
2903 Convert a signed integer constant to the corresponding floating
2904 point constant. TYPE must be a scalar or vector floating point type.
2905 CST must be of scalar or vector integer type. Both CST and TYPE must
2906 be scalars, or vectors of the same number of elements. If the value
2907 won't fit in the floating point type, the results are undefined.
2908 ``ptrtoint (CST to TYPE)``
2909 Convert a pointer typed constant to the corresponding integer
2910 constant. ``TYPE`` must be an integer type. ``CST`` must be of
2911 pointer type. The ``CST`` value is zero extended, truncated, or
2912 unchanged to make it fit in ``TYPE``.
2913 ``inttoptr (CST to TYPE)``
2914 Convert an integer constant to a pointer constant. TYPE must be a
2915 pointer type. CST must be of integer type. The CST value is zero
2916 extended, truncated, or unchanged to make it fit in a pointer size.
2917 This one is *really* dangerous!
2918 ``bitcast (CST to TYPE)``
2919 Convert a constant, CST, to another TYPE. The constraints of the
2920 operands are the same as those for the :ref:`bitcast
2921 instruction <i_bitcast>`.
2922 ``addrspacecast (CST to TYPE)``
2923 Convert a constant pointer or constant vector of pointer, CST, to another
2924 TYPE in a different address space. The constraints of the operands are the
2925 same as those for the :ref:`addrspacecast instruction <i_addrspacecast>`.
2926 ``getelementptr (TY, CSTPTR, IDX0, IDX1, ...)``, ``getelementptr inbounds (TY, CSTPTR, IDX0, IDX1, ...)``
2927 Perform the :ref:`getelementptr operation <i_getelementptr>` on
2928 constants. As with the :ref:`getelementptr <i_getelementptr>`
2929 instruction, the index list may have zero or more indexes, which are
2930 required to make sense for the type of "pointer to TY".
2931 ``select (COND, VAL1, VAL2)``
2932 Perform the :ref:`select operation <i_select>` on constants.
2933 ``icmp COND (VAL1, VAL2)``
2934 Performs the :ref:`icmp operation <i_icmp>` on constants.
2935 ``fcmp COND (VAL1, VAL2)``
2936 Performs the :ref:`fcmp operation <i_fcmp>` on constants.
2937 ``extractelement (VAL, IDX)``
2938 Perform the :ref:`extractelement operation <i_extractelement>` on
2940 ``insertelement (VAL, ELT, IDX)``
2941 Perform the :ref:`insertelement operation <i_insertelement>` on
2943 ``shufflevector (VEC1, VEC2, IDXMASK)``
2944 Perform the :ref:`shufflevector operation <i_shufflevector>` on
2946 ``extractvalue (VAL, IDX0, IDX1, ...)``
2947 Perform the :ref:`extractvalue operation <i_extractvalue>` on
2948 constants. The index list is interpreted in a similar manner as
2949 indices in a ':ref:`getelementptr <i_getelementptr>`' operation. At
2950 least one index value must be specified.
2951 ``insertvalue (VAL, ELT, IDX0, IDX1, ...)``
2952 Perform the :ref:`insertvalue operation <i_insertvalue>` on constants.
2953 The index list is interpreted in a similar manner as indices in a
2954 ':ref:`getelementptr <i_getelementptr>`' operation. At least one index
2955 value must be specified.
2956 ``OPCODE (LHS, RHS)``
2957 Perform the specified operation of the LHS and RHS constants. OPCODE
2958 may be any of the :ref:`binary <binaryops>` or :ref:`bitwise
2959 binary <bitwiseops>` operations. The constraints on operands are
2960 the same as those for the corresponding instruction (e.g. no bitwise
2961 operations on floating point values are allowed).
2968 Inline Assembler Expressions
2969 ----------------------------
2971 LLVM supports inline assembler expressions (as opposed to :ref:`Module-Level
2972 Inline Assembly <moduleasm>`) through the use of a special value. This value
2973 represents the inline assembler as a template string (containing the
2974 instructions to emit), a list of operand constraints (stored as a string), a
2975 flag that indicates whether or not the inline asm expression has side effects,
2976 and a flag indicating whether the function containing the asm needs to align its
2977 stack conservatively.
2979 The template string supports argument substitution of the operands using "``$``"
2980 followed by a number, to indicate substitution of the given register/memory
2981 location, as specified by the constraint string. "``${NUM:MODIFIER}``" may also
2982 be used, where ``MODIFIER`` is a target-specific annotation for how to print the
2983 operand (See :ref:`inline-asm-modifiers`).
2985 A literal "``$``" may be included by using "``$$``" in the template. To include
2986 other special characters into the output, the usual "``\XX``" escapes may be
2987 used, just as in other strings. Note that after template substitution, the
2988 resulting assembly string is parsed by LLVM's integrated assembler unless it is
2989 disabled -- even when emitting a ``.s`` file -- and thus must contain assembly
2990 syntax known to LLVM.
2992 LLVM's support for inline asm is modeled closely on the requirements of Clang's
2993 GCC-compatible inline-asm support. Thus, the feature-set and the constraint and
2994 modifier codes listed here are similar or identical to those in GCC's inline asm
2995 support. However, to be clear, the syntax of the template and constraint strings
2996 described here is *not* the same as the syntax accepted by GCC and Clang, and,
2997 while most constraint letters are passed through as-is by Clang, some get
2998 translated to other codes when converting from the C source to the LLVM
3001 An example inline assembler expression is:
3003 .. code-block:: llvm
3005 i32 (i32) asm "bswap $0", "=r,r"
3007 Inline assembler expressions may **only** be used as the callee operand
3008 of a :ref:`call <i_call>` or an :ref:`invoke <i_invoke>` instruction.
3009 Thus, typically we have:
3011 .. code-block:: llvm
3013 %X = call i32 asm "bswap $0", "=r,r"(i32 %Y)
3015 Inline asms with side effects not visible in the constraint list must be
3016 marked as having side effects. This is done through the use of the
3017 '``sideeffect``' keyword, like so:
3019 .. code-block:: llvm
3021 call void asm sideeffect "eieio", ""()
3023 In some cases inline asms will contain code that will not work unless
3024 the stack is aligned in some way, such as calls or SSE instructions on
3025 x86, yet will not contain code that does that alignment within the asm.
3026 The compiler should make conservative assumptions about what the asm
3027 might contain and should generate its usual stack alignment code in the
3028 prologue if the '``alignstack``' keyword is present:
3030 .. code-block:: llvm
3032 call void asm alignstack "eieio", ""()
3034 Inline asms also support using non-standard assembly dialects. The
3035 assumed dialect is ATT. When the '``inteldialect``' keyword is present,
3036 the inline asm is using the Intel dialect. Currently, ATT and Intel are
3037 the only supported dialects. An example is:
3039 .. code-block:: llvm
3041 call void asm inteldialect "eieio", ""()
3043 If multiple keywords appear the '``sideeffect``' keyword must come
3044 first, the '``alignstack``' keyword second and the '``inteldialect``'
3047 Inline Asm Constraint String
3048 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3050 The constraint list is a comma-separated string, each element containing one or
3051 more constraint codes.
3053 For each element in the constraint list an appropriate register or memory
3054 operand will be chosen, and it will be made available to assembly template
3055 string expansion as ``$0`` for the first constraint in the list, ``$1`` for the
3058 There are three different types of constraints, which are distinguished by a
3059 prefix symbol in front of the constraint code: Output, Input, and Clobber. The
3060 constraints must always be given in that order: outputs first, then inputs, then
3061 clobbers. They cannot be intermingled.
3063 There are also three different categories of constraint codes:
3065 - Register constraint. This is either a register class, or a fixed physical
3066 register. This kind of constraint will allocate a register, and if necessary,
3067 bitcast the argument or result to the appropriate type.
3068 - Memory constraint. This kind of constraint is for use with an instruction
3069 taking a memory operand. Different constraints allow for different addressing
3070 modes used by the target.
3071 - Immediate value constraint. This kind of constraint is for an integer or other
3072 immediate value which can be rendered directly into an instruction. The
3073 various target-specific constraints allow the selection of a value in the
3074 proper range for the instruction you wish to use it with.
3079 Output constraints are specified by an "``=``" prefix (e.g. "``=r``"). This
3080 indicates that the assembly will write to this operand, and the operand will
3081 then be made available as a return value of the ``asm`` expression. Output
3082 constraints do not consume an argument from the call instruction. (Except, see
3083 below about indirect outputs).
3085 Normally, it is expected that no output locations are written to by the assembly
3086 expression until *all* of the inputs have been read. As such, LLVM may assign
3087 the same register to an output and an input. If this is not safe (e.g. if the
3088 assembly contains two instructions, where the first writes to one output, and
3089 the second reads an input and writes to a second output), then the "``&``"
3090 modifier must be used (e.g. "``=&r``") to specify that the output is an
3091 "early-clobber" output. Marking an ouput as "early-clobber" ensures that LLVM
3092 will not use the same register for any inputs (other than an input tied to this
3098 Input constraints do not have a prefix -- just the constraint codes. Each input
3099 constraint will consume one argument from the call instruction. It is not
3100 permitted for the asm to write to any input register or memory location (unless
3101 that input is tied to an output). Note also that multiple inputs may all be
3102 assigned to the same register, if LLVM can determine that they necessarily all
3103 contain the same value.
3105 Instead of providing a Constraint Code, input constraints may also "tie"
3106 themselves to an output constraint, by providing an integer as the constraint
3107 string. Tied inputs still consume an argument from the call instruction, and
3108 take up a position in the asm template numbering as is usual -- they will simply
3109 be constrained to always use the same register as the output they've been tied
3110 to. For example, a constraint string of "``=r,0``" says to assign a register for
3111 output, and use that register as an input as well (it being the 0'th
3114 It is permitted to tie an input to an "early-clobber" output. In that case, no
3115 *other* input may share the same register as the input tied to the early-clobber
3116 (even when the other input has the same value).
3118 You may only tie an input to an output which has a register constraint, not a
3119 memory constraint. Only a single input may be tied to an output.
3121 There is also an "interesting" feature which deserves a bit of explanation: if a
3122 register class constraint allocates a register which is too small for the value
3123 type operand provided as input, the input value will be split into multiple
3124 registers, and all of them passed to the inline asm.
3126 However, this feature is often not as useful as you might think.
3128 Firstly, the registers are *not* guaranteed to be consecutive. So, on those
3129 architectures that have instructions which operate on multiple consecutive
3130 instructions, this is not an appropriate way to support them. (e.g. the 32-bit
3131 SparcV8 has a 64-bit load, which instruction takes a single 32-bit register. The
3132 hardware then loads into both the named register, and the next register. This
3133 feature of inline asm would not be useful to support that.)
3135 A few of the targets provide a template string modifier allowing explicit access
3136 to the second register of a two-register operand (e.g. MIPS ``L``, ``M``, and
3137 ``D``). On such an architecture, you can actually access the second allocated
3138 register (yet, still, not any subsequent ones). But, in that case, you're still
3139 probably better off simply splitting the value into two separate operands, for
3140 clarity. (e.g. see the description of the ``A`` constraint on X86, which,
3141 despite existing only for use with this feature, is not really a good idea to
3144 Indirect inputs and outputs
3145 """""""""""""""""""""""""""
3147 Indirect output or input constraints can be specified by the "``*``" modifier
3148 (which goes after the "``=``" in case of an output). This indicates that the asm
3149 will write to or read from the contents of an *address* provided as an input
3150 argument. (Note that in this way, indirect outputs act more like an *input* than
3151 an output: just like an input, they consume an argument of the call expression,
3152 rather than producing a return value. An indirect output constraint is an
3153 "output" only in that the asm is expected to write to the contents of the input
3154 memory location, instead of just read from it).
3156 This is most typically used for memory constraint, e.g. "``=*m``", to pass the
3157 address of a variable as a value.
3159 It is also possible to use an indirect *register* constraint, but only on output
3160 (e.g. "``=*r``"). This will cause LLVM to allocate a register for an output
3161 value normally, and then, separately emit a store to the address provided as
3162 input, after the provided inline asm. (It's not clear what value this
3163 functionality provides, compared to writing the store explicitly after the asm
3164 statement, and it can only produce worse code, since it bypasses many
3165 optimization passes. I would recommend not using it.)
3171 A clobber constraint is indicated by a "``~``" prefix. A clobber does not
3172 consume an input operand, nor generate an output. Clobbers cannot use any of the
3173 general constraint code letters -- they may use only explicit register
3174 constraints, e.g. "``~{eax}``". The one exception is that a clobber string of
3175 "``~{memory}``" indicates that the assembly writes to arbitrary undeclared
3176 memory locations -- not only the memory pointed to by a declared indirect
3182 After a potential prefix comes constraint code, or codes.
3184 A Constraint Code is either a single letter (e.g. "``r``"), a "``^``" character
3185 followed by two letters (e.g. "``^wc``"), or "``{``" register-name "``}``"
3188 The one and two letter constraint codes are typically chosen to be the same as
3189 GCC's constraint codes.
3191 A single constraint may include one or more than constraint code in it, leaving
3192 it up to LLVM to choose which one to use. This is included mainly for
3193 compatibility with the translation of GCC inline asm coming from clang.
3195 There are two ways to specify alternatives, and either or both may be used in an
3196 inline asm constraint list:
3198 1) Append the codes to each other, making a constraint code set. E.g. "``im``"
3199 or "``{eax}m``". This means "choose any of the options in the set". The
3200 choice of constraint is made independently for each constraint in the
3203 2) Use "``|``" between constraint code sets, creating alternatives. Every
3204 constraint in the constraint list must have the same number of alternative
3205 sets. With this syntax, the same alternative in *all* of the items in the
3206 constraint list will be chosen together.
3208 Putting those together, you might have a two operand constraint string like
3209 ``"rm|r,ri|rm"``. This indicates that if operand 0 is ``r`` or ``m``, then
3210 operand 1 may be one of ``r`` or ``i``. If operand 0 is ``r``, then operand 1
3211 may be one of ``r`` or ``m``. But, operand 0 and 1 cannot both be of type m.
3213 However, the use of either of the alternatives features is *NOT* recommended, as
3214 LLVM is not able to make an intelligent choice about which one to use. (At the
3215 point it currently needs to choose, not enough information is available to do so
3216 in a smart way.) Thus, it simply tries to make a choice that's most likely to
3217 compile, not one that will be optimal performance. (e.g., given "``rm``", it'll
3218 always choose to use memory, not registers). And, if given multiple registers,
3219 or multiple register classes, it will simply choose the first one. (In fact, it
3220 doesn't currently even ensure explicitly specified physical registers are
3221 unique, so specifying multiple physical registers as alternatives, like
3222 ``{r11}{r12},{r11}{r12}``, will assign r11 to both operands, not at all what was
3225 Supported Constraint Code List
3226 """"""""""""""""""""""""""""""
3228 The constraint codes are, in general, expected to behave the same way they do in
3229 GCC. LLVM's support is often implemented on an 'as-needed' basis, to support C
3230 inline asm code which was supported by GCC. A mismatch in behavior between LLVM
3231 and GCC likely indicates a bug in LLVM.
3233 Some constraint codes are typically supported by all targets:
3235 - ``r``: A register in the target's general purpose register class.
3236 - ``m``: A memory address operand. It is target-specific what addressing modes
3237 are supported, typical examples are register, or register + register offset,
3238 or register + immediate offset (of some target-specific size).
3239 - ``i``: An integer constant (of target-specific width). Allows either a simple
3240 immediate, or a relocatable value.
3241 - ``n``: An integer constant -- *not* including relocatable values.
3242 - ``s``: An integer constant, but allowing *only* relocatable values.
3243 - ``X``: Allows an operand of any kind, no constraint whatsoever. Typically
3244 useful to pass a label for an asm branch or call.
3246 .. FIXME: but that surely isn't actually okay to jump out of an asm
3247 block without telling llvm about the control transfer???)
3249 - ``{register-name}``: Requires exactly the named physical register.
3251 Other constraints are target-specific:
3255 - ``z``: An immediate integer 0. Outputs ``WZR`` or ``XZR``, as appropriate.
3256 - ``I``: An immediate integer valid for an ``ADD`` or ``SUB`` instruction,
3257 i.e. 0 to 4095 with optional shift by 12.
3258 - ``J``: An immediate integer that, when negated, is valid for an ``ADD`` or
3259 ``SUB`` instruction, i.e. -1 to -4095 with optional left shift by 12.
3260 - ``K``: An immediate integer that is valid for the 'bitmask immediate 32' of a
3261 logical instruction like ``AND``, ``EOR``, or ``ORR`` with a 32-bit register.
3262 - ``L``: An immediate integer that is valid for the 'bitmask immediate 64' of a
3263 logical instruction like ``AND``, ``EOR``, or ``ORR`` with a 64-bit register.
3264 - ``M``: An immediate integer for use with the ``MOV`` assembly alias on a
3265 32-bit register. This is a superset of ``K``: in addition to the bitmask
3266 immediate, also allows immediate integers which can be loaded with a single
3267 ``MOVZ`` or ``MOVL`` instruction.
3268 - ``N``: An immediate integer for use with the ``MOV`` assembly alias on a
3269 64-bit register. This is a superset of ``L``.
3270 - ``Q``: Memory address operand must be in a single register (no
3271 offsets). (However, LLVM currently does this for the ``m`` constraint as
3273 - ``r``: A 32 or 64-bit integer register (W* or X*).
3274 - ``w``: A 32, 64, or 128-bit floating-point/SIMD register.
3275 - ``x``: A lower 128-bit floating-point/SIMD register (``V0`` to ``V15``).
3279 - ``r``: A 32 or 64-bit integer register.
3280 - ``[0-9]v``: The 32-bit VGPR register, number 0-9.
3281 - ``[0-9]s``: The 32-bit SGPR register, number 0-9.
3286 - ``Q``, ``Um``, ``Un``, ``Uq``, ``Us``, ``Ut``, ``Uv``, ``Uy``: Memory address
3287 operand. Treated the same as operand ``m``, at the moment.
3289 ARM and ARM's Thumb2 mode:
3291 - ``j``: An immediate integer between 0 and 65535 (valid for ``MOVW``)
3292 - ``I``: An immediate integer valid for a data-processing instruction.
3293 - ``J``: An immediate integer between -4095 and 4095.
3294 - ``K``: An immediate integer whose bitwise inverse is valid for a
3295 data-processing instruction. (Can be used with template modifier "``B``" to
3296 print the inverted value).
3297 - ``L``: An immediate integer whose negation is valid for a data-processing
3298 instruction. (Can be used with template modifier "``n``" to print the negated
3300 - ``M``: A power of two or a integer between 0 and 32.
3301 - ``N``: Invalid immediate constraint.
3302 - ``O``: Invalid immediate constraint.
3303 - ``r``: A general-purpose 32-bit integer register (``r0-r15``).
3304 - ``l``: In Thumb2 mode, low 32-bit GPR registers (``r0-r7``). In ARM mode, same
3306 - ``h``: In Thumb2 mode, a high 32-bit GPR register (``r8-r15``). In ARM mode,
3308 - ``w``: A 32, 64, or 128-bit floating-point/SIMD register: ``s0-s31``,
3309 ``d0-d31``, or ``q0-q15``.
3310 - ``x``: A 32, 64, or 128-bit floating-point/SIMD register: ``s0-s15``,
3311 ``d0-d7``, or ``q0-q3``.
3312 - ``t``: A floating-point/SIMD register, only supports 32-bit values:
3317 - ``I``: An immediate integer between 0 and 255.
3318 - ``J``: An immediate integer between -255 and -1.
3319 - ``K``: An immediate integer between 0 and 255, with optional left-shift by
3321 - ``L``: An immediate integer between -7 and 7.
3322 - ``M``: An immediate integer which is a multiple of 4 between 0 and 1020.
3323 - ``N``: An immediate integer between 0 and 31.
3324 - ``O``: An immediate integer which is a multiple of 4 between -508 and 508.
3325 - ``r``: A low 32-bit GPR register (``r0-r7``).
3326 - ``l``: A low 32-bit GPR register (``r0-r7``).
3327 - ``h``: A high GPR register (``r0-r7``).
3328 - ``w``: A 32, 64, or 128-bit floating-point/SIMD register: ``s0-s31``,
3329 ``d0-d31``, or ``q0-q15``.
3330 - ``x``: A 32, 64, or 128-bit floating-point/SIMD register: ``s0-s15``,
3331 ``d0-d7``, or ``q0-q3``.
3332 - ``t``: A floating-point/SIMD register, only supports 32-bit values:
3338 - ``o``, ``v``: A memory address operand, treated the same as constraint ``m``,
3340 - ``r``: A 32 or 64-bit register.
3344 - ``r``: An 8 or 16-bit register.
3348 - ``I``: An immediate signed 16-bit integer.
3349 - ``J``: An immediate integer zero.
3350 - ``K``: An immediate unsigned 16-bit integer.
3351 - ``L``: An immediate 32-bit integer, where the lower 16 bits are 0.
3352 - ``N``: An immediate integer between -65535 and -1.
3353 - ``O``: An immediate signed 15-bit integer.
3354 - ``P``: An immediate integer between 1 and 65535.
3355 - ``m``: A memory address operand. In MIPS-SE mode, allows a base address
3356 register plus 16-bit immediate offset. In MIPS mode, just a base register.
3357 - ``R``: A memory address operand. In MIPS-SE mode, allows a base address
3358 register plus a 9-bit signed offset. In MIPS mode, the same as constraint
3360 - ``ZC``: A memory address operand, suitable for use in a ``pref``, ``ll``, or
3361 ``sc`` instruction on the given subtarget (details vary).
3362 - ``r``, ``d``, ``y``: A 32 or 64-bit GPR register.
3363 - ``f``: A 32 or 64-bit FPU register (``F0-F31``), or a 128-bit MSA register
3364 (``W0-W31``). In the case of MSA registers, it is recommended to use the ``w``
3365 argument modifier for compatibility with GCC.
3366 - ``c``: A 32-bit or 64-bit GPR register suitable for indirect jump (always
3368 - ``l``: The ``lo`` register, 32 or 64-bit.
3373 - ``b``: A 1-bit integer register.
3374 - ``c`` or ``h``: A 16-bit integer register.
3375 - ``r``: A 32-bit integer register.
3376 - ``l`` or ``N``: A 64-bit integer register.
3377 - ``f``: A 32-bit float register.
3378 - ``d``: A 64-bit float register.
3383 - ``I``: An immediate signed 16-bit integer.
3384 - ``J``: An immediate unsigned 16-bit integer, shifted left 16 bits.
3385 - ``K``: An immediate unsigned 16-bit integer.
3386 - ``L``: An immediate signed 16-bit integer, shifted left 16 bits.
3387 - ``M``: An immediate integer greater than 31.
3388 - ``N``: An immediate integer that is an exact power of 2.
3389 - ``O``: The immediate integer constant 0.
3390 - ``P``: An immediate integer constant whose negation is a signed 16-bit
3392 - ``es``, ``o``, ``Q``, ``Z``, ``Zy``: A memory address operand, currently
3393 treated the same as ``m``.
3394 - ``r``: A 32 or 64-bit integer register.
3395 - ``b``: A 32 or 64-bit integer register, excluding ``R0`` (that is:
3397 - ``f``: A 32 or 64-bit float register (``F0-F31``), or when QPX is enabled, a
3398 128 or 256-bit QPX register (``Q0-Q31``; aliases the ``F`` registers).
3399 - ``v``: For ``4 x f32`` or ``4 x f64`` types, when QPX is enabled, a
3400 128 or 256-bit QPX register (``Q0-Q31``), otherwise a 128-bit
3401 altivec vector register (``V0-V31``).
3403 .. FIXME: is this a bug that v accepts QPX registers? I think this
3404 is supposed to only use the altivec vector registers?
3406 - ``y``: Condition register (``CR0-CR7``).
3407 - ``wc``: An individual CR bit in a CR register.
3408 - ``wa``, ``wd``, ``wf``: Any 128-bit VSX vector register, from the full VSX
3409 register set (overlapping both the floating-point and vector register files).
3410 - ``ws``: A 32 or 64-bit floating point register, from the full VSX register
3415 - ``I``: An immediate 13-bit signed integer.
3416 - ``r``: A 32-bit integer register.
3420 - ``I``: An immediate unsigned 8-bit integer.
3421 - ``J``: An immediate unsigned 12-bit integer.
3422 - ``K``: An immediate signed 16-bit integer.
3423 - ``L``: An immediate signed 20-bit integer.
3424 - ``M``: An immediate integer 0x7fffffff.
3425 - ``Q``, ``R``, ``S``, ``T``: A memory address operand, treated the same as
3426 ``m``, at the moment.
3427 - ``r`` or ``d``: A 32, 64, or 128-bit integer register.
3428 - ``a``: A 32, 64, or 128-bit integer address register (excludes R0, which in an
3429 address context evaluates as zero).
3430 - ``h``: A 32-bit value in the high part of a 64bit data register
3432 - ``f``: A 32, 64, or 128-bit floating point register.
3436 - ``I``: An immediate integer between 0 and 31.
3437 - ``J``: An immediate integer between 0 and 64.
3438 - ``K``: An immediate signed 8-bit integer.
3439 - ``L``: An immediate integer, 0xff or 0xffff or (in 64-bit mode only)
3441 - ``M``: An immediate integer between 0 and 3.
3442 - ``N``: An immediate unsigned 8-bit integer.
3443 - ``O``: An immediate integer between 0 and 127.
3444 - ``e``: An immediate 32-bit signed integer.
3445 - ``Z``: An immediate 32-bit unsigned integer.
3446 - ``o``, ``v``: Treated the same as ``m``, at the moment.
3447 - ``q``: An 8, 16, 32, or 64-bit register which can be accessed as an 8-bit
3448 ``l`` integer register. On X86-32, this is the ``a``, ``b``, ``c``, and ``d``
3449 registers, and on X86-64, it is all of the integer registers.
3450 - ``Q``: An 8, 16, 32, or 64-bit register which can be accessed as an 8-bit
3451 ``h`` integer register. This is the ``a``, ``b``, ``c``, and ``d`` registers.
3452 - ``r`` or ``l``: An 8, 16, 32, or 64-bit integer register.
3453 - ``R``: An 8, 16, 32, or 64-bit "legacy" integer register -- one which has
3454 existed since i386, and can be accessed without the REX prefix.
3455 - ``f``: A 32, 64, or 80-bit '387 FPU stack pseudo-register.
3456 - ``y``: A 64-bit MMX register, if MMX is enabled.
3457 - ``x``: If SSE is enabled: a 32 or 64-bit scalar operand, or 128-bit vector
3458 operand in a SSE register. If AVX is also enabled, can also be a 256-bit
3459 vector operand in an AVX register. If AVX-512 is also enabled, can also be a
3460 512-bit vector operand in an AVX512 register, Otherwise, an error.
3461 - ``Y``: The same as ``x``, if *SSE2* is enabled, otherwise an error.
3462 - ``A``: Special case: allocates EAX first, then EDX, for a single operand (in
3463 32-bit mode, a 64-bit integer operand will get split into two registers). It
3464 is not recommended to use this constraint, as in 64-bit mode, the 64-bit
3465 operand will get allocated only to RAX -- if two 32-bit operands are needed,
3466 you're better off splitting it yourself, before passing it to the asm
3471 - ``r``: A 32-bit integer register.
3474 .. _inline-asm-modifiers:
3476 Asm template argument modifiers
3477 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3479 In the asm template string, modifiers can be used on the operand reference, like
3482 The modifiers are, in general, expected to behave the same way they do in
3483 GCC. LLVM's support is often implemented on an 'as-needed' basis, to support C
3484 inline asm code which was supported by GCC. A mismatch in behavior between LLVM
3485 and GCC likely indicates a bug in LLVM.
3489 - ``c``: Print an immediate integer constant unadorned, without
3490 the target-specific immediate punctuation (e.g. no ``$`` prefix).
3491 - ``n``: Negate and print immediate integer constant unadorned, without the
3492 target-specific immediate punctuation (e.g. no ``$`` prefix).
3493 - ``l``: Print as an unadorned label, without the target-specific label
3494 punctuation (e.g. no ``$`` prefix).
3498 - ``w``: Print a GPR register with a ``w*`` name instead of ``x*`` name. E.g.,
3499 instead of ``x30``, print ``w30``.
3500 - ``x``: Print a GPR register with a ``x*`` name. (this is the default, anyhow).
3501 - ``b``, ``h``, ``s``, ``d``, ``q``: Print a floating-point/SIMD register with a
3502 ``b*``, ``h*``, ``s*``, ``d*``, or ``q*`` name, rather than the default of
3511 - ``a``: Print an operand as an address (with ``[`` and ``]`` surrounding a
3515 - ``y``: Print a VFP single-precision register as an indexed double (e.g. print
3516 as ``d4[1]`` instead of ``s9``)
3517 - ``B``: Bitwise invert and print an immediate integer constant without ``#``
3519 - ``L``: Print the low 16-bits of an immediate integer constant.
3520 - ``M``: Print as a register set suitable for ldm/stm. Also prints *all*
3521 register operands subsequent to the specified one (!), so use carefully.
3522 - ``Q``: Print the low-order register of a register-pair, or the low-order
3523 register of a two-register operand.
3524 - ``R``: Print the high-order register of a register-pair, or the high-order
3525 register of a two-register operand.
3526 - ``H``: Print the second register of a register-pair. (On a big-endian system,
3527 ``H`` is equivalent to ``Q``, and on little-endian system, ``H`` is equivalent
3530 .. FIXME: H doesn't currently support printing the second register
3531 of a two-register operand.
3533 - ``e``: Print the low doubleword register of a NEON quad register.
3534 - ``f``: Print the high doubleword register of a NEON quad register.
3535 - ``m``: Print the base register of a memory operand without the ``[`` and ``]``
3540 - ``L``: Print the second register of a two-register operand. Requires that it
3541 has been allocated consecutively to the first.
3543 .. FIXME: why is it restricted to consecutive ones? And there's
3544 nothing that ensures that happens, is there?
3546 - ``I``: Print the letter 'i' if the operand is an integer constant, otherwise
3547 nothing. Used to print 'addi' vs 'add' instructions.
3551 No additional modifiers.
3555 - ``X``: Print an immediate integer as hexadecimal
3556 - ``x``: Print the low 16 bits of an immediate integer as hexadecimal.
3557 - ``d``: Print an immediate integer as decimal.
3558 - ``m``: Subtract one and print an immediate integer as decimal.
3559 - ``z``: Print $0 if an immediate zero, otherwise print normally.
3560 - ``L``: Print the low-order register of a two-register operand, or prints the
3561 address of the low-order word of a double-word memory operand.
3563 .. FIXME: L seems to be missing memory operand support.
3565 - ``M``: Print the high-order register of a two-register operand, or prints the
3566 address of the high-order word of a double-word memory operand.
3568 .. FIXME: M seems to be missing memory operand support.
3570 - ``D``: Print the second register of a two-register operand, or prints the
3571 second word of a double-word memory operand. (On a big-endian system, ``D`` is
3572 equivalent to ``L``, and on little-endian system, ``D`` is equivalent to
3574 - ``w``: No effect. Provided for compatibility with GCC which requires this
3575 modifier in order to print MSA registers (``W0-W31``) with the ``f``
3584 - ``L``: Print the second register of a two-register operand. Requires that it
3585 has been allocated consecutively to the first.
3587 .. FIXME: why is it restricted to consecutive ones? And there's
3588 nothing that ensures that happens, is there?
3590 - ``I``: Print the letter 'i' if the operand is an integer constant, otherwise
3591 nothing. Used to print 'addi' vs 'add' instructions.
3592 - ``y``: For a memory operand, prints formatter for a two-register X-form
3593 instruction. (Currently always prints ``r0,OPERAND``).
3594 - ``U``: Prints 'u' if the memory operand is an update form, and nothing
3595 otherwise. (NOTE: LLVM does not support update form, so this will currently
3596 always print nothing)
3597 - ``X``: Prints 'x' if the memory operand is an indexed form. (NOTE: LLVM does
3598 not support indexed form, so this will currently always print nothing)
3606 SystemZ implements only ``n``, and does *not* support any of the other
3607 target-independent modifiers.
3611 - ``c``: Print an unadorned integer or symbol name. (The latter is
3612 target-specific behavior for this typically target-independent modifier).
3613 - ``A``: Print a register name with a '``*``' before it.
3614 - ``b``: Print an 8-bit register name (e.g. ``al``); do nothing on a memory
3616 - ``h``: Print the upper 8-bit register name (e.g. ``ah``); do nothing on a
3618 - ``w``: Print the 16-bit register name (e.g. ``ax``); do nothing on a memory
3620 - ``k``: Print the 32-bit register name (e.g. ``eax``); do nothing on a memory
3622 - ``q``: Print the 64-bit register name (e.g. ``rax``), if 64-bit registers are
3623 available, otherwise the 32-bit register name; do nothing on a memory operand.
3624 - ``n``: Negate and print an unadorned integer, or, for operands other than an
3625 immediate integer (e.g. a relocatable symbol expression), print a '-' before
3626 the operand. (The behavior for relocatable symbol expressions is a
3627 target-specific behavior for this typically target-independent modifier)
3628 - ``H``: Print a memory reference with additional offset +8.
3629 - ``P``: Print a memory reference or operand for use as the argument of a call
3630 instruction. (E.g. omit ``(rip)``, even though it's PC-relative.)
3634 No additional modifiers.
3640 The call instructions that wrap inline asm nodes may have a
3641 "``!srcloc``" MDNode attached to it that contains a list of constant
3642 integers. If present, the code generator will use the integer as the
3643 location cookie value when report errors through the ``LLVMContext``
3644 error reporting mechanisms. This allows a front-end to correlate backend
3645 errors that occur with inline asm back to the source code that produced
3648 .. code-block:: llvm
3650 call void asm sideeffect "something bad", ""(), !srcloc !42
3652 !42 = !{ i32 1234567 }
3654 It is up to the front-end to make sense of the magic numbers it places
3655 in the IR. If the MDNode contains multiple constants, the code generator
3656 will use the one that corresponds to the line of the asm that the error
3664 LLVM IR allows metadata to be attached to instructions in the program
3665 that can convey extra information about the code to the optimizers and
3666 code generator. One example application of metadata is source-level
3667 debug information. There are two metadata primitives: strings and nodes.
3669 Metadata does not have a type, and is not a value. If referenced from a
3670 ``call`` instruction, it uses the ``metadata`` type.
