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
208 into the object file corresponding to the LLVM module. They exist to
209 allow inlining and other optimizations to take place given knowledge
210 of the definition of the global, which is known to be somewhere
211 outside the module. Globals with ``available_externally`` linkage
212 are allowed to be discarded at will, and are otherwise the same as
213 ``linkonce_odr``. This linkage type is only allowed on definitions,
216 Globals with "``linkonce``" linkage are merged with other globals of
217 the same name when linkage occurs. This can be used to implement
218 some forms of inline functions, templates, or other code which must
219 be generated in each translation unit that uses it, but where the
220 body may be overridden with a more definitive definition later.
221 Unreferenced ``linkonce`` globals are allowed to be discarded. Note
222 that ``linkonce`` linkage does not actually allow the optimizer to
223 inline the body of this function into callers because it doesn't
224 know if this definition of the function is the definitive definition
225 within the program or whether it will be overridden by a stronger
226 definition. To enable inlining and other optimizations, use
227 "``linkonce_odr``" linkage.
229 "``weak``" linkage has the same merging semantics as ``linkonce``
230 linkage, except that unreferenced globals with ``weak`` linkage may
231 not be discarded. This is used for globals that are declared "weak"
234 "``common``" linkage is most similar to "``weak``" linkage, but they
235 are used for tentative definitions in C, such as "``int X;``" at
236 global scope. Symbols with "``common``" linkage are merged in the
237 same way as ``weak symbols``, and they may not be deleted if
238 unreferenced. ``common`` symbols may not have an explicit section,
239 must have a zero initializer, and may not be marked
240 ':ref:`constant <globalvars>`'. Functions and aliases may not have
243 .. _linkage_appending:
246 "``appending``" linkage may only be applied to global variables of
247 pointer to array type. When two global variables with appending
248 linkage are linked together, the two global arrays are appended
249 together. This is the LLVM, typesafe, equivalent of having the
250 system linker append together "sections" with identical names when
253 The semantics of this linkage follow the ELF object file model: the
254 symbol is weak until linked, if not linked, the symbol becomes null
255 instead of being an undefined reference.
256 ``linkonce_odr``, ``weak_odr``
257 Some languages allow differing globals to be merged, such as two
258 functions with different semantics. Other languages, such as
259 ``C++``, ensure that only equivalent globals are ever merged (the
260 "one definition rule" --- "ODR"). Such languages can use the
261 ``linkonce_odr`` and ``weak_odr`` linkage types to indicate that the
262 global will only be merged with equivalent globals. These linkage
263 types are otherwise the same as their non-``odr`` versions.
265 If none of the above identifiers are used, the global is externally
266 visible, meaning that it participates in linkage and can be used to
267 resolve external symbol references.
269 It is illegal for a function *declaration* to have any linkage type
270 other than ``external`` or ``extern_weak``.
277 LLVM :ref:`functions <functionstructure>`, :ref:`calls <i_call>` and
278 :ref:`invokes <i_invoke>` can all have an optional calling convention
279 specified for the call. The calling convention of any pair of dynamic
280 caller/callee must match, or the behavior of the program is undefined.
281 The following calling conventions are supported by LLVM, and more may be
284 "``ccc``" - The C calling convention
285 This calling convention (the default if no other calling convention
286 is specified) matches the target C calling conventions. This calling
287 convention supports varargs function calls and tolerates some
288 mismatch in the declared prototype and implemented declaration of
289 the function (as does normal C).
290 "``fastcc``" - The fast calling convention
291 This calling convention attempts to make calls as fast as possible
292 (e.g. by passing things in registers). This calling convention
293 allows the target to use whatever tricks it wants to produce fast
294 code for the target, without having to conform to an externally
295 specified ABI (Application Binary Interface). `Tail calls can only
296 be optimized when this, the GHC or the HiPE convention is
297 used. <CodeGenerator.html#id80>`_ This calling convention does not
298 support varargs and requires the prototype of all callees to exactly
299 match the prototype of the function definition.
300 "``coldcc``" - The cold calling convention
301 This calling convention attempts to make code in the caller as
302 efficient as possible under the assumption that the call is not
303 commonly executed. As such, these calls often preserve all registers
304 so that the call does not break any live ranges in the caller side.
305 This calling convention does not support varargs and requires the
306 prototype of all callees to exactly match the prototype of the
307 function definition. Furthermore the inliner doesn't consider such function
309 "``cc 10``" - GHC convention
310 This calling convention has been implemented specifically for use by
311 the `Glasgow Haskell Compiler (GHC) <http://www.haskell.org/ghc>`_.
312 It passes everything in registers, going to extremes to achieve this
313 by disabling callee save registers. This calling convention should
314 not be used lightly but only for specific situations such as an
315 alternative to the *register pinning* performance technique often
316 used when implementing functional programming languages. At the
317 moment only X86 supports this convention and it has the following
320 - On *X86-32* only supports up to 4 bit type parameters. No
321 floating point types are supported.
322 - On *X86-64* only supports up to 10 bit type parameters and 6
323 floating point parameters.
325 This calling convention supports `tail call
326 optimization <CodeGenerator.html#id80>`_ but requires both the
327 caller and callee are using it.
328 "``cc 11``" - The HiPE calling convention
329 This calling convention has been implemented specifically for use by
330 the `High-Performance Erlang
331 (HiPE) <http://www.it.uu.se/research/group/hipe/>`_ compiler, *the*
332 native code compiler of the `Ericsson's Open Source Erlang/OTP
333 system <http://www.erlang.org/download.shtml>`_. It uses more
334 registers for argument passing than the ordinary C calling
335 convention and defines no callee-saved registers. The calling
336 convention properly supports `tail call
337 optimization <CodeGenerator.html#id80>`_ but requires that both the
338 caller and the callee use it. It uses a *register pinning*
339 mechanism, similar to GHC's convention, for keeping frequently
340 accessed runtime components pinned to specific hardware registers.
341 At the moment only X86 supports this convention (both 32 and 64
343 "``webkit_jscc``" - WebKit's JavaScript calling convention
344 This calling convention has been implemented for `WebKit FTL JIT
345 <https://trac.webkit.org/wiki/FTLJIT>`_. It passes arguments on the
346 stack right to left (as cdecl does), and returns a value in the
347 platform's customary return register.
348 "``anyregcc``" - Dynamic calling convention for code patching
349 This is a special convention that supports patching an arbitrary code
350 sequence in place of a call site. This convention forces the call
351 arguments into registers but allows them to be dynamically
352 allocated. This can currently only be used with calls to
353 llvm.experimental.patchpoint because only this intrinsic records
354 the location of its arguments in a side table. See :doc:`StackMaps`.
355 "``preserve_mostcc``" - The `PreserveMost` calling convention
356 This calling convention attempts to make the code in the caller as
357 unintrusive as possible. This convention behaves identically to the `C`
358 calling convention on how arguments and return values are passed, but it
359 uses a different set of caller/callee-saved registers. This alleviates the
360 burden of saving and recovering a large register set before and after the
361 call in the caller. If the arguments are passed in callee-saved registers,
362 then they will be preserved by the callee across the call. This doesn't
363 apply for values returned in callee-saved registers.
365 - On X86-64 the callee preserves all general purpose registers, except for
366 R11. R11 can be used as a scratch register. Floating-point registers
367 (XMMs/YMMs) are not preserved and need to be saved by the caller.
369 The idea behind this convention is to support calls to runtime functions
370 that have a hot path and a cold path. The hot path is usually a small piece
371 of code that doesn't use many registers. The cold path might need to call out to
372 another function and therefore only needs to preserve the caller-saved
373 registers, which haven't already been saved by the caller. The
374 `PreserveMost` calling convention is very similar to the `cold` calling
375 convention in terms of caller/callee-saved registers, but they are used for
376 different types of function calls. `coldcc` is for function calls that are
377 rarely executed, whereas `preserve_mostcc` function calls are intended to be
378 on the hot path and definitely executed a lot. Furthermore `preserve_mostcc`
379 doesn't prevent the inliner from inlining the function call.
381 This calling convention will be used by a future version of the ObjectiveC
382 runtime and should therefore still be considered experimental at this time.
383 Although this convention was created to optimize certain runtime calls to
384 the ObjectiveC runtime, it is not limited to this runtime and might be used
385 by other runtimes in the future too. The current implementation only
386 supports X86-64, but the intention is to support more architectures in the
388 "``preserve_allcc``" - The `PreserveAll` calling convention
389 This calling convention attempts to make the code in the caller even less
390 intrusive than the `PreserveMost` calling convention. This calling
391 convention also behaves identical to the `C` calling convention on how
392 arguments and return values are passed, but it uses a different set of
393 caller/callee-saved registers. This removes the burden of saving and
394 recovering a large register set before and after the call in the caller. If
395 the arguments are passed in callee-saved registers, then they will be
396 preserved by the callee across the call. This doesn't apply for values
397 returned in callee-saved registers.
399 - On X86-64 the callee preserves all general purpose registers, except for
400 R11. R11 can be used as a scratch register. Furthermore it also preserves
401 all floating-point registers (XMMs/YMMs).
403 The idea behind this convention is to support calls to runtime functions
404 that don't need to call out to any other functions.
406 This calling convention, like the `PreserveMost` calling convention, will be
407 used by a future version of the ObjectiveC runtime and should be considered
408 experimental at this time.
409 "``cc <n>``" - Numbered convention
410 Any calling convention may be specified by number, allowing
411 target-specific calling conventions to be used. Target specific
412 calling conventions start at 64.
414 More calling conventions can be added/defined on an as-needed basis, to
415 support Pascal conventions or any other well-known target-independent
418 .. _visibilitystyles:
423 All Global Variables and Functions have one of the following visibility
426 "``default``" - Default style
427 On targets that use the ELF object file format, default visibility
428 means that the declaration is visible to other modules and, in
429 shared libraries, means that the declared entity may be overridden.
430 On Darwin, default visibility means that the declaration is visible
431 to other modules. Default visibility corresponds to "external
432 linkage" in the language.
433 "``hidden``" - Hidden style
434 Two declarations of an object with hidden visibility refer to the
435 same object if they are in the same shared object. Usually, hidden
436 visibility indicates that the symbol will not be placed into the
437 dynamic symbol table, so no other module (executable or shared
438 library) can reference it directly.
439 "``protected``" - Protected style
440 On ELF, protected visibility indicates that the symbol will be
441 placed in the dynamic symbol table, but that references within the
442 defining module will bind to the local symbol. That is, the symbol
443 cannot be overridden by another module.
445 A symbol with ``internal`` or ``private`` linkage must have ``default``
453 All Global Variables, Functions and Aliases can have one of the following
457 "``dllimport``" causes the compiler to reference a function or variable via
458 a global pointer to a pointer that is set up by the DLL exporting the
459 symbol. On Microsoft Windows targets, the pointer name is formed by
460 combining ``__imp_`` and the function or variable name.
462 "``dllexport``" causes the compiler to provide a global pointer to a pointer
463 in a DLL, so that it can be referenced with the ``dllimport`` attribute. On
464 Microsoft Windows targets, the pointer name is formed by combining
465 ``__imp_`` and the function or variable name. Since this storage class
466 exists for defining a dll interface, the compiler, assembler and linker know
467 it is externally referenced and must refrain from deleting the symbol.
471 Thread Local Storage Models
472 ---------------------------
474 A variable may be defined as ``thread_local``, which means that it will
475 not be shared by threads (each thread will have a separated copy of the
476 variable). Not all targets support thread-local variables. Optionally, a
477 TLS model may be specified:
480 For variables that are only used within the current shared library.
482 For variables in modules that will not be loaded dynamically.
484 For variables defined in the executable and only used within it.
486 If no explicit model is given, the "general dynamic" model is used.
488 The models correspond to the ELF TLS models; see `ELF Handling For
489 Thread-Local Storage <http://people.redhat.com/drepper/tls.pdf>`_ for
490 more information on under which circumstances the different models may
491 be used. The target may choose a different TLS model if the specified
492 model is not supported, or if a better choice of model can be made.
494 A model can also be specified in an alias, but then it only governs how
495 the alias is accessed. It will not have any effect in the aliasee.
497 For platforms without linker support of ELF TLS model, the -femulated-tls
498 flag can be used to generate GCC compatible emulated TLS code.
505 LLVM IR allows you to specify both "identified" and "literal" :ref:`structure
506 types <t_struct>`. Literal types are uniqued structurally, but identified types
507 are never uniqued. An :ref:`opaque structural type <t_opaque>` can also be used
508 to forward declare a type that is not yet available.
510 An example of an identified structure specification is:
514 %mytype = type { %mytype*, i32 }
516 Prior to the LLVM 3.0 release, identified types were structurally uniqued. Only
517 literal types are uniqued in recent versions of LLVM.
524 Global variables define regions of memory allocated at compilation time
527 Global variable definitions must be initialized.
529 Global variables in other translation units can also be declared, in which
530 case they don't have an initializer.
532 Either global variable definitions or declarations may have an explicit section
533 to be placed in and may have an optional explicit alignment specified.
535 A variable may be defined as a global ``constant``, which indicates that
536 the contents of the variable will **never** be modified (enabling better
537 optimization, allowing the global data to be placed in the read-only
538 section of an executable, etc). Note that variables that need runtime
539 initialization cannot be marked ``constant`` as there is a store to the
542 LLVM explicitly allows *declarations* of global variables to be marked
543 constant, even if the final definition of the global is not. This
544 capability can be used to enable slightly better optimization of the
545 program, but requires the language definition to guarantee that
546 optimizations based on the 'constantness' are valid for the translation
547 units that do not include the definition.
549 As SSA values, global variables define pointer values that are in scope
550 (i.e. they dominate) all basic blocks in the program. Global variables
551 always define a pointer to their "content" type because they describe a
552 region of memory, and all memory objects in LLVM are accessed through
555 Global variables can be marked with ``unnamed_addr`` which indicates
556 that the address is not significant, only the content. Constants marked
557 like this can be merged with other constants if they have the same
558 initializer. Note that a constant with significant address *can* be
559 merged with a ``unnamed_addr`` constant, the result being a constant
560 whose address is significant.
562 A global variable may be declared to reside in a target-specific
563 numbered address space. For targets that support them, address spaces
564 may affect how optimizations are performed and/or what target
565 instructions are used to access the variable. The default address space
566 is zero. The address space qualifier must precede any other attributes.
568 LLVM allows an explicit section to be specified for globals. If the
569 target supports it, it will emit globals to the section specified.
570 Additionally, the global can placed in a comdat if the target has the necessary
573 By default, global initializers are optimized by assuming that global
574 variables defined within the module are not modified from their
575 initial values before the start of the global initializer. This is
576 true even for variables potentially accessible from outside the
577 module, including those with external linkage or appearing in
578 ``@llvm.used`` or dllexported variables. This assumption may be suppressed
579 by marking the variable with ``externally_initialized``.
581 An explicit alignment may be specified for a global, which must be a
582 power of 2. If not present, or if the alignment is set to zero, the
583 alignment of the global is set by the target to whatever it feels
584 convenient. If an explicit alignment is specified, the global is forced
585 to have exactly that alignment. Targets and optimizers are not allowed
586 to over-align the global if the global has an assigned section. In this
587 case, the extra alignment could be observable: for example, code could
588 assume that the globals are densely packed in their section and try to
589 iterate over them as an array, alignment padding would break this
590 iteration. The maximum alignment is ``1 << 29``.
592 Globals can also have a :ref:`DLL storage class <dllstorageclass>`.
594 Variables and aliases can have a
595 :ref:`Thread Local Storage Model <tls_model>`.
599 [@<GlobalVarName> =] [Linkage] [Visibility] [DLLStorageClass] [ThreadLocal]
600 [unnamed_addr] [AddrSpace] [ExternallyInitialized]
601 <global | constant> <Type> [<InitializerConstant>]
602 [, section "name"] [, comdat [($name)]]
603 [, align <Alignment>]
605 For example, the following defines a global in a numbered address space
606 with an initializer, section, and alignment:
610 @G = addrspace(5) constant float 1.0, section "foo", align 4
612 The following example just declares a global variable
616 @G = external global i32
618 The following example defines a thread-local global with the
619 ``initialexec`` TLS model:
623 @G = thread_local(initialexec) global i32 0, align 4
625 .. _functionstructure:
630 LLVM function definitions consist of the "``define``" keyword, an
631 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
632 style <visibility>`, an optional :ref:`DLL storage class <dllstorageclass>`,
633 an optional :ref:`calling convention <callingconv>`,
634 an optional ``unnamed_addr`` attribute, a return type, an optional
635 :ref:`parameter attribute <paramattrs>` for the return type, a function
636 name, a (possibly empty) argument list (each with optional :ref:`parameter
637 attributes <paramattrs>`), optional :ref:`function attributes <fnattrs>`,
638 an optional section, an optional alignment,
639 an optional :ref:`comdat <langref_comdats>`,
640 an optional :ref:`garbage collector name <gc>`, an optional :ref:`prefix <prefixdata>`,
641 an optional :ref:`prologue <prologuedata>`,
642 an optional :ref:`personality <personalityfn>`,
643 an optional list of attached :ref:`metadata <metadata>`,
644 an opening curly brace, a list of basic blocks, and a closing curly brace.