3672 All metadata are identified in syntax by a exclamation point ('``!``').
3674 .. _metadata-string:
3676 Metadata Nodes and Metadata Strings
3677 -----------------------------------
3679 A metadata string is a string surrounded by double quotes. It can
3680 contain any character by escaping non-printable characters with
3681 "``\xx``" where "``xx``" is the two digit hex code. For example:
3684 Metadata nodes are represented with notation similar to structure
3685 constants (a comma separated list of elements, surrounded by braces and
3686 preceded by an exclamation point). Metadata nodes can have any values as
3687 their operand. For example:
3689 .. code-block:: llvm
3691 !{ !"test\00", i32 10}
3693 Metadata nodes that aren't uniqued use the ``distinct`` keyword. For example:
3695 .. code-block:: llvm
3697 !0 = distinct !{!"test\00", i32 10}
3699 ``distinct`` nodes are useful when nodes shouldn't be merged based on their
3700 content. They can also occur when transformations cause uniquing collisions
3701 when metadata operands change.
3703 A :ref:`named metadata <namedmetadatastructure>` is a collection of
3704 metadata nodes, which can be looked up in the module symbol table. For
3707 .. code-block:: llvm
3711 Metadata can be used as function arguments. Here ``llvm.dbg.value``
3712 function is using two metadata arguments:
3714 .. code-block:: llvm
3716 call void @llvm.dbg.value(metadata !24, i64 0, metadata !25)
3718 Metadata can be attached to an instruction. Here metadata ``!21`` is attached
3719 to the ``add`` instruction using the ``!dbg`` identifier:
3721 .. code-block:: llvm
3723 %indvar.next = add i64 %indvar, 1, !dbg !21
3725 Metadata can also be attached to a function definition. Here metadata ``!22``
3726 is attached to the ``foo`` function using the ``!dbg`` identifier:
3728 .. code-block:: llvm
3730 define void @foo() !dbg !22 {
3734 More information about specific metadata nodes recognized by the
3735 optimizers and code generator is found below.
3737 .. _specialized-metadata:
3739 Specialized Metadata Nodes
3740 ^^^^^^^^^^^^^^^^^^^^^^^^^^
3742 Specialized metadata nodes are custom data structures in metadata (as opposed
3743 to generic tuples). Their fields are labelled, and can be specified in any
3746 These aren't inherently debug info centric, but currently all the specialized
3747 metadata nodes are related to debug info.
3754 ``DICompileUnit`` nodes represent a compile unit. The ``enums:``,
3755 ``retainedTypes:``, ``subprograms:``, ``globals:``, ``imports:`` and ``macros:``
3756 fields are tuples containing the debug info to be emitted along with the compile
3757 unit, regardless of code optimizations (some nodes are only emitted if there are
3758 references to them from instructions).
3760 .. code-block:: llvm
3762 !0 = !DICompileUnit(language: DW_LANG_C99, file: !1, producer: "clang",
3763 isOptimized: true, flags: "-O2", runtimeVersion: 2,
3764 splitDebugFilename: "abc.debug", emissionKind: 1,
3765 enums: !2, retainedTypes: !3, subprograms: !4,
3766 globals: !5, imports: !6, macros: !7, dwoId: 0x0abcd)
3768 Compile unit descriptors provide the root scope for objects declared in a
3769 specific compilation unit. File descriptors are defined using this scope.
3770 These descriptors are collected by a named metadata ``!llvm.dbg.cu``. They
3771 keep track of subprograms, global variables, type information, and imported
3772 entities (declarations and namespaces).
3779 ``DIFile`` nodes represent files. The ``filename:`` can include slashes.
3781 .. code-block:: llvm
3783 !0 = !DIFile(filename: "path/to/file", directory: "/path/to/dir")
3785 Files are sometimes used in ``scope:`` fields, and are the only valid target
3786 for ``file:`` fields.
3793 ``DIBasicType`` nodes represent primitive types, such as ``int``, ``bool`` and
3794 ``float``. ``tag:`` defaults to ``DW_TAG_base_type``.
3796 .. code-block:: llvm
3798 !0 = !DIBasicType(name: "unsigned char", size: 8, align: 8,
3799 encoding: DW_ATE_unsigned_char)
3800 !1 = !DIBasicType(tag: DW_TAG_unspecified_type, name: "decltype(nullptr)")
3802 The ``encoding:`` describes the details of the type. Usually it's one of the
3805 .. code-block:: llvm
3811 DW_ATE_signed_char = 6
3813 DW_ATE_unsigned_char = 8
3815 .. _DISubroutineType:
3820 ``DISubroutineType`` nodes represent subroutine types. Their ``types:`` field
3821 refers to a tuple; the first operand is the return type, while the rest are the
3822 types of the formal arguments in order. If the first operand is ``null``, that
3823 represents a function with no return value (such as ``void foo() {}`` in C++).
3825 .. code-block:: llvm
3827 !0 = !BasicType(name: "int", size: 32, align: 32, DW_ATE_signed)
3828 !1 = !BasicType(name: "char", size: 8, align: 8, DW_ATE_signed_char)
3829 !2 = !DISubroutineType(types: !{null, !0, !1}) ; void (int, char)
3836 ``DIDerivedType`` nodes represent types derived from other types, such as
3839 .. code-block:: llvm
3841 !0 = !DIBasicType(name: "unsigned char", size: 8, align: 8,
3842 encoding: DW_ATE_unsigned_char)
3843 !1 = !DIDerivedType(tag: DW_TAG_pointer_type, baseType: !0, size: 32,
3846 The following ``tag:`` values are valid:
3848 .. code-block:: llvm
3850 DW_TAG_formal_parameter = 5
3852 DW_TAG_pointer_type = 15
3853 DW_TAG_reference_type = 16
3855 DW_TAG_ptr_to_member_type = 31
3856 DW_TAG_const_type = 38
3857 DW_TAG_volatile_type = 53
3858 DW_TAG_restrict_type = 55
3860 ``DW_TAG_member`` is used to define a member of a :ref:`composite type
3861 <DICompositeType>` or :ref:`subprogram <DISubprogram>`. The type of the member
3862 is the ``baseType:``. The ``offset:`` is the member's bit offset.
3863 ``DW_TAG_formal_parameter`` is used to define a member which is a formal
3864 argument of a subprogram.
3866 ``DW_TAG_typedef`` is used to provide a name for the ``baseType:``.
3868 ``DW_TAG_pointer_type``, ``DW_TAG_reference_type``, ``DW_TAG_const_type``,
3869 ``DW_TAG_volatile_type`` and ``DW_TAG_restrict_type`` are used to qualify the
3872 Note that the ``void *`` type is expressed as a type derived from NULL.
3874 .. _DICompositeType:
3879 ``DICompositeType`` nodes represent types composed of other types, like
3880 structures and unions. ``elements:`` points to a tuple of the composed types.
3882 If the source language supports ODR, the ``identifier:`` field gives the unique
3883 identifier used for type merging between modules. When specified, other types
3884 can refer to composite types indirectly via a :ref:`metadata string
3885 <metadata-string>` that matches their identifier.
3887 .. code-block:: llvm
3889 !0 = !DIEnumerator(name: "SixKind", value: 7)
3890 !1 = !DIEnumerator(name: "SevenKind", value: 7)
3891 !2 = !DIEnumerator(name: "NegEightKind", value: -8)
3892 !3 = !DICompositeType(tag: DW_TAG_enumeration_type, name: "Enum", file: !12,
3893 line: 2, size: 32, align: 32, identifier: "_M4Enum",
3894 elements: !{!0, !1, !2})
3896 The following ``tag:`` values are valid:
3898 .. code-block:: llvm
3900 DW_TAG_array_type = 1
3901 DW_TAG_class_type = 2
3902 DW_TAG_enumeration_type = 4
3903 DW_TAG_structure_type = 19
3904 DW_TAG_union_type = 23
3905 DW_TAG_subroutine_type = 21
3906 DW_TAG_inheritance = 28
3909 For ``DW_TAG_array_type``, the ``elements:`` should be :ref:`subrange
3910 descriptors <DISubrange>`, each representing the range of subscripts at that
3911 level of indexing. The ``DIFlagVector`` flag to ``flags:`` indicates that an
3912 array type is a native packed vector.
3914 For ``DW_TAG_enumeration_type``, the ``elements:`` should be :ref:`enumerator
3915 descriptors <DIEnumerator>`, each representing the definition of an enumeration
3916 value for the set. All enumeration type descriptors are collected in the
3917 ``enums:`` field of the :ref:`compile unit <DICompileUnit>`.
3919 For ``DW_TAG_structure_type``, ``DW_TAG_class_type``, and
3920 ``DW_TAG_union_type``, the ``elements:`` should be :ref:`derived types
3921 <DIDerivedType>` with ``tag: DW_TAG_member`` or ``tag: DW_TAG_inheritance``.
3928 ``DISubrange`` nodes are the elements for ``DW_TAG_array_type`` variants of
3929 :ref:`DICompositeType`. ``count: -1`` indicates an empty array.
3931 .. code-block:: llvm
3933 !0 = !DISubrange(count: 5, lowerBound: 0) ; array counting from 0
3934 !1 = !DISubrange(count: 5, lowerBound: 1) ; array counting from 1
3935 !2 = !DISubrange(count: -1) ; empty array.
3942 ``DIEnumerator`` nodes are the elements for ``DW_TAG_enumeration_type``
3943 variants of :ref:`DICompositeType`.
3945 .. code-block:: llvm
3947 !0 = !DIEnumerator(name: "SixKind", value: 7)
3948 !1 = !DIEnumerator(name: "SevenKind", value: 7)
3949 !2 = !DIEnumerator(name: "NegEightKind", value: -8)
3951 DITemplateTypeParameter
3952 """""""""""""""""""""""
3954 ``DITemplateTypeParameter`` nodes represent type parameters to generic source
3955 language constructs. They are used (optionally) in :ref:`DICompositeType` and
3956 :ref:`DISubprogram` ``templateParams:`` fields.
3958 .. code-block:: llvm
3960 !0 = !DITemplateTypeParameter(name: "Ty", type: !1)
3962 DITemplateValueParameter
3963 """"""""""""""""""""""""
3965 ``DITemplateValueParameter`` nodes represent value parameters to generic source
3966 language constructs. ``tag:`` defaults to ``DW_TAG_template_value_parameter``,
3967 but if specified can also be set to ``DW_TAG_GNU_template_template_param`` or
3968 ``DW_TAG_GNU_template_param_pack``. They are used (optionally) in
3969 :ref:`DICompositeType` and :ref:`DISubprogram` ``templateParams:`` fields.
3971 .. code-block:: llvm
3973 !0 = !DITemplateValueParameter(name: "Ty", type: !1, value: i32 7)
3978 ``DINamespace`` nodes represent namespaces in the source language.
3980 .. code-block:: llvm
3982 !0 = !DINamespace(name: "myawesomeproject", scope: !1, file: !2, line: 7)
3987 ``DIGlobalVariable`` nodes represent global variables in the source language.
3989 .. code-block:: llvm
3991 !0 = !DIGlobalVariable(name: "foo", linkageName: "foo", scope: !1,
3992 file: !2, line: 7, type: !3, isLocal: true,
3993 isDefinition: false, variable: i32* @foo,
3996 All global variables should be referenced by the `globals:` field of a
3997 :ref:`compile unit <DICompileUnit>`.
4004 ``DISubprogram`` nodes represent functions from the source language. A
4005 ``DISubprogram`` may be attached to a function definition using ``!dbg``
4006 metadata. The ``variables:`` field points at :ref:`variables <DILocalVariable>`
4007 that must be retained, even if their IR counterparts are optimized out of
4008 the IR. The ``type:`` field must point at an :ref:`DISubroutineType`.
4010 .. code-block:: llvm
4012 define void @_Z3foov() !dbg !0 {
4016 !0 = distinct !DISubprogram(name: "foo", linkageName: "_Zfoov", scope: !1,
4017 file: !2, line: 7, type: !3, isLocal: true,
4018 isDefinition: false, scopeLine: 8,
4020 virtuality: DW_VIRTUALITY_pure_virtual,
4021 virtualIndex: 10, flags: DIFlagPrototyped,
4022 isOptimized: true, templateParams: !5,
4023 declaration: !6, variables: !7)
4030 ``DILexicalBlock`` nodes describe nested blocks within a :ref:`subprogram
4031 <DISubprogram>`. The line number and column numbers are used to distinguish
4032 two lexical blocks at same depth. They are valid targets for ``scope:``
4035 .. code-block:: llvm
4037 !0 = distinct !DILexicalBlock(scope: !1, file: !2, line: 7, column: 35)
4039 Usually lexical blocks are ``distinct`` to prevent node merging based on
4042 .. _DILexicalBlockFile:
4047 ``DILexicalBlockFile`` nodes are used to discriminate between sections of a
4048 :ref:`lexical block <DILexicalBlock>`. The ``file:`` field can be changed to
4049 indicate textual inclusion, or the ``discriminator:`` field can be used to
4050 discriminate between control flow within a single block in the source language.
4052 .. code-block:: llvm
4054 !0 = !DILexicalBlock(scope: !3, file: !4, line: 7, column: 35)
4055 !1 = !DILexicalBlockFile(scope: !0, file: !4, discriminator: 0)
4056 !2 = !DILexicalBlockFile(scope: !0, file: !4, discriminator: 1)
4063 ``DILocation`` nodes represent source debug locations. The ``scope:`` field is
4064 mandatory, and points at an :ref:`DILexicalBlockFile`, an
4065 :ref:`DILexicalBlock`, or an :ref:`DISubprogram`.
4067 .. code-block:: llvm
4069 !0 = !DILocation(line: 2900, column: 42, scope: !1, inlinedAt: !2)
4071 .. _DILocalVariable:
4076 ``DILocalVariable`` nodes represent local variables in the source language. If
4077 the ``arg:`` field is set to non-zero, then this variable is a subprogram
4078 parameter, and it will be included in the ``variables:`` field of its
4079 :ref:`DISubprogram`.
4081 .. code-block:: llvm
4083 !0 = !DILocalVariable(name: "this", arg: 1, scope: !3, file: !2, line: 7,
4084 type: !3, flags: DIFlagArtificial)
4085 !1 = !DILocalVariable(name: "x", arg: 2, scope: !4, file: !2, line: 7,
4087 !2 = !DILocalVariable(name: "y", scope: !5, file: !2, line: 7, type: !3)
4092 ``DIExpression`` nodes represent DWARF expression sequences. They are used in
4093 :ref:`debug intrinsics<dbg_intrinsics>` (such as ``llvm.dbg.declare``) to
4094 describe how the referenced LLVM variable relates to the source language
4097 The current supported vocabulary is limited:
4099 - ``DW_OP_deref`` dereferences the working expression.
4100 - ``DW_OP_plus, 93`` adds ``93`` to the working expression.
4101 - ``DW_OP_bit_piece, 16, 8`` specifies the offset and size (``16`` and ``8``
4102 here, respectively) of the variable piece from the working expression.
4104 .. code-block:: llvm
4106 !0 = !DIExpression(DW_OP_deref)
4107 !1 = !DIExpression(DW_OP_plus, 3)
4108 !2 = !DIExpression(DW_OP_bit_piece, 3, 7)
4109 !3 = !DIExpression(DW_OP_deref, DW_OP_plus, 3, DW_OP_bit_piece, 3, 7)
4114 ``DIObjCProperty`` nodes represent Objective-C property nodes.
4116 .. code-block:: llvm
4118 !3 = !DIObjCProperty(name: "foo", file: !1, line: 7, setter: "setFoo",
4119 getter: "getFoo", attributes: 7, type: !2)
4124 ``DIImportedEntity`` nodes represent entities (such as modules) imported into a
4127 .. code-block:: llvm
4129 !2 = !DIImportedEntity(tag: DW_TAG_imported_module, name: "foo", scope: !0,
4130 entity: !1, line: 7)
4135 ``DIMacro`` nodes represent definition or undefinition of a macro identifiers.
4136 The ``name:`` field is the macro identifier, followed by macro parameters when
4137 definining a function-like macro, and the ``value`` field is the token-string
4138 used to expand the macro identifier.
4140 .. code-block:: llvm
4142 !2 = !DIMacro(macinfo: DW_MACINFO_define, line: 7, name: "foo(x)",
4144 !3 = !DIMacro(macinfo: DW_MACINFO_undef, line: 30, name: "foo")
4149 ``DIMacroFile`` nodes represent inclusion of source files.
4150 The ``nodes:`` field is a list of ``DIMacro`` and ``DIMacroFile`` nodes that
4151 appear in the included source file.
4153 .. code-block:: llvm
4155 !2 = !DIMacroFile(macinfo: DW_MACINFO_start_file, line: 7, file: !2,
4161 In LLVM IR, memory does not have types, so LLVM's own type system is not
4162 suitable for doing TBAA. Instead, metadata is added to the IR to
4163 describe a type system of a higher level language. This can be used to
4164 implement typical C/C++ TBAA, but it can also be used to implement
4165 custom alias analysis behavior for other languages.
4167 The current metadata format is very simple. TBAA metadata nodes have up
4168 to three fields, e.g.:
4170 .. code-block:: llvm
4172 !0 = !{ !"an example type tree" }
4173 !1 = !{ !"int", !0 }
4174 !2 = !{ !"float", !0 }
4175 !3 = !{ !"const float", !2, i64 1 }
4177 The first field is an identity field. It can be any value, usually a
4178 metadata string, which uniquely identifies the type. The most important
4179 name in the tree is the name of the root node. Two trees with different
4180 root node names are entirely disjoint, even if they have leaves with
4183 The second field identifies the type's parent node in the tree, or is
4184 null or omitted for a root node. A type is considered to alias all of
4185 its descendants and all of its ancestors in the tree. Also, a type is
4186 considered to alias all types in other trees, so that bitcode produced
4187 from multiple front-ends is handled conservatively.
4189 If the third field is present, it's an integer which if equal to 1
4190 indicates that the type is "constant" (meaning
4191 ``pointsToConstantMemory`` should return true; see `other useful
4192 AliasAnalysis methods <AliasAnalysis.html#OtherItfs>`_).
4194 '``tbaa.struct``' Metadata
4195 ^^^^^^^^^^^^^^^^^^^^^^^^^^
4197 The :ref:`llvm.memcpy <int_memcpy>` is often used to implement
4198 aggregate assignment operations in C and similar languages, however it
4199 is defined to copy a contiguous region of memory, which is more than
4200 strictly necessary for aggregate types which contain holes due to
4201 padding. Also, it doesn't contain any TBAA information about the fields
4204 ``!tbaa.struct`` metadata can describe which memory subregions in a
4205 memcpy are padding and what the TBAA tags of the struct are.
4207 The current metadata format is very simple. ``!tbaa.struct`` metadata
4208 nodes are a list of operands which are in conceptual groups of three.
4209 For each group of three, the first operand gives the byte offset of a
4210 field in bytes, the second gives its size in bytes, and the third gives
4213 .. code-block:: llvm
4215 !4 = !{ i64 0, i64 4, !1, i64 8, i64 4, !2 }
4217 This describes a struct with two fields. The first is at offset 0 bytes
4218 with size 4 bytes, and has tbaa tag !1. The second is at offset 8 bytes
4219 and has size 4 bytes and has tbaa tag !2.
4221 Note that the fields need not be contiguous. In this example, there is a
4222 4 byte gap between the two fields. This gap represents padding which
4223 does not carry useful data and need not be preserved.
4225 '``noalias``' and '``alias.scope``' Metadata
4226 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4228 ``noalias`` and ``alias.scope`` metadata provide the ability to specify generic
4229 noalias memory-access sets. This means that some collection of memory access
4230 instructions (loads, stores, memory-accessing calls, etc.) that carry
4231 ``noalias`` metadata can specifically be specified not to alias with some other
4232 collection of memory access instructions that carry ``alias.scope`` metadata.
4233 Each type of metadata specifies a list of scopes where each scope has an id and
4234 a domain. When evaluating an aliasing query, if for some domain, the set
4235 of scopes with that domain in one instruction's ``alias.scope`` list is a
4236 subset of (or equal to) the set of scopes for that domain in another
4237 instruction's ``noalias`` list, then the two memory accesses are assumed not to
4240 The metadata identifying each domain is itself a list containing one or two
4241 entries. The first entry is the name of the domain. Note that if the name is a
4242 string then it can be combined across functions and translation units. A
4243 self-reference can be used to create globally unique domain names. A
4244 descriptive string may optionally be provided as a second list entry.
4246 The metadata identifying each scope is also itself a list containing two or
4247 three entries. The first entry is the name of the scope. Note that if the name
4248 is a string then it can be combined across functions and translation units. A
4249 self-reference can be used to create globally unique scope names. A metadata
4250 reference to the scope's domain is the second entry. A descriptive string may
4251 optionally be provided as a third list entry.
4255 .. code-block:: llvm
4257 ; Two scope domains:
4261 ; Some scopes in these domains:
4267 !5 = !{!4} ; A list containing only scope !4
4271 ; These two instructions don't alias:
4272 %0 = load float, float* %c, align 4, !alias.scope !5
4273 store float %0, float* %arrayidx.i, align 4, !noalias !5
4275 ; These two instructions also don't alias (for domain !1, the set of scopes
4276 ; in the !alias.scope equals that in the !noalias list):
4277 %2 = load float, float* %c, align 4, !alias.scope !5
4278 store float %2, float* %arrayidx.i2, align 4, !noalias !6
4280 ; These two instructions may alias (for domain !0, the set of scopes in
4281 ; the !noalias list is not a superset of, or equal to, the scopes in the
4282 ; !alias.scope list):
4283 %2 = load float, float* %c, align 4, !alias.scope !6
4284 store float %0, float* %arrayidx.i, align 4, !noalias !7
4286 '``fpmath``' Metadata
4287 ^^^^^^^^^^^^^^^^^^^^^
4289 ``fpmath`` metadata may be attached to any instruction of floating point
4290 type. It can be used to express the maximum acceptable error in the
4291 result of that instruction, in ULPs, thus potentially allowing the
4292 compiler to use a more efficient but less accurate method of computing
4293 it. ULP is defined as follows:
4295 If ``x`` is a real number that lies between two finite consecutive
4296 floating-point numbers ``a`` and ``b``, without being equal to one
4297 of them, then ``ulp(x) = |b - a|``, otherwise ``ulp(x)`` is the
4298 distance between the two non-equal finite floating-point numbers
4299 nearest ``x``. Moreover, ``ulp(NaN)`` is ``NaN``.
4301 The metadata node shall consist of a single positive floating point
4302 number representing the maximum relative error, for example:
4304 .. code-block:: llvm
4306 !0 = !{ float 2.5 } ; maximum acceptable inaccuracy is 2.5 ULPs
4310 '``range``' Metadata
4311 ^^^^^^^^^^^^^^^^^^^^
4313 ``range`` metadata may be attached only to ``load``, ``call`` and ``invoke`` of
4314 integer types. It expresses the possible ranges the loaded value or the value
4315 returned by the called function at this call site is in. The ranges are
4316 represented with a flattened list of integers. The loaded value or the value
4317 returned is known to be in the union of the ranges defined by each consecutive
4318 pair. Each pair has the following properties:
4320 - The type must match the type loaded by the instruction.
4321 - The pair ``a,b`` represents the range ``[a,b)``.
4322 - Both ``a`` and ``b`` are constants.
4323 - The range is allowed to wrap.
4324 - The range should not represent the full or empty set. That is,
4327 In addition, the pairs must be in signed order of the lower bound and
4328 they must be non-contiguous.
4332 .. code-block:: llvm
4334 %a = load i8, i8* %x, align 1, !range !0 ; Can only be 0 or 1
4335 %b = load i8, i8* %y, align 1, !range !1 ; Can only be 255 (-1), 0 or 1
4336 %c = call i8 @foo(), !range !2 ; Can only be 0, 1, 3, 4 or 5
4337 %d = invoke i8 @bar() to label %cont
4338 unwind label %lpad, !range !3 ; Can only be -2, -1, 3, 4 or 5
4340 !0 = !{ i8 0, i8 2 }
4341 !1 = !{ i8 255, i8 2 }
4342 !2 = !{ i8 0, i8 2, i8 3, i8 6 }
4343 !3 = !{ i8 -2, i8 0, i8 3, i8 6 }
4345 '``unpredictable``' Metadata
4346 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4348 ``unpredictable`` metadata may be attached to any branch or switch
4349 instruction. It can be used to express the unpredictability of control
4350 flow. Similar to the llvm.expect intrinsic, it may be used to alter
4351 optimizations related to compare and branch instructions. The metadata
4352 is treated as a boolean value; if it exists, it signals that the branch
4353 or switch that it is attached to is completely unpredictable.
4358 It is sometimes useful to attach information to loop constructs. Currently,
4359 loop metadata is implemented as metadata attached to the branch instruction
4360 in the loop latch block. This type of metadata refer to a metadata node that is
4361 guaranteed to be separate for each loop. The loop identifier metadata is
4362 specified with the name ``llvm.loop``.
4364 The loop identifier metadata is implemented using a metadata that refers to
4365 itself to avoid merging it with any other identifier metadata, e.g.,
4366 during module linkage or function inlining. That is, each loop should refer
4367 to their own identification metadata even if they reside in separate functions.
4368 The following example contains loop identifier metadata for two separate loop
4371 .. code-block:: llvm
4376 The loop identifier metadata can be used to specify additional
4377 per-loop metadata. Any operands after the first operand can be treated
4378 as user-defined metadata. For example the ``llvm.loop.unroll.count``
4379 suggests an unroll factor to the loop unroller:
4381 .. code-block:: llvm
4383 br i1 %exitcond, label %._crit_edge, label %.lr.ph, !llvm.loop !0
4386 !1 = !{!"llvm.loop.unroll.count", i32 4}
4388 '``llvm.loop.vectorize``' and '``llvm.loop.interleave``'
4389 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4391 Metadata prefixed with ``llvm.loop.vectorize`` or ``llvm.loop.interleave`` are
4392 used to control per-loop vectorization and interleaving parameters such as
4393 vectorization width and interleave count. These metadata should be used in
4394 conjunction with ``llvm.loop`` loop identification metadata. The
4395 ``llvm.loop.vectorize`` and ``llvm.loop.interleave`` metadata are only
4396 optimization hints and the optimizer will only interleave and vectorize loops if
4397 it believes it is safe to do so. The ``llvm.mem.parallel_loop_access`` metadata
4398 which contains information about loop-carried memory dependencies can be helpful
4399 in determining the safety of these transformations.
4401 '``llvm.loop.interleave.count``' Metadata
4402 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4404 This metadata suggests an interleave count to the loop interleaver.
4405 The first operand is the string ``llvm.loop.interleave.count`` and the
4406 second operand is an integer specifying the interleave count. For
4409 .. code-block:: llvm
4411 !0 = !{!"llvm.loop.interleave.count", i32 4}
4413 Note that setting ``llvm.loop.interleave.count`` to 1 disables interleaving
4414 multiple iterations of the loop. If ``llvm.loop.interleave.count`` is set to 0
4415 then the interleave count will be determined automatically.
4417 '``llvm.loop.vectorize.enable``' Metadata
4418 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4420 This metadata selectively enables or disables vectorization for the loop. The
4421 first operand is the string ``llvm.loop.vectorize.enable`` and the second operand
4422 is a bit. If the bit operand value is 1 vectorization is enabled. A value of
4423 0 disables vectorization:
4425 .. code-block:: llvm
4427 !0 = !{!"llvm.loop.vectorize.enable", i1 0}
4428 !1 = !{!"llvm.loop.vectorize.enable", i1 1}
4430 '``llvm.loop.vectorize.width``' Metadata
4431 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4433 This metadata sets the target width of the vectorizer. The first
4434 operand is the string ``llvm.loop.vectorize.width`` and the second
4435 operand is an integer specifying the width. For example:
4437 .. code-block:: llvm
4439 !0 = !{!"llvm.loop.vectorize.width", i32 4}
4441 Note that setting ``llvm.loop.vectorize.width`` to 1 disables
4442 vectorization of the loop. If ``llvm.loop.vectorize.width`` is set to
4443 0 or if the loop does not have this metadata the width will be
4444 determined automatically.
4446 '``llvm.loop.unroll``'
4447 ^^^^^^^^^^^^^^^^^^^^^^
4449 Metadata prefixed with ``llvm.loop.unroll`` are loop unrolling
4450 optimization hints such as the unroll factor. ``llvm.loop.unroll``
4451 metadata should be used in conjunction with ``llvm.loop`` loop
4452 identification metadata. The ``llvm.loop.unroll`` metadata are only
4453 optimization hints and the unrolling will only be performed if the
4454 optimizer believes it is safe to do so.
4456 '``llvm.loop.unroll.count``' Metadata
4457 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4459 This metadata suggests an unroll factor to the loop unroller. The
4460 first operand is the string ``llvm.loop.unroll.count`` and the second
4461 operand is a positive integer specifying the unroll factor. For
4464 .. code-block:: llvm
4466 !0 = !{!"llvm.loop.unroll.count", i32 4}
4468 If the trip count of the loop is less than the unroll count the loop
4469 will be partially unrolled.
4471 '``llvm.loop.unroll.disable``' Metadata
4472 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4474 This metadata disables loop unrolling. The metadata has a single operand
4475 which is the string ``llvm.loop.unroll.disable``. For example:
4477 .. code-block:: llvm
4479 !0 = !{!"llvm.loop.unroll.disable"}
4481 '``llvm.loop.unroll.runtime.disable``' Metadata
4482 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4484 This metadata disables runtime loop unrolling. The metadata has a single
4485 operand which is the string ``llvm.loop.unroll.runtime.disable``. For example:
4487 .. code-block:: llvm
4489 !0 = !{!"llvm.loop.unroll.runtime.disable"}
4491 '``llvm.loop.unroll.enable``' Metadata
4492 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4494 This metadata suggests that the loop should be fully unrolled if the trip count
4495 is known at compile time and partially unrolled if the trip count is not known
4496 at compile time. The metadata has a single operand which is the string
4497 ``llvm.loop.unroll.enable``. For example:
4499 .. code-block:: llvm
4501 !0 = !{!"llvm.loop.unroll.enable"}
4503 '``llvm.loop.unroll.full``' Metadata
4504 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4506 This metadata suggests that the loop should be unrolled fully. The
4507 metadata has a single operand which is the string ``llvm.loop.unroll.full``.
4510 .. code-block:: llvm
4512 !0 = !{!"llvm.loop.unroll.full"}
4517 Metadata types used to annotate memory accesses with information helpful
4518 for optimizations are prefixed with ``llvm.mem``.
4520 '``llvm.mem.parallel_loop_access``' Metadata
4521 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4523 The ``llvm.mem.parallel_loop_access`` metadata refers to a loop identifier,
4524 or metadata containing a list of loop identifiers for nested loops.
4525 The metadata is attached to memory accessing instructions and denotes that
4526 no loop carried memory dependence exist between it and other instructions denoted
4527 with the same loop identifier.
4529 Precisely, given two instructions ``m1`` and ``m2`` that both have the
4530 ``llvm.mem.parallel_loop_access`` metadata, with ``L1`` and ``L2`` being the
4531 set of loops associated with that metadata, respectively, then there is no loop
4532 carried dependence between ``m1`` and ``m2`` for loops in both ``L1`` and
4535 As a special case, if all memory accessing instructions in a loop have
4536 ``llvm.mem.parallel_loop_access`` metadata that refers to that loop, then the
4537 loop has no loop carried memory dependences and is considered to be a parallel
4540 Note that if not all memory access instructions have such metadata referring to
4541 the loop, then the loop is considered not being trivially parallel. Additional
4542 memory dependence analysis is required to make that determination. As a fail
4543 safe mechanism, this causes loops that were originally parallel to be considered
4544 sequential (if optimization passes that are unaware of the parallel semantics
4545 insert new memory instructions into the loop body).
4547 Example of a loop that is considered parallel due to its correct use of
4548 both ``llvm.loop`` and ``llvm.mem.parallel_loop_access``
4549 metadata types that refer to the same loop identifier metadata.
4551 .. code-block:: llvm
4555 %val0 = load i32, i32* %arrayidx, !llvm.mem.parallel_loop_access !0
4557 store i32 %val0, i32* %arrayidx1, !llvm.mem.parallel_loop_access !0
4559 br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !0
4565 It is also possible to have nested parallel loops. In that case the
4566 memory accesses refer to a list of loop identifier metadata nodes instead of
4567 the loop identifier metadata node directly:
4569 .. code-block:: llvm
4573 %val1 = load i32, i32* %arrayidx3, !llvm.mem.parallel_loop_access !2
4575 br label %inner.for.body
4579 %val0 = load i32, i32* %arrayidx1, !llvm.mem.parallel_loop_access !0
4581 store i32 %val0, i32* %arrayidx2, !llvm.mem.parallel_loop_access !0
4583 br i1 %exitcond, label %inner.for.end, label %inner.for.body, !llvm.loop !1
4587 store i32 %val1, i32* %arrayidx4, !llvm.mem.parallel_loop_access !2
4589 br i1 %exitcond, label %outer.for.end, label %outer.for.body, !llvm.loop !2
4591 outer.for.end: ; preds = %for.body
4593 !0 = !{!1, !2} ; a list of loop identifiers
4594 !1 = !{!1} ; an identifier for the inner loop
4595 !2 = !{!2} ; an identifier for the outer loop
4600 The ``llvm.bitsets`` global metadata is used to implement
4601 :doc:`bitsets <BitSets>`.
4603 '``invariant.group``' Metadata
4604 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4606 The ``invariant.group`` metadata may be attached to ``load``/``store`` instructions.
4607 The existence of the ``invariant.group`` metadata on the instruction tells
4608 the optimizer that every ``load`` and ``store`` to the same pointer operand
4609 within the same invariant group can be assumed to load or store the same
4610 value (but see the ``llvm.invariant.group.barrier`` intrinsic which affects
4611 when two pointers are considered the same).
4615 .. code-block:: llvm
4617 @unknownPtr = external global i8
4620 store i8 42, i8* %ptr, !invariant.group !0
4621 call void @foo(i8* %ptr)
4623 %a = load i8, i8* %ptr, !invariant.group !0 ; Can assume that value under %ptr didn't change
4624 call void @foo(i8* %ptr)
4625 %b = load i8, i8* %ptr, !invariant.group !1 ; Can't assume anything, because group changed
4627 %newPtr = call i8* @getPointer(i8* %ptr)
4628 %c = load i8, i8* %newPtr, !invariant.group !0 ; Can't assume anything, because we only have information about %ptr
4630 %unknownValue = load i8, i8* @unknownPtr
4631 store i8 %unknownValue, i8* %ptr, !invariant.group !0 ; Can assume that %unknownValue == 42
4633 call void @foo(i8* %ptr)
4634 %newPtr2 = call i8* @llvm.invariant.group.barrier(i8* %ptr)
4635 %d = load i8, i8* %newPtr2, !invariant.group !0 ; Can't step through invariant.group.barrier to get value of %ptr
4638 declare void @foo(i8*)
4639 declare i8* @getPointer(i8*)
4640 declare i8* @llvm.invariant.group.barrier(i8*)
4642 !0 = !{!"magic ptr"}
4643 !1 = !{!"other ptr"}
4647 Module Flags Metadata
4648 =====================
4650 Information about the module as a whole is difficult to convey to LLVM's
4651 subsystems. The LLVM IR isn't sufficient to transmit this information.
4652 The ``llvm.module.flags`` named metadata exists in order to facilitate
4653 this. These flags are in the form of key / value pairs --- much like a
4654 dictionary --- making it easy for any subsystem who cares about a flag to
4657 The ``llvm.module.flags`` metadata contains a list of metadata triplets.
4658 Each triplet has the following form:
4660 - The first element is a *behavior* flag, which specifies the behavior
4661 when two (or more) modules are merged together, and it encounters two
4662 (or more) metadata with the same ID. The supported behaviors are
4664 - The second element is a metadata string that is a unique ID for the
4665 metadata. Each module may only have one flag entry for each unique ID (not
4666 including entries with the **Require** behavior).
4667 - The third element is the value of the flag.
4669 When two (or more) modules are merged together, the resulting
4670 ``llvm.module.flags`` metadata is the union of the modules' flags. That is, for
4671 each unique metadata ID string, there will be exactly one entry in the merged
4672 modules ``llvm.module.flags`` metadata table, and the value for that entry will
4673 be determined by the merge behavior flag, as described below. The only exception
4674 is that entries with the *Require* behavior are always preserved.
4676 The following behaviors are supported:
4687 Emits an error if two values disagree, otherwise the resulting value
4688 is that of the operands.
4692 Emits a warning if two values disagree. The result value will be the
4693 operand for the flag from the first module being linked.
4697 Adds a requirement that another module flag be present and have a
4698 specified value after linking is performed. The value must be a
4699 metadata pair, where the first element of the pair is the ID of the
4700 module flag to be restricted, and the second element of the pair is
4701 the value the module flag should be restricted to. This behavior can
4702 be used to restrict the allowable results (via triggering of an
4703 error) of linking IDs with the **Override** behavior.
4707 Uses the specified value, regardless of the behavior or value of the
4708 other module. If both modules specify **Override**, but the values
4709 differ, an error will be emitted.
4713 Appends the two values, which are required to be metadata nodes.
4717 Appends the two values, which are required to be metadata
4718 nodes. However, duplicate entries in the second list are dropped
4719 during the append operation.
4721 It is an error for a particular unique flag ID to have multiple behaviors,
4722 except in the case of **Require** (which adds restrictions on another metadata
4723 value) or **Override**.
4725 An example of module flags:
4727 .. code-block:: llvm
4729 !0 = !{ i32 1, !"foo", i32 1 }
4730 !1 = !{ i32 4, !"bar", i32 37 }
4731 !2 = !{ i32 2, !"qux", i32 42 }
4732 !3 = !{ i32 3, !"qux",
4737 !llvm.module.flags = !{ !0, !1, !2, !3 }
4739 - Metadata ``!0`` has the ID ``!"foo"`` and the value '1'. The behavior
4740 if two or more ``!"foo"`` flags are seen is to emit an error if their
4741 values are not equal.
4743 - Metadata ``!1`` has the ID ``!"bar"`` and the value '37'. The
4744 behavior if two or more ``!"bar"`` flags are seen is to use the value
4747 - Metadata ``!2`` has the ID ``!"qux"`` and the value '42'. The
4748 behavior if two or more ``!"qux"`` flags are seen is to emit a
4749 warning if their values are not equal.
4751 - Metadata ``!3`` has the ID ``!"qux"`` and the value:
4757 The behavior is to emit an error if the ``llvm.module.flags`` does not
4758 contain a flag with the ID ``!"foo"`` that has the value '1' after linking is
4761 Objective-C Garbage Collection Module Flags Metadata
4762 ----------------------------------------------------
4764 On the Mach-O platform, Objective-C stores metadata about garbage
4765 collection in a special section called "image info". The metadata
4766 consists of a version number and a bitmask specifying what types of
4767 garbage collection are supported (if any) by the file. If two or more
4768 modules are linked together their garbage collection metadata needs to
4769 be merged rather than appended together.
4771 The Objective-C garbage collection module flags metadata consists of the
4772 following key-value pairs:
4781 * - ``Objective-C Version``
4782 - **[Required]** --- The Objective-C ABI version. Valid values are 1 and 2.
4784 * - ``Objective-C Image Info Version``
4785 - **[Required]** --- The version of the image info section. Currently
4788 * - ``Objective-C Image Info Section``
4789 - **[Required]** --- The section to place the metadata. Valid values are
4790 ``"__OBJC, __image_info, regular"`` for Objective-C ABI version 1, and
4791 ``"__DATA,__objc_imageinfo, regular, no_dead_strip"`` for
4792 Objective-C ABI version 2.
4794 * - ``Objective-C Garbage Collection``
4795 - **[Required]** --- Specifies whether garbage collection is supported or
4796 not. Valid values are 0, for no garbage collection, and 2, for garbage
4797 collection supported.
4799 * - ``Objective-C GC Only``
4800 - **[Optional]** --- Specifies that only garbage collection is supported.
4801 If present, its value must be 6. This flag requires that the
4802 ``Objective-C Garbage Collection`` flag have the value 2.
4804 Some important flag interactions:
4806 - If a module with ``Objective-C Garbage Collection`` set to 0 is
4807 merged with a module with ``Objective-C Garbage Collection`` set to
4808 2, then the resulting module has the
4809 ``Objective-C Garbage Collection`` flag set to 0.
4810 - A module with ``Objective-C Garbage Collection`` set to 0 cannot be
4811 merged with a module with ``Objective-C GC Only`` set to 6.
4813 Automatic Linker Flags Module Flags Metadata
4814 --------------------------------------------
4816 Some targets support embedding flags to the linker inside individual object
4817 files. Typically this is used in conjunction with language extensions which
4818 allow source files to explicitly declare the libraries they depend on, and have
4819 these automatically be transmitted to the linker via object files.
4821 These flags are encoded in the IR using metadata in the module flags section,
4822 using the ``Linker Options`` key. The merge behavior for this flag is required
4823 to be ``AppendUnique``, and the value for the key is expected to be a metadata
4824 node which should be a list of other metadata nodes, each of which should be a
4825 list of metadata strings defining linker options.
4827 For example, the following metadata section specifies two separate sets of
4828 linker options, presumably to link against ``libz`` and the ``Cocoa``
4831 !0 = !{ i32 6, !"Linker Options",
4834 !{ !"-framework", !"Cocoa" } } }
4835 !llvm.module.flags = !{ !0 }
4837 The metadata encoding as lists of lists of options, as opposed to a collapsed
4838 list of options, is chosen so that the IR encoding can use multiple option
4839 strings to specify e.g., a single library, while still having that specifier be
4840 preserved as an atomic element that can be recognized by a target specific
4841 assembly writer or object file emitter.
4843 Each individual option is required to be either a valid option for the target's
4844 linker, or an option that is reserved by the target specific assembly writer or
4845 object file emitter. No other aspect of these options is defined by the IR.
4847 C type width Module Flags Metadata
4848 ----------------------------------
4850 The ARM backend emits a section into each generated object file describing the
4851 options that it was compiled with (in a compiler-independent way) to prevent
4852 linking incompatible objects, and to allow automatic library selection. Some
4853 of these options are not visible at the IR level, namely wchar_t width and enum
4856 To pass this information to the backend, these options are encoded in module
4857 flags metadata, using the following key-value pairs:
4867 - * 0 --- sizeof(wchar_t) == 4
4868 * 1 --- sizeof(wchar_t) == 2
4871 - * 0 --- Enums are at least as large as an ``int``.
4872 * 1 --- Enums are stored in the smallest integer type which can
4873 represent all of its values.
4875 For example, the following metadata section specifies that the module was
4876 compiled with a ``wchar_t`` width of 4 bytes, and the underlying type of an
4877 enum is the smallest type which can represent all of its values::
4879 !llvm.module.flags = !{!0, !1}
4880 !0 = !{i32 1, !"short_wchar", i32 1}
4881 !1 = !{i32 1, !"short_enum", i32 0}
4883 .. _intrinsicglobalvariables:
4885 Intrinsic Global Variables
4886 ==========================
4888 LLVM has a number of "magic" global variables that contain data that
4889 affect code generation or other IR semantics. These are documented here.
4890 All globals of this sort should have a section specified as
4891 "``llvm.metadata``". This section and all globals that start with
4892 "``llvm.``" are reserved for use by LLVM.
4896 The '``llvm.used``' Global Variable
4897 -----------------------------------
4899 The ``@llvm.used`` global is an array which has
4900 :ref:`appending linkage <linkage_appending>`. This array contains a list of
4901 pointers to named global variables, functions and aliases which may optionally
4902 have a pointer cast formed of bitcast or getelementptr. For example, a legal
4905 .. code-block:: llvm
4910 @llvm.used = appending global [2 x i8*] [
4912 i8* bitcast (i32* @Y to i8*)
4913 ], section "llvm.metadata"
4915 If a symbol appears in the ``@llvm.used`` list, then the compiler, assembler,
4916 and linker are required to treat the symbol as if there is a reference to the
4917 symbol that it cannot see (which is why they have to be named). For example, if
4918 a variable has internal linkage and no references other than that from the
4919 ``@llvm.used`` list, it cannot be deleted. This is commonly used to represent
4920 references from inline asms and other things the compiler cannot "see", and
4921 corresponds to "``attribute((used))``" in GNU C.
4923 On some targets, the code generator must emit a directive to the
4924 assembler or object file to prevent the assembler and linker from
4925 molesting the symbol.
4927 .. _gv_llvmcompilerused:
4929 The '``llvm.compiler.used``' Global Variable
4930 --------------------------------------------
4932 The ``@llvm.compiler.used`` directive is the same as the ``@llvm.used``
4933 directive, except that it only prevents the compiler from touching the
4934 symbol. On targets that support it, this allows an intelligent linker to
4935 optimize references to the symbol without being impeded as it would be
4938 This is a rare construct that should only be used in rare circumstances,
4939 and should not be exposed to source languages.
4941 .. _gv_llvmglobalctors:
4943 The '``llvm.global_ctors``' Global Variable
4944 -------------------------------------------
4946 .. code-block:: llvm
4948 %0 = type { i32, void ()*, i8* }
4949 @llvm.global_ctors = appending global [1 x %0] [%0 { i32 65535, void ()* @ctor, i8* @data }]
4951 The ``@llvm.global_ctors`` array contains a list of constructor
4952 functions, priorities, and an optional associated global or function.
4953 The functions referenced by this array will be called in ascending order
4954 of priority (i.e. lowest first) when the module is loaded. The order of
4955 functions with the same priority is not defined.
4957 If the third field is present, non-null, and points to a global variable
4958 or function, the initializer function will only run if the associated
4959 data from the current module is not discarded.
4961 .. _llvmglobaldtors:
4963 The '``llvm.global_dtors``' Global Variable
4964 -------------------------------------------
4966 .. code-block:: llvm
4968 %0 = type { i32, void ()*, i8* }
4969 @llvm.global_dtors = appending global [1 x %0] [%0 { i32 65535, void ()* @dtor, i8* @data }]
4971 The ``@llvm.global_dtors`` array contains a list of destructor
4972 functions, priorities, and an optional associated global or function.
4973 The functions referenced by this array will be called in descending
4974 order of priority (i.e. highest first) when the module is unloaded. The
4975 order of functions with the same priority is not defined.
4977 If the third field is present, non-null, and points to a global variable
4978 or function, the destructor function will only run if the associated
4979 data from the current module is not discarded.
4981 Instruction Reference
4982 =====================
4984 The LLVM instruction set consists of several different classifications
4985 of instructions: :ref:`terminator instructions <terminators>`, :ref:`binary
4986 instructions <binaryops>`, :ref:`bitwise binary
4987 instructions <bitwiseops>`, :ref:`memory instructions <memoryops>`, and
4988 :ref:`other instructions <otherops>`.
4992 Terminator Instructions
4993 -----------------------
4995 As mentioned :ref:`previously <functionstructure>`, every basic block in a
4996 program ends with a "Terminator" instruction, which indicates which
4997 block should be executed after the current block is finished. These
4998 terminator instructions typically yield a '``void``' value: they produce
4999 control flow, not values (the one exception being the
5000 ':ref:`invoke <i_invoke>`' instruction).
5002 The terminator instructions are: ':ref:`ret <i_ret>`',
5003 ':ref:`br <i_br>`', ':ref:`switch <i_switch>`',
5004 ':ref:`indirectbr <i_indirectbr>`', ':ref:`invoke <i_invoke>`',
5005 ':ref:`resume <i_resume>`', ':ref:`catchswitch <i_catchswitch>`',
5006 ':ref:`catchret <i_catchret>`',
5007 ':ref:`cleanupret <i_cleanupret>`',
5008 and ':ref:`unreachable <i_unreachable>`'.
5012 '``ret``' Instruction
5013 ^^^^^^^^^^^^^^^^^^^^^
5020 ret <type> <value> ; Return a value from a non-void function
5021 ret void ; Return from void function
5026 The '``ret``' instruction is used to return control flow (and optionally
5027 a value) from a function back to the caller.
5029 There are two forms of the '``ret``' instruction: one that returns a
5030 value and then causes control flow, and one that just causes control
5036 The '``ret``' instruction optionally accepts a single argument, the
5037 return value. The type of the return value must be a ':ref:`first
5038 class <t_firstclass>`' type.
5040 A function is not :ref:`well formed <wellformed>` if it it has a non-void
5041 return type and contains a '``ret``' instruction with no return value or
5042 a return value with a type that does not match its type, or if it has a
5043 void return type and contains a '``ret``' instruction with a return
5049 When the '``ret``' instruction is executed, control flow returns back to
5050 the calling function's context. If the caller is a
5051 ":ref:`call <i_call>`" instruction, execution continues at the
5052 instruction after the call. If the caller was an
5053 ":ref:`invoke <i_invoke>`" instruction, execution continues at the
5054 beginning of the "normal" destination block. If the instruction returns
5055 a value, that value shall set the call or invoke instruction's return
5061 .. code-block:: llvm
5063 ret i32 5 ; Return an integer value of 5
5064 ret void ; Return from a void function
5065 ret { i32, i8 } { i32 4, i8 2 } ; Return a struct of values 4 and 2
5069 '``br``' Instruction
5070 ^^^^^^^^^^^^^^^^^^^^
5077 br i1 <cond>, label <iftrue>, label <iffalse>
5078 br label <dest> ; Unconditional branch
5083 The '``br``' instruction is used to cause control flow to transfer to a
5084 different basic block in the current function. There are two forms of
5085 this instruction, corresponding to a conditional branch and an
5086 unconditional branch.
5091 The conditional branch form of the '``br``' instruction takes a single
5092 '``i1``' value and two '``label``' values. The unconditional form of the
5093 '``br``' instruction takes a single '``label``' value as a target.
5098 Upon execution of a conditional '``br``' instruction, the '``i1``'
5099 argument is evaluated. If the value is ``true``, control flows to the
5100 '``iftrue``' ``label`` argument. If "cond" is ``false``, control flows
5101 to the '``iffalse``' ``label`` argument.
5106 .. code-block:: llvm
5109 %cond = icmp eq i32 %a, %b
5110 br i1 %cond, label %IfEqual, label %IfUnequal
5118 '``switch``' Instruction
5119 ^^^^^^^^^^^^^^^^^^^^^^^^
5126 switch <intty> <value>, label <defaultdest> [ <intty> <val>, label <dest> ... ]
5131 The '``switch``' instruction is used to transfer control flow to one of
5132 several different places. It is a generalization of the '``br``'
5133 instruction, allowing a branch to occur to one of many possible
5139 The '``switch``' instruction uses three parameters: an integer
5140 comparison value '``value``', a default '``label``' destination, and an
5141 array of pairs of comparison value constants and '``label``'s. The table
5142 is not allowed to contain duplicate constant entries.
5147 The ``switch`` instruction specifies a table of values and destinations.
5148 When the '``switch``' instruction is executed, this table is searched
5149 for the given value. If the value is found, control flow is transferred
5150 to the corresponding destination; otherwise, control flow is transferred
5151 to the default destination.
5156 Depending on properties of the target machine and the particular
5157 ``switch`` instruction, this instruction may be code generated in
5158 different ways. For example, it could be generated as a series of
5159 chained conditional branches or with a lookup table.
5164 .. code-block:: llvm
5166 ; Emulate a conditional br instruction
5167 %Val = zext i1 %value to i32
5168 switch i32 %Val, label %truedest [ i32 0, label %falsedest ]
5170 ; Emulate an unconditional br instruction
5171 switch i32 0, label %dest [ ]
5173 ; Implement a jump table:
5174 switch i32 %val, label %otherwise [ i32 0, label %onzero
5176 i32 2, label %ontwo ]
5180 '``indirectbr``' Instruction
5181 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5188 indirectbr <somety>* <address>, [ label <dest1>, label <dest2>, ... ]
5193 The '``indirectbr``' instruction implements an indirect branch to a
5194 label within the current function, whose address is specified by
5195 "``address``". Address must be derived from a
5196 :ref:`blockaddress <blockaddress>` constant.
5201 The '``address``' argument is the address of the label to jump to. The
5202 rest of the arguments indicate the full set of possible destinations
5203 that the address may point to. Blocks are allowed to occur multiple
5204 times in the destination list, though this isn't particularly useful.
5206 This destination list is required so that dataflow analysis has an
5207 accurate understanding of the CFG.
5212 Control transfers to the block specified in the address argument. All
5213 possible destination blocks must be listed in the label list, otherwise
5214 this instruction has undefined behavior. This implies that jumps to
5215 labels defined in other functions have undefined behavior as well.
5220 This is typically implemented with a jump through a register.
5225 .. code-block:: llvm
5227 indirectbr i8* %Addr, [ label %bb1, label %bb2, label %bb3 ]
5231 '``invoke``' Instruction
5232 ^^^^^^^^^^^^^^^^^^^^^^^^
5239 <result> = invoke [cconv] [ret attrs] <ptr to function ty> <function ptr val>(<function args>) [fn attrs]
5240 [operand bundles] to label <normal label> unwind label <exception label>
5245 The '``invoke``' instruction causes control to transfer to a specified
5246 function, with the possibility of control flow transfer to either the
5247 '``normal``' label or the '``exception``' label. If the callee function
5248 returns with the "``ret``" instruction, control flow will return to the
5249 "normal" label. If the callee (or any indirect callees) returns via the
5250 ":ref:`resume <i_resume>`" instruction or other exception handling
5251 mechanism, control is interrupted and continued at the dynamically
5252 nearest "exception" label.
5254 The '``exception``' label is a `landing
5255 pad <ExceptionHandling.html#overview>`_ for the exception. As such,
5256 '``exception``' label is required to have the
5257 ":ref:`landingpad <i_landingpad>`" instruction, which contains the
5258 information about the behavior of the program after unwinding happens,
5259 as its first non-PHI instruction. The restrictions on the
5260 "``landingpad``" instruction's tightly couples it to the "``invoke``"
5261 instruction, so that the important information contained within the
5262 "``landingpad``" instruction can't be lost through normal code motion.
5267 This instruction requires several arguments:
5269 #. The optional "cconv" marker indicates which :ref:`calling
5270 convention <callingconv>` the call should use. If none is
5271 specified, the call defaults to using C calling conventions.
5272 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
5273 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
5275 #. '``ptr to function ty``': shall be the signature of the pointer to
5276 function value being invoked. In most cases, this is a direct
5277 function invocation, but indirect ``invoke``'s are just as possible,
5278 branching off an arbitrary pointer to function value.
5279 #. '``function ptr val``': An LLVM value containing a pointer to a
5280 function to be invoked.
5281 #. '``function args``': argument list whose types match the function
5282 signature argument types and parameter attributes. All arguments must
5283 be of :ref:`first class <t_firstclass>` type. If the function signature
5284 indicates the function accepts a variable number of arguments, the
5285 extra arguments can be specified.
5286 #. '``normal label``': the label reached when the called function
5287 executes a '``ret``' instruction.
5288 #. '``exception label``': the label reached when a callee returns via
5289 the :ref:`resume <i_resume>` instruction or other exception handling
5291 #. The optional :ref:`function attributes <fnattrs>` list. Only
5292 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
5293 attributes are valid here.
5294 #. The optional :ref:`operand bundles <opbundles>` list.
5299 This instruction is designed to operate as a standard '``call``'
5300 instruction in most regards. The primary difference is that it
5301 establishes an association with a label, which is used by the runtime
5302 library to unwind the stack.
5304 This instruction is used in languages with destructors to ensure that
5305 proper cleanup is performed in the case of either a ``longjmp`` or a
5306 thrown exception. Additionally, this is important for implementation of
5307 '``catch``' clauses in high-level languages that support them.
5309 For the purposes of the SSA form, the definition of the value returned
5310 by the '``invoke``' instruction is deemed to occur on the edge from the
5311 current block to the "normal" label. If the callee unwinds then no
5312 return value is available.
5317 .. code-block:: llvm
5319 %retval = invoke i32 @Test(i32 15) to label %Continue
5320 unwind label %TestCleanup ; i32:retval set
5321 %retval = invoke coldcc i32 %Testfnptr(i32 15) to label %Continue
5322 unwind label %TestCleanup ; i32:retval set
5326 '``resume``' Instruction
5327 ^^^^^^^^^^^^^^^^^^^^^^^^
5334 resume <type> <value>
5339 The '``resume``' instruction is a terminator instruction that has no
5345 The '``resume``' instruction requires one argument, which must have the
5346 same type as the result of any '``landingpad``' instruction in the same
5352 The '``resume``' instruction resumes propagation of an existing
5353 (in-flight) exception whose unwinding was interrupted with a
5354 :ref:`landingpad <i_landingpad>` instruction.
5359 .. code-block:: llvm
5361 resume { i8*, i32 } %exn
5365 '``catchswitch``' Instruction
5366 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5373 <resultval> = catchswitch within <parent> [ label <handler1>, label <handler2>, ... ] unwind to caller
5374 <resultval> = catchswitch within <parent> [ label <handler1>, label <handler2>, ... ] unwind label <default>
5379 The '``catchswitch``' instruction is used by `LLVM's exception handling system
5380 <ExceptionHandling.html#overview>`_ to describe the set of possible catch handlers
5381 that may be executed by the :ref:`EH personality routine <personalityfn>`.
5386 The ``parent`` argument is the token of the funclet that contains the
5387 ``catchswitch`` instruction. If the ``catchswitch`` is not inside a funclet,
5388 this operand may be the token ``none``.
5390 The ``default`` argument is the label of another basic block beginning with a
5391 "pad" instruction, one of ``cleanuppad`` or ``catchswitch``.
5393 The ``handlers`` are a list of successor blocks that each begin with a
5394 :ref:`catchpad <i_catchpad>` instruction.
5399 Executing this instruction transfers control to one of the successors in
5400 ``handlers``, if appropriate, or continues to unwind via the unwind label if
5403 The ``catchswitch`` is both a terminator and a "pad" instruction, meaning that
5404 it must be both the first non-phi instruction and last instruction in the basic
5405 block. Therefore, it must be the only non-phi instruction in the block.
5410 .. code-block:: llvm
5413 %cs1 = catchswitch within none [label %handler0, label %handler1] unwind to caller
5415 %cs2 = catchswitch within %parenthandler [label %handler0] unwind label %cleanup
5419 '``catchpad``' Instruction
5420 ^^^^^^^^^^^^^^^^^^^^^^^^^^
5427 <resultval> = catchpad within <catchswitch> [<args>*]
5432 The '``catchpad``' instruction is used by `LLVM's exception handling
5433 system <ExceptionHandling.html#overview>`_ to specify that a basic block
5434 begins a catch handler --- one where a personality routine attempts to transfer
5435 control to catch an exception.
5440 The ``catchswitch`` operand must always be a token produced by a
5441 :ref:`catchswitch <i_catchswitch>` instruction in a predecessor block. This
5442 ensures that each ``catchpad`` has exactly one predecessor block, and it always
5443 terminates in a ``catchswitch``.
5445 The ``args`` correspond to whatever information the personality routine
5446 requires to know if this is an appropriate handler for the exception. Control
5447 will transfer to the ``catchpad`` if this is the first appropriate handler for
5450 The ``resultval`` has the type :ref:`token <t_token>` and is used to match the
5451 ``catchpad`` to corresponding :ref:`catchrets <i_catchret>` and other nested EH
5457 When the call stack is being unwound due to an exception being thrown, the
5458 exception is compared against the ``args``. If it doesn't match, control will
5459 not reach the ``catchpad`` instruction. The representation of ``args`` is
5460 entirely target and personality function-specific.
5462 Like the :ref:`landingpad <i_landingpad>` instruction, the ``catchpad``
5463 instruction must be the first non-phi of its parent basic block.
5465 The meaning of the tokens produced and consumed by ``catchpad`` and other "pad"
5466 instructions is described in the
5467 `Windows exception handling documentation <ExceptionHandling.html#wineh>`.
5469 Executing a ``catchpad`` instruction constitutes "entering" that pad.
5470 The pad may then be "exited" in one of three ways:
5472 1) explicitly via a ``catchret`` that consumes it. Executing such a ``catchret``
5473 is undefined behavior if any descendant pads have been entered but not yet
5475 2) implicitly via a call (which unwinds all the way to the current function's caller),
5476 or via a ``catchswitch`` or a ``cleanupret`` that unwinds to caller.
5477 3) implicitly via an unwind edge whose destination EH pad isn't a descendant of
5478 the ``catchpad``. When the ``catchpad`` is exited in this manner, it is
5479 undefined behavior if the destination EH pad has a parent which is not an
5480 ancestor of the ``catchpad`` being exited.
5485 .. code-block:: llvm
5488 %cs = catchswitch within none [label %handler0] unwind to caller
5489 ;; A catch block which can catch an integer.
5491 %tok = catchpad within %cs [i8** @_ZTIi]
5495 '``catchret``' Instruction
5496 ^^^^^^^^^^^^^^^^^^^^^^^^^^
5503 catchret from <token> to label <normal>
5508 The '``catchret``' instruction is a terminator instruction that has a
5515 The first argument to a '``catchret``' indicates which ``catchpad`` it
5516 exits. It must be a :ref:`catchpad <i_catchpad>`.
5517 The second argument to a '``catchret``' specifies where control will
5523 The '``catchret``' instruction ends an existing (in-flight) exception whose
5524 unwinding was interrupted with a :ref:`catchpad <i_catchpad>` instruction. The
5525 :ref:`personality function <personalityfn>` gets a chance to execute arbitrary
5526 code to, for example, destroy the active exception. Control then transfers to
5529 The ``token`` argument must be a token produced by a dominating ``catchpad``
5530 instruction. The ``catchret`` destroys the physical frame established by
5531 ``catchpad``, so executing multiple returns on the same token without
5532 re-executing the ``catchpad`` will result in undefined behavior.
5533 See :ref:`catchpad <i_catchpad>` for more details.
5538 .. code-block:: llvm
5540 catchret from %catch label %continue
5544 '``cleanupret``' Instruction
5545 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5552 cleanupret from <value> unwind label <continue>
5553 cleanupret from <value> unwind to caller
5558 The '``cleanupret``' instruction is a terminator instruction that has
5559 an optional successor.
5565 The '``cleanupret``' instruction requires one argument, which indicates
5566 which ``cleanuppad`` it exits, and must be a :ref:`cleanuppad <i_cleanuppad>`.
5567 It also has an optional successor, ``continue``.
5572 The '``cleanupret``' instruction indicates to the
5573 :ref:`personality function <personalityfn>` that one
5574 :ref:`cleanuppad <i_cleanuppad>` it transferred control to has ended.
5575 It transfers control to ``continue`` or unwinds out of the function.
5577 The unwind destination ``continue``, if present, must be an EH pad
5578 whose parent is either ``none`` or an ancestor of the ``cleanuppad``
5579 being returned from. This constitutes an exceptional exit from all
5580 ancestors of the completed ``cleanuppad``, up to but not including
5581 the parent of ``continue``.
5582 See :ref:`cleanuppad <i_cleanuppad>` for more details.
5587 .. code-block:: llvm
5589 cleanupret from %cleanup unwind to caller
5590 cleanupret from %cleanup unwind label %continue
5594 '``unreachable``' Instruction
5595 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5607 The '``unreachable``' instruction has no defined semantics. This
5608 instruction is used to inform the optimizer that a particular portion of
5609 the code is not reachable. This can be used to indicate that the code
5610 after a no-return function cannot be reached, and other facts.
5615 The '``unreachable``' instruction has no defined semantics.
5622 Binary operators are used to do most of the computation in a program.
5623 They require two operands of the same type, execute an operation on
5624 them, and produce a single value. The operands might represent multiple
5625 data, as is the case with the :ref:`vector <t_vector>` data type. The
5626 result value has the same type as its operands.
5628 There are several different binary operators:
5632 '``add``' Instruction
5633 ^^^^^^^^^^^^^^^^^^^^^
5640 <result> = add <ty> <op1>, <op2> ; yields ty:result
5641 <result> = add nuw <ty> <op1>, <op2> ; yields ty:result
5642 <result> = add nsw <ty> <op1>, <op2> ; yields ty:result
5643 <result> = add nuw nsw <ty> <op1>, <op2> ; yields ty:result
5648 The '``add``' instruction returns the sum of its two operands.
5653 The two arguments to the '``add``' instruction must be
5654 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
5655 arguments must have identical types.
5660 The value produced is the integer sum of the two operands.
5662 If the sum has unsigned overflow, the result returned is the
5663 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
5666 Because LLVM integers use a two's complement representation, this
5667 instruction is appropriate for both signed and unsigned integers.
5669 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
5670 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
5671 result value of the ``add`` is a :ref:`poison value <poisonvalues>` if
5672 unsigned and/or signed overflow, respectively, occurs.
5677 .. code-block:: llvm
5679 <result> = add i32 4, %var ; yields i32:result = 4 + %var
5683 '``fadd``' Instruction
5684 ^^^^^^^^^^^^^^^^^^^^^^
5691 <result> = fadd [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
5696 The '``fadd``' instruction returns the sum of its two operands.
5701 The two arguments to the '``fadd``' instruction must be :ref:`floating
5702 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
5703 Both arguments must have identical types.
5708 The value produced is the floating point sum of the two operands. This
5709 instruction can also take any number of :ref:`fast-math flags <fastmath>`,
5710 which are optimization hints to enable otherwise unsafe floating point
5716 .. code-block:: llvm
5718 <result> = fadd float 4.0, %var ; yields float:result = 4.0 + %var
5720 '``sub``' Instruction
5721 ^^^^^^^^^^^^^^^^^^^^^
5728 <result> = sub <ty> <op1>, <op2> ; yields ty:result
5729 <result> = sub nuw <ty> <op1>, <op2> ; yields ty:result
5730 <result> = sub nsw <ty> <op1>, <op2> ; yields ty:result
5731 <result> = sub nuw nsw <ty> <op1>, <op2> ; yields ty:result
5736 The '``sub``' instruction returns the difference of its two operands.
5738 Note that the '``sub``' instruction is used to represent the '``neg``'
5739 instruction present in most other intermediate representations.
5744 The two arguments to the '``sub``' instruction must be
5745 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
5746 arguments must have identical types.
5751 The value produced is the integer difference of the two operands.
5753 If the difference has unsigned overflow, the result returned is the
5754 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
5757 Because LLVM integers use a two's complement representation, this
5758 instruction is appropriate for both signed and unsigned integers.
5760 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
5761 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
5762 result value of the ``sub`` is a :ref:`poison value <poisonvalues>` if
5763 unsigned and/or signed overflow, respectively, occurs.
5768 .. code-block:: llvm
5770 <result> = sub i32 4, %var ; yields i32:result = 4 - %var
5771 <result> = sub i32 0, %val ; yields i32:result = -%var
5775 '``fsub``' Instruction
5776 ^^^^^^^^^^^^^^^^^^^^^^
5783 <result> = fsub [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
5788 The '``fsub``' instruction returns the difference of its two operands.
5790 Note that the '``fsub``' instruction is used to represent the '``fneg``'
5791 instruction present in most other intermediate representations.
5796 The two arguments to the '``fsub``' instruction must be :ref:`floating
5797 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
5798 Both arguments must have identical types.
5803 The value produced is the floating point difference of the two operands.
5804 This instruction can also take any number of :ref:`fast-math
5805 flags <fastmath>`, which are optimization hints to enable otherwise
5806 unsafe floating point optimizations:
5811 .. code-block:: llvm
5813 <result> = fsub float 4.0, %var ; yields float:result = 4.0 - %var
5814 <result> = fsub float -0.0, %val ; yields float:result = -%var
5816 '``mul``' Instruction
5817 ^^^^^^^^^^^^^^^^^^^^^
5824 <result> = mul <ty> <op1>, <op2> ; yields ty:result
5825 <result> = mul nuw <ty> <op1>, <op2> ; yields ty:result
5826 <result> = mul nsw <ty> <op1>, <op2> ; yields ty:result
5827 <result> = mul nuw nsw <ty> <op1>, <op2> ; yields ty:result
5832 The '``mul``' instruction returns the product of its two operands.
5837 The two arguments to the '``mul``' instruction must be
5838 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
5839 arguments must have identical types.
5844 The value produced is the integer product of the two operands.
5846 If the result of the multiplication has unsigned overflow, the result
5847 returned is the mathematical result modulo 2\ :sup:`n`\ , where n is the
5848 bit width of the result.
5850 Because LLVM integers use a two's complement representation, and the
5851 result is the same width as the operands, this instruction returns the
5852 correct result for both signed and unsigned integers. If a full product
5853 (e.g. ``i32`` * ``i32`` -> ``i64``) is needed, the operands should be
5854 sign-extended or zero-extended as appropriate to the width of the full
5857 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
5858 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
5859 result value of the ``mul`` is a :ref:`poison value <poisonvalues>` if
5860 unsigned and/or signed overflow, respectively, occurs.