646 LLVM function declarations consist of the "``declare``" keyword, an
647 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
648 style <visibility>`, an optional :ref:`DLL storage class <dllstorageclass>`,
649 an optional :ref:`calling convention <callingconv>`,
650 an optional ``unnamed_addr`` attribute, a return type, an optional
651 :ref:`parameter attribute <paramattrs>` for the return type, a function
652 name, a possibly empty list of arguments, an optional alignment, an optional
653 :ref:`garbage collector name <gc>`, an optional :ref:`prefix <prefixdata>`,
654 and an optional :ref:`prologue <prologuedata>`.
656 A function definition contains a list of basic blocks, forming the CFG (Control
657 Flow Graph) for the function. Each basic block may optionally start with a label
658 (giving the basic block a symbol table entry), contains a list of instructions,
659 and ends with a :ref:`terminator <terminators>` instruction (such as a branch or
660 function return). If an explicit label is not provided, a block is assigned an
661 implicit numbered label, using the next value from the same counter as used for
662 unnamed temporaries (:ref:`see above<identifiers>`). For example, if a function
663 entry block does not have an explicit label, it will be assigned label "%0",
664 then the first unnamed temporary in that block will be "%1", etc.
666 The first basic block in a function is special in two ways: it is
667 immediately executed on entrance to the function, and it is not allowed
668 to have predecessor basic blocks (i.e. there can not be any branches to
669 the entry block of a function). Because the block can have no
670 predecessors, it also cannot have any :ref:`PHI nodes <i_phi>`.
672 LLVM allows an explicit section to be specified for functions. If the
673 target supports it, it will emit functions to the section specified.
674 Additionally, the function can be placed in a COMDAT.
676 An explicit alignment may be specified for a function. If not present,
677 or if the alignment is set to zero, the alignment of the function is set
678 by the target to whatever it feels convenient. If an explicit alignment
679 is specified, the function is forced to have at least that much
680 alignment. All alignments must be a power of 2.
682 If the ``unnamed_addr`` attribute is given, the address is known to not
683 be significant and two identical functions can be merged.
687 define [linkage] [visibility] [DLLStorageClass]
689 <ResultType> @<FunctionName> ([argument list])
690 [unnamed_addr] [fn Attrs] [section "name"] [comdat [($name)]]
691 [align N] [gc] [prefix Constant] [prologue Constant]
692 [personality Constant] (!name !N)* { ... }
694 The argument list is a comma separated sequence of arguments where each
695 argument is of the following form:
699 <type> [parameter Attrs] [name]
707 Aliases, unlike function or variables, don't create any new data. They
708 are just a new symbol and metadata for an existing position.
710 Aliases have a name and an aliasee that is either a global value or a
713 Aliases may have an optional :ref:`linkage type <linkage>`, an optional
714 :ref:`visibility style <visibility>`, an optional :ref:`DLL storage class
715 <dllstorageclass>` and an optional :ref:`tls model <tls_model>`.
719 @<Name> = [Linkage] [Visibility] [DLLStorageClass] [ThreadLocal] [unnamed_addr] alias <AliaseeTy>, <AliaseeTy>* @<Aliasee>
721 The linkage must be one of ``private``, ``internal``, ``linkonce``, ``weak``,
722 ``linkonce_odr``, ``weak_odr``, ``external``. Note that some system linkers
723 might not correctly handle dropping a weak symbol that is aliased.
725 Aliases that are not ``unnamed_addr`` are guaranteed to have the same address as
726 the aliasee expression. ``unnamed_addr`` ones are only guaranteed to point
729 Since aliases are only a second name, some restrictions apply, of which
730 some can only be checked when producing an object file:
732 * The expression defining the aliasee must be computable at assembly
733 time. Since it is just a name, no relocations can be used.
735 * No alias in the expression can be weak as the possibility of the
736 intermediate alias being overridden cannot be represented in an
739 * No global value in the expression can be a declaration, since that
740 would require a relocation, which is not possible.
747 Comdat IR provides access to COFF and ELF object file COMDAT functionality.
749 Comdats have a name which represents the COMDAT key. All global objects that
750 specify this key will only end up in the final object file if the linker chooses
751 that key over some other key. Aliases are placed in the same COMDAT that their
752 aliasee computes to, if any.
754 Comdats have a selection kind to provide input on how the linker should
755 choose between keys in two different object files.
759 $<Name> = comdat SelectionKind
761 The selection kind must be one of the following:
764 The linker may choose any COMDAT key, the choice is arbitrary.
766 The linker may choose any COMDAT key but the sections must contain the
769 The linker will choose the section containing the largest COMDAT key.
771 The linker requires that only section with this COMDAT key exist.
773 The linker may choose any COMDAT key but the sections must contain the
776 Note that the Mach-O platform doesn't support COMDATs and ELF only supports
777 ``any`` as a selection kind.
779 Here is an example of a COMDAT group where a function will only be selected if
780 the COMDAT key's section is the largest:
784 $foo = comdat largest
785 @foo = global i32 2, comdat($foo)
787 define void @bar() comdat($foo) {
791 As a syntactic sugar the ``$name`` can be omitted if the name is the same as
797 @foo = global i32 2, comdat
800 In a COFF object file, this will create a COMDAT section with selection kind
801 ``IMAGE_COMDAT_SELECT_LARGEST`` containing the contents of the ``@foo`` symbol
802 and another COMDAT section with selection kind
803 ``IMAGE_COMDAT_SELECT_ASSOCIATIVE`` which is associated with the first COMDAT
804 section and contains the contents of the ``@bar`` symbol.
806 There are some restrictions on the properties of the global object.
807 It, or an alias to it, must have the same name as the COMDAT group when
809 The contents and size of this object may be used during link-time to determine
810 which COMDAT groups get selected depending on the selection kind.
811 Because the name of the object must match the name of the COMDAT group, the
812 linkage of the global object must not be local; local symbols can get renamed
813 if a collision occurs in the symbol table.
815 The combined use of COMDATS and section attributes may yield surprising results.
822 @g1 = global i32 42, section "sec", comdat($foo)
823 @g2 = global i32 42, section "sec", comdat($bar)
825 From the object file perspective, this requires the creation of two sections
826 with the same name. This is necessary because both globals belong to different
827 COMDAT groups and COMDATs, at the object file level, are represented by
830 Note that certain IR constructs like global variables and functions may
831 create COMDATs in the object file in addition to any which are specified using
832 COMDAT IR. This arises when the code generator is configured to emit globals
833 in individual sections (e.g. when `-data-sections` or `-function-sections`
834 is supplied to `llc`).
836 .. _namedmetadatastructure:
841 Named metadata is a collection of metadata. :ref:`Metadata
842 nodes <metadata>` (but not metadata strings) are the only valid
843 operands for a named metadata.
845 #. Named metadata are represented as a string of characters with the
846 metadata prefix. The rules for metadata names are the same as for
847 identifiers, but quoted names are not allowed. ``"\xx"`` type escapes
848 are still valid, which allows any character to be part of a name.
852 ; Some unnamed metadata nodes, which are referenced by the named metadata.
857 !name = !{!0, !1, !2}
864 The return type and each parameter of a function type may have a set of
865 *parameter attributes* associated with them. Parameter attributes are
866 used to communicate additional information about the result or
867 parameters of a function. Parameter attributes are considered to be part
868 of the function, not of the function type, so functions with different
869 parameter attributes can have the same function type.
871 Parameter attributes are simple keywords that follow the type specified.
872 If multiple parameter attributes are needed, they are space separated.
877 declare i32 @printf(i8* noalias nocapture, ...)
878 declare i32 @atoi(i8 zeroext)
879 declare signext i8 @returns_signed_char()
881 Note that any attributes for the function result (``nounwind``,
882 ``readonly``) come immediately after the argument list.
884 Currently, only the following parameter attributes are defined:
887 This indicates to the code generator that the parameter or return
888 value should be zero-extended to the extent required by the target's
889 ABI (which is usually 32-bits, but is 8-bits for a i1 on x86-64) by
890 the caller (for a parameter) or the callee (for a return value).
892 This indicates to the code generator that the parameter or return
893 value should be sign-extended to the extent required by the target's
894 ABI (which is usually 32-bits) by the caller (for a parameter) or
895 the callee (for a return value).
897 This indicates that this parameter or return value should be treated
898 in a special target-dependent fashion while emitting code for
899 a function call or return (usually, by putting it in a register as
900 opposed to memory, though some targets use it to distinguish between
901 two different kinds of registers). Use of this attribute is
904 This indicates that the pointer parameter should really be passed by
905 value to the function. The attribute implies that a hidden copy of
906 the pointee is made between the caller and the callee, so the callee
907 is unable to modify the value in the caller. This attribute is only
908 valid on LLVM pointer arguments. It is generally used to pass
909 structs and arrays by value, but is also valid on pointers to
910 scalars. The copy is considered to belong to the caller not the
911 callee (for example, ``readonly`` functions should not write to
912 ``byval`` parameters). This is not a valid attribute for return
915 The byval attribute also supports specifying an alignment with the
916 align attribute. It indicates the alignment of the stack slot to
917 form and the known alignment of the pointer specified to the call
918 site. If the alignment is not specified, then the code generator
919 makes a target-specific assumption.
925 The ``inalloca`` argument attribute allows the caller to take the
926 address of outgoing stack arguments. An ``inalloca`` argument must
927 be a pointer to stack memory produced by an ``alloca`` instruction.
928 The alloca, or argument allocation, must also be tagged with the
929 inalloca keyword. Only the last argument may have the ``inalloca``
930 attribute, and that argument is guaranteed to be passed in memory.
932 An argument allocation may be used by a call at most once because
933 the call may deallocate it. The ``inalloca`` attribute cannot be
934 used in conjunction with other attributes that affect argument
935 storage, like ``inreg``, ``nest``, ``sret``, or ``byval``. The
936 ``inalloca`` attribute also disables LLVM's implicit lowering of
937 large aggregate return values, which means that frontend authors
938 must lower them with ``sret`` pointers.
940 When the call site is reached, the argument allocation must have
941 been the most recent stack allocation that is still live, or the
942 results are undefined. It is possible to allocate additional stack
943 space after an argument allocation and before its call site, but it
944 must be cleared off with :ref:`llvm.stackrestore
947 See :doc:`InAlloca` for more information on how to use this
951 This indicates that the pointer parameter specifies the address of a
952 structure that is the return value of the function in the source
953 program. This pointer must be guaranteed by the caller to be valid:
954 loads and stores to the structure may be assumed by the callee
955 not to trap and to be properly aligned. This may only be applied to
956 the first parameter. This is not a valid attribute for return
960 This indicates that the pointer value may be assumed by the optimizer to
961 have the specified alignment.
963 Note that this attribute has additional semantics when combined with the
969 This indicates that objects accessed via pointer values
970 :ref:`based <pointeraliasing>` on the argument or return value are not also
971 accessed, during the execution of the function, via pointer values not
972 *based* on the argument or return value. The attribute on a return value
973 also has additional semantics described below. The caller shares the
974 responsibility with the callee for ensuring that these requirements are met.
975 For further details, please see the discussion of the NoAlias response in
976 :ref:`alias analysis <Must, May, or No>`.
978 Note that this definition of ``noalias`` is intentionally similar
979 to the definition of ``restrict`` in C99 for function arguments.
981 For function return values, C99's ``restrict`` is not meaningful,
982 while LLVM's ``noalias`` is. Furthermore, the semantics of the ``noalias``
983 attribute on return values are stronger than the semantics of the attribute
984 when used on function arguments. On function return values, the ``noalias``
985 attribute indicates that the function acts like a system memory allocation
986 function, returning a pointer to allocated storage disjoint from the
987 storage for any other object accessible to the caller.
990 This indicates that the callee does not make any copies of the
991 pointer that outlive the callee itself. This is not a valid
992 attribute for return values.
997 This indicates that the pointer parameter can be excised using the
998 :ref:`trampoline intrinsics <int_trampoline>`. This is not a valid
999 attribute for return values and can only be applied to one parameter.
1002 This indicates that the function always returns the argument as its return
1003 value. This is an optimization hint to the code generator when generating
1004 the caller, allowing tail call optimization and omission of register saves
1005 and restores in some cases; it is not checked or enforced when generating
1006 the callee. The parameter and the function return type must be valid
1007 operands for the :ref:`bitcast instruction <i_bitcast>`. This is not a
1008 valid attribute for return values and can only be applied to one parameter.
1011 This indicates that the parameter or return pointer is not null. This
1012 attribute may only be applied to pointer typed parameters. This is not
1013 checked or enforced by LLVM, the caller must ensure that the pointer
1014 passed in is non-null, or the callee must ensure that the returned pointer
1017 ``dereferenceable(<n>)``
1018 This indicates that the parameter or return pointer is dereferenceable. This
1019 attribute may only be applied to pointer typed parameters. A pointer that
1020 is dereferenceable can be loaded from speculatively without a risk of
1021 trapping. The number of bytes known to be dereferenceable must be provided
1022 in parentheses. It is legal for the number of bytes to be less than the
1023 size of the pointee type. The ``nonnull`` attribute does not imply
1024 dereferenceability (consider a pointer to one element past the end of an
1025 array), however ``dereferenceable(<n>)`` does imply ``nonnull`` in
1026 ``addrspace(0)`` (which is the default address space).
1028 ``dereferenceable_or_null(<n>)``
1029 This indicates that the parameter or return value isn't both
1030 non-null and non-dereferenceable (up to ``<n>`` bytes) at the same
1031 time. All non-null pointers tagged with
1032 ``dereferenceable_or_null(<n>)`` are ``dereferenceable(<n>)``.
1033 For address space 0 ``dereferenceable_or_null(<n>)`` implies that
1034 a pointer is exactly one of ``dereferenceable(<n>)`` or ``null``,
1035 and in other address spaces ``dereferenceable_or_null(<n>)``
1036 implies that a pointer is at least one of ``dereferenceable(<n>)``
1037 or ``null`` (i.e. it may be both ``null`` and
1038 ``dereferenceable(<n>)``). This attribute may only be applied to
1039 pointer typed parameters.
1043 Garbage Collector Strategy Names
1044 --------------------------------
1046 Each function may specify a garbage collector strategy name, which is simply a
1049 .. code-block:: llvm
1051 define void @f() gc "name" { ... }
1053 The supported values of *name* includes those :ref:`built in to LLVM
1054 <builtin-gc-strategies>` and any provided by loaded plugins. Specifying a GC
1055 strategy will cause the compiler to alter its output in order to support the
1056 named garbage collection algorithm. Note that LLVM itself does not contain a
1057 garbage collector, this functionality is restricted to generating machine code
1058 which can interoperate with a collector provided externally.
1065 Prefix data is data associated with a function which the code
1066 generator will emit immediately before the function's entrypoint.
1067 The purpose of this feature is to allow frontends to associate
1068 language-specific runtime metadata with specific functions and make it
1069 available through the function pointer while still allowing the
1070 function pointer to be called.
1072 To access the data for a given function, a program may bitcast the
1073 function pointer to a pointer to the constant's type and dereference
1074 index -1. This implies that the IR symbol points just past the end of
1075 the prefix data. For instance, take the example of a function annotated
1076 with a single ``i32``,
1078 .. code-block:: llvm
1080 define void @f() prefix i32 123 { ... }
1082 The prefix data can be referenced as,
1084 .. code-block:: llvm
1086 %0 = bitcast void* () @f to i32*
1087 %a = getelementptr inbounds i32, i32* %0, i32 -1
1088 %b = load i32, i32* %a
1090 Prefix data is laid out as if it were an initializer for a global variable
1091 of the prefix data's type. The function will be placed such that the
1092 beginning of the prefix data is aligned. This means that if the size
1093 of the prefix data is not a multiple of the alignment size, the
1094 function's entrypoint will not be aligned. If alignment of the
1095 function's entrypoint is desired, padding must be added to the prefix
1098 A function may have prefix data but no body. This has similar semantics
1099 to the ``available_externally`` linkage in that the data may be used by the
1100 optimizers but will not be emitted in the object file.
1107 The ``prologue`` attribute allows arbitrary code (encoded as bytes) to
1108 be inserted prior to the function body. This can be used for enabling
1109 function hot-patching and instrumentation.
1111 To maintain the semantics of ordinary function calls, the prologue data must
1112 have a particular format. Specifically, it must begin with a sequence of
1113 bytes which decode to a sequence of machine instructions, valid for the
1114 module's target, which transfer control to the point immediately succeeding
1115 the prologue data, without performing any other visible action. This allows
1116 the inliner and other passes to reason about the semantics of the function
1117 definition without needing to reason about the prologue data. Obviously this
1118 makes the format of the prologue data highly target dependent.