5865 .. code-block:: llvm
5867 <result> = mul i32 4, %var ; yields i32:result = 4 * %var
5871 '``fmul``' Instruction
5872 ^^^^^^^^^^^^^^^^^^^^^^
5879 <result> = fmul [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
5884 The '``fmul``' instruction returns the product of its two operands.
5889 The two arguments to the '``fmul``' instruction must be :ref:`floating
5890 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
5891 Both arguments must have identical types.
5896 The value produced is the floating point product of the two operands.
5897 This instruction can also take any number of :ref:`fast-math
5898 flags <fastmath>`, which are optimization hints to enable otherwise
5899 unsafe floating point optimizations:
5904 .. code-block:: llvm
5906 <result> = fmul float 4.0, %var ; yields float:result = 4.0 * %var
5908 '``udiv``' Instruction
5909 ^^^^^^^^^^^^^^^^^^^^^^
5916 <result> = udiv <ty> <op1>, <op2> ; yields ty:result
5917 <result> = udiv exact <ty> <op1>, <op2> ; yields ty:result
5922 The '``udiv``' instruction returns the quotient of its two operands.
5927 The two arguments to the '``udiv``' instruction must be
5928 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
5929 arguments must have identical types.
5934 The value produced is the unsigned integer quotient of the two operands.
5936 Note that unsigned integer division and signed integer division are
5937 distinct operations; for signed integer division, use '``sdiv``'.
5939 Division by zero leads to undefined behavior.
5941 If the ``exact`` keyword is present, the result value of the ``udiv`` is
5942 a :ref:`poison value <poisonvalues>` if %op1 is not a multiple of %op2 (as
5943 such, "((a udiv exact b) mul b) == a").
5948 .. code-block:: llvm
5950 <result> = udiv i32 4, %var ; yields i32:result = 4 / %var
5952 '``sdiv``' Instruction
5953 ^^^^^^^^^^^^^^^^^^^^^^
5960 <result> = sdiv <ty> <op1>, <op2> ; yields ty:result
5961 <result> = sdiv exact <ty> <op1>, <op2> ; yields ty:result
5966 The '``sdiv``' instruction returns the quotient of its two operands.
5971 The two arguments to the '``sdiv``' instruction must be
5972 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
5973 arguments must have identical types.
5978 The value produced is the signed integer quotient of the two operands
5979 rounded towards zero.
5981 Note that signed integer division and unsigned integer division are
5982 distinct operations; for unsigned integer division, use '``udiv``'.
5984 Division by zero leads to undefined behavior. Overflow also leads to
5985 undefined behavior; this is a rare case, but can occur, for example, by
5986 doing a 32-bit division of -2147483648 by -1.
5988 If the ``exact`` keyword is present, the result value of the ``sdiv`` is
5989 a :ref:`poison value <poisonvalues>` if the result would be rounded.
5994 .. code-block:: llvm
5996 <result> = sdiv i32 4, %var ; yields i32:result = 4 / %var
6000 '``fdiv``' Instruction
6001 ^^^^^^^^^^^^^^^^^^^^^^
6008 <result> = fdiv [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
6013 The '``fdiv``' instruction returns the quotient of its two operands.
6018 The two arguments to the '``fdiv``' instruction must be :ref:`floating
6019 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
6020 Both arguments must have identical types.
6025 The value produced is the floating point quotient of the two operands.
6026 This instruction can also take any number of :ref:`fast-math
6027 flags <fastmath>`, which are optimization hints to enable otherwise
6028 unsafe floating point optimizations:
6033 .. code-block:: llvm
6035 <result> = fdiv float 4.0, %var ; yields float:result = 4.0 / %var
6037 '``urem``' Instruction
6038 ^^^^^^^^^^^^^^^^^^^^^^
6045 <result> = urem <ty> <op1>, <op2> ; yields ty:result
6050 The '``urem``' instruction returns the remainder from the unsigned
6051 division of its two arguments.
6056 The two arguments to the '``urem``' instruction must be
6057 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6058 arguments must have identical types.
6063 This instruction returns the unsigned integer *remainder* of a division.
6064 This instruction always performs an unsigned division to get the
6067 Note that unsigned integer remainder and signed integer remainder are
6068 distinct operations; for signed integer remainder, use '``srem``'.
6070 Taking the remainder of a division by zero leads to undefined behavior.
6075 .. code-block:: llvm
6077 <result> = urem i32 4, %var ; yields i32:result = 4 % %var
6079 '``srem``' Instruction
6080 ^^^^^^^^^^^^^^^^^^^^^^
6087 <result> = srem <ty> <op1>, <op2> ; yields ty:result
6092 The '``srem``' instruction returns the remainder from the signed
6093 division of its two operands. This instruction can also take
6094 :ref:`vector <t_vector>` versions of the values in which case the elements
6100 The two arguments to the '``srem``' instruction must be
6101 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6102 arguments must have identical types.
6107 This instruction returns the *remainder* of a division (where the result
6108 is either zero or has the same sign as the dividend, ``op1``), not the
6109 *modulo* operator (where the result is either zero or has the same sign
6110 as the divisor, ``op2``) of a value. For more information about the
6111 difference, see `The Math
6112 Forum <http://mathforum.org/dr.math/problems/anne.4.28.99.html>`_. For a
6113 table of how this is implemented in various languages, please see
6115 operation <http://en.wikipedia.org/wiki/Modulo_operation>`_.
6117 Note that signed integer remainder and unsigned integer remainder are
6118 distinct operations; for unsigned integer remainder, use '``urem``'.
6120 Taking the remainder of a division by zero leads to undefined behavior.
6121 Overflow also leads to undefined behavior; this is a rare case, but can
6122 occur, for example, by taking the remainder of a 32-bit division of
6123 -2147483648 by -1. (The remainder doesn't actually overflow, but this
6124 rule lets srem be implemented using instructions that return both the
6125 result of the division and the remainder.)
6130 .. code-block:: llvm
6132 <result> = srem i32 4, %var ; yields i32:result = 4 % %var
6136 '``frem``' Instruction
6137 ^^^^^^^^^^^^^^^^^^^^^^
6144 <result> = frem [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
6149 The '``frem``' instruction returns the remainder from the division of
6155 The two arguments to the '``frem``' instruction must be :ref:`floating
6156 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
6157 Both arguments must have identical types.
6162 This instruction returns the *remainder* of a division. The remainder
6163 has the same sign as the dividend. This instruction can also take any
6164 number of :ref:`fast-math flags <fastmath>`, which are optimization hints
6165 to enable otherwise unsafe floating point optimizations:
6170 .. code-block:: llvm
6172 <result> = frem float 4.0, %var ; yields float:result = 4.0 % %var
6176 Bitwise Binary Operations
6177 -------------------------
6179 Bitwise binary operators are used to do various forms of bit-twiddling
6180 in a program. They are generally very efficient instructions and can
6181 commonly be strength reduced from other instructions. They require two
6182 operands of the same type, execute an operation on them, and produce a
6183 single value. The resulting value is the same type as its operands.
6185 '``shl``' Instruction
6186 ^^^^^^^^^^^^^^^^^^^^^
6193 <result> = shl <ty> <op1>, <op2> ; yields ty:result
6194 <result> = shl nuw <ty> <op1>, <op2> ; yields ty:result
6195 <result> = shl nsw <ty> <op1>, <op2> ; yields ty:result
6196 <result> = shl nuw nsw <ty> <op1>, <op2> ; yields ty:result
6201 The '``shl``' instruction returns the first operand shifted to the left
6202 a specified number of bits.
6207 Both arguments to the '``shl``' instruction must be the same
6208 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
6209 '``op2``' is treated as an unsigned value.
6214 The value produced is ``op1`` \* 2\ :sup:`op2` mod 2\ :sup:`n`,
6215 where ``n`` is the width of the result. If ``op2`` is (statically or
6216 dynamically) equal to or larger than the number of bits in
6217 ``op1``, the result is undefined. If the arguments are vectors, each
6218 vector element of ``op1`` is shifted by the corresponding shift amount
6221 If the ``nuw`` keyword is present, then the shift produces a :ref:`poison
6222 value <poisonvalues>` if it shifts out any non-zero bits. If the
6223 ``nsw`` keyword is present, then the shift produces a :ref:`poison
6224 value <poisonvalues>` if it shifts out any bits that disagree with the
6225 resultant sign bit. As such, NUW/NSW have the same semantics as they
6226 would if the shift were expressed as a mul instruction with the same
6227 nsw/nuw bits in (mul %op1, (shl 1, %op2)).
6232 .. code-block:: llvm
6234 <result> = shl i32 4, %var ; yields i32: 4 << %var
6235 <result> = shl i32 4, 2 ; yields i32: 16
6236 <result> = shl i32 1, 10 ; yields i32: 1024
6237 <result> = shl i32 1, 32 ; undefined
6238 <result> = shl <2 x i32> < i32 1, i32 1>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 2, i32 4>
6240 '``lshr``' Instruction
6241 ^^^^^^^^^^^^^^^^^^^^^^
6248 <result> = lshr <ty> <op1>, <op2> ; yields ty:result
6249 <result> = lshr exact <ty> <op1>, <op2> ; yields ty:result
6254 The '``lshr``' instruction (logical shift right) returns the first
6255 operand shifted to the right a specified number of bits with zero fill.
6260 Both arguments to the '``lshr``' instruction must be the same
6261 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
6262 '``op2``' is treated as an unsigned value.
6267 This instruction always performs a logical shift right operation. The
6268 most significant bits of the result will be filled with zero bits after
6269 the shift. If ``op2`` is (statically or dynamically) equal to or larger
6270 than the number of bits in ``op1``, the result is undefined. If the
6271 arguments are vectors, each vector element of ``op1`` is shifted by the
6272 corresponding shift amount in ``op2``.
6274 If the ``exact`` keyword is present, the result value of the ``lshr`` is
6275 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
6281 .. code-block:: llvm
6283 <result> = lshr i32 4, 1 ; yields i32:result = 2
6284 <result> = lshr i32 4, 2 ; yields i32:result = 1
6285 <result> = lshr i8 4, 3 ; yields i8:result = 0
6286 <result> = lshr i8 -2, 1 ; yields i8:result = 0x7F
6287 <result> = lshr i32 1, 32 ; undefined
6288 <result> = lshr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 0x7FFFFFFF, i32 1>
6290 '``ashr``' Instruction
6291 ^^^^^^^^^^^^^^^^^^^^^^
6298 <result> = ashr <ty> <op1>, <op2> ; yields ty:result
6299 <result> = ashr exact <ty> <op1>, <op2> ; yields ty:result
6304 The '``ashr``' instruction (arithmetic shift right) returns the first
6305 operand shifted to the right a specified number of bits with sign
6311 Both arguments to the '``ashr``' instruction must be the same
6312 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
6313 '``op2``' is treated as an unsigned value.
6318 This instruction always performs an arithmetic shift right operation,
6319 The most significant bits of the result will be filled with the sign bit
6320 of ``op1``. If ``op2`` is (statically or dynamically) equal to or larger
6321 than the number of bits in ``op1``, the result is undefined. If the
6322 arguments are vectors, each vector element of ``op1`` is shifted by the
6323 corresponding shift amount in ``op2``.
6325 If the ``exact`` keyword is present, the result value of the ``ashr`` is
6326 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
6332 .. code-block:: llvm
6334 <result> = ashr i32 4, 1 ; yields i32:result = 2
6335 <result> = ashr i32 4, 2 ; yields i32:result = 1
6336 <result> = ashr i8 4, 3 ; yields i8:result = 0
6337 <result> = ashr i8 -2, 1 ; yields i8:result = -1
6338 <result> = ashr i32 1, 32 ; undefined
6339 <result> = ashr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 3> ; yields: result=<2 x i32> < i32 -1, i32 0>
6341 '``and``' Instruction
6342 ^^^^^^^^^^^^^^^^^^^^^
6349 <result> = and <ty> <op1>, <op2> ; yields ty:result
6354 The '``and``' instruction returns the bitwise logical and of its two
6360 The two arguments to the '``and``' instruction must be
6361 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6362 arguments must have identical types.
6367 The truth table used for the '``and``' instruction is:
6384 .. code-block:: llvm
6386 <result> = and i32 4, %var ; yields i32:result = 4 & %var
6387 <result> = and i32 15, 40 ; yields i32:result = 8
6388 <result> = and i32 4, 8 ; yields i32:result = 0
6390 '``or``' Instruction
6391 ^^^^^^^^^^^^^^^^^^^^
6398 <result> = or <ty> <op1>, <op2> ; yields ty:result
6403 The '``or``' instruction returns the bitwise logical inclusive or of its
6409 The two arguments to the '``or``' instruction must be
6410 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6411 arguments must have identical types.
6416 The truth table used for the '``or``' instruction is:
6435 <result> = or i32 4, %var ; yields i32:result = 4 | %var
6436 <result> = or i32 15, 40 ; yields i32:result = 47
6437 <result> = or i32 4, 8 ; yields i32:result = 12
6439 '``xor``' Instruction
6440 ^^^^^^^^^^^^^^^^^^^^^
6447 <result> = xor <ty> <op1>, <op2> ; yields ty:result
6452 The '``xor``' instruction returns the bitwise logical exclusive or of
6453 its two operands. The ``xor`` is used to implement the "one's
6454 complement" operation, which is the "~" operator in C.
6459 The two arguments to the '``xor``' instruction must be
6460 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6461 arguments must have identical types.
6466 The truth table used for the '``xor``' instruction is:
6483 .. code-block:: llvm
6485 <result> = xor i32 4, %var ; yields i32:result = 4 ^ %var
6486 <result> = xor i32 15, 40 ; yields i32:result = 39
6487 <result> = xor i32 4, 8 ; yields i32:result = 12
6488 <result> = xor i32 %V, -1 ; yields i32:result = ~%V
6493 LLVM supports several instructions to represent vector operations in a
6494 target-independent manner. These instructions cover the element-access
6495 and vector-specific operations needed to process vectors effectively.
6496 While LLVM does directly support these vector operations, many
6497 sophisticated algorithms will want to use target-specific intrinsics to
6498 take full advantage of a specific target.
6500 .. _i_extractelement:
6502 '``extractelement``' Instruction
6503 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6510 <result> = extractelement <n x <ty>> <val>, <ty2> <idx> ; yields <ty>
6515 The '``extractelement``' instruction extracts a single scalar element
6516 from a vector at a specified index.
6521 The first operand of an '``extractelement``' instruction is a value of
6522 :ref:`vector <t_vector>` type. The second operand is an index indicating
6523 the position from which to extract the element. The index may be a
6524 variable of any integer type.
6529 The result is a scalar of the same type as the element type of ``val``.
6530 Its value is the value at position ``idx`` of ``val``. If ``idx``
6531 exceeds the length of ``val``, the results are undefined.
6536 .. code-block:: llvm
6538 <result> = extractelement <4 x i32> %vec, i32 0 ; yields i32
6540 .. _i_insertelement:
6542 '``insertelement``' Instruction
6543 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6550 <result> = insertelement <n x <ty>> <val>, <ty> <elt>, <ty2> <idx> ; yields <n x <ty>>
6555 The '``insertelement``' instruction inserts a scalar element into a
6556 vector at a specified index.
6561 The first operand of an '``insertelement``' instruction is a value of
6562 :ref:`vector <t_vector>` type. The second operand is a scalar value whose
6563 type must equal the element type of the first operand. The third operand
6564 is an index indicating the position at which to insert the value. The
6565 index may be a variable of any integer type.
6570 The result is a vector of the same type as ``val``. Its element values
6571 are those of ``val`` except at position ``idx``, where it gets the value
6572 ``elt``. If ``idx`` exceeds the length of ``val``, the results are
6578 .. code-block:: llvm
6580 <result> = insertelement <4 x i32> %vec, i32 1, i32 0 ; yields <4 x i32>
6582 .. _i_shufflevector:
6584 '``shufflevector``' Instruction
6585 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6592 <result> = shufflevector <n x <ty>> <v1>, <n x <ty>> <v2>, <m x i32> <mask> ; yields <m x <ty>>
6597 The '``shufflevector``' instruction constructs a permutation of elements
6598 from two input vectors, returning a vector with the same element type as
6599 the input and length that is the same as the shuffle mask.
6604 The first two operands of a '``shufflevector``' instruction are vectors
6605 with the same type. The third argument is a shuffle mask whose element
6606 type is always 'i32'. The result of the instruction is a vector whose
6607 length is the same as the shuffle mask and whose element type is the
6608 same as the element type of the first two operands.
6610 The shuffle mask operand is required to be a constant vector with either
6611 constant integer or undef values.
6616 The elements of the two input vectors are numbered from left to right
6617 across both of the vectors. The shuffle mask operand specifies, for each
6618 element of the result vector, which element of the two input vectors the
6619 result element gets. The element selector may be undef (meaning "don't
6620 care") and the second operand may be undef if performing a shuffle from
6626 .. code-block:: llvm
6628 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
6629 <4 x i32> <i32 0, i32 4, i32 1, i32 5> ; yields <4 x i32>
6630 <result> = shufflevector <4 x i32> %v1, <4 x i32> undef,
6631 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32> - Identity shuffle.
6632 <result> = shufflevector <8 x i32> %v1, <8 x i32> undef,
6633 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32>
6634 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
6635 <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7 > ; yields <8 x i32>
6637 Aggregate Operations
6638 --------------------
6640 LLVM supports several instructions for working with
6641 :ref:`aggregate <t_aggregate>` values.
6645 '``extractvalue``' Instruction
6646 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6653 <result> = extractvalue <aggregate type> <val>, <idx>{, <idx>}*
6658 The '``extractvalue``' instruction extracts the value of a member field
6659 from an :ref:`aggregate <t_aggregate>` value.
6664 The first operand of an '``extractvalue``' instruction is a value of
6665 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The other operands are
6666 constant indices to specify which value to extract in a similar manner
6667 as indices in a '``getelementptr``' instruction.
6669 The major differences to ``getelementptr`` indexing are:
6671 - Since the value being indexed is not a pointer, the first index is
6672 omitted and assumed to be zero.
6673 - At least one index must be specified.
6674 - Not only struct indices but also array indices must be in bounds.
6679 The result is the value at the position in the aggregate specified by
6685 .. code-block:: llvm
6687 <result> = extractvalue {i32, float} %agg, 0 ; yields i32
6691 '``insertvalue``' Instruction
6692 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6699 <result> = insertvalue <aggregate type> <val>, <ty> <elt>, <idx>{, <idx>}* ; yields <aggregate type>
6704 The '``insertvalue``' instruction inserts a value into a member field in
6705 an :ref:`aggregate <t_aggregate>` value.
6710 The first operand of an '``insertvalue``' instruction is a value of
6711 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The second operand is
6712 a first-class value to insert. The following operands are constant
6713 indices indicating the position at which to insert the value in a
6714 similar manner as indices in a '``extractvalue``' instruction. The value
6715 to insert must have the same type as the value identified by the
6721 The result is an aggregate of the same type as ``val``. Its value is
6722 that of ``val`` except that the value at the position specified by the
6723 indices is that of ``elt``.
6728 .. code-block:: llvm
6730 %agg1 = insertvalue {i32, float} undef, i32 1, 0 ; yields {i32 1, float undef}
6731 %agg2 = insertvalue {i32, float} %agg1, float %val, 1 ; yields {i32 1, float %val}
6732 %agg3 = insertvalue {i32, {float}} undef, float %val, 1, 0 ; yields {i32 undef, {float %val}}
6736 Memory Access and Addressing Operations
6737 ---------------------------------------
6739 A key design point of an SSA-based representation is how it represents
6740 memory. In LLVM, no memory locations are in SSA form, which makes things
6741 very simple. This section describes how to read, write, and allocate
6746 '``alloca``' Instruction
6747 ^^^^^^^^^^^^^^^^^^^^^^^^
6754 <result> = alloca [inalloca] <type> [, <ty> <NumElements>] [, align <alignment>] ; yields type*:result
6759 The '``alloca``' instruction allocates memory on the stack frame of the
6760 currently executing function, to be automatically released when this
6761 function returns to its caller. The object is always allocated in the
6762 generic address space (address space zero).
6767 The '``alloca``' instruction allocates ``sizeof(<type>)*NumElements``
6768 bytes of memory on the runtime stack, returning a pointer of the
6769 appropriate type to the program. If "NumElements" is specified, it is
6770 the number of elements allocated, otherwise "NumElements" is defaulted
6771 to be one. If a constant alignment is specified, the value result of the
6772 allocation is guaranteed to be aligned to at least that boundary. The
6773 alignment may not be greater than ``1 << 29``. If not specified, or if
6774 zero, the target can choose to align the allocation on any convenient
6775 boundary compatible with the type.
6777 '``type``' may be any sized type.
6782 Memory is allocated; a pointer is returned. The operation is undefined
6783 if there is insufficient stack space for the allocation. '``alloca``'d
6784 memory is automatically released when the function returns. The
6785 '``alloca``' instruction is commonly used to represent automatic
6786 variables that must have an address available. When the function returns
6787 (either with the ``ret`` or ``resume`` instructions), the memory is
6788 reclaimed. Allocating zero bytes is legal, but the result is undefined.
6789 The order in which memory is allocated (ie., which way the stack grows)
6795 .. code-block:: llvm
6797 %ptr = alloca i32 ; yields i32*:ptr
6798 %ptr = alloca i32, i32 4 ; yields i32*:ptr
6799 %ptr = alloca i32, i32 4, align 1024 ; yields i32*:ptr
6800 %ptr = alloca i32, align 1024 ; yields i32*:ptr
6804 '``load``' Instruction
6805 ^^^^^^^^^^^^^^^^^^^^^^
6812 <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>]
6813 <result> = load atomic [volatile] <ty>* <pointer> [singlethread] <ordering>, align <alignment> [, !invariant.group !<index>]
6814 !<index> = !{ i32 1 }
6815 !<deref_bytes_node> = !{i64 <dereferenceable_bytes>}
6816 !<align_node> = !{ i64 <value_alignment> }
6821 The '``load``' instruction is used to read from memory.
6826 The argument to the ``load`` instruction specifies the memory address
6827 from which to load. The type specified must be a :ref:`first
6828 class <t_firstclass>` type. If the ``load`` is marked as ``volatile``,
6829 then the optimizer is not allowed to modify the number or order of
6830 execution of this ``load`` with other :ref:`volatile
6831 operations <volatile>`.
6833 If the ``load`` is marked as ``atomic``, it takes an extra
6834 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
6835 ``release`` and ``acq_rel`` orderings are not valid on ``load``
6836 instructions. Atomic loads produce :ref:`defined <memmodel>` results
6837 when they may see multiple atomic stores. The type of the pointee must
6838 be an integer type whose bit width is a power of two greater than or
6839 equal to eight and less than or equal to a target-specific size limit.
6840 ``align`` must be explicitly specified on atomic loads, and the load has
6841 undefined behavior if the alignment is not set to a value which is at
6842 least the size in bytes of the pointee. ``!nontemporal`` does not have
6843 any defined semantics for atomic loads.
6845 The optional constant ``align`` argument specifies the alignment of the
6846 operation (that is, the alignment of the memory address). A value of 0
6847 or an omitted ``align`` argument means that the operation has the ABI
6848 alignment for the target. It is the responsibility of the code emitter
6849 to ensure that the alignment information is correct. Overestimating the
6850 alignment results in undefined behavior. Underestimating the alignment
6851 may produce less efficient code. An alignment of 1 is always safe. The
6852 maximum possible alignment is ``1 << 29``.
6854 The optional ``!nontemporal`` metadata must reference a single
6855 metadata name ``<index>`` corresponding to a metadata node with one
6856 ``i32`` entry of value 1. The existence of the ``!nontemporal``
6857 metadata on the instruction tells the optimizer and code generator
6858 that this load is not expected to be reused in the cache. The code
6859 generator may select special instructions to save cache bandwidth, such
6860 as the ``MOVNT`` instruction on x86.
6862 The optional ``!invariant.load`` metadata must reference a single
6863 metadata name ``<index>`` corresponding to a metadata node with no
6864 entries. The existence of the ``!invariant.load`` metadata on the
6865 instruction tells the optimizer and code generator that the address
6866 operand to this load points to memory which can be assumed unchanged.
6867 Being invariant does not imply that a location is dereferenceable,
6868 but it does imply that once the location is known dereferenceable
6869 its value is henceforth unchanging.
6871 The optional ``!invariant.group`` metadata must reference a single metadata name
6872 ``<index>`` corresponding to a metadata node. See ``invariant.group`` metadata.
6874 The optional ``!nonnull`` metadata must reference a single
6875 metadata name ``<index>`` corresponding to a metadata node with no
6876 entries. The existence of the ``!nonnull`` metadata on the
6877 instruction tells the optimizer that the value loaded is known to
6878 never be null. This is analogous to the ``nonnull`` attribute
6879 on parameters and return values. This metadata can only be applied
6880 to loads of a pointer type.
6882 The optional ``!dereferenceable`` metadata must reference a single metadata
6883 name ``<deref_bytes_node>`` corresponding to a metadata node with one ``i64``
6884 entry. The existence of the ``!dereferenceable`` metadata on the instruction
6885 tells the optimizer that the value loaded is known to be dereferenceable.
6886 The number of bytes known to be dereferenceable is specified by the integer
6887 value in the metadata node. This is analogous to the ''dereferenceable''
6888 attribute on parameters and return values. This metadata can only be applied
6889 to loads of a pointer type.
6891 The optional ``!dereferenceable_or_null`` metadata must reference a single
6892 metadata name ``<deref_bytes_node>`` corresponding to a metadata node with one
6893 ``i64`` entry. The existence of the ``!dereferenceable_or_null`` metadata on the
6894 instruction tells the optimizer that the value loaded is known to be either
6895 dereferenceable or null.
6896 The number of bytes known to be dereferenceable is specified by the integer
6897 value in the metadata node. This is analogous to the ''dereferenceable_or_null''
6898 attribute on parameters and return values. This metadata can only be applied
6899 to loads of a pointer type.
6901 The optional ``!align`` metadata must reference a single metadata name
6902 ``<align_node>`` corresponding to a metadata node with one ``i64`` entry.
6903 The existence of the ``!align`` metadata on the instruction tells the
6904 optimizer that the value loaded is known to be aligned to a boundary specified
6905 by the integer value in the metadata node. The alignment must be a power of 2.
6906 This is analogous to the ''align'' attribute on parameters and return values.
6907 This metadata can only be applied to loads of a pointer type.
6912 The location of memory pointed to is loaded. If the value being loaded
6913 is of scalar type then the number of bytes read does not exceed the
6914 minimum number of bytes needed to hold all bits of the type. For
6915 example, loading an ``i24`` reads at most three bytes. When loading a
6916 value of a type like ``i20`` with a size that is not an integral number
6917 of bytes, the result is undefined if the value was not originally
6918 written using a store of the same type.
6923 .. code-block:: llvm
6925 %ptr = alloca i32 ; yields i32*:ptr
6926 store i32 3, i32* %ptr ; yields void
6927 %val = load i32, i32* %ptr ; yields i32:val = i32 3
6931 '``store``' Instruction
6932 ^^^^^^^^^^^^^^^^^^^^^^^
6939 store [volatile] <ty> <value>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>][, !invariant.group !<index>] ; yields void
6940 store atomic [volatile] <ty> <value>, <ty>* <pointer> [singlethread] <ordering>, align <alignment> [, !invariant.group !<index>] ; yields void
6945 The '``store``' instruction is used to write to memory.
6950 There are two arguments to the ``store`` instruction: a value to store
6951 and an address at which to store it. The type of the ``<pointer>``
6952 operand must be a pointer to the :ref:`first class <t_firstclass>` type of
6953 the ``<value>`` operand. If the ``store`` is marked as ``volatile``,
6954 then the optimizer is not allowed to modify the number or order of
6955 execution of this ``store`` with other :ref:`volatile
6956 operations <volatile>`.
6958 If the ``store`` is marked as ``atomic``, it takes an extra
6959 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
6960 ``acquire`` and ``acq_rel`` orderings aren't valid on ``store``
6961 instructions. Atomic loads produce :ref:`defined <memmodel>` results
6962 when they may see multiple atomic stores. The type of the pointee must
6963 be an integer type whose bit width is a power of two greater than or
6964 equal to eight and less than or equal to a target-specific size limit.
6965 ``align`` must be explicitly specified on atomic stores, and the store
6966 has undefined behavior if the alignment is not set to a value which is
6967 at least the size in bytes of the pointee. ``!nontemporal`` does not
6968 have any defined semantics for atomic stores.
6970 The optional constant ``align`` argument specifies the alignment of the
6971 operation (that is, the alignment of the memory address). A value of 0
6972 or an omitted ``align`` argument means that the operation has the ABI
6973 alignment for the target. It is the responsibility of the code emitter
6974 to ensure that the alignment information is correct. Overestimating the
6975 alignment results in undefined behavior. Underestimating the
6976 alignment may produce less efficient code. An alignment of 1 is always
6977 safe. The maximum possible alignment is ``1 << 29``.
6979 The optional ``!nontemporal`` metadata must reference a single metadata
6980 name ``<index>`` corresponding to a metadata node with one ``i32`` entry of
6981 value 1. The existence of the ``!nontemporal`` metadata on the instruction
6982 tells the optimizer and code generator that this load is not expected to
6983 be reused in the cache. The code generator may select special
6984 instructions to save cache bandwidth, such as the MOVNT instruction on
6987 The optional ``!invariant.group`` metadata must reference a
6988 single metadata name ``<index>``. See ``invariant.group`` metadata.
6993 The contents of memory are updated to contain ``<value>`` at the
6994 location specified by the ``<pointer>`` operand. If ``<value>`` is
6995 of scalar type then the number of bytes written does not exceed the
6996 minimum number of bytes needed to hold all bits of the type. For
6997 example, storing an ``i24`` writes at most three bytes. When writing a
6998 value of a type like ``i20`` with a size that is not an integral number
6999 of bytes, it is unspecified what happens to the extra bits that do not
7000 belong to the type, but they will typically be overwritten.
7005 .. code-block:: llvm
7007 %ptr = alloca i32 ; yields i32*:ptr
7008 store i32 3, i32* %ptr ; yields void
7009 %val = load i32, i32* %ptr ; yields i32:val = i32 3
7013 '``fence``' Instruction
7014 ^^^^^^^^^^^^^^^^^^^^^^^
7021 fence [singlethread] <ordering> ; yields void
7026 The '``fence``' instruction is used to introduce happens-before edges
7032 '``fence``' instructions take an :ref:`ordering <ordering>` argument which
7033 defines what *synchronizes-with* edges they add. They can only be given
7034 ``acquire``, ``release``, ``acq_rel``, and ``seq_cst`` orderings.
7039 A fence A which has (at least) ``release`` ordering semantics
7040 *synchronizes with* a fence B with (at least) ``acquire`` ordering
7041 semantics if and only if there exist atomic operations X and Y, both
7042 operating on some atomic object M, such that A is sequenced before X, X
7043 modifies M (either directly or through some side effect of a sequence
7044 headed by X), Y is sequenced before B, and Y observes M. This provides a
7045 *happens-before* dependency between A and B. Rather than an explicit
7046 ``fence``, one (but not both) of the atomic operations X or Y might
7047 provide a ``release`` or ``acquire`` (resp.) ordering constraint and
7048 still *synchronize-with* the explicit ``fence`` and establish the
7049 *happens-before* edge.
7051 A ``fence`` which has ``seq_cst`` ordering, in addition to having both
7052 ``acquire`` and ``release`` semantics specified above, participates in
7053 the global program order of other ``seq_cst`` operations and/or fences.
7055 The optional ":ref:`singlethread <singlethread>`" argument specifies
7056 that the fence only synchronizes with other fences in the same thread.
7057 (This is useful for interacting with signal handlers.)
7062 .. code-block:: llvm
7064 fence acquire ; yields void
7065 fence singlethread seq_cst ; yields void
7069 '``cmpxchg``' Instruction
7070 ^^^^^^^^^^^^^^^^^^^^^^^^^
7077 cmpxchg [weak] [volatile] <ty>* <pointer>, <ty> <cmp>, <ty> <new> [singlethread] <success ordering> <failure ordering> ; yields { ty, i1 }
7082 The '``cmpxchg``' instruction is used to atomically modify memory. It
7083 loads a value in memory and compares it to a given value. If they are
7084 equal, it tries to store a new value into the memory.
7089 There are three arguments to the '``cmpxchg``' instruction: an address
7090 to operate on, a value to compare to the value currently be at that
7091 address, and a new value to place at that address if the compared values
7092 are equal. The type of '<cmp>' must be an integer type whose bit width
7093 is a power of two greater than or equal to eight and less than or equal
7094 to a target-specific size limit. '<cmp>' and '<new>' must have the same
7095 type, and the type of '<pointer>' must be a pointer to that type. If the
7096 ``cmpxchg`` is marked as ``volatile``, then the optimizer is not allowed
7097 to modify the number or order of execution of this ``cmpxchg`` with
7098 other :ref:`volatile operations <volatile>`.
7100 The success and failure :ref:`ordering <ordering>` arguments specify how this
7101 ``cmpxchg`` synchronizes with other atomic operations. Both ordering parameters
7102 must be at least ``monotonic``, the ordering constraint on failure must be no
7103 stronger than that on success, and the failure ordering cannot be either
7104 ``release`` or ``acq_rel``.
7106 The optional "``singlethread``" argument declares that the ``cmpxchg``
7107 is only atomic with respect to code (usually signal handlers) running in
7108 the same thread as the ``cmpxchg``. Otherwise the cmpxchg is atomic with
7109 respect to all other code in the system.
7111 The pointer passed into cmpxchg must have alignment greater than or
7112 equal to the size in memory of the operand.
7117 The contents of memory at the location specified by the '``<pointer>``' operand
7118 is read and compared to '``<cmp>``'; if the read value is the equal, the
7119 '``<new>``' is written. The original value at the location is returned, together
7120 with a flag indicating success (true) or failure (false).
7122 If the cmpxchg operation is marked as ``weak`` then a spurious failure is
7123 permitted: the operation may not write ``<new>`` even if the comparison
7126 If the cmpxchg operation is strong (the default), the i1 value is 1 if and only
7127 if the value loaded equals ``cmp``.
7129 A successful ``cmpxchg`` is a read-modify-write instruction for the purpose of
7130 identifying release sequences. A failed ``cmpxchg`` is equivalent to an atomic
7131 load with an ordering parameter determined the second ordering parameter.