1120 A trivial example of valid prologue data for the x86 architecture is ``i8 144``,
1121 which encodes the ``nop`` instruction:
1123 .. code-block:: llvm
1125 define void @f() prologue i8 144 { ... }
1127 Generally prologue data can be formed by encoding a relative branch instruction
1128 which skips the metadata, as in this example of valid prologue data for the
1129 x86_64 architecture, where the first two bytes encode ``jmp .+10``:
1131 .. code-block:: llvm
1133 %0 = type <{ i8, i8, i8* }>
1135 define void @f() prologue %0 <{ i8 235, i8 8, i8* @md}> { ... }
1137 A function may have prologue data but no body. This has similar semantics
1138 to the ``available_externally`` linkage in that the data may be used by the
1139 optimizers but will not be emitted in the object file.
1143 Personality Function
1144 --------------------
1146 The ``personality`` attribute permits functions to specify what function
1147 to use for exception handling.
1154 Attribute groups are groups of attributes that are referenced by objects within
1155 the IR. They are important for keeping ``.ll`` files readable, because a lot of
1156 functions will use the same set of attributes. In the degenerative case of a
1157 ``.ll`` file that corresponds to a single ``.c`` file, the single attribute
1158 group will capture the important command line flags used to build that file.
1160 An attribute group is a module-level object. To use an attribute group, an
1161 object references the attribute group's ID (e.g. ``#37``). An object may refer
1162 to more than one attribute group. In that situation, the attributes from the
1163 different groups are merged.
1165 Here is an example of attribute groups for a function that should always be
1166 inlined, has a stack alignment of 4, and which shouldn't use SSE instructions:
1168 .. code-block:: llvm
1170 ; Target-independent attributes:
1171 attributes #0 = { alwaysinline alignstack=4 }
1173 ; Target-dependent attributes:
1174 attributes #1 = { "no-sse" }
1176 ; Function @f has attributes: alwaysinline, alignstack=4, and "no-sse".
1177 define void @f() #0 #1 { ... }
1184 Function attributes are set to communicate additional information about
1185 a function. Function attributes are considered to be part of the
1186 function, not of the function type, so functions with different function
1187 attributes can have the same function type.
1189 Function attributes are simple keywords that follow the type specified.
1190 If multiple attributes are needed, they are space separated. For
1193 .. code-block:: llvm
1195 define void @f() noinline { ... }
1196 define void @f() alwaysinline { ... }
1197 define void @f() alwaysinline optsize { ... }
1198 define void @f() optsize { ... }
1201 This attribute indicates that, when emitting the prologue and
1202 epilogue, the backend should forcibly align the stack pointer.
1203 Specify the desired alignment, which must be a power of two, in
1206 This attribute indicates that the inliner should attempt to inline
1207 this function into callers whenever possible, ignoring any active
1208 inlining size threshold for this caller.
1210 This indicates that the callee function at a call site should be
1211 recognized as a built-in function, even though the function's declaration
1212 uses the ``nobuiltin`` attribute. This is only valid at call sites for
1213 direct calls to functions that are declared with the ``nobuiltin``
1216 This attribute indicates that this function is rarely called. When
1217 computing edge weights, basic blocks post-dominated by a cold
1218 function call are also considered to be cold; and, thus, given low
1221 This attribute indicates that the callee is dependent on a convergent
1222 thread execution pattern under certain parallel execution models.
1223 Transformations that are execution model agnostic may not make the execution
1224 of a convergent operation control dependent on any additional values.
1226 This attribute indicates that the source code contained a hint that
1227 inlining this function is desirable (such as the "inline" keyword in
1228 C/C++). It is just a hint; it imposes no requirements on the
1231 This attribute indicates that the function should be added to a
1232 jump-instruction table at code-generation time, and that all address-taken
1233 references to this function should be replaced with a reference to the
1234 appropriate jump-instruction-table function pointer. Note that this creates
1235 a new pointer for the original function, which means that code that depends
1236 on function-pointer identity can break. So, any function annotated with
1237 ``jumptable`` must also be ``unnamed_addr``.
1239 This attribute suggests that optimization passes and code generator
1240 passes make choices that keep the code size of this function as small
1241 as possible and perform optimizations that may sacrifice runtime
1242 performance in order to minimize the size of the generated code.
1244 This attribute disables prologue / epilogue emission for the
1245 function. This can have very system-specific consequences.
1247 This indicates that the callee function at a call site is not recognized as
1248 a built-in function. LLVM will retain the original call and not replace it
1249 with equivalent code based on the semantics of the built-in function, unless
1250 the call site uses the ``builtin`` attribute. This is valid at call sites
1251 and on function declarations and definitions.
1253 This attribute indicates that calls to the function cannot be
1254 duplicated. A call to a ``noduplicate`` function may be moved
1255 within its parent function, but may not be duplicated within
1256 its parent function.
1258 A function containing a ``noduplicate`` call may still
1259 be an inlining candidate, provided that the call is not
1260 duplicated by inlining. That implies that the function has
1261 internal linkage and only has one call site, so the original
1262 call is dead after inlining.
1264 This attributes disables implicit floating point instructions.
1266 This attribute indicates that the inliner should never inline this
1267 function in any situation. This attribute may not be used together
1268 with the ``alwaysinline`` attribute.
1270 This attribute suppresses lazy symbol binding for the function. This
1271 may make calls to the function faster, at the cost of extra program
1272 startup time if the function is not called during program startup.
1274 This attribute indicates that the code generator should not use a
1275 red zone, even if the target-specific ABI normally permits it.
1277 This function attribute indicates that the function never returns
1278 normally. This produces undefined behavior at runtime if the
1279 function ever does dynamically return.
1281 This function attribute indicates that the function does not call itself
1282 either directly or indirectly down any possible call path. This produces
1283 undefined behavior at runtime if the function ever does recurse.
1285 This function attribute indicates that the function never raises an
1286 exception. If the function does raise an exception, its runtime
1287 behavior is undefined. However, functions marked nounwind may still
1288 trap or generate asynchronous exceptions. Exception handling schemes
1289 that are recognized by LLVM to handle asynchronous exceptions, such
1290 as SEH, will still provide their implementation defined semantics.
1292 This function attribute indicates that most optimization passes will skip
1293 this function, with the exception of interprocedural optimization passes.
1294 Code generation defaults to the "fast" instruction selector.
1295 This attribute cannot be used together with the ``alwaysinline``
1296 attribute; this attribute is also incompatible
1297 with the ``minsize`` attribute and the ``optsize`` attribute.
1299 This attribute requires the ``noinline`` attribute to be specified on
1300 the function as well, so the function is never inlined into any caller.
1301 Only functions with the ``alwaysinline`` attribute are valid
1302 candidates for inlining into the body of this function.
1304 This attribute suggests that optimization passes and code generator
1305 passes make choices that keep the code size of this function low,
1306 and otherwise do optimizations specifically to reduce code size as
1307 long as they do not significantly impact runtime performance.
1309 On a function, this attribute indicates that the function computes its
1310 result (or decides to unwind an exception) based strictly on its arguments,
1311 without dereferencing any pointer arguments or otherwise accessing
1312 any mutable state (e.g. memory, control registers, etc) visible to
1313 caller functions. It does not write through any pointer arguments
1314 (including ``byval`` arguments) and never changes any state visible
1315 to callers. This means that it cannot unwind exceptions by calling
1316 the ``C++`` exception throwing methods.
1318 On an argument, this attribute indicates that the function does not
1319 dereference that pointer argument, even though it may read or write the
1320 memory that the pointer points to if accessed through other pointers.
1322 On a function, this attribute indicates that the function does not write
1323 through any pointer arguments (including ``byval`` arguments) or otherwise
1324 modify any state (e.g. memory, control registers, etc) visible to
1325 caller functions. It may dereference pointer arguments and read
1326 state that may be set in the caller. A readonly function always
1327 returns the same value (or unwinds an exception identically) when
1328 called with the same set of arguments and global state. It cannot
1329 unwind an exception by calling the ``C++`` exception throwing
1332 On an argument, this attribute indicates that the function does not write
1333 through this pointer argument, even though it may write to the memory that
1334 the pointer points to.
1336 This attribute indicates that the only memory accesses inside function are
1337 loads and stores from objects pointed to by its pointer-typed arguments,
1338 with arbitrary offsets. Or in other words, all memory operations in the
1339 function can refer to memory only using pointers based on its function
1341 Note that ``argmemonly`` can be used together with ``readonly`` attribute
1342 in order to specify that function reads only from its arguments.
1344 This attribute indicates that this function can return twice. The C
1345 ``setjmp`` is an example of such a function. The compiler disables
1346 some optimizations (like tail calls) in the caller of these
1349 This attribute indicates that
1350 `SafeStack <http://clang.llvm.org/docs/SafeStack.html>`_
1351 protection is enabled for this function.
1353 If a function that has a ``safestack`` attribute is inlined into a
1354 function that doesn't have a ``safestack`` attribute or which has an
1355 ``ssp``, ``sspstrong`` or ``sspreq`` attribute, then the resulting
1356 function will have a ``safestack`` attribute.
1357 ``sanitize_address``
1358 This attribute indicates that AddressSanitizer checks
1359 (dynamic address safety analysis) are enabled for this function.
1361 This attribute indicates that MemorySanitizer checks (dynamic detection
1362 of accesses to uninitialized memory) are enabled for this function.
1364 This attribute indicates that ThreadSanitizer checks
1365 (dynamic thread safety analysis) are enabled for this function.
1367 This attribute indicates that the function should emit a stack
1368 smashing protector. It is in the form of a "canary" --- a random value
1369 placed on the stack before the local variables that's checked upon
1370 return from the function to see if it has been overwritten. A
1371 heuristic is used to determine if a function needs stack protectors
1372 or not. The heuristic used will enable protectors for functions with:
1374 - Character arrays larger than ``ssp-buffer-size`` (default 8).
1375 - Aggregates containing character arrays larger than ``ssp-buffer-size``.
1376 - Calls to alloca() with variable sizes or constant sizes greater than
1377 ``ssp-buffer-size``.
1379 Variables that are identified as requiring a protector will be arranged
1380 on the stack such that they are adjacent to the stack protector guard.
1382 If a function that has an ``ssp`` attribute is inlined into a
1383 function that doesn't have an ``ssp`` attribute, then the resulting
1384 function will have an ``ssp`` attribute.
1386 This attribute indicates that the function should *always* emit a
1387 stack smashing protector. This overrides the ``ssp`` function
1390 Variables that are identified as requiring a protector will be arranged
1391 on the stack such that they are adjacent to the stack protector guard.
1392 The specific layout rules are:
1394 #. Large arrays and structures containing large arrays
1395 (``>= ssp-buffer-size``) are closest to the stack protector.
1396 #. Small arrays and structures containing small arrays
1397 (``< ssp-buffer-size``) are 2nd closest to the protector.
1398 #. Variables that have had their address taken are 3rd closest to the
1401 If a function that has an ``sspreq`` attribute is inlined into a
1402 function that doesn't have an ``sspreq`` attribute or which has an
1403 ``ssp`` or ``sspstrong`` attribute, then the resulting function will have
1404 an ``sspreq`` attribute.
1406 This attribute indicates that the function should emit a stack smashing
1407 protector. This attribute causes a strong heuristic to be used when
1408 determining if a function needs stack protectors. The strong heuristic
1409 will enable protectors for functions with:
1411 - Arrays of any size and type
1412 - Aggregates containing an array of any size and type.
1413 - Calls to alloca().
1414 - Local variables that have had their address taken.
1416 Variables that are identified as requiring a protector will be arranged
1417 on the stack such that they are adjacent to the stack protector guard.
1418 The specific layout rules are:
1420 #. Large arrays and structures containing large arrays
1421 (``>= ssp-buffer-size``) are closest to the stack protector.
1422 #. Small arrays and structures containing small arrays
1423 (``< ssp-buffer-size``) are 2nd closest to the protector.
1424 #. Variables that have had their address taken are 3rd closest to the
1427 This overrides the ``ssp`` function attribute.
1429 If a function that has an ``sspstrong`` attribute is inlined into a
1430 function that doesn't have an ``sspstrong`` attribute, then the
1431 resulting function will have an ``sspstrong`` attribute.
1433 This attribute indicates that the function will delegate to some other
1434 function with a tail call. The prototype of a thunk should not be used for
1435 optimization purposes. The caller is expected to cast the thunk prototype to
1436 match the thunk target prototype.
1438 This attribute indicates that the ABI being targeted requires that
1439 an unwind table entry be produced for this function even if we can
1440 show that no exceptions passes by it. This is normally the case for
1441 the ELF x86-64 abi, but it can be disabled for some compilation
1450 Note: operand bundles are a work in progress, and they should be
1451 considered experimental at this time.
1453 Operand bundles are tagged sets of SSA values that can be associated
1454 with certain LLVM instructions (currently only ``call`` s and
1455 ``invoke`` s). In a way they are like metadata, but dropping them is
1456 incorrect and will change program semantics.
1460 operand bundle set ::= '[' operand bundle (, operand bundle )* ']'
1461 operand bundle ::= tag '(' [ bundle operand ] (, bundle operand )* ')'
1462 bundle operand ::= SSA value
1463 tag ::= string constant
1465 Operand bundles are **not** part of a function's signature, and a
1466 given function may be called from multiple places with different kinds
1467 of operand bundles. This reflects the fact that the operand bundles
1468 are conceptually a part of the ``call`` (or ``invoke``), not the
1469 callee being dispatched to.
1471 Operand bundles are a generic mechanism intended to support
1472 runtime-introspection-like functionality for managed languages. While
1473 the exact semantics of an operand bundle depend on the bundle tag,
1474 there are certain limitations to how much the presence of an operand
1475 bundle can influence the semantics of a program. These restrictions
1476 are described as the semantics of an "unknown" operand bundle. As
1477 long as the behavior of an operand bundle is describable within these
1478 restrictions, LLVM does not need to have special knowledge of the
1479 operand bundle to not miscompile programs containing it.
1481 - The bundle operands for an unknown operand bundle escape in unknown
1482 ways before control is transferred to the callee or invokee.
1483 - Calls and invokes with operand bundles have unknown read / write
1484 effect on the heap on entry and exit (even if the call target is
1485 ``readnone`` or ``readonly``), unless they're overriden with
1486 callsite specific attributes.
1487 - An operand bundle at a call site cannot change the implementation
1488 of the called function. Inter-procedural optimizations work as
1489 usual as long as they take into account the first two properties.
1491 More specific types of operand bundles are described below.
1493 Deoptimization Operand Bundles
1494 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1496 Deoptimization operand bundles are characterized by the ``"deopt"``
1497 operand bundle tag. These operand bundles represent an alternate
1498 "safe" continuation for the call site they're attached to, and can be
1499 used by a suitable runtime to deoptimize the compiled frame at the
1500 specified call site. There can be at most one ``"deopt"`` operand
1501 bundle attached to a call site. Exact details of deoptimization is
1502 out of scope for the language reference, but it usually involves
1503 rewriting a compiled frame into a set of interpreted frames.
1505 From the compiler's perspective, deoptimization operand bundles make
1506 the call sites they're attached to at least ``readonly``. They read
1507 through all of their pointer typed operands (even if they're not
1508 otherwise escaped) and the entire visible heap. Deoptimization
1509 operand bundles do not capture their operands except during
1510 deoptimization, in which case control will not be returned to the
1513 The inliner knows how to inline through calls that have deoptimization
1514 operand bundles. Just like inlining through a normal call site
1515 involves composing the normal and exceptional continuations, inlining
1516 through a call site with a deoptimization operand bundle needs to
1517 appropriately compose the "safe" deoptimization continuation. The
1518 inliner does this by prepending the parent's deoptimization
1519 continuation to every deoptimization continuation in the inlined body.
1520 E.g. inlining ``@f`` into ``@g`` in the following example
1522 .. code-block:: llvm
1525 call void @x() ;; no deopt state
1526 call void @y() [ "deopt"(i32 10) ]
1527 call void @y() [ "deopt"(i32 10), "unknown"(i8* null) ]
1532 call void @f() [ "deopt"(i32 20) ]
1538 .. code-block:: llvm
1541 call void @x() ;; still no deopt state
1542 call void @y() [ "deopt"(i32 20, i32 10) ]
1543 call void @y() [ "deopt"(i32 20, i32 10), "unknown"(i8* null) ]
1547 It is the frontend's responsibility to structure or encode the
1548 deoptimization state in a way that syntactically prepending the
1549 caller's deoptimization state to the callee's deoptimization state is
1550 semantically equivalent to composing the caller's deoptimization
1551 continuation after the callee's deoptimization continuation.
1555 Module-Level Inline Assembly
1556 ----------------------------
1558 Modules may contain "module-level inline asm" blocks, which corresponds
1559 to the GCC "file scope inline asm" blocks. These blocks are internally
1560 concatenated by LLVM and treated as a single unit, but may be separated
1561 in the ``.ll`` file if desired. The syntax is very simple:
1563 .. code-block:: llvm
1565 module asm "inline asm code goes here"
1566 module asm "more can go here"
1568 The strings can contain any character by escaping non-printable
1569 characters. The escape sequence used is simply "\\xx" where "xx" is the
1570 two digit hex code for the number.
1572 Note that the assembly string *must* be parseable by LLVM's integrated assembler
1573 (unless it is disabled), even when emitting a ``.s`` file.
1575 .. _langref_datalayout:
1580 A module may specify a target specific data layout string that specifies
1581 how data is to be laid out in memory. The syntax for the data layout is
1584 .. code-block:: llvm
1586 target datalayout = "layout specification"
1588 The *layout specification* consists of a list of specifications
1589 separated by the minus sign character ('-'). Each specification starts
1590 with a letter and may include other information after the letter to
1591 define some aspect of the data layout. The specifications accepted are
1595 Specifies that the target lays out data in big-endian form. That is,
1596 the bits with the most significance have the lowest address
1599 Specifies that the target lays out data in little-endian form. That
1600 is, the bits with the least significance have the lowest address
1603 Specifies the natural alignment of the stack in bits. Alignment
1604 promotion of stack variables is limited to the natural stack
1605 alignment to avoid dynamic stack realignment. The stack alignment
1606 must be a multiple of 8-bits. If omitted, the natural stack
1607 alignment defaults to "unspecified", which does not prevent any
1608 alignment promotions.
1609 ``p[n]:<size>:<abi>:<pref>``
1610 This specifies the *size* of a pointer and its ``<abi>`` and
1611 ``<pref>``\erred alignments for address space ``n``. All sizes are in
1612 bits. The address space, ``n``, is optional, and if not specified,
1613 denotes the default address space 0. The value of ``n`` must be
1614 in the range [1,2^23).
1615 ``i<size>:<abi>:<pref>``
1616 This specifies the alignment for an integer type of a given bit
1617 ``<size>``. The value of ``<size>`` must be in the range [1,2^23).
1618 ``v<size>:<abi>:<pref>``
1619 This specifies the alignment for a vector type of a given bit
1621 ``f<size>:<abi>:<pref>``
1622 This specifies the alignment for a floating point type of a given bit
1623 ``<size>``. Only values of ``<size>`` that are supported by the target
1624 will work. 32 (float) and 64 (double) are supported on all targets; 80
1625 or 128 (different flavors of long double) are also supported on some
1628 This specifies the alignment for an object of aggregate type.
1630 If present, specifies that llvm names are mangled in the output. The
1633 * ``e``: ELF mangling: Private symbols get a ``.L`` prefix.
1634 * ``m``: Mips mangling: Private symbols get a ``$`` prefix.
1635 * ``o``: Mach-O mangling: Private symbols get ``L`` prefix. Other
1636 symbols get a ``_`` prefix.
1637 * ``w``: Windows COFF prefix: Similar to Mach-O, but stdcall and fastcall
1638 functions also get a suffix based on the frame size.
1639 * ``x``: Windows x86 COFF prefix: Similar to Windows COFF, but use a ``_``
1640 prefix for ``__cdecl`` functions.
1641 ``n<size1>:<size2>:<size3>...``
1642 This specifies a set of native integer widths for the target CPU in
1643 bits. For example, it might contain ``n32`` for 32-bit PowerPC,
1644 ``n32:64`` for PowerPC 64, or ``n8:16:32:64`` for X86-64. Elements of
1645 this set are considered to support most general arithmetic operations
1648 On every specification that takes a ``<abi>:<pref>``, specifying the
1649 ``<pref>`` alignment is optional. If omitted, the preceding ``:``
1650 should be omitted too and ``<pref>`` will be equal to ``<abi>``.
1652 When constructing the data layout for a given target, LLVM starts with a
1653 default set of specifications which are then (possibly) overridden by
1654 the specifications in the ``datalayout`` keyword. The default
1655 specifications are given in this list:
1657 - ``E`` - big endian
1658 - ``p:64:64:64`` - 64-bit pointers with 64-bit alignment.
1659 - ``p[n]:64:64:64`` - Other address spaces are assumed to be the
1660 same as the default address space.
1661 - ``S0`` - natural stack alignment is unspecified
1662 - ``i1:8:8`` - i1 is 8-bit (byte) aligned
1663 - ``i8:8:8`` - i8 is 8-bit (byte) aligned
1664 - ``i16:16:16`` - i16 is 16-bit aligned
1665 - ``i32:32:32`` - i32 is 32-bit aligned
1666 - ``i64:32:64`` - i64 has ABI alignment of 32-bits but preferred
1667 alignment of 64-bits
1668 - ``f16:16:16`` - half is 16-bit aligned
1669 - ``f32:32:32`` - float is 32-bit aligned
1670 - ``f64:64:64`` - double is 64-bit aligned
1671 - ``f128:128:128`` - quad is 128-bit aligned
1672 - ``v64:64:64`` - 64-bit vector is 64-bit aligned
1673 - ``v128:128:128`` - 128-bit vector is 128-bit aligned
1674 - ``a:0:64`` - aggregates are 64-bit aligned
1676 When LLVM is determining the alignment for a given type, it uses the
1679 #. If the type sought is an exact match for one of the specifications,
1680 that specification is used.
1681 #. If no match is found, and the type sought is an integer type, then
1682 the smallest integer type that is larger than the bitwidth of the
1683 sought type is used. If none of the specifications are larger than
1684 the bitwidth then the largest integer type is used. For example,
1685 given the default specifications above, the i7 type will use the
1686 alignment of i8 (next largest) while both i65 and i256 will use the
1687 alignment of i64 (largest specified).
1688 #. If no match is found, and the type sought is a vector type, then the
1689 largest vector type that is smaller than the sought vector type will
1690 be used as a fall back. This happens because <128 x double> can be
1691 implemented in terms of 64 <2 x double>, for example.
1693 The function of the data layout string may not be what you expect.
1694 Notably, this is not a specification from the frontend of what alignment
1695 the code generator should use.
1697 Instead, if specified, the target data layout is required to match what
1698 the ultimate *code generator* expects. This string is used by the
1699 mid-level optimizers to improve code, and this only works if it matches
1700 what the ultimate code generator uses. There is no way to generate IR
1701 that does not embed this target-specific detail into the IR. If you
1702 don't specify the string, the default specifications will be used to
1703 generate a Data Layout and the optimization phases will operate
1704 accordingly and introduce target specificity into the IR with respect to
1705 these default specifications.
1712 A module may specify a target triple string that describes the target
1713 host. The syntax for the target triple is simply:
1715 .. code-block:: llvm
1717 target triple = "x86_64-apple-macosx10.7.0"
1719 The *target triple* string consists of a series of identifiers delimited
1720 by the minus sign character ('-'). The canonical forms are:
1724 ARCHITECTURE-VENDOR-OPERATING_SYSTEM
1725 ARCHITECTURE-VENDOR-OPERATING_SYSTEM-ENVIRONMENT
1727 This information is passed along to the backend so that it generates
1728 code for the proper architecture. It's possible to override this on the
1729 command line with the ``-mtriple`` command line option.
1731 .. _pointeraliasing:
1733 Pointer Aliasing Rules
1734 ----------------------
1736 Any memory access must be done through a pointer value associated with
1737 an address range of the memory access, otherwise the behavior is
1738 undefined. Pointer values are associated with address ranges according
1739 to the following rules:
1741 - A pointer value is associated with the addresses associated with any
1742 value it is *based* on.
1743 - An address of a global variable is associated with the address range
1744 of the variable's storage.
1745 - The result value of an allocation instruction is associated with the
1746 address range of the allocated storage.
1747 - A null pointer in the default address-space is associated with no
1749 - An integer constant other than zero or a pointer value returned from
1750 a function not defined within LLVM may be associated with address
1751 ranges allocated through mechanisms other than those provided by
1752 LLVM. Such ranges shall not overlap with any ranges of addresses
1753 allocated by mechanisms provided by LLVM.
1755 A pointer value is *based* on another pointer value according to the
1758 - A pointer value formed from a ``getelementptr`` operation is *based*
1759 on the first value operand of the ``getelementptr``.
1760 - The result value of a ``bitcast`` is *based* on the operand of the
1762 - A pointer value formed by an ``inttoptr`` is *based* on all pointer
1763 values that contribute (directly or indirectly) to the computation of
1764 the pointer's value.
1765 - The "*based* on" relationship is transitive.
1767 Note that this definition of *"based"* is intentionally similar to the
1768 definition of *"based"* in C99, though it is slightly weaker.
1770 LLVM IR does not associate types with memory. The result type of a
1771 ``load`` merely indicates the size and alignment of the memory from
1772 which to load, as well as the interpretation of the value. The first
1773 operand type of a ``store`` similarly only indicates the size and
1774 alignment of the store.
1776 Consequently, type-based alias analysis, aka TBAA, aka
1777 ``-fstrict-aliasing``, is not applicable to general unadorned LLVM IR.
1778 :ref:`Metadata <metadata>` may be used to encode additional information
1779 which specialized optimization passes may use to implement type-based
1784 Volatile Memory Accesses
1785 ------------------------
1787 Certain memory accesses, such as :ref:`load <i_load>`'s,
1788 :ref:`store <i_store>`'s, and :ref:`llvm.memcpy <int_memcpy>`'s may be
1789 marked ``volatile``. The optimizers must not change the number of
1790 volatile operations or change their order of execution relative to other
1791 volatile operations. The optimizers *may* change the order of volatile
1792 operations relative to non-volatile operations. This is not Java's
1793 "volatile" and has no cross-thread synchronization behavior.
1795 IR-level volatile loads and stores cannot safely be optimized into
1796 llvm.memcpy or llvm.memmove intrinsics even when those intrinsics are
1797 flagged volatile. Likewise, the backend should never split or merge
1798 target-legal volatile load/store instructions.
1800 .. admonition:: Rationale
1802 Platforms may rely on volatile loads and stores of natively supported
1803 data width to be executed as single instruction. For example, in C
1804 this holds for an l-value of volatile primitive type with native
1805 hardware support, but not necessarily for aggregate types. The
1806 frontend upholds these expectations, which are intentionally
1807 unspecified in the IR. The rules above ensure that IR transformations
1808 do not violate the frontend's contract with the language.
1812 Memory Model for Concurrent Operations
1813 --------------------------------------
1815 The LLVM IR does not define any way to start parallel threads of
1816 execution or to register signal handlers. Nonetheless, there are
1817 platform-specific ways to create them, and we define LLVM IR's behavior
1818 in their presence. This model is inspired by the C++0x memory model.
1820 For a more informal introduction to this model, see the :doc:`Atomics`.
1822 We define a *happens-before* partial order as the least partial order
1825 - Is a superset of single-thread program order, and
1826 - When a *synchronizes-with* ``b``, includes an edge from ``a`` to
1827 ``b``. *Synchronizes-with* pairs are introduced by platform-specific
1828 techniques, like pthread locks, thread creation, thread joining,
1829 etc., and by atomic instructions. (See also :ref:`Atomic Memory Ordering
1830 Constraints <ordering>`).
1832 Note that program order does not introduce *happens-before* edges
1833 between a thread and signals executing inside that thread.
1835 Every (defined) read operation (load instructions, memcpy, atomic
1836 loads/read-modify-writes, etc.) R reads a series of bytes written by
1837 (defined) write operations (store instructions, atomic
1838 stores/read-modify-writes, memcpy, etc.). For the purposes of this
1839 section, initialized globals are considered to have a write of the
1840 initializer which is atomic and happens before any other read or write
1841 of the memory in question. For each byte of a read R, R\ :sub:`byte`
1842 may see any write to the same byte, except:
1844 - If write\ :sub:`1` happens before write\ :sub:`2`, and
1845 write\ :sub:`2` happens before R\ :sub:`byte`, then
1846 R\ :sub:`byte` does not see write\ :sub:`1`.
1847 - If R\ :sub:`byte` happens before write\ :sub:`3`, then
1848 R\ :sub:`byte` does not see write\ :sub:`3`.
1850 Given that definition, R\ :sub:`byte` is defined as follows:
1852 - If R is volatile, the result is target-dependent. (Volatile is
1853 supposed to give guarantees which can support ``sig_atomic_t`` in
1854 C/C++, and may be used for accesses to addresses that do not behave
1855 like normal memory. It does not generally provide cross-thread
1857 - Otherwise, if there is no write to the same byte that happens before
1858 R\ :sub:`byte`, R\ :sub:`byte` returns ``undef`` for that byte.
1859 - Otherwise, if R\ :sub:`byte` may see exactly one write,
1860 R\ :sub:`byte` returns the value written by that write.
1861 - Otherwise, if R is atomic, and all the writes R\ :sub:`byte` may
1862 see are atomic, it chooses one of the values written. See the :ref:`Atomic
1863 Memory Ordering Constraints <ordering>` section for additional
1864 constraints on how the choice is made.
1865 - Otherwise R\ :sub:`byte` returns ``undef``.
1867 R returns the value composed of the series of bytes it read. This
1868 implies that some bytes within the value may be ``undef`` **without**
1869 the entire value being ``undef``. Note that this only defines the
1870 semantics of the operation; it doesn't mean that targets will emit more
1871 than one instruction to read the series of bytes.
1873 Note that in cases where none of the atomic intrinsics are used, this
1874 model places only one restriction on IR transformations on top of what
1875 is required for single-threaded execution: introducing a store to a byte
1876 which might not otherwise be stored is not allowed in general.
1877 (Specifically, in the case where another thread might write to and read
1878 from an address, introducing a store can change a load that may see
1879 exactly one write into a load that may see multiple writes.)
1883 Atomic Memory Ordering Constraints
1884 ----------------------------------
1886 Atomic instructions (:ref:`cmpxchg <i_cmpxchg>`,
1887 :ref:`atomicrmw <i_atomicrmw>`, :ref:`fence <i_fence>`,
1888 :ref:`atomic load <i_load>`, and :ref:`atomic store <i_store>`) take
1889 ordering parameters that determine which other atomic instructions on
1890 the same address they *synchronize with*. These semantics are borrowed
1891 from Java and C++0x, but are somewhat more colloquial. If these
1892 descriptions aren't precise enough, check those specs (see spec
1893 references in the :doc:`atomics guide <Atomics>`).
1894 :ref:`fence <i_fence>` instructions treat these orderings somewhat
1895 differently since they don't take an address. See that instruction's
1896 documentation for details.
1898 For a simpler introduction to the ordering constraints, see the
1902 The set of values that can be read is governed by the happens-before
1903 partial order. A value cannot be read unless some operation wrote
1904 it. This is intended to provide a guarantee strong enough to model
1905 Java's non-volatile shared variables. This ordering cannot be
1906 specified for read-modify-write operations; it is not strong enough
1907 to make them atomic in any interesting way.
1909 In addition to the guarantees of ``unordered``, there is a single
1910 total order for modifications by ``monotonic`` operations on each
1911 address. All modification orders must be compatible with the
1912 happens-before order. There is no guarantee that the modification
1913 orders can be combined to a global total order for the whole program
1914 (and this often will not be possible). The read in an atomic
1915 read-modify-write operation (:ref:`cmpxchg <i_cmpxchg>` and
1916 :ref:`atomicrmw <i_atomicrmw>`) reads the value in the modification
1917 order immediately before the value it writes. If one atomic read
1918 happens before another atomic read of the same address, the later
1919 read must see the same value or a later value in the address's
1920 modification order. This disallows reordering of ``monotonic`` (or
1921 stronger) operations on the same address. If an address is written
1922 ``monotonic``-ally by one thread, and other threads ``monotonic``-ally
1923 read that address repeatedly, the other threads must eventually see
1924 the write. This corresponds to the C++0x/C1x
1925 ``memory_order_relaxed``.
1927 In addition to the guarantees of ``monotonic``, a
1928 *synchronizes-with* edge may be formed with a ``release`` operation.
1929 This is intended to model C++'s ``memory_order_acquire``.
1931 In addition to the guarantees of ``monotonic``, if this operation
1932 writes a value which is subsequently read by an ``acquire``
1933 operation, it *synchronizes-with* that operation. (This isn't a
1934 complete description; see the C++0x definition of a release
1935 sequence.) This corresponds to the C++0x/C1x
1936 ``memory_order_release``.
1937 ``acq_rel`` (acquire+release)
1938 Acts as both an ``acquire`` and ``release`` operation on its
1939 address. This corresponds to the C++0x/C1x ``memory_order_acq_rel``.