7136 .. code-block:: llvm
7139 %orig = atomic load i32, i32* %ptr unordered ; yields i32
7143 %cmp = phi i32 [ %orig, %entry ], [%old, %loop]
7144 %squared = mul i32 %cmp, %cmp
7145 %val_success = cmpxchg i32* %ptr, i32 %cmp, i32 %squared acq_rel monotonic ; yields { i32, i1 }
7146 %value_loaded = extractvalue { i32, i1 } %val_success, 0
7147 %success = extractvalue { i32, i1 } %val_success, 1
7148 br i1 %success, label %done, label %loop
7155 '``atomicrmw``' Instruction
7156 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7163 atomicrmw [volatile] <operation> <ty>* <pointer>, <ty> <value> [singlethread] <ordering> ; yields ty
7168 The '``atomicrmw``' instruction is used to atomically modify memory.
7173 There are three arguments to the '``atomicrmw``' instruction: an
7174 operation to apply, an address whose value to modify, an argument to the
7175 operation. The operation must be one of the following keywords:
7189 The type of '<value>' must be an integer type whose bit width is a power
7190 of two greater than or equal to eight and less than or equal to a
7191 target-specific size limit. The type of the '``<pointer>``' operand must
7192 be a pointer to that type. If the ``atomicrmw`` is marked as
7193 ``volatile``, then the optimizer is not allowed to modify the number or
7194 order of execution of this ``atomicrmw`` with other :ref:`volatile
7195 operations <volatile>`.
7200 The contents of memory at the location specified by the '``<pointer>``'
7201 operand are atomically read, modified, and written back. The original
7202 value at the location is returned. The modification is specified by the
7205 - xchg: ``*ptr = val``
7206 - add: ``*ptr = *ptr + val``
7207 - sub: ``*ptr = *ptr - val``
7208 - and: ``*ptr = *ptr & val``
7209 - nand: ``*ptr = ~(*ptr & val)``
7210 - or: ``*ptr = *ptr | val``
7211 - xor: ``*ptr = *ptr ^ val``
7212 - max: ``*ptr = *ptr > val ? *ptr : val`` (using a signed comparison)
7213 - min: ``*ptr = *ptr < val ? *ptr : val`` (using a signed comparison)
7214 - umax: ``*ptr = *ptr > val ? *ptr : val`` (using an unsigned
7216 - umin: ``*ptr = *ptr < val ? *ptr : val`` (using an unsigned
7222 .. code-block:: llvm
7224 %old = atomicrmw add i32* %ptr, i32 1 acquire ; yields i32
7226 .. _i_getelementptr:
7228 '``getelementptr``' Instruction
7229 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7236 <result> = getelementptr <ty>, <ty>* <ptrval>{, <ty> <idx>}*
7237 <result> = getelementptr inbounds <ty>, <ty>* <ptrval>{, <ty> <idx>}*
7238 <result> = getelementptr <ty>, <ptr vector> <ptrval>, <vector index type> <idx>
7243 The '``getelementptr``' instruction is used to get the address of a
7244 subelement of an :ref:`aggregate <t_aggregate>` data structure. It performs
7245 address calculation only and does not access memory. The instruction can also
7246 be used to calculate a vector of such addresses.
7251 The first argument is always a type used as the basis for the calculations.
7252 The second argument is always a pointer or a vector of pointers, and is the
7253 base address to start from. The remaining arguments are indices
7254 that indicate which of the elements of the aggregate object are indexed.
7255 The interpretation of each index is dependent on the type being indexed
7256 into. The first index always indexes the pointer value given as the
7257 first argument, the second index indexes a value of the type pointed to
7258 (not necessarily the value directly pointed to, since the first index
7259 can be non-zero), etc. The first type indexed into must be a pointer
7260 value, subsequent types can be arrays, vectors, and structs. Note that
7261 subsequent types being indexed into can never be pointers, since that
7262 would require loading the pointer before continuing calculation.
7264 The type of each index argument depends on the type it is indexing into.
7265 When indexing into a (optionally packed) structure, only ``i32`` integer
7266 **constants** are allowed (when using a vector of indices they must all
7267 be the **same** ``i32`` integer constant). When indexing into an array,
7268 pointer or vector, integers of any width are allowed, and they are not
7269 required to be constant. These integers are treated as signed values
7272 For example, let's consider a C code fragment and how it gets compiled
7288 int *foo(struct ST *s) {
7289 return &s[1].Z.B[5][13];
7292 The LLVM code generated by Clang is:
7294 .. code-block:: llvm
7296 %struct.RT = type { i8, [10 x [20 x i32]], i8 }
7297 %struct.ST = type { i32, double, %struct.RT }
7299 define i32* @foo(%struct.ST* %s) nounwind uwtable readnone optsize ssp {
7301 %arrayidx = getelementptr inbounds %struct.ST, %struct.ST* %s, i64 1, i32 2, i32 1, i64 5, i64 13
7308 In the example above, the first index is indexing into the
7309 '``%struct.ST*``' type, which is a pointer, yielding a '``%struct.ST``'
7310 = '``{ i32, double, %struct.RT }``' type, a structure. The second index
7311 indexes into the third element of the structure, yielding a
7312 '``%struct.RT``' = '``{ i8 , [10 x [20 x i32]], i8 }``' type, another
7313 structure. The third index indexes into the second element of the
7314 structure, yielding a '``[10 x [20 x i32]]``' type, an array. The two
7315 dimensions of the array are subscripted into, yielding an '``i32``'
7316 type. The '``getelementptr``' instruction returns a pointer to this
7317 element, thus computing a value of '``i32*``' type.
7319 Note that it is perfectly legal to index partially through a structure,
7320 returning a pointer to an inner element. Because of this, the LLVM code
7321 for the given testcase is equivalent to:
7323 .. code-block:: llvm
7325 define i32* @foo(%struct.ST* %s) {
7326 %t1 = getelementptr %struct.ST, %struct.ST* %s, i32 1 ; yields %struct.ST*:%t1
7327 %t2 = getelementptr %struct.ST, %struct.ST* %t1, i32 0, i32 2 ; yields %struct.RT*:%t2
7328 %t3 = getelementptr %struct.RT, %struct.RT* %t2, i32 0, i32 1 ; yields [10 x [20 x i32]]*:%t3
7329 %t4 = getelementptr [10 x [20 x i32]], [10 x [20 x i32]]* %t3, i32 0, i32 5 ; yields [20 x i32]*:%t4
7330 %t5 = getelementptr [20 x i32], [20 x i32]* %t4, i32 0, i32 13 ; yields i32*:%t5
7334 If the ``inbounds`` keyword is present, the result value of the
7335 ``getelementptr`` is a :ref:`poison value <poisonvalues>` if the base
7336 pointer is not an *in bounds* address of an allocated object, or if any
7337 of the addresses that would be formed by successive addition of the
7338 offsets implied by the indices to the base address with infinitely
7339 precise signed arithmetic are not an *in bounds* address of that
7340 allocated object. The *in bounds* addresses for an allocated object are
7341 all the addresses that point into the object, plus the address one byte
7342 past the end. In cases where the base is a vector of pointers the
7343 ``inbounds`` keyword applies to each of the computations element-wise.
7345 If the ``inbounds`` keyword is not present, the offsets are added to the
7346 base address with silently-wrapping two's complement arithmetic. If the
7347 offsets have a different width from the pointer, they are sign-extended
7348 or truncated to the width of the pointer. The result value of the
7349 ``getelementptr`` may be outside the object pointed to by the base
7350 pointer. The result value may not necessarily be used to access memory
7351 though, even if it happens to point into allocated storage. See the
7352 :ref:`Pointer Aliasing Rules <pointeraliasing>` section for more
7355 The getelementptr instruction is often confusing. For some more insight
7356 into how it works, see :doc:`the getelementptr FAQ <GetElementPtr>`.
7361 .. code-block:: llvm
7363 ; yields [12 x i8]*:aptr
7364 %aptr = getelementptr {i32, [12 x i8]}, {i32, [12 x i8]}* %saptr, i64 0, i32 1
7366 %vptr = getelementptr {i32, <2 x i8>}, {i32, <2 x i8>}* %svptr, i64 0, i32 1, i32 1
7368 %eptr = getelementptr [12 x i8], [12 x i8]* %aptr, i64 0, i32 1
7370 %iptr = getelementptr [10 x i32], [10 x i32]* @arr, i16 0, i16 0
7375 The ``getelementptr`` returns a vector of pointers, instead of a single address,
7376 when one or more of its arguments is a vector. In such cases, all vector
7377 arguments should have the same number of elements, and every scalar argument
7378 will be effectively broadcast into a vector during address calculation.
7380 .. code-block:: llvm
7382 ; All arguments are vectors:
7383 ; A[i] = ptrs[i] + offsets[i]*sizeof(i8)
7384 %A = getelementptr i8, <4 x i8*> %ptrs, <4 x i64> %offsets
7386 ; Add the same scalar offset to each pointer of a vector:
7387 ; A[i] = ptrs[i] + offset*sizeof(i8)
7388 %A = getelementptr i8, <4 x i8*> %ptrs, i64 %offset
7390 ; Add distinct offsets to the same pointer:
7391 ; A[i] = ptr + offsets[i]*sizeof(i8)
7392 %A = getelementptr i8, i8* %ptr, <4 x i64> %offsets
7394 ; In all cases described above the type of the result is <4 x i8*>
7396 The two following instructions are equivalent:
7398 .. code-block:: llvm
7400 getelementptr %struct.ST, <4 x %struct.ST*> %s, <4 x i64> %ind1,
7401 <4 x i32> <i32 2, i32 2, i32 2, i32 2>,
7402 <4 x i32> <i32 1, i32 1, i32 1, i32 1>,
7404 <4 x i64> <i64 13, i64 13, i64 13, i64 13>
7406 getelementptr %struct.ST, <4 x %struct.ST*> %s, <4 x i64> %ind1,
7407 i32 2, i32 1, <4 x i32> %ind4, i64 13
7409 Let's look at the C code, where the vector version of ``getelementptr``
7414 // Let's assume that we vectorize the following loop:
7415 double *A, B; int *C;
7416 for (int i = 0; i < size; ++i) {
7420 .. code-block:: llvm
7422 ; get pointers for 8 elements from array B
7423 %ptrs = getelementptr double, double* %B, <8 x i32> %C
7424 ; load 8 elements from array B into A
7425 %A = call <8 x double> @llvm.masked.gather.v8f64(<8 x double*> %ptrs,
7426 i32 8, <8 x i1> %mask, <8 x double> %passthru)
7428 Conversion Operations
7429 ---------------------
7431 The instructions in this category are the conversion instructions
7432 (casting) which all take a single operand and a type. They perform
7433 various bit conversions on the operand.
7435 '``trunc .. to``' Instruction
7436 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7443 <result> = trunc <ty> <value> to <ty2> ; yields ty2
7448 The '``trunc``' instruction truncates its operand to the type ``ty2``.
7453 The '``trunc``' instruction takes a value to trunc, and a type to trunc
7454 it to. Both types must be of :ref:`integer <t_integer>` types, or vectors
7455 of the same number of integers. The bit size of the ``value`` must be
7456 larger than the bit size of the destination type, ``ty2``. Equal sized
7457 types are not allowed.
7462 The '``trunc``' instruction truncates the high order bits in ``value``
7463 and converts the remaining bits to ``ty2``. Since the source size must
7464 be larger than the destination size, ``trunc`` cannot be a *no-op cast*.
7465 It will always truncate bits.
7470 .. code-block:: llvm
7472 %X = trunc i32 257 to i8 ; yields i8:1
7473 %Y = trunc i32 123 to i1 ; yields i1:true
7474 %Z = trunc i32 122 to i1 ; yields i1:false
7475 %W = trunc <2 x i16> <i16 8, i16 7> to <2 x i8> ; yields <i8 8, i8 7>
7477 '``zext .. to``' Instruction
7478 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7485 <result> = zext <ty> <value> to <ty2> ; yields ty2
7490 The '``zext``' instruction zero extends its operand to type ``ty2``.
7495 The '``zext``' instruction takes a value to cast, and a type to cast it
7496 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
7497 the same number of integers. The bit size of the ``value`` must be
7498 smaller than the bit size of the destination type, ``ty2``.
7503 The ``zext`` fills the high order bits of the ``value`` with zero bits
7504 until it reaches the size of the destination type, ``ty2``.
7506 When zero extending from i1, the result will always be either 0 or 1.
7511 .. code-block:: llvm
7513 %X = zext i32 257 to i64 ; yields i64:257
7514 %Y = zext i1 true to i32 ; yields i32:1
7515 %Z = zext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
7517 '``sext .. to``' Instruction
7518 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7525 <result> = sext <ty> <value> to <ty2> ; yields ty2
7530 The '``sext``' sign extends ``value`` to the type ``ty2``.
7535 The '``sext``' instruction takes a value to cast, and a type to cast it
7536 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
7537 the same number of integers. The bit size of the ``value`` must be
7538 smaller than the bit size of the destination type, ``ty2``.
7543 The '``sext``' instruction performs a sign extension by copying the sign
7544 bit (highest order bit) of the ``value`` until it reaches the bit size
7545 of the type ``ty2``.
7547 When sign extending from i1, the extension always results in -1 or 0.
7552 .. code-block:: llvm
7554 %X = sext i8 -1 to i16 ; yields i16 :65535
7555 %Y = sext i1 true to i32 ; yields i32:-1
7556 %Z = sext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
7558 '``fptrunc .. to``' Instruction
7559 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7566 <result> = fptrunc <ty> <value> to <ty2> ; yields ty2
7571 The '``fptrunc``' instruction truncates ``value`` to type ``ty2``.
7576 The '``fptrunc``' instruction takes a :ref:`floating point <t_floating>`
7577 value to cast and a :ref:`floating point <t_floating>` type to cast it to.
7578 The size of ``value`` must be larger than the size of ``ty2``. This
7579 implies that ``fptrunc`` cannot be used to make a *no-op cast*.
7584 The '``fptrunc``' instruction casts a ``value`` from a larger
7585 :ref:`floating point <t_floating>` type to a smaller :ref:`floating
7586 point <t_floating>` type. If the value cannot fit (i.e. overflows) within the
7587 destination type, ``ty2``, then the results are undefined. If the cast produces
7588 an inexact result, how rounding is performed (e.g. truncation, also known as
7589 round to zero) is undefined.
7594 .. code-block:: llvm
7596 %X = fptrunc double 123.0 to float ; yields float:123.0
7597 %Y = fptrunc double 1.0E+300 to float ; yields undefined
7599 '``fpext .. to``' Instruction
7600 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7607 <result> = fpext <ty> <value> to <ty2> ; yields ty2
7612 The '``fpext``' extends a floating point ``value`` to a larger floating
7618 The '``fpext``' instruction takes a :ref:`floating point <t_floating>`
7619 ``value`` to cast, and a :ref:`floating point <t_floating>` type to cast it
7620 to. The source type must be smaller than the destination type.
7625 The '``fpext``' instruction extends the ``value`` from a smaller
7626 :ref:`floating point <t_floating>` type to a larger :ref:`floating
7627 point <t_floating>` type. The ``fpext`` cannot be used to make a
7628 *no-op cast* because it always changes bits. Use ``bitcast`` to make a
7629 *no-op cast* for a floating point cast.
7634 .. code-block:: llvm
7636 %X = fpext float 3.125 to double ; yields double:3.125000e+00
7637 %Y = fpext double %X to fp128 ; yields fp128:0xL00000000000000004000900000000000
7639 '``fptoui .. to``' Instruction
7640 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7647 <result> = fptoui <ty> <value> to <ty2> ; yields ty2
7652 The '``fptoui``' converts a floating point ``value`` to its unsigned
7653 integer equivalent of type ``ty2``.
7658 The '``fptoui``' instruction takes a value to cast, which must be a
7659 scalar or vector :ref:`floating point <t_floating>` value, and a type to
7660 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
7661 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
7662 type with the same number of elements as ``ty``
7667 The '``fptoui``' instruction converts its :ref:`floating
7668 point <t_floating>` operand into the nearest (rounding towards zero)
7669 unsigned integer value. If the value cannot fit in ``ty2``, the results
7675 .. code-block:: llvm
7677 %X = fptoui double 123.0 to i32 ; yields i32:123
7678 %Y = fptoui float 1.0E+300 to i1 ; yields undefined:1
7679 %Z = fptoui float 1.04E+17 to i8 ; yields undefined:1
7681 '``fptosi .. to``' Instruction
7682 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7689 <result> = fptosi <ty> <value> to <ty2> ; yields ty2
7694 The '``fptosi``' instruction converts :ref:`floating point <t_floating>`
7695 ``value`` to type ``ty2``.
7700 The '``fptosi``' instruction takes a value to cast, which must be a
7701 scalar or vector :ref:`floating point <t_floating>` value, and a type to
7702 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
7703 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
7704 type with the same number of elements as ``ty``
7709 The '``fptosi``' instruction converts its :ref:`floating
7710 point <t_floating>` operand into the nearest (rounding towards zero)
7711 signed integer value. If the value cannot fit in ``ty2``, the results
7717 .. code-block:: llvm
7719 %X = fptosi double -123.0 to i32 ; yields i32:-123
7720 %Y = fptosi float 1.0E-247 to i1 ; yields undefined:1
7721 %Z = fptosi float 1.04E+17 to i8 ; yields undefined:1
7723 '``uitofp .. to``' Instruction
7724 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7731 <result> = uitofp <ty> <value> to <ty2> ; yields ty2
7736 The '``uitofp``' instruction regards ``value`` as an unsigned integer
7737 and converts that value to the ``ty2`` type.
7742 The '``uitofp``' instruction takes a value to cast, which must be a
7743 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
7744 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
7745 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
7746 type with the same number of elements as ``ty``
7751 The '``uitofp``' instruction interprets its operand as an unsigned
7752 integer quantity and converts it to the corresponding floating point
7753 value. If the value cannot fit in the floating point value, the results
7759 .. code-block:: llvm
7761 %X = uitofp i32 257 to float ; yields float:257.0
7762 %Y = uitofp i8 -1 to double ; yields double:255.0
7764 '``sitofp .. to``' Instruction
7765 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7772 <result> = sitofp <ty> <value> to <ty2> ; yields ty2
7777 The '``sitofp``' instruction regards ``value`` as a signed integer and
7778 converts that value to the ``ty2`` type.
7783 The '``sitofp``' instruction takes a value to cast, which must be a
7784 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
7785 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
7786 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
7787 type with the same number of elements as ``ty``
7792 The '``sitofp``' instruction interprets its operand as a signed integer
7793 quantity and converts it to the corresponding floating point value. If
7794 the value cannot fit in the floating point value, the results are
7800 .. code-block:: llvm
7802 %X = sitofp i32 257 to float ; yields float:257.0
7803 %Y = sitofp i8 -1 to double ; yields double:-1.0
7807 '``ptrtoint .. to``' Instruction
7808 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7815 <result> = ptrtoint <ty> <value> to <ty2> ; yields ty2
7820 The '``ptrtoint``' instruction converts the pointer or a vector of
7821 pointers ``value`` to the integer (or vector of integers) type ``ty2``.
7826 The '``ptrtoint``' instruction takes a ``value`` to cast, which must be
7827 a value of type :ref:`pointer <t_pointer>` or a vector of pointers, and a
7828 type to cast it to ``ty2``, which must be an :ref:`integer <t_integer>` or
7829 a vector of integers type.
7834 The '``ptrtoint``' instruction converts ``value`` to integer type
7835 ``ty2`` by interpreting the pointer value as an integer and either
7836 truncating or zero extending that value to the size of the integer type.
7837 If ``value`` is smaller than ``ty2`` then a zero extension is done. If
7838 ``value`` is larger than ``ty2`` then a truncation is done. If they are
7839 the same size, then nothing is done (*no-op cast*) other than a type
7845 .. code-block:: llvm
7847 %X = ptrtoint i32* %P to i8 ; yields truncation on 32-bit architecture
7848 %Y = ptrtoint i32* %P to i64 ; yields zero extension on 32-bit architecture
7849 %Z = ptrtoint <4 x i32*> %P to <4 x i64>; yields vector zero extension for a vector of addresses on 32-bit architecture
7853 '``inttoptr .. to``' Instruction
7854 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7861 <result> = inttoptr <ty> <value> to <ty2> ; yields ty2
7866 The '``inttoptr``' instruction converts an integer ``value`` to a
7867 pointer type, ``ty2``.
7872 The '``inttoptr``' instruction takes an :ref:`integer <t_integer>` value to
7873 cast, and a type to cast it to, which must be a :ref:`pointer <t_pointer>`
7879 The '``inttoptr``' instruction converts ``value`` to type ``ty2`` by
7880 applying either a zero extension or a truncation depending on the size
7881 of the integer ``value``. If ``value`` is larger than the size of a
7882 pointer then a truncation is done. If ``value`` is smaller than the size
7883 of a pointer then a zero extension is done. If they are the same size,
7884 nothing is done (*no-op cast*).
7889 .. code-block:: llvm
7891 %X = inttoptr i32 255 to i32* ; yields zero extension on 64-bit architecture
7892 %Y = inttoptr i32 255 to i32* ; yields no-op on 32-bit architecture
7893 %Z = inttoptr i64 0 to i32* ; yields truncation on 32-bit architecture
7894 %Z = inttoptr <4 x i32> %G to <4 x i8*>; yields truncation of vector G to four pointers
7898 '``bitcast .. to``' Instruction
7899 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7906 <result> = bitcast <ty> <value> to <ty2> ; yields ty2
7911 The '``bitcast``' instruction converts ``value`` to type ``ty2`` without
7917 The '``bitcast``' instruction takes a value to cast, which must be a
7918 non-aggregate first class value, and a type to cast it to, which must
7919 also be a non-aggregate :ref:`first class <t_firstclass>` type. The
7920 bit sizes of ``value`` and the destination type, ``ty2``, must be
7921 identical. If the source type is a pointer, the destination type must
7922 also be a pointer of the same size. This instruction supports bitwise
7923 conversion of vectors to integers and to vectors of other types (as
7924 long as they have the same size).
7929 The '``bitcast``' instruction converts ``value`` to type ``ty2``. It
7930 is always a *no-op cast* because no bits change with this
7931 conversion. The conversion is done as if the ``value`` had been stored
7932 to memory and read back as type ``ty2``. Pointer (or vector of
7933 pointers) types may only be converted to other pointer (or vector of
7934 pointers) types with the same address space through this instruction.
7935 To convert pointers to other types, use the :ref:`inttoptr <i_inttoptr>`
7936 or :ref:`ptrtoint <i_ptrtoint>` instructions first.
7941 .. code-block:: llvm
7943 %X = bitcast i8 255 to i8 ; yields i8 :-1
7944 %Y = bitcast i32* %x to sint* ; yields sint*:%x
7945 %Z = bitcast <2 x int> %V to i64; ; yields i64: %V
7946 %Z = bitcast <2 x i32*> %V to <2 x i64*> ; yields <2 x i64*>
7948 .. _i_addrspacecast:
7950 '``addrspacecast .. to``' Instruction
7951 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7958 <result> = addrspacecast <pty> <ptrval> to <pty2> ; yields pty2
7963 The '``addrspacecast``' instruction converts ``ptrval`` from ``pty`` in
7964 address space ``n`` to type ``pty2`` in address space ``m``.
7969 The '``addrspacecast``' instruction takes a pointer or vector of pointer value
7970 to cast and a pointer type to cast it to, which must have a different
7976 The '``addrspacecast``' instruction converts the pointer value
7977 ``ptrval`` to type ``pty2``. It can be a *no-op cast* or a complex
7978 value modification, depending on the target and the address space
7979 pair. Pointer conversions within the same address space must be
7980 performed with the ``bitcast`` instruction. Note that if the address space
7981 conversion is legal then both result and operand refer to the same memory
7987 .. code-block:: llvm
7989 %X = addrspacecast i32* %x to i32 addrspace(1)* ; yields i32 addrspace(1)*:%x
7990 %Y = addrspacecast i32 addrspace(1)* %y to i64 addrspace(2)* ; yields i64 addrspace(2)*:%y
7991 %Z = addrspacecast <4 x i32*> %z to <4 x float addrspace(3)*> ; yields <4 x float addrspace(3)*>:%z
7998 The instructions in this category are the "miscellaneous" instructions,
7999 which defy better classification.
8003 '``icmp``' Instruction
8004 ^^^^^^^^^^^^^^^^^^^^^^
8011 <result> = icmp <cond> <ty> <op1>, <op2> ; yields i1 or <N x i1>:result
8016 The '``icmp``' instruction returns a boolean value or a vector of
8017 boolean values based on comparison of its two integer, integer vector,
8018 pointer, or pointer vector operands.
8023 The '``icmp``' instruction takes three operands. The first operand is
8024 the condition code indicating the kind of comparison to perform. It is
8025 not a value, just a keyword. The possible condition code are:
8028 #. ``ne``: not equal
8029 #. ``ugt``: unsigned greater than
8030 #. ``uge``: unsigned greater or equal
8031 #. ``ult``: unsigned less than
8032 #. ``ule``: unsigned less or equal
8033 #. ``sgt``: signed greater than
8034 #. ``sge``: signed greater or equal
8035 #. ``slt``: signed less than
8036 #. ``sle``: signed less or equal
8038 The remaining two arguments must be :ref:`integer <t_integer>` or
8039 :ref:`pointer <t_pointer>` or integer :ref:`vector <t_vector>` typed. They
8040 must also be identical types.
8045 The '``icmp``' compares ``op1`` and ``op2`` according to the condition
8046 code given as ``cond``. The comparison performed always yields either an
8047 :ref:`i1 <t_integer>` or vector of ``i1`` result, as follows:
8049 #. ``eq``: yields ``true`` if the operands are equal, ``false``
8050 otherwise. No sign interpretation is necessary or performed.
8051 #. ``ne``: yields ``true`` if the operands are unequal, ``false``
8052 otherwise. No sign interpretation is necessary or performed.
8053 #. ``ugt``: interprets the operands as unsigned values and yields
8054 ``true`` if ``op1`` is greater than ``op2``.
8055 #. ``uge``: interprets the operands as unsigned values and yields
8056 ``true`` if ``op1`` is greater than or equal to ``op2``.
8057 #. ``ult``: interprets the operands as unsigned values and yields
8058 ``true`` if ``op1`` is less than ``op2``.
8059 #. ``ule``: interprets the operands as unsigned values and yields
8060 ``true`` if ``op1`` is less than or equal to ``op2``.
8061 #. ``sgt``: interprets the operands as signed values and yields ``true``
8062 if ``op1`` is greater than ``op2``.
8063 #. ``sge``: interprets the operands as signed values and yields ``true``
8064 if ``op1`` is greater than or equal to ``op2``.
8065 #. ``slt``: interprets the operands as signed values and yields ``true``
8066 if ``op1`` is less than ``op2``.
8067 #. ``sle``: interprets the operands as signed values and yields ``true``
8068 if ``op1`` is less than or equal to ``op2``.
8070 If the operands are :ref:`pointer <t_pointer>` typed, the pointer values
8071 are compared as if they were integers.
8073 If the operands are integer vectors, then they are compared element by
8074 element. The result is an ``i1`` vector with the same number of elements
8075 as the values being compared. Otherwise, the result is an ``i1``.
8080 .. code-block:: llvm
8082 <result> = icmp eq i32 4, 5 ; yields: result=false
8083 <result> = icmp ne float* %X, %X ; yields: result=false
8084 <result> = icmp ult i16 4, 5 ; yields: result=true
8085 <result> = icmp sgt i16 4, 5 ; yields: result=false
8086 <result> = icmp ule i16 -4, 5 ; yields: result=false
8087 <result> = icmp sge i16 4, 5 ; yields: result=false
8089 Note that the code generator does not yet support vector types with the
8090 ``icmp`` instruction.
8094 '``fcmp``' Instruction
8095 ^^^^^^^^^^^^^^^^^^^^^^
8102 <result> = fcmp [fast-math flags]* <cond> <ty> <op1>, <op2> ; yields i1 or <N x i1>:result
8107 The '``fcmp``' instruction returns a boolean value or vector of boolean
8108 values based on comparison of its operands.
8110 If the operands are floating point scalars, then the result type is a
8111 boolean (:ref:`i1 <t_integer>`).
8113 If the operands are floating point vectors, then the result type is a
8114 vector of boolean with the same number of elements as the operands being
8120 The '``fcmp``' instruction takes three operands. The first operand is
8121 the condition code indicating the kind of comparison to perform. It is
8122 not a value, just a keyword. The possible condition code are:
8124 #. ``false``: no comparison, always returns false
8125 #. ``oeq``: ordered and equal
8126 #. ``ogt``: ordered and greater than
8127 #. ``oge``: ordered and greater than or equal
8128 #. ``olt``: ordered and less than
8129 #. ``ole``: ordered and less than or equal
8130 #. ``one``: ordered and not equal
8131 #. ``ord``: ordered (no nans)
8132 #. ``ueq``: unordered or equal
8133 #. ``ugt``: unordered or greater than
8134 #. ``uge``: unordered or greater than or equal
8135 #. ``ult``: unordered or less than
8136 #. ``ule``: unordered or less than or equal
8137 #. ``une``: unordered or not equal
8138 #. ``uno``: unordered (either nans)
8139 #. ``true``: no comparison, always returns true
8141 *Ordered* means that neither operand is a QNAN while *unordered* means
8142 that either operand may be a QNAN.
8144 Each of ``val1`` and ``val2`` arguments must be either a :ref:`floating
8145 point <t_floating>` type or a :ref:`vector <t_vector>` of floating point
8146 type. They must have identical types.
8151 The '``fcmp``' instruction compares ``op1`` and ``op2`` according to the
8152 condition code given as ``cond``. If the operands are vectors, then the
8153 vectors are compared element by element. Each comparison performed
8154 always yields an :ref:`i1 <t_integer>` result, as follows:
8156 #. ``false``: always yields ``false``, regardless of operands.
8157 #. ``oeq``: yields ``true`` if both operands are not a QNAN and ``op1``
8158 is equal to ``op2``.
8159 #. ``ogt``: yields ``true`` if both operands are not a QNAN and ``op1``
8160 is greater than ``op2``.
8161 #. ``oge``: yields ``true`` if both operands are not a QNAN and ``op1``
8162 is greater than or equal to ``op2``.
8163 #. ``olt``: yields ``true`` if both operands are not a QNAN and ``op1``
8164 is less than ``op2``.
8165 #. ``ole``: yields ``true`` if both operands are not a QNAN and ``op1``
8166 is less than or equal to ``op2``.
8167 #. ``one``: yields ``true`` if both operands are not a QNAN and ``op1``
8168 is not equal to ``op2``.
8169 #. ``ord``: yields ``true`` if both operands are not a QNAN.
8170 #. ``ueq``: yields ``true`` if either operand is a QNAN or ``op1`` is
8172 #. ``ugt``: yields ``true`` if either operand is a QNAN or ``op1`` is
8173 greater than ``op2``.
8174 #. ``uge``: yields ``true`` if either operand is a QNAN or ``op1`` is
8175 greater than or equal to ``op2``.
8176 #. ``ult``: yields ``true`` if either operand is a QNAN or ``op1`` is
8178 #. ``ule``: yields ``true`` if either operand is a QNAN or ``op1`` is
8179 less than or equal to ``op2``.
8180 #. ``une``: yields ``true`` if either operand is a QNAN or ``op1`` is
8181 not equal to ``op2``.
8182 #. ``uno``: yields ``true`` if either operand is a QNAN.
8183 #. ``true``: always yields ``true``, regardless of operands.
8185 The ``fcmp`` instruction can also optionally take any number of
8186 :ref:`fast-math flags <fastmath>`, which are optimization hints to enable
8187 otherwise unsafe floating point optimizations.
8189 Any set of fast-math flags are legal on an ``fcmp`` instruction, but the
8190 only flags that have any effect on its semantics are those that allow
8191 assumptions to be made about the values of input arguments; namely
8192 ``nnan``, ``ninf``, and ``nsz``. See :ref:`fastmath` for more information.
8197 .. code-block:: llvm
8199 <result> = fcmp oeq float 4.0, 5.0 ; yields: result=false
8200 <result> = fcmp one float 4.0, 5.0 ; yields: result=true
8201 <result> = fcmp olt float 4.0, 5.0 ; yields: result=true
8202 <result> = fcmp ueq double 1.0, 2.0 ; yields: result=false
8204 Note that the code generator does not yet support vector types with the
8205 ``fcmp`` instruction.
8209 '``phi``' Instruction
8210 ^^^^^^^^^^^^^^^^^^^^^
8217 <result> = phi <ty> [ <val0>, <label0>], ...
8222 The '``phi``' instruction is used to implement the φ node in the SSA
8223 graph representing the function.
8228 The type of the incoming values is specified with the first type field.
8229 After this, the '``phi``' instruction takes a list of pairs as
8230 arguments, with one pair for each predecessor basic block of the current
8231 block. Only values of :ref:`first class <t_firstclass>` type may be used as
8232 the value arguments to the PHI node. Only labels may be used as the
8235 There must be no non-phi instructions between the start of a basic block
8236 and the PHI instructions: i.e. PHI instructions must be first in a basic
8239 For the purposes of the SSA form, the use of each incoming value is
8240 deemed to occur on the edge from the corresponding predecessor block to
8241 the current block (but after any definition of an '``invoke``'
8242 instruction's return value on the same edge).
8247 At runtime, the '``phi``' instruction logically takes on the value
8248 specified by the pair corresponding to the predecessor basic block that
8249 executed just prior to the current block.
8254 .. code-block:: llvm
8256 Loop: ; Infinite loop that counts from 0 on up...
8257 %indvar = phi i32 [ 0, %LoopHeader ], [ %nextindvar, %Loop ]
8258 %nextindvar = add i32 %indvar, 1
8263 '``select``' Instruction
8264 ^^^^^^^^^^^^^^^^^^^^^^^^
8271 <result> = select selty <cond>, <ty> <val1>, <ty> <val2> ; yields ty
8273 selty is either i1 or {<N x i1>}
8278 The '``select``' instruction is used to choose one value based on a
8279 condition, without IR-level branching.
8284 The '``select``' instruction requires an 'i1' value or a vector of 'i1'
8285 values indicating the condition, and two values of the same :ref:`first
8286 class <t_firstclass>` type.
8291 If the condition is an i1 and it evaluates to 1, the instruction returns
8292 the first value argument; otherwise, it returns the second value
8295 If the condition is a vector of i1, then the value arguments must be
8296 vectors of the same size, and the selection is done element by element.