1940 ``seq_cst`` (sequentially consistent)
1941 In addition to the guarantees of ``acq_rel`` (``acquire`` for an
1942 operation that only reads, ``release`` for an operation that only
1943 writes), there is a global total order on all
1944 sequentially-consistent operations on all addresses, which is
1945 consistent with the *happens-before* partial order and with the
1946 modification orders of all the affected addresses. Each
1947 sequentially-consistent read sees the last preceding write to the
1948 same address in this global order. This corresponds to the C++0x/C1x
1949 ``memory_order_seq_cst`` and Java volatile.
1953 If an atomic operation is marked ``singlethread``, it only *synchronizes
1954 with* or participates in modification and seq\_cst total orderings with
1955 other operations running in the same thread (for example, in signal
1963 LLVM IR floating-point binary ops (:ref:`fadd <i_fadd>`,
1964 :ref:`fsub <i_fsub>`, :ref:`fmul <i_fmul>`, :ref:`fdiv <i_fdiv>`,
1965 :ref:`frem <i_frem>`, :ref:`fcmp <i_fcmp>`) have the following flags that can
1966 be set to enable otherwise unsafe floating point operations
1969 No NaNs - Allow optimizations to assume the arguments and result are not
1970 NaN. Such optimizations are required to retain defined behavior over
1971 NaNs, but the value of the result is undefined.
1974 No Infs - Allow optimizations to assume the arguments and result are not
1975 +/-Inf. Such optimizations are required to retain defined behavior over
1976 +/-Inf, but the value of the result is undefined.
1979 No Signed Zeros - Allow optimizations to treat the sign of a zero
1980 argument or result as insignificant.
1983 Allow Reciprocal - Allow optimizations to use the reciprocal of an
1984 argument rather than perform division.
1987 Fast - Allow algebraically equivalent transformations that may
1988 dramatically change results in floating point (e.g. reassociate). This
1989 flag implies all the others.
1993 Use-list Order Directives
1994 -------------------------
1996 Use-list directives encode the in-memory order of each use-list, allowing the
1997 order to be recreated. ``<order-indexes>`` is a comma-separated list of
1998 indexes that are assigned to the referenced value's uses. The referenced
1999 value's use-list is immediately sorted by these indexes.
2001 Use-list directives may appear at function scope or global scope. They are not
2002 instructions, and have no effect on the semantics of the IR. When they're at
2003 function scope, they must appear after the terminator of the final basic block.
2005 If basic blocks have their address taken via ``blockaddress()`` expressions,
2006 ``uselistorder_bb`` can be used to reorder their use-lists from outside their
2013 uselistorder <ty> <value>, { <order-indexes> }
2014 uselistorder_bb @function, %block { <order-indexes> }
2020 define void @foo(i32 %arg1, i32 %arg2) {
2022 ; ... instructions ...
2024 ; ... instructions ...
2026 ; At function scope.
2027 uselistorder i32 %arg1, { 1, 0, 2 }
2028 uselistorder label %bb, { 1, 0 }
2032 uselistorder i32* @global, { 1, 2, 0 }
2033 uselistorder i32 7, { 1, 0 }
2034 uselistorder i32 (i32) @bar, { 1, 0 }
2035 uselistorder_bb @foo, %bb, { 5, 1, 3, 2, 0, 4 }
2042 The LLVM type system is one of the most important features of the
2043 intermediate representation. Being typed enables a number of
2044 optimizations to be performed on the intermediate representation
2045 directly, without having to do extra analyses on the side before the
2046 transformation. A strong type system makes it easier to read the
2047 generated code and enables novel analyses and transformations that are
2048 not feasible to perform on normal three address code representations.
2058 The void type does not represent any value and has no size.
2076 The function type can be thought of as a function signature. It consists of a
2077 return type and a list of formal parameter types. The return type of a function
2078 type is a void type or first class type --- except for :ref:`label <t_label>`
2079 and :ref:`metadata <t_metadata>` types.
2085 <returntype> (<parameter list>)
2087 ...where '``<parameter list>``' is a comma-separated list of type
2088 specifiers. Optionally, the parameter list may include a type ``...``, which
2089 indicates that the function takes a variable number of arguments. Variable
2090 argument functions can access their arguments with the :ref:`variable argument
2091 handling intrinsic <int_varargs>` functions. '``<returntype>``' is any type
2092 except :ref:`label <t_label>` and :ref:`metadata <t_metadata>`.
2096 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2097 | ``i32 (i32)`` | function taking an ``i32``, returning an ``i32`` |
2098 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2099 | ``float (i16, i32 *) *`` | :ref:`Pointer <t_pointer>` to a function that takes an ``i16`` and a :ref:`pointer <t_pointer>` to ``i32``, returning ``float``. |
2100 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2101 | ``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. |
2102 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2103 | ``{i32, i32} (i32)`` | A function taking an ``i32``, returning a :ref:`structure <t_struct>` containing two ``i32`` values |
2104 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2111 The :ref:`first class <t_firstclass>` types are perhaps the most important.
2112 Values of these types are the only ones which can be produced by
2120 These are the types that are valid in registers from CodeGen's perspective.
2129 The integer type is a very simple type that simply specifies an
2130 arbitrary bit width for the integer type desired. Any bit width from 1
2131 bit to 2\ :sup:`23`\ -1 (about 8 million) can be specified.
2139 The number of bits the integer will occupy is specified by the ``N``
2145 +----------------+------------------------------------------------+
2146 | ``i1`` | a single-bit integer. |
2147 +----------------+------------------------------------------------+
2148 | ``i32`` | a 32-bit integer. |
2149 +----------------+------------------------------------------------+
2150 | ``i1942652`` | a really big integer of over 1 million bits. |
2151 +----------------+------------------------------------------------+
2155 Floating Point Types
2156 """"""""""""""""""""
2165 - 16-bit floating point value
2168 - 32-bit floating point value
2171 - 64-bit floating point value
2174 - 128-bit floating point value (112-bit mantissa)
2177 - 80-bit floating point value (X87)
2180 - 128-bit floating point value (two 64-bits)
2187 The x86_mmx type represents a value held in an MMX register on an x86
2188 machine. The operations allowed on it are quite limited: parameters and
2189 return values, load and store, and bitcast. User-specified MMX
2190 instructions are represented as intrinsic or asm calls with arguments
2191 and/or results of this type. There are no arrays, vectors or constants
2208 The pointer type is used to specify memory locations. Pointers are
2209 commonly used to reference objects in memory.
2211 Pointer types may have an optional address space attribute defining the
2212 numbered address space where the pointed-to object resides. The default
2213 address space is number zero. The semantics of non-zero address spaces
2214 are target-specific.
2216 Note that LLVM does not permit pointers to void (``void*``) nor does it
2217 permit pointers to labels (``label*``). Use ``i8*`` instead.
2227 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2228 | ``[4 x i32]*`` | A :ref:`pointer <t_pointer>` to :ref:`array <t_array>` of four ``i32`` values. |
2229 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2230 | ``i32 (i32*) *`` | A :ref:`pointer <t_pointer>` to a :ref:`function <t_function>` that takes an ``i32*``, returning an ``i32``. |
2231 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2232 | ``i32 addrspace(5)*`` | A :ref:`pointer <t_pointer>` to an ``i32`` value that resides in address space #5. |
2233 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2242 A vector type is a simple derived type that represents a vector of
2243 elements. Vector types are used when multiple primitive data are
2244 operated in parallel using a single instruction (SIMD). A vector type
2245 requires a size (number of elements) and an underlying primitive data
2246 type. Vector types are considered :ref:`first class <t_firstclass>`.
2252 < <# elements> x <elementtype> >
2254 The number of elements is a constant integer value larger than 0;
2255 elementtype may be any integer, floating point or pointer type. Vectors
2256 of size zero are not allowed.
2260 +-------------------+--------------------------------------------------+
2261 | ``<4 x i32>`` | Vector of 4 32-bit integer values. |
2262 +-------------------+--------------------------------------------------+
2263 | ``<8 x float>`` | Vector of 8 32-bit floating-point values. |
2264 +-------------------+--------------------------------------------------+
2265 | ``<2 x i64>`` | Vector of 2 64-bit integer values. |
2266 +-------------------+--------------------------------------------------+
2267 | ``<4 x i64*>`` | Vector of 4 pointers to 64-bit integer values. |
2268 +-------------------+--------------------------------------------------+
2277 The label type represents code labels.
2292 The token type is used when a value is associated with an instruction
2293 but all uses of the value must not attempt to introspect or obscure it.
2294 As such, it is not appropriate to have a :ref:`phi <i_phi>` or
2295 :ref:`select <i_select>` of type token.
2312 The metadata type represents embedded metadata. No derived types may be
2313 created from metadata except for :ref:`function <t_function>` arguments.
2326 Aggregate Types are a subset of derived types that can contain multiple
2327 member types. :ref:`Arrays <t_array>` and :ref:`structs <t_struct>` are
2328 aggregate types. :ref:`Vectors <t_vector>` are not considered to be
2338 The array type is a very simple derived type that arranges elements
2339 sequentially in memory. The array type requires a size (number of
2340 elements) and an underlying data type.
2346 [<# elements> x <elementtype>]
2348 The number of elements is a constant integer value; ``elementtype`` may
2349 be any type with a size.
2353 +------------------+--------------------------------------+
2354 | ``[40 x i32]`` | Array of 40 32-bit integer values. |
2355 +------------------+--------------------------------------+
2356 | ``[41 x i32]`` | Array of 41 32-bit integer values. |
2357 +------------------+--------------------------------------+
2358 | ``[4 x i8]`` | Array of 4 8-bit integer values. |
2359 +------------------+--------------------------------------+
2361 Here are some examples of multidimensional arrays:
2363 +-----------------------------+----------------------------------------------------------+
2364 | ``[3 x [4 x i32]]`` | 3x4 array of 32-bit integer values. |
2365 +-----------------------------+----------------------------------------------------------+
2366 | ``[12 x [10 x float]]`` | 12x10 array of single precision floating point values. |
2367 +-----------------------------+----------------------------------------------------------+
2368 | ``[2 x [3 x [4 x i16]]]`` | 2x3x4 array of 16-bit integer values. |
2369 +-----------------------------+----------------------------------------------------------+
2371 There is no restriction on indexing beyond the end of the array implied
2372 by a static type (though there are restrictions on indexing beyond the
2373 bounds of an allocated object in some cases). This means that
2374 single-dimension 'variable sized array' addressing can be implemented in
2375 LLVM with a zero length array type. An implementation of 'pascal style
2376 arrays' in LLVM could use the type "``{ i32, [0 x float]}``", for
2386 The structure type is used to represent a collection of data members
2387 together in memory. The elements of a structure may be any type that has
2390 Structures in memory are accessed using '``load``' and '``store``' by
2391 getting a pointer to a field with the '``getelementptr``' instruction.
2392 Structures in registers are accessed using the '``extractvalue``' and
2393 '``insertvalue``' instructions.
2395 Structures may optionally be "packed" structures, which indicate that
2396 the alignment of the struct is one byte, and that there is no padding
2397 between the elements. In non-packed structs, padding between field types
2398 is inserted as defined by the DataLayout string in the module, which is
2399 required to match what the underlying code generator expects.
2401 Structures can either be "literal" or "identified". A literal structure
2402 is defined inline with other types (e.g. ``{i32, i32}*``) whereas
2403 identified types are always defined at the top level with a name.
2404 Literal types are uniqued by their contents and can never be recursive
2405 or opaque since there is no way to write one. Identified types can be
2406 recursive, can be opaqued, and are never uniqued.
2412 %T1 = type { <type list> } ; Identified normal struct type
2413 %T2 = type <{ <type list> }> ; Identified packed struct type
2417 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2418 | ``{ i32, i32, i32 }`` | A triple of three ``i32`` values |
2419 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2420 | ``{ 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``. |
2421 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2422 | ``<{ i8, i32 }>`` | A packed struct known to be 5 bytes in size. |
2423 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2427 Opaque Structure Types
2428 """"""""""""""""""""""
2432 Opaque structure types are used to represent named structure types that
2433 do not have a body specified. This corresponds (for example) to the C
2434 notion of a forward declared structure.
2445 +--------------+-------------------+
2446 | ``opaque`` | An opaque type. |
2447 +--------------+-------------------+
2454 LLVM has several different basic types of constants. This section
2455 describes them all and their syntax.
2460 **Boolean constants**
2461 The two strings '``true``' and '``false``' are both valid constants
2463 **Integer constants**
2464 Standard integers (such as '4') are constants of the
2465 :ref:`integer <t_integer>` type. Negative numbers may be used with
2467 **Floating point constants**
2468 Floating point constants use standard decimal notation (e.g.
2469 123.421), exponential notation (e.g. 1.23421e+2), or a more precise
2470 hexadecimal notation (see below). The assembler requires the exact
2471 decimal value of a floating-point constant. For example, the
2472 assembler accepts 1.25 but rejects 1.3 because 1.3 is a repeating
2473 decimal in binary. Floating point constants must have a :ref:`floating
2474 point <t_floating>` type.
2475 **Null pointer constants**
2476 The identifier '``null``' is recognized as a null pointer constant
2477 and must be of :ref:`pointer type <t_pointer>`.
2479 The identifier '``none``' is recognized as an empty token constant
2480 and must be of :ref:`token type <t_token>`.
2482 The one non-intuitive notation for constants is the hexadecimal form of
2483 floating point constants. For example, the form
2484 '``double 0x432ff973cafa8000``' is equivalent to (but harder to read
2485 than) '``double 4.5e+15``'. The only time hexadecimal floating point
2486 constants are required (and the only time that they are generated by the
2487 disassembler) is when a floating point constant must be emitted but it
2488 cannot be represented as a decimal floating point number in a reasonable
2489 number of digits. For example, NaN's, infinities, and other special
2490 values are represented in their IEEE hexadecimal format so that assembly
2491 and disassembly do not cause any bits to change in the constants.
2493 When using the hexadecimal form, constants of types half, float, and
2494 double are represented using the 16-digit form shown above (which
2495 matches the IEEE754 representation for double); half and float values
2496 must, however, be exactly representable as IEEE 754 half and single
2497 precision, respectively. Hexadecimal format is always used for long
2498 double, and there are three forms of long double. The 80-bit format used
2499 by x86 is represented as ``0xK`` followed by 20 hexadecimal digits. The
2500 128-bit format used by PowerPC (two adjacent doubles) is represented by
2501 ``0xM`` followed by 32 hexadecimal digits. The IEEE 128-bit format is
2502 represented by ``0xL`` followed by 32 hexadecimal digits. Long doubles
2503 will only work if they match the long double format on your target.
2504 The IEEE 16-bit format (half precision) is represented by ``0xH``
2505 followed by 4 hexadecimal digits. All hexadecimal formats are big-endian
2506 (sign bit at the left).
2508 There are no constants of type x86_mmx.
2510 .. _complexconstants:
2515 Complex constants are a (potentially recursive) combination of simple
2516 constants and smaller complex constants.
2518 **Structure constants**
2519 Structure constants are represented with notation similar to
2520 structure type definitions (a comma separated list of elements,
2521 surrounded by braces (``{}``)). For example:
2522 "``{ i32 4, float 17.0, i32* @G }``", where "``@G``" is declared as
2523 "``@G = external global i32``". Structure constants must have
2524 :ref:`structure type <t_struct>`, and the number and types of elements
2525 must match those specified by the type.
2527 Array constants are represented with notation similar to array type
2528 definitions (a comma separated list of elements, surrounded by
2529 square brackets (``[]``)). For example:
2530 "``[ i32 42, i32 11, i32 74 ]``". Array constants must have
2531 :ref:`array type <t_array>`, and the number and types of elements must
2532 match those specified by the type. As a special case, character array
2533 constants may also be represented as a double-quoted string using the ``c``
2534 prefix. For example: "``c"Hello World\0A\00"``".
2535 **Vector constants**
2536 Vector constants are represented with notation similar to vector
2537 type definitions (a comma separated list of elements, surrounded by
2538 less-than/greater-than's (``<>``)). For example:
2539 "``< i32 42, i32 11, i32 74, i32 100 >``". Vector constants
2540 must have :ref:`vector type <t_vector>`, and the number and types of
2541 elements must match those specified by the type.
2542 **Zero initialization**
2543 The string '``zeroinitializer``' can be used to zero initialize a
2544 value to zero of *any* type, including scalar and
2545 :ref:`aggregate <t_aggregate>` types. This is often used to avoid
2546 having to print large zero initializers (e.g. for large arrays) and
2547 is always exactly equivalent to using explicit zero initializers.
2549 A metadata node is a constant tuple without types. For example:
2550 "``!{!0, !{!2, !0}, !"test"}``". Metadata can reference constant values,
2551 for example: "``!{!0, i32 0, i8* @global, i64 (i64)* @function, !"str"}``".
2552 Unlike other typed constants that are meant to be interpreted as part of
2553 the instruction stream, metadata is a place to attach additional
2554 information such as debug info.
2556 Global Variable and Function Addresses
2557 --------------------------------------
2559 The addresses of :ref:`global variables <globalvars>` and
2560 :ref:`functions <functionstructure>` are always implicitly valid
2561 (link-time) constants. These constants are explicitly referenced when
2562 the :ref:`identifier for the global <identifiers>` is used and always have
2563 :ref:`pointer <t_pointer>` type. For example, the following is a legal LLVM
2566 .. code-block:: llvm
2570 @Z = global [2 x i32*] [ i32* @X, i32* @Y ]
2577 The string '``undef``' can be used anywhere a constant is expected, and
2578 indicates that the user of the value may receive an unspecified
2579 bit-pattern. Undefined values may be of any type (other than '``label``'
2580 or '``void``') and be used anywhere a constant is permitted.
2582 Undefined values are useful because they indicate to the compiler that
2583 the program is well defined no matter what value is used. This gives the
2584 compiler more freedom to optimize. Here are some examples of
2585 (potentially surprising) transformations that are valid (in pseudo IR):
2587 .. code-block:: llvm
2597 This is safe because all of the output bits are affected by the undef
2598 bits. Any output bit can have a zero or one depending on the input bits.
2600 .. code-block:: llvm
2611 These logical operations have bits that are not always affected by the
2612 input. For example, if ``%X`` has a zero bit, then the output of the
2613 '``and``' operation will always be a zero for that bit, no matter what
2614 the corresponding bit from the '``undef``' is. As such, it is unsafe to
2615 optimize or assume that the result of the '``and``' is '``undef``'.
2616 However, it is safe to assume that all bits of the '``undef``' could be
2617 0, and optimize the '``and``' to 0. Likewise, it is safe to assume that
2618 all the bits of the '``undef``' operand to the '``or``' could be set,
2619 allowing the '``or``' to be folded to -1.
2621 .. code-block:: llvm
2623 %A = select undef, %X, %Y
2624 %B = select undef, 42, %Y
2625 %C = select %X, %Y, undef
2635 This set of examples shows that undefined '``select``' (and conditional
2636 branch) conditions can go *either way*, but they have to come from one
2637 of the two operands. In the ``%A`` example, if ``%X`` and ``%Y`` were
2638 both known to have a clear low bit, then ``%A`` would have to have a
2639 cleared low bit. However, in the ``%C`` example, the optimizer is
2640 allowed to assume that the '``undef``' operand could be the same as
2641 ``%Y``, allowing the whole '``select``' to be eliminated.
2643 .. code-block:: llvm
2645 %A = xor undef, undef
2662 This example points out that two '``undef``' operands are not
2663 necessarily the same. This can be surprising to people (and also matches
2664 C semantics) where they assume that "``X^X``" is always zero, even if
2665 ``X`` is undefined. This isn't true for a number of reasons, but the
2666 short answer is that an '``undef``' "variable" can arbitrarily change
2667 its value over its "live range". This is true because the variable
2668 doesn't actually *have a live range*. Instead, the value is logically
2669 read from arbitrary registers that happen to be around when needed, so
2670 the value is not necessarily consistent over time. In fact, ``%A`` and
2671 ``%C`` need to have the same semantics or the core LLVM "replace all
2672 uses with" concept would not hold.
2674 .. code-block:: llvm
2682 These examples show the crucial difference between an *undefined value*
2683 and *undefined behavior*. An undefined value (like '``undef``') is
2684 allowed to have an arbitrary bit-pattern. This means that the ``%A``
2685 operation can be constant folded to '``undef``', because the '``undef``'
2686 could be an SNaN, and ``fdiv`` is not (currently) defined on SNaN's.
2687 However, in the second example, we can make a more aggressive
2688 assumption: because the ``undef`` is allowed to be an arbitrary value,
2689 we are allowed to assume that it could be zero. Since a divide by zero
2690 has *undefined behavior*, we are allowed to assume that the operation
2691 does not execute at all. This allows us to delete the divide and all
2692 code after it. Because the undefined operation "can't happen", the
2693 optimizer can assume that it occurs in dead code.
2695 .. code-block:: llvm
2697 a: store undef -> %X
2698 b: store %X -> undef
2703 These examples reiterate the ``fdiv`` example: a store *of* an undefined
2704 value can be assumed to not have any effect; we can assume that the
2705 value is overwritten with bits that happen to match what was already
2706 there. However, a store *to* an undefined location could clobber
2707 arbitrary memory, therefore, it has undefined behavior.
2714 Poison values are similar to :ref:`undef values <undefvalues>`, however
2715 they also represent the fact that an instruction or constant expression
2716 that cannot evoke side effects has nevertheless detected a condition
2717 that results in undefined behavior.
2719 There is currently no way of representing a poison value in the IR; they
2720 only exist when produced by operations such as :ref:`add <i_add>` with
2723 Poison value behavior is defined in terms of value *dependence*:
2725 - Values other than :ref:`phi <i_phi>` nodes depend on their operands.
2726 - :ref:`Phi <i_phi>` nodes depend on the operand corresponding to
2727 their dynamic predecessor basic block.
2728 - Function arguments depend on the corresponding actual argument values
2729 in the dynamic callers of their functions.
2730 - :ref:`Call <i_call>` instructions depend on the :ref:`ret <i_ret>`
2731 instructions that dynamically transfer control back to them.
2732 - :ref:`Invoke <i_invoke>` instructions depend on the
2733 :ref:`ret <i_ret>`, :ref:`resume <i_resume>`, or exception-throwing
2734 call instructions that dynamically transfer control back to them.
2735 - Non-volatile loads and stores depend on the most recent stores to all
2736 of the referenced memory addresses, following the order in the IR
2737 (including loads and stores implied by intrinsics such as
2738 :ref:`@llvm.memcpy <int_memcpy>`.)
2739 - An instruction with externally visible side effects depends on the
2740 most recent preceding instruction with externally visible side
2741 effects, following the order in the IR. (This includes :ref:`volatile
2742 operations <volatile>`.)
2743 - An instruction *control-depends* on a :ref:`terminator
2744 instruction <terminators>` if the terminator instruction has
2745 multiple successors and the instruction is always executed when
2746 control transfers to one of the successors, and may not be executed
2747 when control is transferred to another.
2748 - Additionally, an instruction also *control-depends* on a terminator
2749 instruction if the set of instructions it otherwise depends on would
2750 be different if the terminator had transferred control to a different
2752 - Dependence is transitive.
2754 Poison values have the same behavior as :ref:`undef values <undefvalues>`,
2755 with the additional effect that any instruction that has a *dependence*
2756 on a poison value has undefined behavior.
2758 Here are some examples:
2760 .. code-block:: llvm
2763 %poison = sub nuw i32 0, 1 ; Results in a poison value.
2764 %still_poison = and i32 %poison, 0 ; 0, but also poison.
2765 %poison_yet_again = getelementptr i32, i32* @h, i32 %still_poison
2766 store i32 0, i32* %poison_yet_again ; memory at @h[0] is poisoned
2768 store i32 %poison, i32* @g ; Poison value stored to memory.
2769 %poison2 = load i32, i32* @g ; Poison value loaded back from memory.
2771 store volatile i32 %poison, i32* @g ; External observation; undefined behavior.
2773 %narrowaddr = bitcast i32* @g to i16*
2774 %wideaddr = bitcast i32* @g to i64*
2775 %poison3 = load i16, i16* %narrowaddr ; Returns a poison value.
2776 %poison4 = load i64, i64* %wideaddr ; Returns a poison value.
2778 %cmp = icmp slt i32 %poison, 0 ; Returns a poison value.
2779 br i1 %cmp, label %true, label %end ; Branch to either destination.
2782 store volatile i32 0, i32* @g ; This is control-dependent on %cmp, so
2783 ; it has undefined behavior.
2787 %p = phi i32 [ 0, %entry ], [ 1, %true ]
2788 ; Both edges into this PHI are
2789 ; control-dependent on %cmp, so this
2790 ; always results in a poison value.
2792 store volatile i32 0, i32* @g ; This would depend on the store in %true
2793 ; if %cmp is true, or the store in %entry
2794 ; otherwise, so this is undefined behavior.
2796 br i1 %cmp, label %second_true, label %second_end
2797 ; The same branch again, but this time the
2798 ; true block doesn't have side effects.
2805 store volatile i32 0, i32* @g ; This time, the instruction always depends
2806 ; on the store in %end. Also, it is
2807 ; control-equivalent to %end, so this is
2808 ; well-defined (ignoring earlier undefined
2809 ; behavior in this example).
2813 Addresses of Basic Blocks
2814 -------------------------
2816 ``blockaddress(@function, %block)``
2818 The '``blockaddress``' constant computes the address of the specified
2819 basic block in the specified function, and always has an ``i8*`` type.
2820 Taking the address of the entry block is illegal.
2822 This value only has defined behavior when used as an operand to the
2823 ':ref:`indirectbr <i_indirectbr>`' instruction, or for comparisons
2824 against null. Pointer equality tests between labels addresses results in
2825 undefined behavior --- though, again, comparison against null is ok, and
2826 no label is equal to the null pointer. This may be passed around as an
2827 opaque pointer sized value as long as the bits are not inspected. This
2828 allows ``ptrtoint`` and arithmetic to be performed on these values so
2829 long as the original value is reconstituted before the ``indirectbr``
2832 Finally, some targets may provide defined semantics when using the value
2833 as the operand to an inline assembly, but that is target specific.
2837 Constant Expressions
2838 --------------------
2840 Constant expressions are used to allow expressions involving other
2841 constants to be used as constants. Constant expressions may be of any
2842 :ref:`first class <t_firstclass>` type and may involve any LLVM operation
2843 that does not have side effects (e.g. load and call are not supported).
2844 The following is the syntax for constant expressions:
2846 ``trunc (CST to TYPE)``
2847 Truncate a constant to another type. The bit size of CST must be
2848 larger than the bit size of TYPE. Both types must be integers.
2849 ``zext (CST to TYPE)``
2850 Zero extend a constant to another type. The bit size of CST must be
2851 smaller than the bit size of TYPE. Both types must be integers.
2852 ``sext (CST to TYPE)``
2853 Sign extend a constant to another type. The bit size of CST must be
2854 smaller than the bit size of TYPE. Both types must be integers.
2855 ``fptrunc (CST to TYPE)``
2856 Truncate a floating point constant to another floating point type.
2857 The size of CST must be larger than the size of TYPE. Both types
2858 must be floating point.
2859 ``fpext (CST to TYPE)``
2860 Floating point extend a constant to another type. The size of CST
2861 must be smaller or equal to the size of TYPE. Both types must be
2863 ``fptoui (CST to TYPE)``
2864 Convert a floating point constant to the corresponding unsigned
2865 integer constant. TYPE must be a scalar or vector integer type. CST
2866 must be of scalar or vector floating point type. Both CST and TYPE
2867 must be scalars, or vectors of the same number of elements. If the
2868 value won't fit in the integer type, the results are undefined.
2869 ``fptosi (CST to TYPE)``
2870 Convert a floating point constant to the corresponding signed
2871 integer constant. TYPE must be a scalar or vector integer type. CST
2872 must be of scalar or vector floating point type. Both CST and TYPE
2873 must be scalars, or vectors of the same number of elements. If the
2874 value won't fit in the integer type, the results are undefined.
2875 ``uitofp (CST to TYPE)``
2876 Convert an unsigned integer constant to the corresponding floating
2877 point constant. TYPE must be a scalar or vector floating point type.
2878 CST must be of scalar or vector integer type. Both CST and TYPE must
2879 be scalars, or vectors of the same number of elements. If the value
2880 won't fit in the floating point type, the results are undefined.
2881 ``sitofp (CST to TYPE)``
2882 Convert a signed integer constant to the corresponding floating
2883 point constant. TYPE must be a scalar or vector floating point type.
2884 CST must be of scalar or vector integer type. Both CST and TYPE must
2885 be scalars, or vectors of the same number of elements. If the value
2886 won't fit in the floating point type, the results are undefined.
2887 ``ptrtoint (CST to TYPE)``
2888 Convert a pointer typed constant to the corresponding integer
2889 constant. ``TYPE`` must be an integer type. ``CST`` must be of
2890 pointer type. The ``CST`` value is zero extended, truncated, or
2891 unchanged to make it fit in ``TYPE``.
2892 ``inttoptr (CST to TYPE)``
2893 Convert an integer constant to a pointer constant. TYPE must be a
2894 pointer type. CST must be of integer type. The CST value is zero
2895 extended, truncated, or unchanged to make it fit in a pointer size.
2896 This one is *really* dangerous!
2897 ``bitcast (CST to TYPE)``
2898 Convert a constant, CST, to another TYPE. The constraints of the
2899 operands are the same as those for the :ref:`bitcast
2900 instruction <i_bitcast>`.
2901 ``addrspacecast (CST to TYPE)``
2902 Convert a constant pointer or constant vector of pointer, CST, to another
2903 TYPE in a different address space. The constraints of the operands are the
2904 same as those for the :ref:`addrspacecast instruction <i_addrspacecast>`.
2905 ``getelementptr (TY, CSTPTR, IDX0, IDX1, ...)``, ``getelementptr inbounds (TY, CSTPTR, IDX0, IDX1, ...)``
2906 Perform the :ref:`getelementptr operation <i_getelementptr>` on
2907 constants. As with the :ref:`getelementptr <i_getelementptr>`
2908 instruction, the index list may have zero or more indexes, which are
2909 required to make sense for the type of "pointer to TY".
2910 ``select (COND, VAL1, VAL2)``
2911 Perform the :ref:`select operation <i_select>` on constants.
2912 ``icmp COND (VAL1, VAL2)``
2913 Performs the :ref:`icmp operation <i_icmp>` on constants.
2914 ``fcmp COND (VAL1, VAL2)``
2915 Performs the :ref:`fcmp operation <i_fcmp>` on constants.
2916 ``extractelement (VAL, IDX)``
2917 Perform the :ref:`extractelement operation <i_extractelement>` on
2919 ``insertelement (VAL, ELT, IDX)``
2920 Perform the :ref:`insertelement operation <i_insertelement>` on
2922 ``shufflevector (VEC1, VEC2, IDXMASK)``
2923 Perform the :ref:`shufflevector operation <i_shufflevector>` on
2925 ``extractvalue (VAL, IDX0, IDX1, ...)``
2926 Perform the :ref:`extractvalue operation <i_extractvalue>` on
2927 constants. The index list is interpreted in a similar manner as
2928 indices in a ':ref:`getelementptr <i_getelementptr>`' operation. At
2929 least one index value must be specified.
2930 ``insertvalue (VAL, ELT, IDX0, IDX1, ...)``
2931 Perform the :ref:`insertvalue operation <i_insertvalue>` on constants.
2932 The index list is interpreted in a similar manner as indices in a
2933 ':ref:`getelementptr <i_getelementptr>`' operation. At least one index
2934 value must be specified.
2935 ``OPCODE (LHS, RHS)``
2936 Perform the specified operation of the LHS and RHS constants. OPCODE
2937 may be any of the :ref:`binary <binaryops>` or :ref:`bitwise
2938 binary <bitwiseops>` operations. The constraints on operands are
2939 the same as those for the corresponding instruction (e.g. no bitwise
2940 operations on floating point values are allowed).
2947 Inline Assembler Expressions
2948 ----------------------------
2950 LLVM supports inline assembler expressions (as opposed to :ref:`Module-Level
2951 Inline Assembly <moduleasm>`) through the use of a special value. This value
2952 represents the inline assembler as a template string (containing the
2953 instructions to emit), a list of operand constraints (stored as a string), a
2954 flag that indicates whether or not the inline asm expression has side effects,
2955 and a flag indicating whether the function containing the asm needs to align its
2956 stack conservatively.
2958 The template string supports argument substitution of the operands using "``$``"
2959 followed by a number, to indicate substitution of the given register/memory
2960 location, as specified by the constraint string. "``${NUM:MODIFIER}``" may also
2961 be used, where ``MODIFIER`` is a target-specific annotation for how to print the
2962 operand (See :ref:`inline-asm-modifiers`).
2964 A literal "``$``" may be included by using "``$$``" in the template. To include
2965 other special characters into the output, the usual "``\XX``" escapes may be
2966 used, just as in other strings. Note that after template substitution, the
2967 resulting assembly string is parsed by LLVM's integrated assembler unless it is
2968 disabled -- even when emitting a ``.s`` file -- and thus must contain assembly
2969 syntax known to LLVM.