8298 If the condition is an i1 and the value arguments are vectors of the
8299 same size, then an entire vector is selected.
8304 .. code-block:: llvm
8306 %X = select i1 true, i8 17, i8 42 ; yields i8:17
8310 '``call``' Instruction
8311 ^^^^^^^^^^^^^^^^^^^^^^
8318 <result> = [tail | musttail | notail ] call [fast-math flags] [cconv] [ret attrs] <ty> [<fnty>*] <fnptrval>(<function args>) [fn attrs]
8324 The '``call``' instruction represents a simple function call.
8329 This instruction requires several arguments:
8331 #. The optional ``tail`` and ``musttail`` markers indicate that the optimizers
8332 should perform tail call optimization. The ``tail`` marker is a hint that
8333 `can be ignored <CodeGenerator.html#sibcallopt>`_. The ``musttail`` marker
8334 means that the call must be tail call optimized in order for the program to
8335 be correct. The ``musttail`` marker provides these guarantees:
8337 #. The call will not cause unbounded stack growth if it is part of a
8338 recursive cycle in the call graph.
8339 #. Arguments with the :ref:`inalloca <attr_inalloca>` attribute are
8342 Both markers imply that the callee does not access allocas or varargs from
8343 the caller. Calls marked ``musttail`` must obey the following additional
8346 - The call must immediately precede a :ref:`ret <i_ret>` instruction,
8347 or a pointer bitcast followed by a ret instruction.
8348 - The ret instruction must return the (possibly bitcasted) value
8349 produced by the call or void.
8350 - The caller and callee prototypes must match. Pointer types of
8351 parameters or return types may differ in pointee type, but not
8353 - The calling conventions of the caller and callee must match.
8354 - All ABI-impacting function attributes, such as sret, byval, inreg,
8355 returned, and inalloca, must match.
8356 - The callee must be varargs iff the caller is varargs. Bitcasting a
8357 non-varargs function to the appropriate varargs type is legal so
8358 long as the non-varargs prefixes obey the other rules.
8360 Tail call optimization for calls marked ``tail`` is guaranteed to occur if
8361 the following conditions are met:
8363 - Caller and callee both have the calling convention ``fastcc``.
8364 - The call is in tail position (ret immediately follows call and ret
8365 uses value of call or is void).
8366 - Option ``-tailcallopt`` is enabled, or
8367 ``llvm::GuaranteedTailCallOpt`` is ``true``.
8368 - `Platform-specific constraints are
8369 met. <CodeGenerator.html#tailcallopt>`_
8371 #. The optional ``notail`` marker indicates that the optimizers should not add
8372 ``tail`` or ``musttail`` markers to the call. It is used to prevent tail
8373 call optimization from being performed on the call.
8375 #. The optional ``fast-math flags`` marker indicates that the call has one or more
8376 :ref:`fast-math flags <fastmath>`, which are optimization hints to enable
8377 otherwise unsafe floating-point optimizations. Fast-math flags are only valid
8378 for calls that return a floating-point scalar or vector type.
8380 #. The optional "cconv" marker indicates which :ref:`calling
8381 convention <callingconv>` the call should use. If none is
8382 specified, the call defaults to using C calling conventions. The
8383 calling convention of the call must match the calling convention of
8384 the target function, or else the behavior is undefined.
8385 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
8386 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
8388 #. '``ty``': the type of the call instruction itself which is also the
8389 type of the return value. Functions that return no value are marked
8391 #. '``fnty``': shall be the signature of the pointer to function value
8392 being invoked. The argument types must match the types implied by
8393 this signature. This type can be omitted if the function is not
8394 varargs and if the function type does not return a pointer to a
8396 #. '``fnptrval``': An LLVM value containing a pointer to a function to
8397 be invoked. In most cases, this is a direct function invocation, but
8398 indirect ``call``'s are just as possible, calling an arbitrary pointer
8400 #. '``function args``': argument list whose types match the function
8401 signature argument types and parameter attributes. All arguments must
8402 be of :ref:`first class <t_firstclass>` type. If the function signature
8403 indicates the function accepts a variable number of arguments, the
8404 extra arguments can be specified.
8405 #. The optional :ref:`function attributes <fnattrs>` list. Only
8406 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
8407 attributes are valid here.
8408 #. The optional :ref:`operand bundles <opbundles>` list.
8413 The '``call``' instruction is used to cause control flow to transfer to
8414 a specified function, with its incoming arguments bound to the specified
8415 values. Upon a '``ret``' instruction in the called function, control
8416 flow continues with the instruction after the function call, and the
8417 return value of the function is bound to the result argument.
8422 .. code-block:: llvm
8424 %retval = call i32 @test(i32 %argc)
8425 call i32 (i8*, ...)* @printf(i8* %msg, i32 12, i8 42) ; yields i32
8426 %X = tail call i32 @foo() ; yields i32
8427 %Y = tail call fastcc i32 @foo() ; yields i32
8428 call void %foo(i8 97 signext)
8430 %struct.A = type { i32, i8 }
8431 %r = call %struct.A @foo() ; yields { i32, i8 }
8432 %gr = extractvalue %struct.A %r, 0 ; yields i32
8433 %gr1 = extractvalue %struct.A %r, 1 ; yields i8
8434 %Z = call void @foo() noreturn ; indicates that %foo never returns normally
8435 %ZZ = call zeroext i32 @bar() ; Return value is %zero extended
8437 llvm treats calls to some functions with names and arguments that match
8438 the standard C99 library as being the C99 library functions, and may
8439 perform optimizations or generate code for them under that assumption.
8440 This is something we'd like to change in the future to provide better
8441 support for freestanding environments and non-C-based languages.
8445 '``va_arg``' Instruction
8446 ^^^^^^^^^^^^^^^^^^^^^^^^
8453 <resultval> = va_arg <va_list*> <arglist>, <argty>
8458 The '``va_arg``' instruction is used to access arguments passed through
8459 the "variable argument" area of a function call. It is used to implement
8460 the ``va_arg`` macro in C.
8465 This instruction takes a ``va_list*`` value and the type of the
8466 argument. It returns a value of the specified argument type and
8467 increments the ``va_list`` to point to the next argument. The actual
8468 type of ``va_list`` is target specific.
8473 The '``va_arg``' instruction loads an argument of the specified type
8474 from the specified ``va_list`` and causes the ``va_list`` to point to
8475 the next argument. For more information, see the variable argument
8476 handling :ref:`Intrinsic Functions <int_varargs>`.
8478 It is legal for this instruction to be called in a function which does
8479 not take a variable number of arguments, for example, the ``vfprintf``
8482 ``va_arg`` is an LLVM instruction instead of an :ref:`intrinsic
8483 function <intrinsics>` because it takes a type as an argument.
8488 See the :ref:`variable argument processing <int_varargs>` section.
8490 Note that the code generator does not yet fully support va\_arg on many
8491 targets. Also, it does not currently support va\_arg with aggregate
8492 types on any target.
8496 '``landingpad``' Instruction
8497 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8504 <resultval> = landingpad <resultty> <clause>+
8505 <resultval> = landingpad <resultty> cleanup <clause>*
8507 <clause> := catch <type> <value>
8508 <clause> := filter <array constant type> <array constant>
8513 The '``landingpad``' instruction is used by `LLVM's exception handling
8514 system <ExceptionHandling.html#overview>`_ to specify that a basic block
8515 is a landing pad --- one where the exception lands, and corresponds to the
8516 code found in the ``catch`` portion of a ``try``/``catch`` sequence. It
8517 defines values supplied by the :ref:`personality function <personalityfn>` upon
8518 re-entry to the function. The ``resultval`` has the type ``resultty``.
8524 ``cleanup`` flag indicates that the landing pad block is a cleanup.
8526 A ``clause`` begins with the clause type --- ``catch`` or ``filter`` --- and
8527 contains the global variable representing the "type" that may be caught
8528 or filtered respectively. Unlike the ``catch`` clause, the ``filter``
8529 clause takes an array constant as its argument. Use
8530 "``[0 x i8**] undef``" for a filter which cannot throw. The
8531 '``landingpad``' instruction must contain *at least* one ``clause`` or
8532 the ``cleanup`` flag.
8537 The '``landingpad``' instruction defines the values which are set by the
8538 :ref:`personality function <personalityfn>` upon re-entry to the function, and
8539 therefore the "result type" of the ``landingpad`` instruction. As with
8540 calling conventions, how the personality function results are
8541 represented in LLVM IR is target specific.
8543 The clauses are applied in order from top to bottom. If two
8544 ``landingpad`` instructions are merged together through inlining, the
8545 clauses from the calling function are appended to the list of clauses.
8546 When the call stack is being unwound due to an exception being thrown,
8547 the exception is compared against each ``clause`` in turn. If it doesn't
8548 match any of the clauses, and the ``cleanup`` flag is not set, then
8549 unwinding continues further up the call stack.
8551 The ``landingpad`` instruction has several restrictions:
8553 - A landing pad block is a basic block which is the unwind destination
8554 of an '``invoke``' instruction.
8555 - A landing pad block must have a '``landingpad``' instruction as its
8556 first non-PHI instruction.
8557 - There can be only one '``landingpad``' instruction within the landing
8559 - A basic block that is not a landing pad block may not include a
8560 '``landingpad``' instruction.
8565 .. code-block:: llvm
8567 ;; A landing pad which can catch an integer.
8568 %res = landingpad { i8*, i32 }
8570 ;; A landing pad that is a cleanup.
8571 %res = landingpad { i8*, i32 }
8573 ;; A landing pad which can catch an integer and can only throw a double.
8574 %res = landingpad { i8*, i32 }
8576 filter [1 x i8**] [@_ZTId]
8580 '``cleanuppad``' Instruction
8581 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8588 <resultval> = cleanuppad within <parent> [<args>*]
8593 The '``cleanuppad``' instruction is used by `LLVM's exception handling
8594 system <ExceptionHandling.html#overview>`_ to specify that a basic block
8595 is a cleanup block --- one where a personality routine attempts to
8596 transfer control to run cleanup actions.
8597 The ``args`` correspond to whatever additional
8598 information the :ref:`personality function <personalityfn>` requires to
8599 execute the cleanup.
8600 The ``resultval`` has the type :ref:`token <t_token>` and is used to
8601 match the ``cleanuppad`` to corresponding :ref:`cleanuprets <i_cleanupret>`.
8602 The ``parent`` argument is the token of the funclet that contains the
8603 ``cleanuppad`` instruction. If the ``cleanuppad`` is not inside a funclet,
8604 this operand may be the token ``none``.
8609 The instruction takes a list of arbitrary values which are interpreted
8610 by the :ref:`personality function <personalityfn>`.
8615 When the call stack is being unwound due to an exception being thrown,
8616 the :ref:`personality function <personalityfn>` transfers control to the
8617 ``cleanuppad`` with the aid of the personality-specific arguments.
8618 As with calling conventions, how the personality function results are
8619 represented in LLVM IR is target specific.
8621 The ``cleanuppad`` instruction has several restrictions:
8623 - A cleanup block is a basic block which is the unwind destination of
8624 an exceptional instruction.
8625 - A cleanup block must have a '``cleanuppad``' instruction as its
8626 first non-PHI instruction.
8627 - There can be only one '``cleanuppad``' instruction within the
8629 - A basic block that is not a cleanup block may not include a
8630 '``cleanuppad``' instruction.
8632 Executing a ``cleanuppad`` instruction constitutes "entering" that pad.
8633 The pad may then be "exited" in one of three ways:
8635 1) explicitly via a ``cleanupret`` that consumes it. Executing such a ``cleanupret``
8636 is undefined behavior if any descendant pads have been entered but not yet
8638 2) implicitly via a call (which unwinds all the way to the current function's caller),
8639 or via a ``catchswitch`` or a ``cleanupret`` that unwinds to caller.
8640 3) implicitly via an unwind edge whose destination EH pad isn't a descendant of
8641 the ``cleanuppad``. When the ``cleanuppad`` is exited in this manner, it is
8642 undefined behavior if the destination EH pad has a parent which is not an
8643 ancestor of the ``cleanuppad`` being exited.
8645 It is undefined behavior for the ``cleanuppad`` to exit via an unwind edge which
8646 does not transitively unwind to the same destination as a constituent
8652 .. code-block:: llvm
8654 %tok = cleanuppad within %cs []
8661 LLVM supports the notion of an "intrinsic function". These functions
8662 have well known names and semantics and are required to follow certain
8663 restrictions. Overall, these intrinsics represent an extension mechanism
8664 for the LLVM language that does not require changing all of the
8665 transformations in LLVM when adding to the language (or the bitcode
8666 reader/writer, the parser, etc...).
8668 Intrinsic function names must all start with an "``llvm.``" prefix. This
8669 prefix is reserved in LLVM for intrinsic names; thus, function names may
8670 not begin with this prefix. Intrinsic functions must always be external
8671 functions: you cannot define the body of intrinsic functions. Intrinsic
8672 functions may only be used in call or invoke instructions: it is illegal
8673 to take the address of an intrinsic function. Additionally, because
8674 intrinsic functions are part of the LLVM language, it is required if any
8675 are added that they be documented here.
8677 Some intrinsic functions can be overloaded, i.e., the intrinsic
8678 represents a family of functions that perform the same operation but on
8679 different data types. Because LLVM can represent over 8 million
8680 different integer types, overloading is used commonly to allow an
8681 intrinsic function to operate on any integer type. One or more of the
8682 argument types or the result type can be overloaded to accept any
8683 integer type. Argument types may also be defined as exactly matching a
8684 previous argument's type or the result type. This allows an intrinsic
8685 function which accepts multiple arguments, but needs all of them to be
8686 of the same type, to only be overloaded with respect to a single
8687 argument or the result.
8689 Overloaded intrinsics will have the names of its overloaded argument
8690 types encoded into its function name, each preceded by a period. Only
8691 those types which are overloaded result in a name suffix. Arguments
8692 whose type is matched against another type do not. For example, the
8693 ``llvm.ctpop`` function can take an integer of any width and returns an
8694 integer of exactly the same integer width. This leads to a family of
8695 functions such as ``i8 @llvm.ctpop.i8(i8 %val)`` and
8696 ``i29 @llvm.ctpop.i29(i29 %val)``. Only one type, the return type, is
8697 overloaded, and only one type suffix is required. Because the argument's
8698 type is matched against the return type, it does not require its own
8701 To learn how to add an intrinsic function, please see the `Extending
8702 LLVM Guide <ExtendingLLVM.html>`_.
8706 Variable Argument Handling Intrinsics
8707 -------------------------------------
8709 Variable argument support is defined in LLVM with the
8710 :ref:`va_arg <i_va_arg>` instruction and these three intrinsic
8711 functions. These functions are related to the similarly named macros
8712 defined in the ``<stdarg.h>`` header file.
8714 All of these functions operate on arguments that use a target-specific
8715 value type "``va_list``". The LLVM assembly language reference manual
8716 does not define what this type is, so all transformations should be
8717 prepared to handle these functions regardless of the type used.
8719 This example shows how the :ref:`va_arg <i_va_arg>` instruction and the
8720 variable argument handling intrinsic functions are used.
8722 .. code-block:: llvm
8724 ; This struct is different for every platform. For most platforms,
8725 ; it is merely an i8*.
8726 %struct.va_list = type { i8* }
8728 ; For Unix x86_64 platforms, va_list is the following struct:
8729 ; %struct.va_list = type { i32, i32, i8*, i8* }
8731 define i32 @test(i32 %X, ...) {
8732 ; Initialize variable argument processing
8733 %ap = alloca %struct.va_list
8734 %ap2 = bitcast %struct.va_list* %ap to i8*
8735 call void @llvm.va_start(i8* %ap2)
8737 ; Read a single integer argument
8738 %tmp = va_arg i8* %ap2, i32
8740 ; Demonstrate usage of llvm.va_copy and llvm.va_end
8742 %aq2 = bitcast i8** %aq to i8*
8743 call void @llvm.va_copy(i8* %aq2, i8* %ap2)
8744 call void @llvm.va_end(i8* %aq2)
8746 ; Stop processing of arguments.
8747 call void @llvm.va_end(i8* %ap2)
8751 declare void @llvm.va_start(i8*)
8752 declare void @llvm.va_copy(i8*, i8*)
8753 declare void @llvm.va_end(i8*)
8757 '``llvm.va_start``' Intrinsic
8758 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8765 declare void @llvm.va_start(i8* <arglist>)
8770 The '``llvm.va_start``' intrinsic initializes ``*<arglist>`` for
8771 subsequent use by ``va_arg``.
8776 The argument is a pointer to a ``va_list`` element to initialize.
8781 The '``llvm.va_start``' intrinsic works just like the ``va_start`` macro
8782 available in C. In a target-dependent way, it initializes the
8783 ``va_list`` element to which the argument points, so that the next call
8784 to ``va_arg`` will produce the first variable argument passed to the
8785 function. Unlike the C ``va_start`` macro, this intrinsic does not need
8786 to know the last argument of the function as the compiler can figure
8789 '``llvm.va_end``' Intrinsic
8790 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8797 declare void @llvm.va_end(i8* <arglist>)
8802 The '``llvm.va_end``' intrinsic destroys ``*<arglist>``, which has been
8803 initialized previously with ``llvm.va_start`` or ``llvm.va_copy``.
8808 The argument is a pointer to a ``va_list`` to destroy.
8813 The '``llvm.va_end``' intrinsic works just like the ``va_end`` macro
8814 available in C. In a target-dependent way, it destroys the ``va_list``
8815 element to which the argument points. Calls to
8816 :ref:`llvm.va_start <int_va_start>` and
8817 :ref:`llvm.va_copy <int_va_copy>` must be matched exactly with calls to
8822 '``llvm.va_copy``' Intrinsic
8823 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8830 declare void @llvm.va_copy(i8* <destarglist>, i8* <srcarglist>)
8835 The '``llvm.va_copy``' intrinsic copies the current argument position
8836 from the source argument list to the destination argument list.
8841 The first argument is a pointer to a ``va_list`` element to initialize.
8842 The second argument is a pointer to a ``va_list`` element to copy from.
8847 The '``llvm.va_copy``' intrinsic works just like the ``va_copy`` macro
8848 available in C. In a target-dependent way, it copies the source
8849 ``va_list`` element into the destination ``va_list`` element. This
8850 intrinsic is necessary because the `` llvm.va_start`` intrinsic may be
8851 arbitrarily complex and require, for example, memory allocation.
8853 Accurate Garbage Collection Intrinsics
8854 --------------------------------------
8856 LLVM's support for `Accurate Garbage Collection <GarbageCollection.html>`_
8857 (GC) requires the frontend to generate code containing appropriate intrinsic
8858 calls and select an appropriate GC strategy which knows how to lower these
8859 intrinsics in a manner which is appropriate for the target collector.
8861 These intrinsics allow identification of :ref:`GC roots on the
8862 stack <int_gcroot>`, as well as garbage collector implementations that
8863 require :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers.
8864 Frontends for type-safe garbage collected languages should generate
8865 these intrinsics to make use of the LLVM garbage collectors. For more
8866 details, see `Garbage Collection with LLVM <GarbageCollection.html>`_.
8868 Experimental Statepoint Intrinsics
8869 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8871 LLVM provides an second experimental set of intrinsics for describing garbage
8872 collection safepoints in compiled code. These intrinsics are an alternative
8873 to the ``llvm.gcroot`` intrinsics, but are compatible with the ones for
8874 :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers. The
8875 differences in approach are covered in the `Garbage Collection with LLVM
8876 <GarbageCollection.html>`_ documentation. The intrinsics themselves are
8877 described in :doc:`Statepoints`.
8881 '``llvm.gcroot``' Intrinsic
8882 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8889 declare void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
8894 The '``llvm.gcroot``' intrinsic declares the existence of a GC root to
8895 the code generator, and allows some metadata to be associated with it.
8900 The first argument specifies the address of a stack object that contains
8901 the root pointer. The second pointer (which must be either a constant or
8902 a global value address) contains the meta-data to be associated with the
8908 At runtime, a call to this intrinsic stores a null pointer into the
8909 "ptrloc" location. At compile-time, the code generator generates
8910 information to allow the runtime to find the pointer at GC safe points.
8911 The '``llvm.gcroot``' intrinsic may only be used in a function which
8912 :ref:`specifies a GC algorithm <gc>`.
8916 '``llvm.gcread``' Intrinsic
8917 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8924 declare i8* @llvm.gcread(i8* %ObjPtr, i8** %Ptr)
8929 The '``llvm.gcread``' intrinsic identifies reads of references from heap
8930 locations, allowing garbage collector implementations that require read
8936 The second argument is the address to read from, which should be an
8937 address allocated from the garbage collector. The first object is a
8938 pointer to the start of the referenced object, if needed by the language
8939 runtime (otherwise null).
8944 The '``llvm.gcread``' intrinsic has the same semantics as a load
8945 instruction, but may be replaced with substantially more complex code by
8946 the garbage collector runtime, as needed. The '``llvm.gcread``'
8947 intrinsic may only be used in a function which :ref:`specifies a GC
8952 '``llvm.gcwrite``' Intrinsic
8953 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8960 declare void @llvm.gcwrite(i8* %P1, i8* %Obj, i8** %P2)
8965 The '``llvm.gcwrite``' intrinsic identifies writes of references to heap
8966 locations, allowing garbage collector implementations that require write
8967 barriers (such as generational or reference counting collectors).
8972 The first argument is the reference to store, the second is the start of
8973 the object to store it to, and the third is the address of the field of
8974 Obj to store to. If the runtime does not require a pointer to the
8975 object, Obj may be null.
8980 The '``llvm.gcwrite``' intrinsic has the same semantics as a store
8981 instruction, but may be replaced with substantially more complex code by
8982 the garbage collector runtime, as needed. The '``llvm.gcwrite``'
8983 intrinsic may only be used in a function which :ref:`specifies a GC
8986 Code Generator Intrinsics
8987 -------------------------
8989 These intrinsics are provided by LLVM to expose special features that
8990 may only be implemented with code generator support.
8992 '``llvm.returnaddress``' Intrinsic
8993 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9000 declare i8 *@llvm.returnaddress(i32 <level>)
9005 The '``llvm.returnaddress``' intrinsic attempts to compute a
9006 target-specific value indicating the return address of the current
9007 function or one of its callers.
9012 The argument to this intrinsic indicates which function to return the
9013 address for. Zero indicates the calling function, one indicates its
9014 caller, etc. The argument is **required** to be a constant integer
9020 The '``llvm.returnaddress``' intrinsic either returns a pointer
9021 indicating the return address of the specified call frame, or zero if it
9022 cannot be identified. The value returned by this intrinsic is likely to
9023 be incorrect or 0 for arguments other than zero, so it should only be
9024 used for debugging purposes.
9026 Note that calling this intrinsic does not prevent function inlining or
9027 other aggressive transformations, so the value returned may not be that
9028 of the obvious source-language caller.
9030 '``llvm.frameaddress``' Intrinsic
9031 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9038 declare i8* @llvm.frameaddress(i32 <level>)
9043 The '``llvm.frameaddress``' intrinsic attempts to return the
9044 target-specific frame pointer value for the specified stack frame.
9049 The argument to this intrinsic indicates which function to return the
9050 frame pointer for. Zero indicates the calling function, one indicates
9051 its caller, etc. The argument is **required** to be a constant integer
9057 The '``llvm.frameaddress``' intrinsic either returns a pointer
9058 indicating the frame address of the specified call frame, or zero if it
9059 cannot be identified. The value returned by this intrinsic is likely to
9060 be incorrect or 0 for arguments other than zero, so it should only be
9061 used for debugging purposes.
9063 Note that calling this intrinsic does not prevent function inlining or
9064 other aggressive transformations, so the value returned may not be that
9065 of the obvious source-language caller.
9067 '``llvm.localescape``' and '``llvm.localrecover``' Intrinsics
9068 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9075 declare void @llvm.localescape(...)
9076 declare i8* @llvm.localrecover(i8* %func, i8* %fp, i32 %idx)
9081 The '``llvm.localescape``' intrinsic escapes offsets of a collection of static
9082 allocas, and the '``llvm.localrecover``' intrinsic applies those offsets to a
9083 live frame pointer to recover the address of the allocation. The offset is
9084 computed during frame layout of the caller of ``llvm.localescape``.
9089 All arguments to '``llvm.localescape``' must be pointers to static allocas or
9090 casts of static allocas. Each function can only call '``llvm.localescape``'
9091 once, and it can only do so from the entry block.
9093 The ``func`` argument to '``llvm.localrecover``' must be a constant
9094 bitcasted pointer to a function defined in the current module. The code
9095 generator cannot determine the frame allocation offset of functions defined in
9098 The ``fp`` argument to '``llvm.localrecover``' must be a frame pointer of a
9099 call frame that is currently live. The return value of '``llvm.localaddress``'
9100 is one way to produce such a value, but various runtimes also expose a suitable
9101 pointer in platform-specific ways.
9103 The ``idx`` argument to '``llvm.localrecover``' indicates which alloca passed to
9104 '``llvm.localescape``' to recover. It is zero-indexed.
9109 These intrinsics allow a group of functions to share access to a set of local
9110 stack allocations of a one parent function. The parent function may call the
9111 '``llvm.localescape``' intrinsic once from the function entry block, and the
9112 child functions can use '``llvm.localrecover``' to access the escaped allocas.
9113 The '``llvm.localescape``' intrinsic blocks inlining, as inlining changes where
9114 the escaped allocas are allocated, which would break attempts to use
9115 '``llvm.localrecover``'.
9117 .. _int_read_register:
9118 .. _int_write_register:
9120 '``llvm.read_register``' and '``llvm.write_register``' Intrinsics
9121 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9128 declare i32 @llvm.read_register.i32(metadata)
9129 declare i64 @llvm.read_register.i64(metadata)
9130 declare void @llvm.write_register.i32(metadata, i32 @value)
9131 declare void @llvm.write_register.i64(metadata, i64 @value)
9137 The '``llvm.read_register``' and '``llvm.write_register``' intrinsics
9138 provides access to the named register. The register must be valid on
9139 the architecture being compiled to. The type needs to be compatible
9140 with the register being read.
9145 The '``llvm.read_register``' intrinsic returns the current value of the
9146 register, where possible. The '``llvm.write_register``' intrinsic sets
9147 the current value of the register, where possible.
9149 This is useful to implement named register global variables that need
9150 to always be mapped to a specific register, as is common practice on
9151 bare-metal programs including OS kernels.
9153 The compiler doesn't check for register availability or use of the used
9154 register in surrounding code, including inline assembly. Because of that,
9155 allocatable registers are not supported.
9157 Warning: So far it only works with the stack pointer on selected
9158 architectures (ARM, AArch64, PowerPC and x86_64). Significant amount of
9159 work is needed to support other registers and even more so, allocatable
9164 '``llvm.stacksave``' Intrinsic
9165 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9172 declare i8* @llvm.stacksave()
9177 The '``llvm.stacksave``' intrinsic is used to remember the current state
9178 of the function stack, for use with
9179 :ref:`llvm.stackrestore <int_stackrestore>`. This is useful for
9180 implementing language features like scoped automatic variable sized
9186 This intrinsic returns a opaque pointer value that can be passed to
9187 :ref:`llvm.stackrestore <int_stackrestore>`. When an
9188 ``llvm.stackrestore`` intrinsic is executed with a value saved from
9189 ``llvm.stacksave``, it effectively restores the state of the stack to
9190 the state it was in when the ``llvm.stacksave`` intrinsic executed. In
9191 practice, this pops any :ref:`alloca <i_alloca>` blocks from the stack that
9192 were allocated after the ``llvm.stacksave`` was executed.
9194 .. _int_stackrestore:
9196 '``llvm.stackrestore``' Intrinsic
9197 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9204 declare void @llvm.stackrestore(i8* %ptr)
9209 The '``llvm.stackrestore``' intrinsic is used to restore the state of
9210 the function stack to the state it was in when the corresponding
9211 :ref:`llvm.stacksave <int_stacksave>` intrinsic executed. This is
9212 useful for implementing language features like scoped automatic variable
9213 sized arrays in C99.
9218 See the description for :ref:`llvm.stacksave <int_stacksave>`.
9220 .. _int_get_dynamic_area_offset:
9222 '``llvm.get.dynamic.area.offset``' Intrinsic
9223 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9230 declare i32 @llvm.get.dynamic.area.offset.i32()
9231 declare i64 @llvm.get.dynamic.area.offset.i64()
9236 The '``llvm.get.dynamic.area.offset.*``' intrinsic family is used to
9237 get the offset from native stack pointer to the address of the most
9238 recent dynamic alloca on the caller's stack. These intrinsics are
9239 intendend for use in combination with
9240 :ref:`llvm.stacksave <int_stacksave>` to get a
9241 pointer to the most recent dynamic alloca. This is useful, for example,
9242 for AddressSanitizer's stack unpoisoning routines.
9247 These intrinsics return a non-negative integer value that can be used to
9248 get the address of the most recent dynamic alloca, allocated by :ref:`alloca <i_alloca>`
9249 on the caller's stack. In particular, for targets where stack grows downwards,
9250 adding this offset to the native stack pointer would get the address of the most
9251 recent dynamic alloca. For targets where stack grows upwards, the situation is a bit more
9252 complicated, because substracting this value from stack pointer would get the address
9253 one past the end of the most recent dynamic alloca.
9255 Although for most targets `llvm.get.dynamic.area.offset <int_get_dynamic_area_offset>`
9256 returns just a zero, for others, such as PowerPC and PowerPC64, it returns a
9257 compile-time-known constant value.
9259 The return value type of :ref:`llvm.get.dynamic.area.offset <int_get_dynamic_area_offset>`
9260 must match the target's generic address space's (address space 0) pointer type.
9262 '``llvm.prefetch``' Intrinsic
9263 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9270 declare void @llvm.prefetch(i8* <address>, i32 <rw>, i32 <locality>, i32 <cache type>)
9275 The '``llvm.prefetch``' intrinsic is a hint to the code generator to
9276 insert a prefetch instruction if supported; otherwise, it is a noop.
9277 Prefetches have no effect on the behavior of the program but can change
9278 its performance characteristics.
9283 ``address`` is the address to be prefetched, ``rw`` is the specifier
9284 determining if the fetch should be for a read (0) or write (1), and
9285 ``locality`` is a temporal locality specifier ranging from (0) - no
9286 locality, to (3) - extremely local keep in cache. The ``cache type``
9287 specifies whether the prefetch is performed on the data (1) or
9288 instruction (0) cache. The ``rw``, ``locality`` and ``cache type``
9289 arguments must be constant integers.
9294 This intrinsic does not modify the behavior of the program. In
9295 particular, prefetches cannot trap and do not produce a value. On
9296 targets that support this intrinsic, the prefetch can provide hints to
9297 the processor cache for better performance.
9299 '``llvm.pcmarker``' Intrinsic
9300 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9307 declare void @llvm.pcmarker(i32 <id>)
9312 The '``llvm.pcmarker``' intrinsic is a method to export a Program
9313 Counter (PC) in a region of code to simulators and other tools. The
9314 method is target specific, but it is expected that the marker will use
9315 exported symbols to transmit the PC of the marker. The marker makes no
9316 guarantees that it will remain with any specific instruction after
9317 optimizations. It is possible that the presence of a marker will inhibit
9318 optimizations. The intended use is to be inserted after optimizations to
9319 allow correlations of simulation runs.
9324 ``id`` is a numerical id identifying the marker.
9329 This intrinsic does not modify the behavior of the program. Backends
9330 that do not support this intrinsic may ignore it.
9332 '``llvm.readcyclecounter``' Intrinsic
9333 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9340 declare i64 @llvm.readcyclecounter()
9345 The '``llvm.readcyclecounter``' intrinsic provides access to the cycle
9346 counter register (or similar low latency, high accuracy clocks) on those
9347 targets that support it. On X86, it should map to RDTSC. On Alpha, it
9348 should map to RPCC. As the backing counters overflow quickly (on the
9349 order of 9 seconds on alpha), this should only be used for small
9355 When directly supported, reading the cycle counter should not modify any
9356 memory. Implementations are allowed to either return a application
9357 specific value or a system wide value. On backends without support, this
9358 is lowered to a constant 0.
9360 Note that runtime support may be conditional on the privilege-level code is
9361 running at and the host platform.
9363 '``llvm.clear_cache``' Intrinsic
9364 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9371 declare void @llvm.clear_cache(i8*, i8*)
9376 The '``llvm.clear_cache``' intrinsic ensures visibility of modifications
9377 in the specified range to the execution unit of the processor. On
9378 targets with non-unified instruction and data cache, the implementation
9379 flushes the instruction cache.
9384 On platforms with coherent instruction and data caches (e.g. x86), this
9385 intrinsic is a nop. On platforms with non-coherent instruction and data
9386 cache (e.g. ARM, MIPS), the intrinsic is lowered either to appropriate
9387 instructions or a system call, if cache flushing requires special
9390 The default behavior is to emit a call to ``__clear_cache`` from the run
9393 This instrinsic does *not* empty the instruction pipeline. Modifications
9394 of the current function are outside the scope of the intrinsic.
9396 '``llvm.instrprof_increment``' Intrinsic
9397 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9404 declare void @llvm.instrprof_increment(i8* <name>, i64 <hash>,
9405 i32 <num-counters>, i32 <index>)
9410 The '``llvm.instrprof_increment``' intrinsic can be emitted by a
9411 frontend for use with instrumentation based profiling. These will be
9412 lowered by the ``-instrprof`` pass to generate execution counts of a
9418 The first argument is a pointer to a global variable containing the
9419 name of the entity being instrumented. This should generally be the
9420 (mangled) function name for a set of counters.
9422 The second argument is a hash value that can be used by the consumer
9423 of the profile data to detect changes to the instrumented source, and
9424 the third is the number of counters associated with ``name``. It is an
9425 error if ``hash`` or ``num-counters`` differ between two instances of
9426 ``instrprof_increment`` that refer to the same name.
9428 The last argument refers to which of the counters for ``name`` should
9429 be incremented. It should be a value between 0 and ``num-counters``.
9434 This intrinsic represents an increment of a profiling counter. It will
9435 cause the ``-instrprof`` pass to generate the appropriate data
9436 structures and the code to increment the appropriate value, in a
9437 format that can be written out by a compiler runtime and consumed via
9438 the ``llvm-profdata`` tool.
9440 '``llvm.instrprof_value_profile``' Intrinsic
9441 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9448 declare void @llvm.instrprof_value_profile(i8* <name>, i64 <hash>,
9449 i64 <value>, i32 <value_kind>,
9455 The '``llvm.instrprof_value_profile``' intrinsic can be emitted by a
9456 frontend for use with instrumentation based profiling. This will be
9457 lowered by the ``-instrprof`` pass to find out the target values,
9458 instrumented expressions take in a program at runtime.
9463 The first argument is a pointer to a global variable containing the
9464 name of the entity being instrumented. ``name`` should generally be the
9465 (mangled) function name for a set of counters.
9467 The second argument is a hash value that can be used by the consumer
9468 of the profile data to detect changes to the instrumented source. It
9469 is an error if ``hash`` differs between two instances of
9470 ``llvm.instrprof_*`` that refer to the same name.
9472 The third argument is the value of the expression being profiled. The profiled
9473 expression's value should be representable as an unsigned 64-bit value. The
9474 fourth argument represents the kind of value profiling that is being done. The
9475 supported value profiling kinds are enumerated through the
9476 ``InstrProfValueKind`` type declared in the
9477 ``<include/llvm/ProfileData/InstrProf.h>`` header file. The last argument is the
9478 index of the instrumented expression within ``name``. It should be >= 0.
9483 This intrinsic represents the point where a call to a runtime routine
9484 should be inserted for value profiling of target expressions. ``-instrprof``
9485 pass will generate the appropriate data structures and replace the
9486 ``llvm.instrprof_value_profile`` intrinsic with the call to the profile
9487 runtime library with proper arguments.
9489 Standard C Library Intrinsics
9490 -----------------------------
9492 LLVM provides intrinsics for a few important standard C library
9493 functions. These intrinsics allow source-language front-ends to pass
9494 information about the alignment of the pointer arguments to the code
9495 generator, providing opportunity for more efficient code generation.
9499 '``llvm.memcpy``' Intrinsic
9500 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9505 This is an overloaded intrinsic. You can use ``llvm.memcpy`` on any
9506 integer bit width and for different address spaces. Not all targets
9507 support all bit widths however.
9511 declare void @llvm.memcpy.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
9512 i32 <len>, i32 <align>, i1 <isvolatile>)
9513 declare void @llvm.memcpy.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
9514 i64 <len>, i32 <align>, i1 <isvolatile>)
9519 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
9520 source location to the destination location.
9522 Note that, unlike the standard libc function, the ``llvm.memcpy.*``
9523 intrinsics do not return a value, takes extra alignment/isvolatile
9524 arguments and the pointers can be in specified address spaces.
9529 The first argument is a pointer to the destination, the second is a
9530 pointer to the source. The third argument is an integer argument
9531 specifying the number of bytes to copy, the fourth argument is the
9532 alignment of the source and destination locations, and the fifth is a
9533 boolean indicating a volatile access.
9535 If the call to this intrinsic has an alignment value that is not 0 or 1,
9536 then the caller guarantees that both the source and destination pointers
9537 are aligned to that boundary.
9539 If the ``isvolatile`` parameter is ``true``, the ``llvm.memcpy`` call is
9540 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
9541 very cleanly specified and it is unwise to depend on it.
9546 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
9547 source location to the destination location, which are not allowed to
9548 overlap. It copies "len" bytes of memory over. If the argument is known
9549 to be aligned to some boundary, this can be specified as the fourth
9550 argument, otherwise it should be set to 0 or 1 (both meaning no alignment).
9552 '``llvm.memmove``' Intrinsic
9553 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9558 This is an overloaded intrinsic. You can use llvm.memmove on any integer
9559 bit width and for different address space. Not all targets support all
9564 declare void @llvm.memmove.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
9565 i32 <len>, i32 <align>, i1 <isvolatile>)
9566 declare void @llvm.memmove.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
9567 i64 <len>, i32 <align>, i1 <isvolatile>)
9572 The '``llvm.memmove.*``' intrinsics move a block of memory from the
9573 source location to the destination location. It is similar to the
9574 '``llvm.memcpy``' intrinsic but allows the two memory locations to
9577 Note that, unlike the standard libc function, the ``llvm.memmove.*``
9578 intrinsics do not return a value, takes extra alignment/isvolatile
9579 arguments and the pointers can be in specified address spaces.
9584 The first argument is a pointer to the destination, the second is a
9585 pointer to the source. The third argument is an integer argument
9586 specifying the number of bytes to copy, the fourth argument is the
9587 alignment of the source and destination locations, and the fifth is a
9588 boolean indicating a volatile access.
9590 If the call to this intrinsic has an alignment value that is not 0 or 1,
9591 then the caller guarantees that the source and destination pointers are
9592 aligned to that boundary.
9594 If the ``isvolatile`` parameter is ``true``, the ``llvm.memmove`` call
9595 is a :ref:`volatile operation <volatile>`. The detailed access behavior is
9596 not very cleanly specified and it is unwise to depend on it.
9601 The '``llvm.memmove.*``' intrinsics copy a block of memory from the
9602 source location to the destination location, which may overlap. It
9603 copies "len" bytes of memory over. If the argument is known to be
9604 aligned to some boundary, this can be specified as the fourth argument,
9605 otherwise it should be set to 0 or 1 (both meaning no alignment).
9607 '``llvm.memset.*``' Intrinsics
9608 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9613 This is an overloaded intrinsic. You can use llvm.memset on any integer
9614 bit width and for different address spaces. However, not all targets
9615 support all bit widths.
9619 declare void @llvm.memset.p0i8.i32(i8* <dest>, i8 <val>,
9620 i32 <len>, i32 <align>, i1 <isvolatile>)
9621 declare void @llvm.memset.p0i8.i64(i8* <dest>, i8 <val>,
9622 i64 <len>, i32 <align>, i1 <isvolatile>)
9627 The '``llvm.memset.*``' intrinsics fill a block of memory with a
9628 particular byte value.
9630 Note that, unlike the standard libc function, the ``llvm.memset``
9631 intrinsic does not return a value and takes extra alignment/volatile
9632 arguments. Also, the destination can be in an arbitrary address space.
9637 The first argument is a pointer to the destination to fill, the second
9638 is the byte value with which to fill it, the third argument is an
9639 integer argument specifying the number of bytes to fill, and the fourth
9640 argument is the known alignment of the destination location.
9642 If the call to this intrinsic has an alignment value that is not 0 or 1,
9643 then the caller guarantees that the destination pointer is aligned to
9646 If the ``isvolatile`` parameter is ``true``, the ``llvm.memset`` call is
9647 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
9648 very cleanly specified and it is unwise to depend on it.
9653 The '``llvm.memset.*``' intrinsics fill "len" bytes of memory starting
9654 at the destination location. If the argument is known to be aligned to
9655 some boundary, this can be specified as the fourth argument, otherwise
9656 it should be set to 0 or 1 (both meaning no alignment).
9658 '``llvm.sqrt.*``' Intrinsic
9659 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9664 This is an overloaded intrinsic. You can use ``llvm.sqrt`` on any
9665 floating point or vector of floating point type. Not all targets support
9670 declare float @llvm.sqrt.f32(float %Val)
9671 declare double @llvm.sqrt.f64(double %Val)
9672 declare x86_fp80 @llvm.sqrt.f80(x86_fp80 %Val)
9673 declare fp128 @llvm.sqrt.f128(fp128 %Val)
9674 declare ppc_fp128 @llvm.sqrt.ppcf128(ppc_fp128 %Val)
9679 The '``llvm.sqrt``' intrinsics return the sqrt of the specified operand,
9680 returning the same value as the libm '``sqrt``' functions would. Unlike
9681 ``sqrt`` in libm, however, ``llvm.sqrt`` has undefined behavior for
9682 negative numbers other than -0.0 (which allows for better optimization,
9683 because there is no need to worry about errno being set).
9684 ``llvm.sqrt(-0.0)`` is defined to return -0.0 like IEEE sqrt.
9689 The argument and return value are floating point numbers of the same
9695 This function returns the sqrt of the specified operand if it is a
9696 nonnegative floating point number.
9698 '``llvm.powi.*``' Intrinsic
9699 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9704 This is an overloaded intrinsic. You can use ``llvm.powi`` on any
9705 floating point or vector of floating point type. Not all targets support
9710 declare float @llvm.powi.f32(float %Val, i32 %power)
9711 declare double @llvm.powi.f64(double %Val, i32 %power)
9712 declare x86_fp80 @llvm.powi.f80(x86_fp80 %Val, i32 %power)
9713 declare fp128 @llvm.powi.f128(fp128 %Val, i32 %power)
9714 declare ppc_fp128 @llvm.powi.ppcf128(ppc_fp128 %Val, i32 %power)
9719 The '``llvm.powi.*``' intrinsics return the first operand raised to the
9720 specified (positive or negative) power. The order of evaluation of
9721 multiplications is not defined. When a vector of floating point type is
9722 used, the second argument remains a scalar integer value.
9727 The second argument is an integer power, and the first is a value to
9728 raise to that power.
9733 This function returns the first value raised to the second power with an
9734 unspecified sequence of rounding operations.
9736 '``llvm.sin.*``' Intrinsic
9737 ^^^^^^^^^^^^^^^^^^^^^^^^^^
9742 This is an overloaded intrinsic. You can use ``llvm.sin`` on any
9743 floating point or vector of floating point type. Not all targets support
9748 declare float @llvm.sin.f32(float %Val)
9749 declare double @llvm.sin.f64(double %Val)
9750 declare x86_fp80 @llvm.sin.f80(x86_fp80 %Val)
9751 declare fp128 @llvm.sin.f128(fp128 %Val)
9752 declare ppc_fp128 @llvm.sin.ppcf128(ppc_fp128 %Val)
9757 The '``llvm.sin.*``' intrinsics return the sine of the operand.
9762 The argument and return value are floating point numbers of the same
9768 This function returns the sine of the specified operand, returning the
9769 same values as the libm ``sin`` functions would, and handles error
9770 conditions in the same way.
9772 '``llvm.cos.*``' Intrinsic
9773 ^^^^^^^^^^^^^^^^^^^^^^^^^^
9778 This is an overloaded intrinsic. You can use ``llvm.cos`` on any
9779 floating point or vector of floating point type. Not all targets support
9784 declare float @llvm.cos.f32(float %Val)
9785 declare double @llvm.cos.f64(double %Val)
9786 declare x86_fp80 @llvm.cos.f80(x86_fp80 %Val)
9787 declare fp128 @llvm.cos.f128(fp128 %Val)
9788 declare ppc_fp128 @llvm.cos.ppcf128(ppc_fp128 %Val)
9793 The '``llvm.cos.*``' intrinsics return the cosine of the operand.
9798 The argument and return value are floating point numbers of the same
9804 This function returns the cosine of the specified operand, returning the
9805 same values as the libm ``cos`` functions would, and handles error
9806 conditions in the same way.
9808 '``llvm.pow.*``' Intrinsic
9809 ^^^^^^^^^^^^^^^^^^^^^^^^^^
9814 This is an overloaded intrinsic. You can use ``llvm.pow`` on any
9815 floating point or vector of floating point type. Not all targets support
9820 declare float @llvm.pow.f32(float %Val, float %Power)
9821 declare double @llvm.pow.f64(double %Val, double %Power)
9822 declare x86_fp80 @llvm.pow.f80(x86_fp80 %Val, x86_fp80 %Power)
9823 declare fp128 @llvm.pow.f128(fp128 %Val, fp128 %Power)
9824 declare ppc_fp128 @llvm.pow.ppcf128(ppc_fp128 %Val, ppc_fp128 Power)
9829 The '``llvm.pow.*``' intrinsics return the first operand raised to the
9830 specified (positive or negative) power.
9835 The second argument is a floating point power, and the first is a value
9836 to raise to that power.
9841 This function returns the first value raised to the second power,
9842 returning the same values as the libm ``pow`` functions would, and
9843 handles error conditions in the same way.
9845 '``llvm.exp.*``' Intrinsic
9846 ^^^^^^^^^^^^^^^^^^^^^^^^^^
9851 This is an overloaded intrinsic. You can use ``llvm.exp`` on any
9852 floating point or vector of floating point type. Not all targets support
9857 declare float @llvm.exp.f32(float %Val)
9858 declare double @llvm.exp.f64(double %Val)
9859 declare x86_fp80 @llvm.exp.f80(x86_fp80 %Val)
9860 declare fp128 @llvm.exp.f128(fp128 %Val)
9861 declare ppc_fp128 @llvm.exp.ppcf128(ppc_fp128 %Val)
9866 The '``llvm.exp.*``' intrinsics perform the exp function.
9871 The argument and return value are floating point numbers of the same
9877 This function returns the same values as the libm ``exp`` functions
9878 would, and handles error conditions in the same way.
9880 '``llvm.exp2.*``' Intrinsic
9881 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9886 This is an overloaded intrinsic. You can use ``llvm.exp2`` on any
9887 floating point or vector of floating point type. Not all targets support
9892 declare float @llvm.exp2.f32(float %Val)
9893 declare double @llvm.exp2.f64(double %Val)
9894 declare x86_fp80 @llvm.exp2.f80(x86_fp80 %Val)
9895 declare fp128 @llvm.exp2.f128(fp128 %Val)
9896 declare ppc_fp128 @llvm.exp2.ppcf128(ppc_fp128 %Val)
9901 The '``llvm.exp2.*``' intrinsics perform the exp2 function.
9906 The argument and return value are floating point numbers of the same
9912 This function returns the same values as the libm ``exp2`` functions
9913 would, and handles error conditions in the same way.
9915 '``llvm.log.*``' Intrinsic
9916 ^^^^^^^^^^^^^^^^^^^^^^^^^^
9921 This is an overloaded intrinsic. You can use ``llvm.log`` on any
9922 floating point or vector of floating point type. Not all targets support
9927 declare float @llvm.log.f32(float %Val)
9928 declare double @llvm.log.f64(double %Val)
9929 declare x86_fp80 @llvm.log.f80(x86_fp80 %Val)
9930 declare fp128 @llvm.log.f128(fp128 %Val)
9931 declare ppc_fp128 @llvm.log.ppcf128(ppc_fp128 %Val)
9936 The '``llvm.log.*``' intrinsics perform the log function.
9941 The argument and return value are floating point numbers of the same
9947 This function returns the same values as the libm ``log`` functions
9948 would, and handles error conditions in the same way.
9950 '``llvm.log10.*``' Intrinsic
9951 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9956 This is an overloaded intrinsic. You can use ``llvm.log10`` on any
9957 floating point or vector of floating point type. Not all targets support
9962 declare float @llvm.log10.f32(float %Val)
9963 declare double @llvm.log10.f64(double %Val)
9964 declare x86_fp80 @llvm.log10.f80(x86_fp80 %Val)
9965 declare fp128 @llvm.log10.f128(fp128 %Val)
9966 declare ppc_fp128 @llvm.log10.ppcf128(ppc_fp128 %Val)
9971 The '``llvm.log10.*``' intrinsics perform the log10 function.
9976 The argument and return value are floating point numbers of the same
9982 This function returns the same values as the libm ``log10`` functions
9983 would, and handles error conditions in the same way.
9985 '``llvm.log2.*``' Intrinsic
9986 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9991 This is an overloaded intrinsic. You can use ``llvm.log2`` on any
9992 floating point or vector of floating point type. Not all targets support
9997 declare float @llvm.log2.f32(float %Val)
9998 declare double @llvm.log2.f64(double %Val)
9999 declare x86_fp80 @llvm.log2.f80(x86_fp80 %Val)
10000 declare fp128 @llvm.log2.f128(fp128 %Val)
10001 declare ppc_fp128 @llvm.log2.ppcf128(ppc_fp128 %Val)
10006 The '``llvm.log2.*``' intrinsics perform the log2 function.
10011 The argument and return value are floating point numbers of the same
10017 This function returns the same values as the libm ``log2`` functions
10018 would, and handles error conditions in the same way.
10020 '``llvm.fma.*``' Intrinsic
10021 ^^^^^^^^^^^^^^^^^^^^^^^^^^
10026 This is an overloaded intrinsic. You can use ``llvm.fma`` on any
10027 floating point or vector of floating point type. Not all targets support
10032 declare float @llvm.fma.f32(float %a, float %b, float %c)
10033 declare double @llvm.fma.f64(double %a, double %b, double %c)
10034 declare x86_fp80 @llvm.fma.f80(x86_fp80 %a, x86_fp80 %b, x86_fp80 %c)
10035 declare fp128 @llvm.fma.f128(fp128 %a, fp128 %b, fp128 %c)
10036 declare ppc_fp128 @llvm.fma.ppcf128(ppc_fp128 %a, ppc_fp128 %b, ppc_fp128 %c)
10041 The '``llvm.fma.*``' intrinsics perform the fused multiply-add
10047 The argument and return value are floating point numbers of the same
10053 This function returns the same values as the libm ``fma`` functions
10054 would, and does not set errno.
10056 '``llvm.fabs.*``' Intrinsic
10057 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10062 This is an overloaded intrinsic. You can use ``llvm.fabs`` on any
10063 floating point or vector of floating point type. Not all targets support
10068 declare float @llvm.fabs.f32(float %Val)
10069 declare double @llvm.fabs.f64(double %Val)
10070 declare x86_fp80 @llvm.fabs.f80(x86_fp80 %Val)
10071 declare fp128 @llvm.fabs.f128(fp128 %Val)
10072 declare ppc_fp128 @llvm.fabs.ppcf128(ppc_fp128 %Val)
10077 The '``llvm.fabs.*``' intrinsics return the absolute value of the
10083 The argument and return value are floating point numbers of the same
10089 This function returns the same values as the libm ``fabs`` functions
10090 would, and handles error conditions in the same way.
10092 '``llvm.minnum.*``' Intrinsic
10093 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10098 This is an overloaded intrinsic. You can use ``llvm.minnum`` on any
10099 floating point or vector of floating point type. Not all targets support
10104 declare float @llvm.minnum.f32(float %Val0, float %Val1)
10105 declare double @llvm.minnum.f64(double %Val0, double %Val1)
10106 declare x86_fp80 @llvm.minnum.f80(x86_fp80 %Val0, x86_fp80 %Val1)
10107 declare fp128 @llvm.minnum.f128(fp128 %Val0, fp128 %Val1)
10108 declare ppc_fp128 @llvm.minnum.ppcf128(ppc_fp128 %Val0, ppc_fp128 %Val1)
10113 The '``llvm.minnum.*``' intrinsics return the minimum of the two
10120 The arguments and return value are floating point numbers of the same
10126 Follows the IEEE-754 semantics for minNum, which also match for libm's
10129 If either operand is a NaN, returns the other non-NaN operand. Returns
10130 NaN only if both operands are NaN. If the operands compare equal,
10131 returns a value that compares equal to both operands. This means that
10132 fmin(+/-0.0, +/-0.0) could return either -0.0 or 0.0.
10134 '``llvm.maxnum.*``' Intrinsic
10135 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10140 This is an overloaded intrinsic. You can use ``llvm.maxnum`` on any
10141 floating point or vector of floating point type. Not all targets support
10146 declare float @llvm.maxnum.f32(float %Val0, float %Val1l)
10147 declare double @llvm.maxnum.f64(double %Val0, double %Val1)
10148 declare x86_fp80 @llvm.maxnum.f80(x86_fp80 %Val0, x86_fp80 %Val1)
10149 declare fp128 @llvm.maxnum.f128(fp128 %Val0, fp128 %Val1)
10150 declare ppc_fp128 @llvm.maxnum.ppcf128(ppc_fp128 %Val0, ppc_fp128 %Val1)
10155 The '``llvm.maxnum.*``' intrinsics return the maximum of the two
10162 The arguments and return value are floating point numbers of the same
10167 Follows the IEEE-754 semantics for maxNum, which also match for libm's
10170 If either operand is a NaN, returns the other non-NaN operand. Returns
10171 NaN only if both operands are NaN. If the operands compare equal,
10172 returns a value that compares equal to both operands. This means that
10173 fmax(+/-0.0, +/-0.0) could return either -0.0 or 0.0.
10175 '``llvm.copysign.*``' Intrinsic
10176 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10181 This is an overloaded intrinsic. You can use ``llvm.copysign`` on any
10182 floating point or vector of floating point type. Not all targets support
10187 declare float @llvm.copysign.f32(float %Mag, float %Sgn)
10188 declare double @llvm.copysign.f64(double %Mag, double %Sgn)
10189 declare x86_fp80 @llvm.copysign.f80(x86_fp80 %Mag, x86_fp80 %Sgn)
10190 declare fp128 @llvm.copysign.f128(fp128 %Mag, fp128 %Sgn)
10191 declare ppc_fp128 @llvm.copysign.ppcf128(ppc_fp128 %Mag, ppc_fp128 %Sgn)
10196 The '``llvm.copysign.*``' intrinsics return a value with the magnitude of the
10197 first operand and the sign of the second operand.
10202 The arguments and return value are floating point numbers of the same
10208 This function returns the same values as the libm ``copysign``
10209 functions would, and handles error conditions in the same way.
10211 '``llvm.floor.*``' Intrinsic
10212 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10217 This is an overloaded intrinsic. You can use ``llvm.floor`` on any
10218 floating point or vector of floating point type. Not all targets support
10223 declare float @llvm.floor.f32(float %Val)
10224 declare double @llvm.floor.f64(double %Val)
10225 declare x86_fp80 @llvm.floor.f80(x86_fp80 %Val)
10226 declare fp128 @llvm.floor.f128(fp128 %Val)
10227 declare ppc_fp128 @llvm.floor.ppcf128(ppc_fp128 %Val)
10232 The '``llvm.floor.*``' intrinsics return the floor of the operand.
10237 The argument and return value are floating point numbers of the same
10243 This function returns the same values as the libm ``floor`` functions
10244 would, and handles error conditions in the same way.
10246 '``llvm.ceil.*``' Intrinsic
10247 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10252 This is an overloaded intrinsic. You can use ``llvm.ceil`` on any
10253 floating point or vector of floating point type. Not all targets support
10258 declare float @llvm.ceil.f32(float %Val)
10259 declare double @llvm.ceil.f64(double %Val)
10260 declare x86_fp80 @llvm.ceil.f80(x86_fp80 %Val)
10261 declare fp128 @llvm.ceil.f128(fp128 %Val)
10262 declare ppc_fp128 @llvm.ceil.ppcf128(ppc_fp128 %Val)
10267 The '``llvm.ceil.*``' intrinsics return the ceiling of the operand.
10272 The argument and return value are floating point numbers of the same
10278 This function returns the same values as the libm ``ceil`` functions
10279 would, and handles error conditions in the same way.
10281 '``llvm.trunc.*``' Intrinsic
10282 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10287 This is an overloaded intrinsic. You can use ``llvm.trunc`` on any
10288 floating point or vector of floating point type. Not all targets support
10293 declare float @llvm.trunc.f32(float %Val)
10294 declare double @llvm.trunc.f64(double %Val)
10295 declare x86_fp80 @llvm.trunc.f80(x86_fp80 %Val)
10296 declare fp128 @llvm.trunc.f128(fp128 %Val)
10297 declare ppc_fp128 @llvm.trunc.ppcf128(ppc_fp128 %Val)
10302 The '``llvm.trunc.*``' intrinsics returns the operand rounded to the
10303 nearest integer not larger in magnitude than the operand.
10308 The argument and return value are floating point numbers of the same
10314 This function returns the same values as the libm ``trunc`` functions
10315 would, and handles error conditions in the same way.
10317 '``llvm.rint.*``' Intrinsic
10318 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10323 This is an overloaded intrinsic. You can use ``llvm.rint`` on any
10324 floating point or vector of floating point type. Not all targets support
10329 declare float @llvm.rint.f32(float %Val)
10330 declare double @llvm.rint.f64(double %Val)
10331 declare x86_fp80 @llvm.rint.f80(x86_fp80 %Val)
10332 declare fp128 @llvm.rint.f128(fp128 %Val)
10333 declare ppc_fp128 @llvm.rint.ppcf128(ppc_fp128 %Val)
10338 The '``llvm.rint.*``' intrinsics returns the operand rounded to the
10339 nearest integer. It may raise an inexact floating-point exception if the
10340 operand isn't an integer.
10345 The argument and return value are floating point numbers of the same
10351 This function returns the same values as the libm ``rint`` functions
10352 would, and handles error conditions in the same way.
10354 '``llvm.nearbyint.*``' Intrinsic
10355 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10360 This is an overloaded intrinsic. You can use ``llvm.nearbyint`` on any
10361 floating point or vector of floating point type. Not all targets support
10366 declare float @llvm.nearbyint.f32(float %Val)
10367 declare double @llvm.nearbyint.f64(double %Val)
10368 declare x86_fp80 @llvm.nearbyint.f80(x86_fp80 %Val)
10369 declare fp128 @llvm.nearbyint.f128(fp128 %Val)
10370 declare ppc_fp128 @llvm.nearbyint.ppcf128(ppc_fp128 %Val)
10375 The '``llvm.nearbyint.*``' intrinsics returns the operand rounded to the
10381 The argument and return value are floating point numbers of the same
10387 This function returns the same values as the libm ``nearbyint``
10388 functions would, and handles error conditions in the same way.
10390 '``llvm.round.*``' Intrinsic
10391 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10396 This is an overloaded intrinsic. You can use ``llvm.round`` on any
10397 floating point or vector of floating point type. Not all targets support
10402 declare float @llvm.round.f32(float %Val)
10403 declare double @llvm.round.f64(double %Val)
10404 declare x86_fp80 @llvm.round.f80(x86_fp80 %Val)
10405 declare fp128 @llvm.round.f128(fp128 %Val)
10406 declare ppc_fp128 @llvm.round.ppcf128(ppc_fp128 %Val)
10411 The '``llvm.round.*``' intrinsics returns the operand rounded to the
10417 The argument and return value are floating point numbers of the same
10423 This function returns the same values as the libm ``round``
10424 functions would, and handles error conditions in the same way.
10426 Bit Manipulation Intrinsics
10427 ---------------------------
10429 LLVM provides intrinsics for a few important bit manipulation
10430 operations. These allow efficient code generation for some algorithms.
10432 '``llvm.bitreverse.*``' Intrinsics
10433 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10438 This is an overloaded intrinsic function. You can use bitreverse on any
10443 declare i16 @llvm.bitreverse.i16(i16 <id>)
10444 declare i32 @llvm.bitreverse.i32(i32 <id>)
10445 declare i64 @llvm.bitreverse.i64(i64 <id>)
10450 The '``llvm.bitreverse``' family of intrinsics is used to reverse the
10451 bitpattern of an integer value; for example ``0b1234567`` becomes
10457 The ``llvm.bitreverse.iN`` intrinsic returns an i16 value that has bit
10458 ``M`` in the input moved to bit ``N-M`` in the output.
10460 '``llvm.bswap.*``' Intrinsics
10461 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10466 This is an overloaded intrinsic function. You can use bswap on any
10467 integer type that is an even number of bytes (i.e. BitWidth % 16 == 0).
10471 declare i16 @llvm.bswap.i16(i16 <id>)
10472 declare i32 @llvm.bswap.i32(i32 <id>)
10473 declare i64 @llvm.bswap.i64(i64 <id>)
10478 The '``llvm.bswap``' family of intrinsics is used to byte swap integer
10479 values with an even number of bytes (positive multiple of 16 bits).
10480 These are useful for performing operations on data that is not in the
10481 target's native byte order.
10486 The ``llvm.bswap.i16`` intrinsic returns an i16 value that has the high
10487 and low byte of the input i16 swapped. Similarly, the ``llvm.bswap.i32``
10488 intrinsic returns an i32 value that has the four bytes of the input i32
10489 swapped, so that if the input bytes are numbered 0, 1, 2, 3 then the
10490 returned i32 will have its bytes in 3, 2, 1, 0 order. The
10491 ``llvm.bswap.i48``, ``llvm.bswap.i64`` and other intrinsics extend this
10492 concept to additional even-byte lengths (6 bytes, 8 bytes and more,
10495 '``llvm.ctpop.*``' Intrinsic
10496 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10501 This is an overloaded intrinsic. You can use llvm.ctpop on any integer
10502 bit width, or on any vector with integer elements. Not all targets
10503 support all bit widths or vector types, however.
10507 declare i8 @llvm.ctpop.i8(i8 <src>)
10508 declare i16 @llvm.ctpop.i16(i16 <src>)
10509 declare i32 @llvm.ctpop.i32(i32 <src>)
10510 declare i64 @llvm.ctpop.i64(i64 <src>)
10511 declare i256 @llvm.ctpop.i256(i256 <src>)
10512 declare <2 x i32> @llvm.ctpop.v2i32(<2 x i32> <src>)
10517 The '``llvm.ctpop``' family of intrinsics counts the number of bits set
10523 The only argument is the value to be counted. The argument may be of any
10524 integer type, or a vector with integer elements. The return type must
10525 match the argument type.
10530 The '``llvm.ctpop``' intrinsic counts the 1's in a variable, or within
10531 each element of a vector.
10533 '``llvm.ctlz.*``' Intrinsic
10534 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10539 This is an overloaded intrinsic. You can use ``llvm.ctlz`` on any
10540 integer bit width, or any vector whose elements are integers. Not all
10541 targets support all bit widths or vector types, however.
10545 declare i8 @llvm.ctlz.i8 (i8 <src>, i1 <is_zero_undef>)
10546 declare i16 @llvm.ctlz.i16 (i16 <src>, i1 <is_zero_undef>)
10547 declare i32 @llvm.ctlz.i32 (i32 <src>, i1 <is_zero_undef>)
10548 declare i64 @llvm.ctlz.i64 (i64 <src>, i1 <is_zero_undef>)
10549 declare i256 @llvm.ctlz.i256(i256 <src>, i1 <is_zero_undef>)
10550 declase <2 x i32> @llvm.ctlz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
10555 The '``llvm.ctlz``' family of intrinsic functions counts the number of
10556 leading zeros in a variable.
10561 The first argument is the value to be counted. This argument may be of
10562 any integer type, or a vector with integer element type. The return
10563 type must match the first argument type.
10565 The second argument must be a constant and is a flag to indicate whether
10566 the intrinsic should ensure that a zero as the first argument produces a
10567 defined result. Historically some architectures did not provide a
10568 defined result for zero values as efficiently, and many algorithms are
10569 now predicated on avoiding zero-value inputs.
10574 The '``llvm.ctlz``' intrinsic counts the leading (most significant)
10575 zeros in a variable, or within each element of the vector. If
10576 ``src == 0`` then the result is the size in bits of the type of ``src``
10577 if ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
10578 ``llvm.ctlz(i32 2) = 30``.
10580 '``llvm.cttz.*``' Intrinsic
10581 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10586 This is an overloaded intrinsic. You can use ``llvm.cttz`` on any
10587 integer bit width, or any vector of integer elements. Not all targets
10588 support all bit widths or vector types, however.
10592 declare i8 @llvm.cttz.i8 (i8 <src>, i1 <is_zero_undef>)
10593 declare i16 @llvm.cttz.i16 (i16 <src>, i1 <is_zero_undef>)
10594 declare i32 @llvm.cttz.i32 (i32 <src>, i1 <is_zero_undef>)
10595 declare i64 @llvm.cttz.i64 (i64 <src>, i1 <is_zero_undef>)
10596 declare i256 @llvm.cttz.i256(i256 <src>, i1 <is_zero_undef>)
10597 declase <2 x i32> @llvm.cttz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
10602 The '``llvm.cttz``' family of intrinsic functions counts the number of
10608 The first argument is the value to be counted. This argument may be of
10609 any integer type, or a vector with integer element type. The return
10610 type must match the first argument type.
10612 The second argument must be a constant and is a flag to indicate whether
10613 the intrinsic should ensure that a zero as the first argument produces a
10614 defined result. Historically some architectures did not provide a
10615 defined result for zero values as efficiently, and many algorithms are
10616 now predicated on avoiding zero-value inputs.
10621 The '``llvm.cttz``' intrinsic counts the trailing (least significant)
10622 zeros in a variable, or within each element of a vector. If ``src == 0``
10623 then the result is the size in bits of the type of ``src`` if
10624 ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
10625 ``llvm.cttz(2) = 1``.
10629 Arithmetic with Overflow Intrinsics
10630 -----------------------------------
10632 LLVM provides intrinsics for some arithmetic with overflow operations.
10634 '``llvm.sadd.with.overflow.*``' Intrinsics
10635 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10640 This is an overloaded intrinsic. You can use ``llvm.sadd.with.overflow``
10641 on any integer bit width.
10645 declare {i16, i1} @llvm.sadd.with.overflow.i16(i16 %a, i16 %b)
10646 declare {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
10647 declare {i64, i1} @llvm.sadd.with.overflow.i64(i64 %a, i64 %b)
10652 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
10653 a signed addition of the two arguments, and indicate whether an overflow
10654 occurred during the signed summation.
10659 The arguments (%a and %b) and the first element of the result structure
10660 may be of integer types of any bit width, but they must have the same
10661 bit width. The second element of the result structure must be of type
10662 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
10668 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
10669 a signed addition of the two variables. They return a structure --- the
10670 first element of which is the signed summation, and the second element
10671 of which is a bit specifying if the signed summation resulted in an
10677 .. code-block:: llvm
10679 %res = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
10680 %sum = extractvalue {i32, i1} %res, 0
10681 %obit = extractvalue {i32, i1} %res, 1
10682 br i1 %obit, label %overflow, label %normal
10684 '``llvm.uadd.with.overflow.*``' Intrinsics
10685 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10690 This is an overloaded intrinsic. You can use ``llvm.uadd.with.overflow``
10691 on any integer bit width.
10695 declare {i16, i1} @llvm.uadd.with.overflow.i16(i16 %a, i16 %b)
10696 declare {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
10697 declare {i64, i1} @llvm.uadd.with.overflow.i64(i64 %a, i64 %b)
10702 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
10703 an unsigned addition of the two arguments, and indicate whether a carry
10704 occurred during the unsigned summation.
10709 The arguments (%a and %b) and the first element of the result structure
10710 may be of integer types of any bit width, but they must have the same
10711 bit width. The second element of the result structure must be of type
10712 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
10718 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
10719 an unsigned addition of the two arguments. They return a structure --- the
10720 first element of which is the sum, and the second element of which is a
10721 bit specifying if the unsigned summation resulted in a carry.