2971 LLVM's support for inline asm is modeled closely on the requirements of Clang's
2972 GCC-compatible inline-asm support. Thus, the feature-set and the constraint and
2973 modifier codes listed here are similar or identical to those in GCC's inline asm
2974 support. However, to be clear, the syntax of the template and constraint strings
2975 described here is *not* the same as the syntax accepted by GCC and Clang, and,
2976 while most constraint letters are passed through as-is by Clang, some get
2977 translated to other codes when converting from the C source to the LLVM
2980 An example inline assembler expression is:
2982 .. code-block:: llvm
2984 i32 (i32) asm "bswap $0", "=r,r"
2986 Inline assembler expressions may **only** be used as the callee operand
2987 of a :ref:`call <i_call>` or an :ref:`invoke <i_invoke>` instruction.
2988 Thus, typically we have:
2990 .. code-block:: llvm
2992 %X = call i32 asm "bswap $0", "=r,r"(i32 %Y)
2994 Inline asms with side effects not visible in the constraint list must be
2995 marked as having side effects. This is done through the use of the
2996 '``sideeffect``' keyword, like so:
2998 .. code-block:: llvm
3000 call void asm sideeffect "eieio", ""()
3002 In some cases inline asms will contain code that will not work unless
3003 the stack is aligned in some way, such as calls or SSE instructions on
3004 x86, yet will not contain code that does that alignment within the asm.
3005 The compiler should make conservative assumptions about what the asm
3006 might contain and should generate its usual stack alignment code in the
3007 prologue if the '``alignstack``' keyword is present:
3009 .. code-block:: llvm
3011 call void asm alignstack "eieio", ""()
3013 Inline asms also support using non-standard assembly dialects. The
3014 assumed dialect is ATT. When the '``inteldialect``' keyword is present,
3015 the inline asm is using the Intel dialect. Currently, ATT and Intel are
3016 the only supported dialects. An example is:
3018 .. code-block:: llvm
3020 call void asm inteldialect "eieio", ""()
3022 If multiple keywords appear the '``sideeffect``' keyword must come
3023 first, the '``alignstack``' keyword second and the '``inteldialect``'
3026 Inline Asm Constraint String
3027 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3029 The constraint list is a comma-separated string, each element containing one or
3030 more constraint codes.
3032 For each element in the constraint list an appropriate register or memory
3033 operand will be chosen, and it will be made available to assembly template
3034 string expansion as ``$0`` for the first constraint in the list, ``$1`` for the
3037 There are three different types of constraints, which are distinguished by a
3038 prefix symbol in front of the constraint code: Output, Input, and Clobber. The
3039 constraints must always be given in that order: outputs first, then inputs, then
3040 clobbers. They cannot be intermingled.
3042 There are also three different categories of constraint codes:
3044 - Register constraint. This is either a register class, or a fixed physical
3045 register. This kind of constraint will allocate a register, and if necessary,
3046 bitcast the argument or result to the appropriate type.
3047 - Memory constraint. This kind of constraint is for use with an instruction
3048 taking a memory operand. Different constraints allow for different addressing
3049 modes used by the target.
3050 - Immediate value constraint. This kind of constraint is for an integer or other
3051 immediate value which can be rendered directly into an instruction. The
3052 various target-specific constraints allow the selection of a value in the
3053 proper range for the instruction you wish to use it with.
3058 Output constraints are specified by an "``=``" prefix (e.g. "``=r``"). This
3059 indicates that the assembly will write to this operand, and the operand will
3060 then be made available as a return value of the ``asm`` expression. Output
3061 constraints do not consume an argument from the call instruction. (Except, see
3062 below about indirect outputs).
3064 Normally, it is expected that no output locations are written to by the assembly
3065 expression until *all* of the inputs have been read. As such, LLVM may assign
3066 the same register to an output and an input. If this is not safe (e.g. if the
3067 assembly contains two instructions, where the first writes to one output, and
3068 the second reads an input and writes to a second output), then the "``&``"
3069 modifier must be used (e.g. "``=&r``") to specify that the output is an
3070 "early-clobber" output. Marking an ouput as "early-clobber" ensures that LLVM
3071 will not use the same register for any inputs (other than an input tied to this
3077 Input constraints do not have a prefix -- just the constraint codes. Each input
3078 constraint will consume one argument from the call instruction. It is not
3079 permitted for the asm to write to any input register or memory location (unless
3080 that input is tied to an output). Note also that multiple inputs may all be
3081 assigned to the same register, if LLVM can determine that they necessarily all
3082 contain the same value.
3084 Instead of providing a Constraint Code, input constraints may also "tie"
3085 themselves to an output constraint, by providing an integer as the constraint
3086 string. Tied inputs still consume an argument from the call instruction, and
3087 take up a position in the asm template numbering as is usual -- they will simply
3088 be constrained to always use the same register as the output they've been tied
3089 to. For example, a constraint string of "``=r,0``" says to assign a register for
3090 output, and use that register as an input as well (it being the 0'th
3093 It is permitted to tie an input to an "early-clobber" output. In that case, no
3094 *other* input may share the same register as the input tied to the early-clobber
3095 (even when the other input has the same value).
3097 You may only tie an input to an output which has a register constraint, not a
3098 memory constraint. Only a single input may be tied to an output.
3100 There is also an "interesting" feature which deserves a bit of explanation: if a
3101 register class constraint allocates a register which is too small for the value
3102 type operand provided as input, the input value will be split into multiple
3103 registers, and all of them passed to the inline asm.
3105 However, this feature is often not as useful as you might think.
3107 Firstly, the registers are *not* guaranteed to be consecutive. So, on those
3108 architectures that have instructions which operate on multiple consecutive
3109 instructions, this is not an appropriate way to support them. (e.g. the 32-bit
3110 SparcV8 has a 64-bit load, which instruction takes a single 32-bit register. The
3111 hardware then loads into both the named register, and the next register. This
3112 feature of inline asm would not be useful to support that.)
3114 A few of the targets provide a template string modifier allowing explicit access
3115 to the second register of a two-register operand (e.g. MIPS ``L``, ``M``, and
3116 ``D``). On such an architecture, you can actually access the second allocated
3117 register (yet, still, not any subsequent ones). But, in that case, you're still
3118 probably better off simply splitting the value into two separate operands, for
3119 clarity. (e.g. see the description of the ``A`` constraint on X86, which,
3120 despite existing only for use with this feature, is not really a good idea to
3123 Indirect inputs and outputs
3124 """""""""""""""""""""""""""
3126 Indirect output or input constraints can be specified by the "``*``" modifier
3127 (which goes after the "``=``" in case of an output). This indicates that the asm
3128 will write to or read from the contents of an *address* provided as an input
3129 argument. (Note that in this way, indirect outputs act more like an *input* than
3130 an output: just like an input, they consume an argument of the call expression,
3131 rather than producing a return value. An indirect output constraint is an
3132 "output" only in that the asm is expected to write to the contents of the input
3133 memory location, instead of just read from it).
3135 This is most typically used for memory constraint, e.g. "``=*m``", to pass the
3136 address of a variable as a value.
3138 It is also possible to use an indirect *register* constraint, but only on output
3139 (e.g. "``=*r``"). This will cause LLVM to allocate a register for an output
3140 value normally, and then, separately emit a store to the address provided as
3141 input, after the provided inline asm. (It's not clear what value this
3142 functionality provides, compared to writing the store explicitly after the asm
3143 statement, and it can only produce worse code, since it bypasses many
3144 optimization passes. I would recommend not using it.)
3150 A clobber constraint is indicated by a "``~``" prefix. A clobber does not
3151 consume an input operand, nor generate an output. Clobbers cannot use any of the
3152 general constraint code letters -- they may use only explicit register
3153 constraints, e.g. "``~{eax}``". The one exception is that a clobber string of
3154 "``~{memory}``" indicates that the assembly writes to arbitrary undeclared
3155 memory locations -- not only the memory pointed to by a declared indirect
3161 After a potential prefix comes constraint code, or codes.
3163 A Constraint Code is either a single letter (e.g. "``r``"), a "``^``" character
3164 followed by two letters (e.g. "``^wc``"), or "``{``" register-name "``}``"
3167 The one and two letter constraint codes are typically chosen to be the same as
3168 GCC's constraint codes.
3170 A single constraint may include one or more than constraint code in it, leaving
3171 it up to LLVM to choose which one to use. This is included mainly for
3172 compatibility with the translation of GCC inline asm coming from clang.
3174 There are two ways to specify alternatives, and either or both may be used in an
3175 inline asm constraint list:
3177 1) Append the codes to each other, making a constraint code set. E.g. "``im``"
3178 or "``{eax}m``". This means "choose any of the options in the set". The
3179 choice of constraint is made independently for each constraint in the
3182 2) Use "``|``" between constraint code sets, creating alternatives. Every
3183 constraint in the constraint list must have the same number of alternative
3184 sets. With this syntax, the same alternative in *all* of the items in the
3185 constraint list will be chosen together.
3187 Putting those together, you might have a two operand constraint string like
3188 ``"rm|r,ri|rm"``. This indicates that if operand 0 is ``r`` or ``m``, then
3189 operand 1 may be one of ``r`` or ``i``. If operand 0 is ``r``, then operand 1
3190 may be one of ``r`` or ``m``. But, operand 0 and 1 cannot both be of type m.
3192 However, the use of either of the alternatives features is *NOT* recommended, as
3193 LLVM is not able to make an intelligent choice about which one to use. (At the
3194 point it currently needs to choose, not enough information is available to do so
3195 in a smart way.) Thus, it simply tries to make a choice that's most likely to
3196 compile, not one that will be optimal performance. (e.g., given "``rm``", it'll
3197 always choose to use memory, not registers). And, if given multiple registers,
3198 or multiple register classes, it will simply choose the first one. (In fact, it
3199 doesn't currently even ensure explicitly specified physical registers are
3200 unique, so specifying multiple physical registers as alternatives, like
3201 ``{r11}{r12},{r11}{r12}``, will assign r11 to both operands, not at all what was
3204 Supported Constraint Code List
3205 """"""""""""""""""""""""""""""
3207 The constraint codes are, in general, expected to behave the same way they do in
3208 GCC. LLVM's support is often implemented on an 'as-needed' basis, to support C
3209 inline asm code which was supported by GCC. A mismatch in behavior between LLVM
3210 and GCC likely indicates a bug in LLVM.
3212 Some constraint codes are typically supported by all targets:
3214 - ``r``: A register in the target's general purpose register class.
3215 - ``m``: A memory address operand. It is target-specific what addressing modes
3216 are supported, typical examples are register, or register + register offset,
3217 or register + immediate offset (of some target-specific size).
3218 - ``i``: An integer constant (of target-specific width). Allows either a simple
3219 immediate, or a relocatable value.
3220 - ``n``: An integer constant -- *not* including relocatable values.
3221 - ``s``: An integer constant, but allowing *only* relocatable values.
3222 - ``X``: Allows an operand of any kind, no constraint whatsoever. Typically
3223 useful to pass a label for an asm branch or call.
3225 .. FIXME: but that surely isn't actually okay to jump out of an asm
3226 block without telling llvm about the control transfer???)
3228 - ``{register-name}``: Requires exactly the named physical register.
3230 Other constraints are target-specific:
3234 - ``z``: An immediate integer 0. Outputs ``WZR`` or ``XZR``, as appropriate.
3235 - ``I``: An immediate integer valid for an ``ADD`` or ``SUB`` instruction,
3236 i.e. 0 to 4095 with optional shift by 12.
3237 - ``J``: An immediate integer that, when negated, is valid for an ``ADD`` or
3238 ``SUB`` instruction, i.e. -1 to -4095 with optional left shift by 12.
3239 - ``K``: An immediate integer that is valid for the 'bitmask immediate 32' of a
3240 logical instruction like ``AND``, ``EOR``, or ``ORR`` with a 32-bit register.
3241 - ``L``: An immediate integer that is valid for the 'bitmask immediate 64' of a
3242 logical instruction like ``AND``, ``EOR``, or ``ORR`` with a 64-bit register.
3243 - ``M``: An immediate integer for use with the ``MOV`` assembly alias on a
3244 32-bit register. This is a superset of ``K``: in addition to the bitmask
3245 immediate, also allows immediate integers which can be loaded with a single
3246 ``MOVZ`` or ``MOVL`` instruction.
3247 - ``N``: An immediate integer for use with the ``MOV`` assembly alias on a
3248 64-bit register. This is a superset of ``L``.
3249 - ``Q``: Memory address operand must be in a single register (no
3250 offsets). (However, LLVM currently does this for the ``m`` constraint as
3252 - ``r``: A 32 or 64-bit integer register (W* or X*).
3253 - ``w``: A 32, 64, or 128-bit floating-point/SIMD register.
3254 - ``x``: A lower 128-bit floating-point/SIMD register (``V0`` to ``V15``).
3258 - ``r``: A 32 or 64-bit integer register.
3259 - ``[0-9]v``: The 32-bit VGPR register, number 0-9.
3260 - ``[0-9]s``: The 32-bit SGPR register, number 0-9.
3265 - ``Q``, ``Um``, ``Un``, ``Uq``, ``Us``, ``Ut``, ``Uv``, ``Uy``: Memory address
3266 operand. Treated the same as operand ``m``, at the moment.
3268 ARM and ARM's Thumb2 mode:
3270 - ``j``: An immediate integer between 0 and 65535 (valid for ``MOVW``)
3271 - ``I``: An immediate integer valid for a data-processing instruction.
3272 - ``J``: An immediate integer between -4095 and 4095.
3273 - ``K``: An immediate integer whose bitwise inverse is valid for a
3274 data-processing instruction. (Can be used with template modifier "``B``" to
3275 print the inverted value).
3276 - ``L``: An immediate integer whose negation is valid for a data-processing
3277 instruction. (Can be used with template modifier "``n``" to print the negated
3279 - ``M``: A power of two or a integer between 0 and 32.
3280 - ``N``: Invalid immediate constraint.
3281 - ``O``: Invalid immediate constraint.
3282 - ``r``: A general-purpose 32-bit integer register (``r0-r15``).
3283 - ``l``: In Thumb2 mode, low 32-bit GPR registers (``r0-r7``). In ARM mode, same
3285 - ``h``: In Thumb2 mode, a high 32-bit GPR register (``r8-r15``). In ARM mode,
3287 - ``w``: A 32, 64, or 128-bit floating-point/SIMD register: ``s0-s31``,
3288 ``d0-d31``, or ``q0-q15``.
3289 - ``x``: A 32, 64, or 128-bit floating-point/SIMD register: ``s0-s15``,
3290 ``d0-d7``, or ``q0-q3``.
3291 - ``t``: A floating-point/SIMD register, only supports 32-bit values:
3296 - ``I``: An immediate integer between 0 and 255.
3297 - ``J``: An immediate integer between -255 and -1.
3298 - ``K``: An immediate integer between 0 and 255, with optional left-shift by
3300 - ``L``: An immediate integer between -7 and 7.
3301 - ``M``: An immediate integer which is a multiple of 4 between 0 and 1020.
3302 - ``N``: An immediate integer between 0 and 31.
3303 - ``O``: An immediate integer which is a multiple of 4 between -508 and 508.
3304 - ``r``: A low 32-bit GPR register (``r0-r7``).
3305 - ``l``: A low 32-bit GPR register (``r0-r7``).
3306 - ``h``: A high GPR register (``r0-r7``).
3307 - ``w``: A 32, 64, or 128-bit floating-point/SIMD register: ``s0-s31``,
3308 ``d0-d31``, or ``q0-q15``.
3309 - ``x``: A 32, 64, or 128-bit floating-point/SIMD register: ``s0-s15``,
3310 ``d0-d7``, or ``q0-q3``.
3311 - ``t``: A floating-point/SIMD register, only supports 32-bit values:
3317 - ``o``, ``v``: A memory address operand, treated the same as constraint ``m``,
3319 - ``r``: A 32 or 64-bit register.
3323 - ``r``: An 8 or 16-bit register.
3327 - ``I``: An immediate signed 16-bit integer.
3328 - ``J``: An immediate integer zero.
3329 - ``K``: An immediate unsigned 16-bit integer.
3330 - ``L``: An immediate 32-bit integer, where the lower 16 bits are 0.
3331 - ``N``: An immediate integer between -65535 and -1.
3332 - ``O``: An immediate signed 15-bit integer.
3333 - ``P``: An immediate integer between 1 and 65535.
3334 - ``m``: A memory address operand. In MIPS-SE mode, allows a base address
3335 register plus 16-bit immediate offset. In MIPS mode, just a base register.
3336 - ``R``: A memory address operand. In MIPS-SE mode, allows a base address
3337 register plus a 9-bit signed offset. In MIPS mode, the same as constraint
3339 - ``ZC``: A memory address operand, suitable for use in a ``pref``, ``ll``, or
3340 ``sc`` instruction on the given subtarget (details vary).
3341 - ``r``, ``d``, ``y``: A 32 or 64-bit GPR register.
3342 - ``f``: A 32 or 64-bit FPU register (``F0-F31``), or a 128-bit MSA register
3343 (``W0-W31``). In the case of MSA registers, it is recommended to use the ``w``
3344 argument modifier for compatibility with GCC.