10726 .. code-block:: llvm
10728 %res = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
10729 %sum = extractvalue {i32, i1} %res, 0
10730 %obit = extractvalue {i32, i1} %res, 1
10731 br i1 %obit, label %carry, label %normal
10733 '``llvm.ssub.with.overflow.*``' Intrinsics
10734 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10739 This is an overloaded intrinsic. You can use ``llvm.ssub.with.overflow``
10740 on any integer bit width.
10744 declare {i16, i1} @llvm.ssub.with.overflow.i16(i16 %a, i16 %b)
10745 declare {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
10746 declare {i64, i1} @llvm.ssub.with.overflow.i64(i64 %a, i64 %b)
10751 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
10752 a signed subtraction of the two arguments, and indicate whether an
10753 overflow occurred during the signed subtraction.
10758 The arguments (%a and %b) and the first element of the result structure
10759 may be of integer types of any bit width, but they must have the same
10760 bit width. The second element of the result structure must be of type
10761 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
10767 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
10768 a signed subtraction of the two arguments. They return a structure --- the
10769 first element of which is the subtraction, and the second element of
10770 which is a bit specifying if the signed subtraction resulted in an
10776 .. code-block:: llvm
10778 %res = call {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
10779 %sum = extractvalue {i32, i1} %res, 0
10780 %obit = extractvalue {i32, i1} %res, 1
10781 br i1 %obit, label %overflow, label %normal
10783 '``llvm.usub.with.overflow.*``' Intrinsics
10784 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10789 This is an overloaded intrinsic. You can use ``llvm.usub.with.overflow``
10790 on any integer bit width.
10794 declare {i16, i1} @llvm.usub.with.overflow.i16(i16 %a, i16 %b)
10795 declare {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
10796 declare {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
10801 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
10802 an unsigned subtraction of the two arguments, and indicate whether an
10803 overflow occurred during the unsigned subtraction.
10808 The arguments (%a and %b) and the first element of the result structure
10809 may be of integer types of any bit width, but they must have the same
10810 bit width. The second element of the result structure must be of type
10811 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
10817 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
10818 an unsigned subtraction of the two arguments. They return a structure ---
10819 the first element of which is the subtraction, and the second element of
10820 which is a bit specifying if the unsigned subtraction resulted in an
10826 .. code-block:: llvm
10828 %res = call {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
10829 %sum = extractvalue {i32, i1} %res, 0
10830 %obit = extractvalue {i32, i1} %res, 1
10831 br i1 %obit, label %overflow, label %normal
10833 '``llvm.smul.with.overflow.*``' Intrinsics
10834 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10839 This is an overloaded intrinsic. You can use ``llvm.smul.with.overflow``
10840 on any integer bit width.
10844 declare {i16, i1} @llvm.smul.with.overflow.i16(i16 %a, i16 %b)
10845 declare {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
10846 declare {i64, i1} @llvm.smul.with.overflow.i64(i64 %a, i64 %b)
10851 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
10852 a signed multiplication of the two arguments, and indicate whether an
10853 overflow occurred during the signed multiplication.
10858 The arguments (%a and %b) and the first element of the result structure
10859 may be of integer types of any bit width, but they must have the same
10860 bit width. The second element of the result structure must be of type
10861 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
10867 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
10868 a signed multiplication of the two arguments. They return a structure ---
10869 the first element of which is the multiplication, and the second element
10870 of which is a bit specifying if the signed multiplication resulted in an
10876 .. code-block:: llvm
10878 %res = call {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
10879 %sum = extractvalue {i32, i1} %res, 0
10880 %obit = extractvalue {i32, i1} %res, 1
10881 br i1 %obit, label %overflow, label %normal
10883 '``llvm.umul.with.overflow.*``' Intrinsics
10884 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10889 This is an overloaded intrinsic. You can use ``llvm.umul.with.overflow``
10890 on any integer bit width.
10894 declare {i16, i1} @llvm.umul.with.overflow.i16(i16 %a, i16 %b)
10895 declare {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
10896 declare {i64, i1} @llvm.umul.with.overflow.i64(i64 %a, i64 %b)
10901 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
10902 a unsigned multiplication of the two arguments, and indicate whether an
10903 overflow occurred during the unsigned multiplication.
10908 The arguments (%a and %b) and the first element of the result structure
10909 may be of integer types of any bit width, but they must have the same
10910 bit width. The second element of the result structure must be of type
10911 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
10917 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
10918 an unsigned multiplication of the two arguments. They return a structure ---
10919 the first element of which is the multiplication, and the second
10920 element of which is a bit specifying if the unsigned multiplication
10921 resulted in an overflow.
10926 .. code-block:: llvm
10928 %res = call {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
10929 %sum = extractvalue {i32, i1} %res, 0
10930 %obit = extractvalue {i32, i1} %res, 1
10931 br i1 %obit, label %overflow, label %normal
10933 Specialised Arithmetic Intrinsics
10934 ---------------------------------
10936 '``llvm.canonicalize.*``' Intrinsic
10937 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10944 declare float @llvm.canonicalize.f32(float %a)
10945 declare double @llvm.canonicalize.f64(double %b)
10950 The '``llvm.canonicalize.*``' intrinsic returns the platform specific canonical
10951 encoding of a floating point number. This canonicalization is useful for
10952 implementing certain numeric primitives such as frexp. The canonical encoding is
10953 defined by IEEE-754-2008 to be:
10957 2.1.8 canonical encoding: The preferred encoding of a floating-point
10958 representation in a format. Applied to declets, significands of finite
10959 numbers, infinities, and NaNs, especially in decimal formats.
10961 This operation can also be considered equivalent to the IEEE-754-2008
10962 conversion of a floating-point value to the same format. NaNs are handled
10963 according to section 6.2.
10965 Examples of non-canonical encodings:
10967 - x87 pseudo denormals, pseudo NaNs, pseudo Infinity, Unnormals. These are
10968 converted to a canonical representation per hardware-specific protocol.
10969 - Many normal decimal floating point numbers have non-canonical alternative
10971 - Some machines, like GPUs or ARMv7 NEON, do not support subnormal values.
10972 These are treated as non-canonical encodings of zero and with be flushed to
10973 a zero of the same sign by this operation.
10975 Note that per IEEE-754-2008 6.2, systems that support signaling NaNs with
10976 default exception handling must signal an invalid exception, and produce a
10979 This function should always be implementable as multiplication by 1.0, provided
10980 that the compiler does not constant fold the operation. Likewise, division by
10981 1.0 and ``llvm.minnum(x, x)`` are possible implementations. Addition with
10982 -0.0 is also sufficient provided that the rounding mode is not -Infinity.
10984 ``@llvm.canonicalize`` must preserve the equality relation. That is:
10986 - ``(@llvm.canonicalize(x) == x)`` is equivalent to ``(x == x)``
10987 - ``(@llvm.canonicalize(x) == @llvm.canonicalize(y))`` is equivalent to
10990 Additionally, the sign of zero must be conserved:
10991 ``@llvm.canonicalize(-0.0) = -0.0`` and ``@llvm.canonicalize(+0.0) = +0.0``
10993 The payload bits of a NaN must be conserved, with two exceptions.
10994 First, environments which use only a single canonical representation of NaN
10995 must perform said canonicalization. Second, SNaNs must be quieted per the
10998 The canonicalization operation may be optimized away if:
11000 - The input is known to be canonical. For example, it was produced by a
11001 floating-point operation that is required by the standard to be canonical.
11002 - The result is consumed only by (or fused with) other floating-point
11003 operations. That is, the bits of the floating point value are not examined.
11005 '``llvm.fmuladd.*``' Intrinsic
11006 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11013 declare float @llvm.fmuladd.f32(float %a, float %b, float %c)
11014 declare double @llvm.fmuladd.f64(double %a, double %b, double %c)
11019 The '``llvm.fmuladd.*``' intrinsic functions represent multiply-add
11020 expressions that can be fused if the code generator determines that (a) the
11021 target instruction set has support for a fused operation, and (b) that the
11022 fused operation is more efficient than the equivalent, separate pair of mul
11023 and add instructions.
11028 The '``llvm.fmuladd.*``' intrinsics each take three arguments: two
11029 multiplicands, a and b, and an addend c.
11038 %0 = call float @llvm.fmuladd.f32(%a, %b, %c)
11040 is equivalent to the expression a \* b + c, except that rounding will
11041 not be performed between the multiplication and addition steps if the
11042 code generator fuses the operations. Fusion is not guaranteed, even if
11043 the target platform supports it. If a fused multiply-add is required the
11044 corresponding llvm.fma.\* intrinsic function should be used
11045 instead. This never sets errno, just as '``llvm.fma.*``'.
11050 .. code-block:: llvm
11052 %r2 = call float @llvm.fmuladd.f32(float %a, float %b, float %c) ; yields float:r2 = (a * b) + c
11054 Half Precision Floating Point Intrinsics
11055 ----------------------------------------
11057 For most target platforms, half precision floating point is a
11058 storage-only format. This means that it is a dense encoding (in memory)
11059 but does not support computation in the format.
11061 This means that code must first load the half-precision floating point
11062 value as an i16, then convert it to float with
11063 :ref:`llvm.convert.from.fp16 <int_convert_from_fp16>`. Computation can
11064 then be performed on the float value (including extending to double
11065 etc). To store the value back to memory, it is first converted to float
11066 if needed, then converted to i16 with
11067 :ref:`llvm.convert.to.fp16 <int_convert_to_fp16>`, then storing as an
11070 .. _int_convert_to_fp16:
11072 '``llvm.convert.to.fp16``' Intrinsic
11073 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11080 declare i16 @llvm.convert.to.fp16.f32(float %a)
11081 declare i16 @llvm.convert.to.fp16.f64(double %a)
11086 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion from a
11087 conventional floating point type to half precision floating point format.
11092 The intrinsic function contains single argument - the value to be
11098 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion from a
11099 conventional floating point format to half precision floating point format. The
11100 return value is an ``i16`` which contains the converted number.
11105 .. code-block:: llvm
11107 %res = call i16 @llvm.convert.to.fp16.f32(float %a)
11108 store i16 %res, i16* @x, align 2
11110 .. _int_convert_from_fp16:
11112 '``llvm.convert.from.fp16``' Intrinsic
11113 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11120 declare float @llvm.convert.from.fp16.f32(i16 %a)
11121 declare double @llvm.convert.from.fp16.f64(i16 %a)
11126 The '``llvm.convert.from.fp16``' intrinsic function performs a
11127 conversion from half precision floating point format to single precision
11128 floating point format.
11133 The intrinsic function contains single argument - the value to be
11139 The '``llvm.convert.from.fp16``' intrinsic function performs a
11140 conversion from half single precision floating point format to single
11141 precision floating point format. The input half-float value is
11142 represented by an ``i16`` value.
11147 .. code-block:: llvm
11149 %a = load i16, i16* @x, align 2
11150 %res = call float @llvm.convert.from.fp16(i16 %a)
11152 .. _dbg_intrinsics:
11154 Debugger Intrinsics
11155 -------------------
11157 The LLVM debugger intrinsics (which all start with ``llvm.dbg.``
11158 prefix), are described in the `LLVM Source Level
11159 Debugging <SourceLevelDebugging.html#format_common_intrinsics>`_
11162 Exception Handling Intrinsics
11163 -----------------------------
11165 The LLVM exception handling intrinsics (which all start with
11166 ``llvm.eh.`` prefix), are described in the `LLVM Exception
11167 Handling <ExceptionHandling.html#format_common_intrinsics>`_ document.
11169 .. _int_trampoline:
11171 Trampoline Intrinsics
11172 ---------------------
11174 These intrinsics make it possible to excise one parameter, marked with
11175 the :ref:`nest <nest>` attribute, from a function. The result is a
11176 callable function pointer lacking the nest parameter - the caller does
11177 not need to provide a value for it. Instead, the value to use is stored
11178 in advance in a "trampoline", a block of memory usually allocated on the
11179 stack, which also contains code to splice the nest value into the
11180 argument list. This is used to implement the GCC nested function address
11183 For example, if the function is ``i32 f(i8* nest %c, i32 %x, i32 %y)``
11184 then the resulting function pointer has signature ``i32 (i32, i32)*``.
11185 It can be created as follows:
11187 .. code-block:: llvm
11189 %tramp = alloca [10 x i8], align 4 ; size and alignment only correct for X86
11190 %tramp1 = getelementptr [10 x i8], [10 x i8]* %tramp, i32 0, i32 0
11191 call i8* @llvm.init.trampoline(i8* %tramp1, i8* bitcast (i32 (i8*, i32, i32)* @f to i8*), i8* %nval)
11192 %p = call i8* @llvm.adjust.trampoline(i8* %tramp1)
11193 %fp = bitcast i8* %p to i32 (i32, i32)*
11195 The call ``%val = call i32 %fp(i32 %x, i32 %y)`` is then equivalent to
11196 ``%val = call i32 %f(i8* %nval, i32 %x, i32 %y)``.
11200 '``llvm.init.trampoline``' Intrinsic
11201 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11208 declare void @llvm.init.trampoline(i8* <tramp>, i8* <func>, i8* <nval>)
11213 This fills the memory pointed to by ``tramp`` with executable code,
11214 turning it into a trampoline.
11219 The ``llvm.init.trampoline`` intrinsic takes three arguments, all
11220 pointers. The ``tramp`` argument must point to a sufficiently large and
11221 sufficiently aligned block of memory; this memory is written to by the
11222 intrinsic. Note that the size and the alignment are target-specific -
11223 LLVM currently provides no portable way of determining them, so a
11224 front-end that generates this intrinsic needs to have some
11225 target-specific knowledge. The ``func`` argument must hold a function
11226 bitcast to an ``i8*``.
11231 The block of memory pointed to by ``tramp`` is filled with target
11232 dependent code, turning it into a function. Then ``tramp`` needs to be
11233 passed to :ref:`llvm.adjust.trampoline <int_at>` to get a pointer which can
11234 be :ref:`bitcast (to a new function) and called <int_trampoline>`. The new
11235 function's signature is the same as that of ``func`` with any arguments
11236 marked with the ``nest`` attribute removed. At most one such ``nest``
11237 argument is allowed, and it must be of pointer type. Calling the new
11238 function is equivalent to calling ``func`` with the same argument list,
11239 but with ``nval`` used for the missing ``nest`` argument. If, after
11240 calling ``llvm.init.trampoline``, the memory pointed to by ``tramp`` is
11241 modified, then the effect of any later call to the returned function
11242 pointer is undefined.
11246 '``llvm.adjust.trampoline``' Intrinsic
11247 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11254 declare i8* @llvm.adjust.trampoline(i8* <tramp>)
11259 This performs any required machine-specific adjustment to the address of
11260 a trampoline (passed as ``tramp``).
11265 ``tramp`` must point to a block of memory which already has trampoline
11266 code filled in by a previous call to
11267 :ref:`llvm.init.trampoline <int_it>`.
11272 On some architectures the address of the code to be executed needs to be
11273 different than the address where the trampoline is actually stored. This
11274 intrinsic returns the executable address corresponding to ``tramp``
11275 after performing the required machine specific adjustments. The pointer
11276 returned can then be :ref:`bitcast and executed <int_trampoline>`.
11278 .. _int_mload_mstore:
11280 Masked Vector Load and Store Intrinsics
11281 ---------------------------------------
11283 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.
11287 '``llvm.masked.load.*``' Intrinsics
11288 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11292 This is an overloaded intrinsic. The loaded data is a vector of any integer, floating point or pointer data type.
11296 declare <16 x float> @llvm.masked.load.v16f32 (<16 x float>* <ptr>, i32 <alignment>, <16 x i1> <mask>, <16 x float> <passthru>)
11297 declare <2 x double> @llvm.masked.load.v2f64 (<2 x double>* <ptr>, i32 <alignment>, <2 x i1> <mask>, <2 x double> <passthru>)
11298 ;; The data is a vector of pointers to double
11299 declare <8 x double*> @llvm.masked.load.v8p0f64 (<8 x double*>* <ptr>, i32 <alignment>, <8 x i1> <mask>, <8 x double*> <passthru>)
11300 ;; The data is a vector of function pointers
11301 declare <8 x i32 ()*> @llvm.masked.load.v8p0f_i32f (<8 x i32 ()*>* <ptr>, i32 <alignment>, <8 x i1> <mask>, <8 x i32 ()*> <passthru>)
11306 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.
11312 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.
11318 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.
11319 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.
11324 %res = call <16 x float> @llvm.masked.load.v16f32 (<16 x float>* %ptr, i32 4, <16 x i1>%mask, <16 x float> %passthru)
11326 ;; The result of the two following instructions is identical aside from potential memory access exception
11327 %loadlal = load <16 x float>, <16 x float>* %ptr, align 4
11328 %res = select <16 x i1> %mask, <16 x float> %loadlal, <16 x float> %passthru
11332 '``llvm.masked.store.*``' Intrinsics
11333 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11337 This is an overloaded intrinsic. The data stored in memory is a vector of any integer, floating point or pointer data type.
11341 declare void @llvm.masked.store.v8i32 (<8 x i32> <value>, <8 x i32>* <ptr>, i32 <alignment>, <8 x i1> <mask>)
11342 declare void @llvm.masked.store.v16f32 (<16 x float> <value>, <16 x float>* <ptr>, i32 <alignment>, <16 x i1> <mask>)
11343 ;; The data is a vector of pointers to double
11344 declare void @llvm.masked.store.v8p0f64 (<8 x double*> <value>, <8 x double*>* <ptr>, i32 <alignment>, <8 x i1> <mask>)
11345 ;; The data is a vector of function pointers
11346 declare void @llvm.masked.store.v4p0f_i32f (<4 x i32 ()*> <value>, <4 x i32 ()*>* <ptr>, i32 <alignment>, <4 x i1> <mask>)
11351 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.
11356 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.
11362 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.
11363 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.
11367 call void @llvm.masked.store.v16f32(<16 x float> %value, <16 x float>* %ptr, i32 4, <16 x i1> %mask)
11369 ;; The result of the following instructions is identical aside from potential data races and memory access exceptions
11370 %oldval = load <16 x float>, <16 x float>* %ptr, align 4
11371 %res = select <16 x i1> %mask, <16 x float> %value, <16 x float> %oldval
11372 store <16 x float> %res, <16 x float>* %ptr, align 4
11375 Masked Vector Gather and Scatter Intrinsics
11376 -------------------------------------------
11378 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.
11382 '``llvm.masked.gather.*``' Intrinsics
11383 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11387 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.
11391 declare <16 x float> @llvm.masked.gather.v16f32 (<16 x float*> <ptrs>, i32 <alignment>, <16 x i1> <mask>, <16 x float> <passthru>)
11392 declare <2 x double> @llvm.masked.gather.v2f64 (<2 x double*> <ptrs>, i32 <alignment>, <2 x i1> <mask>, <2 x double> <passthru>)
11393 declare <8 x float*> @llvm.masked.gather.v8p0f32 (<8 x float**> <ptrs>, i32 <alignment>, <8 x i1> <mask>, <8 x float*> <passthru>)
11398 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.
11404 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.
11410 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.
11411 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.
11416 %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>)
11418 ;; The gather with all-true mask is equivalent to the following instruction sequence
11419 %ptr0 = extractelement <4 x double*> %ptrs, i32 0
11420 %ptr1 = extractelement <4 x double*> %ptrs, i32 1
11421 %ptr2 = extractelement <4 x double*> %ptrs, i32 2
11422 %ptr3 = extractelement <4 x double*> %ptrs, i32 3
11424 %val0 = load double, double* %ptr0, align 8
11425 %val1 = load double, double* %ptr1, align 8
11426 %val2 = load double, double* %ptr2, align 8
11427 %val3 = load double, double* %ptr3, align 8
11429 %vec0 = insertelement <4 x double>undef, %val0, 0
11430 %vec01 = insertelement <4 x double>%vec0, %val1, 1
11431 %vec012 = insertelement <4 x double>%vec01, %val2, 2
11432 %vec0123 = insertelement <4 x double>%vec012, %val3, 3
11436 '``llvm.masked.scatter.*``' Intrinsics
11437 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11441 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.
11445 declare void @llvm.masked.scatter.v8i32 (<8 x i32> <value>, <8 x i32*> <ptrs>, i32 <alignment>, <8 x i1> <mask>)
11446 declare void @llvm.masked.scatter.v16f32 (<16 x float> <value>, <16 x float*> <ptrs>, i32 <alignment>, <16 x i1> <mask>)
11447 declare void @llvm.masked.scatter.v4p0f64 (<4 x double*> <value>, <4 x double**> <ptrs>, i32 <alignment>, <4 x i1> <mask>)
11452 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.
11457 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.
11463 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.
11467 ;; This instruction unconditionaly stores data vector in multiple addresses
11468 call @llvm.masked.scatter.v8i32 (<8 x i32> %value, <8 x i32*> %ptrs, i32 4, <8 x i1> <true, true, .. true>)
11470 ;; It is equivalent to a list of scalar stores
11471 %val0 = extractelement <8 x i32> %value, i32 0
11472 %val1 = extractelement <8 x i32> %value, i32 1
11474 %val7 = extractelement <8 x i32> %value, i32 7
11475 %ptr0 = extractelement <8 x i32*> %ptrs, i32 0
11476 %ptr1 = extractelement <8 x i32*> %ptrs, i32 1
11478 %ptr7 = extractelement <8 x i32*> %ptrs, i32 7
11479 ;; Note: the order of the following stores is important when they overlap:
11480 store i32 %val0, i32* %ptr0, align 4
11481 store i32 %val1, i32* %ptr1, align 4
11483 store i32 %val7, i32* %ptr7, align 4
11489 This class of intrinsics provides information about the lifetime of
11490 memory objects and ranges where variables are immutable.
11494 '``llvm.lifetime.start``' Intrinsic
11495 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11502 declare void @llvm.lifetime.start(i64 <size>, i8* nocapture <ptr>)
11507 The '``llvm.lifetime.start``' intrinsic specifies the start of a memory
11513 The first argument is a constant integer representing the size of the
11514 object, or -1 if it is variable sized. The second argument is a pointer
11520 This intrinsic indicates that before this point in the code, the value
11521 of the memory pointed to by ``ptr`` is dead. This means that it is known
11522 to never be used and has an undefined value. A load from the pointer
11523 that precedes this intrinsic can be replaced with ``'undef'``.
11527 '``llvm.lifetime.end``' Intrinsic
11528 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11535 declare void @llvm.lifetime.end(i64 <size>, i8* nocapture <ptr>)
11540 The '``llvm.lifetime.end``' intrinsic specifies the end of a memory
11546 The first argument is a constant integer representing the size of the
11547 object, or -1 if it is variable sized. The second argument is a pointer
11553 This intrinsic indicates that after this point in the code, the value of
11554 the memory pointed to by ``ptr`` is dead. This means that it is known to
11555 never be used and has an undefined value. Any stores into the memory
11556 object following this intrinsic may be removed as dead.
11558 '``llvm.invariant.start``' Intrinsic
11559 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11566 declare {}* @llvm.invariant.start(i64 <size>, i8* nocapture <ptr>)
11571 The '``llvm.invariant.start``' intrinsic specifies that the contents of
11572 a memory object will not change.
11577 The first argument is a constant integer representing the size of the
11578 object, or -1 if it is variable sized. The second argument is a pointer
11584 This intrinsic indicates that until an ``llvm.invariant.end`` that uses
11585 the return value, the referenced memory location is constant and
11588 '``llvm.invariant.end``' Intrinsic
11589 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11596 declare void @llvm.invariant.end({}* <start>, i64 <size>, i8* nocapture <ptr>)
11601 The '``llvm.invariant.end``' intrinsic specifies that the contents of a
11602 memory object are mutable.
11607 The first argument is the matching ``llvm.invariant.start`` intrinsic.
11608 The second argument is a constant integer representing the size of the
11609 object, or -1 if it is variable sized and the third argument is a
11610 pointer to the object.
11615 This intrinsic indicates that the memory is mutable again.
11617 '``llvm.invariant.group.barrier``' Intrinsic
11618 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11625 declare i8* @llvm.invariant.group.barrier(i8* <ptr>)
11630 The '``llvm.invariant.group.barrier``' intrinsic can be used when an invariant
11631 established by invariant.group metadata no longer holds, to obtain a new pointer
11632 value that does not carry the invariant information.
11638 The ``llvm.invariant.group.barrier`` takes only one argument, which is
11639 the pointer to the memory for which the ``invariant.group`` no longer holds.
11644 Returns another pointer that aliases its argument but which is considered different
11645 for the purposes of ``load``/``store`` ``invariant.group`` metadata.
11650 This class of intrinsics is designed to be generic and has no specific
11653 '``llvm.var.annotation``' Intrinsic
11654 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11661 declare void @llvm.var.annotation(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
11666 The '``llvm.var.annotation``' intrinsic.
11671 The first argument is a pointer to a value, the second is a pointer to a
11672 global string, the third is a pointer to a global string which is the
11673 source file name, and the last argument is the line number.
11678 This intrinsic allows annotation of local variables with arbitrary
11679 strings. This can be useful for special purpose optimizations that want
11680 to look for these annotations. These have no other defined use; they are
11681 ignored by code generation and optimization.
11683 '``llvm.ptr.annotation.*``' Intrinsic
11684 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11689 This is an overloaded intrinsic. You can use '``llvm.ptr.annotation``' on a
11690 pointer to an integer of any width. *NOTE* you must specify an address space for
11691 the pointer. The identifier for the default address space is the integer
11696 declare i8* @llvm.ptr.annotation.p<address space>i8(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
11697 declare i16* @llvm.ptr.annotation.p<address space>i16(i16* <val>, i8* <str>, i8* <str>, i32 <int>)
11698 declare i32* @llvm.ptr.annotation.p<address space>i32(i32* <val>, i8* <str>, i8* <str>, i32 <int>)
11699 declare i64* @llvm.ptr.annotation.p<address space>i64(i64* <val>, i8* <str>, i8* <str>, i32 <int>)
11700 declare i256* @llvm.ptr.annotation.p<address space>i256(i256* <val>, i8* <str>, i8* <str>, i32 <int>)
11705 The '``llvm.ptr.annotation``' intrinsic.
11710 The first argument is a pointer to an integer value of arbitrary bitwidth
11711 (result of some expression), the second is a pointer to a global string, the
11712 third is a pointer to a global string which is the source file name, and the
11713 last argument is the line number. It returns the value of the first argument.
11718 This intrinsic allows annotation of a pointer to an integer with arbitrary
11719 strings. This can be useful for special purpose optimizations that want to look
11720 for these annotations. These have no other defined use; they are ignored by code
11721 generation and optimization.
11723 '``llvm.annotation.*``' Intrinsic
11724 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11729 This is an overloaded intrinsic. You can use '``llvm.annotation``' on
11730 any integer bit width.
11734 declare i8 @llvm.annotation.i8(i8 <val>, i8* <str>, i8* <str>, i32 <int>)
11735 declare i16 @llvm.annotation.i16(i16 <val>, i8* <str>, i8* <str>, i32 <int>)
11736 declare i32 @llvm.annotation.i32(i32 <val>, i8* <str>, i8* <str>, i32 <int>)
11737 declare i64 @llvm.annotation.i64(i64 <val>, i8* <str>, i8* <str>, i32 <int>)
11738 declare i256 @llvm.annotation.i256(i256 <val>, i8* <str>, i8* <str>, i32 <int>)
11743 The '``llvm.annotation``' intrinsic.
11748 The first argument is an integer value (result of some expression), the
11749 second is a pointer to a global string, the third is a pointer to a
11750 global string which is the source file name, and the last argument is
11751 the line number. It returns the value of the first argument.
11756 This intrinsic allows annotations to be put on arbitrary expressions
11757 with arbitrary strings. This can be useful for special purpose
11758 optimizations that want to look for these annotations. These have no
11759 other defined use; they are ignored by code generation and optimization.
11761 '``llvm.trap``' Intrinsic
11762 ^^^^^^^^^^^^^^^^^^^^^^^^^
11769 declare void @llvm.trap() noreturn nounwind
11774 The '``llvm.trap``' intrinsic.
11784 This intrinsic is lowered to the target dependent trap instruction. If
11785 the target does not have a trap instruction, this intrinsic will be
11786 lowered to a call of the ``abort()`` function.
11788 '``llvm.debugtrap``' Intrinsic
11789 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11796 declare void @llvm.debugtrap() nounwind
11801 The '``llvm.debugtrap``' intrinsic.
11811 This intrinsic is lowered to code which is intended to cause an
11812 execution trap with the intention of requesting the attention of a
11815 '``llvm.stackprotector``' Intrinsic
11816 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11823 declare void @llvm.stackprotector(i8* <guard>, i8** <slot>)
11828 The ``llvm.stackprotector`` intrinsic takes the ``guard`` and stores it
11829 onto the stack at ``slot``. The stack slot is adjusted to ensure that it
11830 is placed on the stack before local variables.
11835 The ``llvm.stackprotector`` intrinsic requires two pointer arguments.
11836 The first argument is the value loaded from the stack guard
11837 ``@__stack_chk_guard``. The second variable is an ``alloca`` that has
11838 enough space to hold the value of the guard.
11843 This intrinsic causes the prologue/epilogue inserter to force the position of
11844 the ``AllocaInst`` stack slot to be before local variables on the stack. This is
11845 to ensure that if a local variable on the stack is overwritten, it will destroy
11846 the value of the guard. When the function exits, the guard on the stack is
11847 checked against the original guard by ``llvm.stackprotectorcheck``. If they are
11848 different, then ``llvm.stackprotectorcheck`` causes the program to abort by
11849 calling the ``__stack_chk_fail()`` function.
11851 '``llvm.stackprotectorcheck``' Intrinsic
11852 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11859 declare void @llvm.stackprotectorcheck(i8** <guard>)
11864 The ``llvm.stackprotectorcheck`` intrinsic compares ``guard`` against an already
11865 created stack protector and if they are not equal calls the
11866 ``__stack_chk_fail()`` function.
11871 The ``llvm.stackprotectorcheck`` intrinsic requires one pointer argument, the
11872 the variable ``@__stack_chk_guard``.
11877 This intrinsic is provided to perform the stack protector check by comparing
11878 ``guard`` with the stack slot created by ``llvm.stackprotector`` and if the
11879 values do not match call the ``__stack_chk_fail()`` function.
11881 The reason to provide this as an IR level intrinsic instead of implementing it
11882 via other IR operations is that in order to perform this operation at the IR
11883 level without an intrinsic, one would need to create additional basic blocks to
11884 handle the success/failure cases. This makes it difficult to stop the stack
11885 protector check from disrupting sibling tail calls in Codegen. With this
11886 intrinsic, we are able to generate the stack protector basic blocks late in
11887 codegen after the tail call decision has occurred.
11889 '``llvm.objectsize``' Intrinsic
11890 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11897 declare i32 @llvm.objectsize.i32(i8* <object>, i1 <min>)
11898 declare i64 @llvm.objectsize.i64(i8* <object>, i1 <min>)
11903 The ``llvm.objectsize`` intrinsic is designed to provide information to
11904 the optimizers to determine at compile time whether a) an operation
11905 (like memcpy) will overflow a buffer that corresponds to an object, or
11906 b) that a runtime check for overflow isn't necessary. An object in this
11907 context means an allocation of a specific class, structure, array, or
11913 The ``llvm.objectsize`` intrinsic takes two arguments. The first
11914 argument is a pointer to or into the ``object``. The second argument is
11915 a boolean and determines whether ``llvm.objectsize`` returns 0 (if true)
11916 or -1 (if false) when the object size is unknown. The second argument
11917 only accepts constants.
11922 The ``llvm.objectsize`` intrinsic is lowered to a constant representing
11923 the size of the object concerned. If the size cannot be determined at
11924 compile time, ``llvm.objectsize`` returns ``i32/i64 -1 or 0`` (depending
11925 on the ``min`` argument).
11927 '``llvm.expect``' Intrinsic
11928 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
11933 This is an overloaded intrinsic. You can use ``llvm.expect`` on any
11938 declare i1 @llvm.expect.i1(i1 <val>, i1 <expected_val>)
11939 declare i32 @llvm.expect.i32(i32 <val>, i32 <expected_val>)
11940 declare i64 @llvm.expect.i64(i64 <val>, i64 <expected_val>)
11945 The ``llvm.expect`` intrinsic provides information about expected (the
11946 most probable) value of ``val``, which can be used by optimizers.
11951 The ``llvm.expect`` intrinsic takes two arguments. The first argument is
11952 a value. The second argument is an expected value, this needs to be a
11953 constant value, variables are not allowed.
11958 This intrinsic is lowered to the ``val``.
11962 '``llvm.assume``' Intrinsic
11963 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11970 declare void @llvm.assume(i1 %cond)
11975 The ``llvm.assume`` allows the optimizer to assume that the provided
11976 condition is true. This information can then be used in simplifying other parts
11982 The condition which the optimizer may assume is always true.
11987 The intrinsic allows the optimizer to assume that the provided condition is
11988 always true whenever the control flow reaches the intrinsic call. No code is
11989 generated for this intrinsic, and instructions that contribute only to the
11990 provided condition are not used for code generation. If the condition is
11991 violated during execution, the behavior is undefined.
11993 Note that the optimizer might limit the transformations performed on values
11994 used by the ``llvm.assume`` intrinsic in order to preserve the instructions
11995 only used to form the intrinsic's input argument. This might prove undesirable
11996 if the extra information provided by the ``llvm.assume`` intrinsic does not cause
11997 sufficient overall improvement in code quality. For this reason,
11998 ``llvm.assume`` should not be used to document basic mathematical invariants
11999 that the optimizer can otherwise deduce or facts that are of little use to the
12004 '``llvm.bitset.test``' Intrinsic
12005 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12012 declare i1 @llvm.bitset.test(i8* %ptr, metadata %bitset) nounwind readnone
12018 The first argument is a pointer to be tested. The second argument is a
12019 metadata object representing an identifier for a :doc:`bitset <BitSets>`.
12024 The ``llvm.bitset.test`` intrinsic tests whether the given pointer is a
12025 member of the given bitset.
12027 '``llvm.donothing``' Intrinsic
12028 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12035 declare void @llvm.donothing() nounwind readnone
12040 The ``llvm.donothing`` intrinsic doesn't perform any operation. It's one of only
12041 two intrinsics (besides ``llvm.experimental.patchpoint``) that can be called
12042 with an invoke instruction.
12052 This intrinsic does nothing, and it's removed by optimizers and ignored
12055 Stack Map Intrinsics
12056 --------------------
12058 LLVM provides experimental intrinsics to support runtime patching
12059 mechanisms commonly desired in dynamic language JITs. These intrinsics
12060 are described in :doc:`StackMaps`.