3345 - ``c``: A 32-bit or 64-bit GPR register suitable for indirect jump (always
3347 - ``l``: The ``lo`` register, 32 or 64-bit.
3352 - ``b``: A 1-bit integer register.
3353 - ``c`` or ``h``: A 16-bit integer register.
3354 - ``r``: A 32-bit integer register.
3355 - ``l`` or ``N``: A 64-bit integer register.
3356 - ``f``: A 32-bit float register.
3357 - ``d``: A 64-bit float register.
3362 - ``I``: An immediate signed 16-bit integer.
3363 - ``J``: An immediate unsigned 16-bit integer, shifted left 16 bits.
3364 - ``K``: An immediate unsigned 16-bit integer.
3365 - ``L``: An immediate signed 16-bit integer, shifted left 16 bits.
3366 - ``M``: An immediate integer greater than 31.
3367 - ``N``: An immediate integer that is an exact power of 2.
3368 - ``O``: The immediate integer constant 0.
3369 - ``P``: An immediate integer constant whose negation is a signed 16-bit
3371 - ``es``, ``o``, ``Q``, ``Z``, ``Zy``: A memory address operand, currently
3372 treated the same as ``m``.
3373 - ``r``: A 32 or 64-bit integer register.
3374 - ``b``: A 32 or 64-bit integer register, excluding ``R0`` (that is:
3376 - ``f``: A 32 or 64-bit float register (``F0-F31``), or when QPX is enabled, a
3377 128 or 256-bit QPX register (``Q0-Q31``; aliases the ``F`` registers).
3378 - ``v``: For ``4 x f32`` or ``4 x f64`` types, when QPX is enabled, a
3379 128 or 256-bit QPX register (``Q0-Q31``), otherwise a 128-bit
3380 altivec vector register (``V0-V31``).
3382 .. FIXME: is this a bug that v accepts QPX registers? I think this
3383 is supposed to only use the altivec vector registers?
3385 - ``y``: Condition register (``CR0-CR7``).
3386 - ``wc``: An individual CR bit in a CR register.
3387 - ``wa``, ``wd``, ``wf``: Any 128-bit VSX vector register, from the full VSX
3388 register set (overlapping both the floating-point and vector register files).
3389 - ``ws``: A 32 or 64-bit floating point register, from the full VSX register
3394 - ``I``: An immediate 13-bit signed integer.
3395 - ``r``: A 32-bit integer register.
3399 - ``I``: An immediate unsigned 8-bit integer.
3400 - ``J``: An immediate unsigned 12-bit integer.
3401 - ``K``: An immediate signed 16-bit integer.
3402 - ``L``: An immediate signed 20-bit integer.
3403 - ``M``: An immediate integer 0x7fffffff.
3404 - ``Q``, ``R``, ``S``, ``T``: A memory address operand, treated the same as
3405 ``m``, at the moment.
3406 - ``r`` or ``d``: A 32, 64, or 128-bit integer register.
3407 - ``a``: A 32, 64, or 128-bit integer address register (excludes R0, which in an
3408 address context evaluates as zero).
3409 - ``h``: A 32-bit value in the high part of a 64bit data register
3411 - ``f``: A 32, 64, or 128-bit floating point register.
3415 - ``I``: An immediate integer between 0 and 31.
3416 - ``J``: An immediate integer between 0 and 64.
3417 - ``K``: An immediate signed 8-bit integer.
3418 - ``L``: An immediate integer, 0xff or 0xffff or (in 64-bit mode only)
3420 - ``M``: An immediate integer between 0 and 3.
3421 - ``N``: An immediate unsigned 8-bit integer.
3422 - ``O``: An immediate integer between 0 and 127.
3423 - ``e``: An immediate 32-bit signed integer.
3424 - ``Z``: An immediate 32-bit unsigned integer.
3425 - ``o``, ``v``: Treated the same as ``m``, at the moment.
3426 - ``q``: An 8, 16, 32, or 64-bit register which can be accessed as an 8-bit
3427 ``l`` integer register. On X86-32, this is the ``a``, ``b``, ``c``, and ``d``
3428 registers, and on X86-64, it is all of the integer registers.
3429 - ``Q``: An 8, 16, 32, or 64-bit register which can be accessed as an 8-bit
3430 ``h`` integer register. This is the ``a``, ``b``, ``c``, and ``d`` registers.
3431 - ``r`` or ``l``: An 8, 16, 32, or 64-bit integer register.
3432 - ``R``: An 8, 16, 32, or 64-bit "legacy" integer register -- one which has
3433 existed since i386, and can be accessed without the REX prefix.
3434 - ``f``: A 32, 64, or 80-bit '387 FPU stack pseudo-register.
3435 - ``y``: A 64-bit MMX register, if MMX is enabled.
3436 - ``x``: If SSE is enabled: a 32 or 64-bit scalar operand, or 128-bit vector
3437 operand in a SSE register. If AVX is also enabled, can also be a 256-bit
3438 vector operand in an AVX register. If AVX-512 is also enabled, can also be a
3439 512-bit vector operand in an AVX512 register, Otherwise, an error.
3440 - ``Y``: The same as ``x``, if *SSE2* is enabled, otherwise an error.
3441 - ``A``: Special case: allocates EAX first, then EDX, for a single operand (in
3442 32-bit mode, a 64-bit integer operand will get split into two registers). It
3443 is not recommended to use this constraint, as in 64-bit mode, the 64-bit
3444 operand will get allocated only to RAX -- if two 32-bit operands are needed,
3445 you're better off splitting it yourself, before passing it to the asm
3450 - ``r``: A 32-bit integer register.
3453 .. _inline-asm-modifiers:
3455 Asm template argument modifiers
3456 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3458 In the asm template string, modifiers can be used on the operand reference, like
3461 The modifiers are, in general, expected to behave the same way they do in
3462 GCC. LLVM's support is often implemented on an 'as-needed' basis, to support C
3463 inline asm code which was supported by GCC. A mismatch in behavior between LLVM
3464 and GCC likely indicates a bug in LLVM.
3468 - ``c``: Print an immediate integer constant unadorned, without
3469 the target-specific immediate punctuation (e.g. no ``$`` prefix).
3470 - ``n``: Negate and print immediate integer constant unadorned, without the
3471 target-specific immediate punctuation (e.g. no ``$`` prefix).
3472 - ``l``: Print as an unadorned label, without the target-specific label
3473 punctuation (e.g. no ``$`` prefix).
3477 - ``w``: Print a GPR register with a ``w*`` name instead of ``x*`` name. E.g.,
3478 instead of ``x30``, print ``w30``.
3479 - ``x``: Print a GPR register with a ``x*`` name. (this is the default, anyhow).
3480 - ``b``, ``h``, ``s``, ``d``, ``q``: Print a floating-point/SIMD register with a
3481 ``b*``, ``h*``, ``s*``, ``d*``, or ``q*`` name, rather than the default of
3490 - ``a``: Print an operand as an address (with ``[`` and ``]`` surrounding a
3494 - ``y``: Print a VFP single-precision register as an indexed double (e.g. print
3495 as ``d4[1]`` instead of ``s9``)
3496 - ``B``: Bitwise invert and print an immediate integer constant without ``#``
3498 - ``L``: Print the low 16-bits of an immediate integer constant.
3499 - ``M``: Print as a register set suitable for ldm/stm. Also prints *all*
3500 register operands subsequent to the specified one (!), so use carefully.
3501 - ``Q``: Print the low-order register of a register-pair, or the low-order
3502 register of a two-register operand.
3503 - ``R``: Print the high-order register of a register-pair, or the high-order
3504 register of a two-register operand.
3505 - ``H``: Print the second register of a register-pair. (On a big-endian system,
3506 ``H`` is equivalent to ``Q``, and on little-endian system, ``H`` is equivalent
3509 .. FIXME: H doesn't currently support printing the second register
3510 of a two-register operand.
3512 - ``e``: Print the low doubleword register of a NEON quad register.
3513 - ``f``: Print the high doubleword register of a NEON quad register.
3514 - ``m``: Print the base register of a memory operand without the ``[`` and ``]``
3519 - ``L``: Print the second register of a two-register operand. Requires that it
3520 has been allocated consecutively to the first.
3522 .. FIXME: why is it restricted to consecutive ones? And there's
3523 nothing that ensures that happens, is there?
3525 - ``I``: Print the letter 'i' if the operand is an integer constant, otherwise
3526 nothing. Used to print 'addi' vs 'add' instructions.
3530 No additional modifiers.
3534 - ``X``: Print an immediate integer as hexadecimal
3535 - ``x``: Print the low 16 bits of an immediate integer as hexadecimal.
3536 - ``d``: Print an immediate integer as decimal.
3537 - ``m``: Subtract one and print an immediate integer as decimal.
3538 - ``z``: Print $0 if an immediate zero, otherwise print normally.
3539 - ``L``: Print the low-order register of a two-register operand, or prints the
3540 address of the low-order word of a double-word memory operand.
3542 .. FIXME: L seems to be missing memory operand support.
3544 - ``M``: Print the high-order register of a two-register operand, or prints the
3545 address of the high-order word of a double-word memory operand.
3547 .. FIXME: M seems to be missing memory operand support.
3549 - ``D``: Print the second register of a two-register operand, or prints the
3550 second word of a double-word memory operand. (On a big-endian system, ``D`` is
3551 equivalent to ``L``, and on little-endian system, ``D`` is equivalent to
3553 - ``w``: No effect. Provided for compatibility with GCC which requires this
3554 modifier in order to print MSA registers (``W0-W31``) with the ``f``
3563 - ``L``: Print the second register of a two-register operand. Requires that it
3564 has been allocated consecutively to the first.
3566 .. FIXME: why is it restricted to consecutive ones? And there's
3567 nothing that ensures that happens, is there?
3569 - ``I``: Print the letter 'i' if the operand is an integer constant, otherwise
3570 nothing. Used to print 'addi' vs 'add' instructions.
3571 - ``y``: For a memory operand, prints formatter for a two-register X-form
3572 instruction. (Currently always prints ``r0,OPERAND``).
3573 - ``U``: Prints 'u' if the memory operand is an update form, and nothing
3574 otherwise. (NOTE: LLVM does not support update form, so this will currently
3575 always print nothing)
3576 - ``X``: Prints 'x' if the memory operand is an indexed form. (NOTE: LLVM does
3577 not support indexed form, so this will currently always print nothing)
3585 SystemZ implements only ``n``, and does *not* support any of the other
3586 target-independent modifiers.
3590 - ``c``: Print an unadorned integer or symbol name. (The latter is
3591 target-specific behavior for this typically target-independent modifier).
3592 - ``A``: Print a register name with a '``*``' before it.
3593 - ``b``: Print an 8-bit register name (e.g. ``al``); do nothing on a memory
3595 - ``h``: Print the upper 8-bit register name (e.g. ``ah``); do nothing on a
3597 - ``w``: Print the 16-bit register name (e.g. ``ax``); do nothing on a memory
3599 - ``k``: Print the 32-bit register name (e.g. ``eax``); do nothing on a memory
3601 - ``q``: Print the 64-bit register name (e.g. ``rax``), if 64-bit registers are
3602 available, otherwise the 32-bit register name; do nothing on a memory operand.
3603 - ``n``: Negate and print an unadorned integer, or, for operands other than an
3604 immediate integer (e.g. a relocatable symbol expression), print a '-' before
3605 the operand. (The behavior for relocatable symbol expressions is a
3606 target-specific behavior for this typically target-independent modifier)
3607 - ``H``: Print a memory reference with additional offset +8.
3608 - ``P``: Print a memory reference or operand for use as the argument of a call
3609 instruction. (E.g. omit ``(rip)``, even though it's PC-relative.)
3613 No additional modifiers.
3619 The call instructions that wrap inline asm nodes may have a
3620 "``!srcloc``" MDNode attached to it that contains a list of constant
3621 integers. If present, the code generator will use the integer as the
3622 location cookie value when report errors through the ``LLVMContext``
3623 error reporting mechanisms. This allows a front-end to correlate backend
3624 errors that occur with inline asm back to the source code that produced
3627 .. code-block:: llvm
3629 call void asm sideeffect "something bad", ""(), !srcloc !42
3631 !42 = !{ i32 1234567 }
3633 It is up to the front-end to make sense of the magic numbers it places
3634 in the IR. If the MDNode contains multiple constants, the code generator
3635 will use the one that corresponds to the line of the asm that the error
3643 LLVM IR allows metadata to be attached to instructions in the program
3644 that can convey extra information about the code to the optimizers and
3645 code generator. One example application of metadata is source-level
3646 debug information. There are two metadata primitives: strings and nodes.
3648 Metadata does not have a type, and is not a value. If referenced from a
3649 ``call`` instruction, it uses the ``metadata`` type.
3651 All metadata are identified in syntax by a exclamation point ('``!``').
3653 .. _metadata-string:
3655 Metadata Nodes and Metadata Strings
3656 -----------------------------------
3658 A metadata string is a string surrounded by double quotes. It can
3659 contain any character by escaping non-printable characters with
3660 "``\xx``" where "``xx``" is the two digit hex code. For example:
3663 Metadata nodes are represented with notation similar to structure
3664 constants (a comma separated list of elements, surrounded by braces and
3665 preceded by an exclamation point). Metadata nodes can have any values as
3666 their operand. For example:
3668 .. code-block:: llvm
3670 !{ !"test\00", i32 10}
3672 Metadata nodes that aren't uniqued use the ``distinct`` keyword. For example:
3674 .. code-block:: llvm
3676 !0 = distinct !{!"test\00", i32 10}
3678 ``distinct`` nodes are useful when nodes shouldn't be merged based on their
3679 content. They can also occur when transformations cause uniquing collisions
3680 when metadata operands change.
3682 A :ref:`named metadata <namedmetadatastructure>` is a collection of
3683 metadata nodes, which can be looked up in the module symbol table. For
3686 .. code-block:: llvm
3690 Metadata can be used as function arguments. Here ``llvm.dbg.value``
3691 function is using two metadata arguments:
3693 .. code-block:: llvm
3695 call void @llvm.dbg.value(metadata !24, i64 0, metadata !25)
3697 Metadata can be attached to an instruction. Here metadata ``!21`` is attached
3698 to the ``add`` instruction using the ``!dbg`` identifier:
3700 .. code-block:: llvm
3702 %indvar.next = add i64 %indvar, 1, !dbg !21
3704 Metadata can also be attached to a function definition. Here metadata ``!22``
3705 is attached to the ``foo`` function using the ``!dbg`` identifier:
3707 .. code-block:: llvm
3709 define void @foo() !dbg !22 {
3713 More information about specific metadata nodes recognized by the
3714 optimizers and code generator is found below.
3716 .. _specialized-metadata:
3718 Specialized Metadata Nodes
3719 ^^^^^^^^^^^^^^^^^^^^^^^^^^
3721 Specialized metadata nodes are custom data structures in metadata (as opposed
3722 to generic tuples). Their fields are labelled, and can be specified in any
3725 These aren't inherently debug info centric, but currently all the specialized
3726 metadata nodes are related to debug info.
3733 ``DICompileUnit`` nodes represent a compile unit. The ``enums:``,
3734 ``retainedTypes:``, ``subprograms:``, ``globals:`` and ``imports:`` fields are
3735 tuples containing the debug info to be emitted along with the compile unit,
3736 regardless of code optimizations (some nodes are only emitted if there are
3737 references to them from instructions).
3739 .. code-block:: llvm
3741 !0 = !DICompileUnit(language: DW_LANG_C99, file: !1, producer: "clang",
3742 isOptimized: true, flags: "-O2", runtimeVersion: 2,
3743 splitDebugFilename: "abc.debug", emissionKind: 1,
3744 enums: !2, retainedTypes: !3, subprograms: !4,
3745 globals: !5, imports: !6)
3747 Compile unit descriptors provide the root scope for objects declared in a
3748 specific compilation unit. File descriptors are defined using this scope.
3749 These descriptors are collected by a named metadata ``!llvm.dbg.cu``. They
3750 keep track of subprograms, global variables, type information, and imported
3751 entities (declarations and namespaces).
3758 ``DIFile`` nodes represent files. The ``filename:`` can include slashes.
3760 .. code-block:: llvm
3762 !0 = !DIFile(filename: "path/to/file", directory: "/path/to/dir")
3764 Files are sometimes used in ``scope:`` fields, and are the only valid target
3765 for ``file:`` fields.
3772 ``DIBasicType`` nodes represent primitive types, such as ``int``, ``bool`` and
3773 ``float``. ``tag:`` defaults to ``DW_TAG_base_type``.
3775 .. code-block:: llvm
3777 !0 = !DIBasicType(name: "unsigned char", size: 8, align: 8,
3778 encoding: DW_ATE_unsigned_char)
3779 !1 = !DIBasicType(tag: DW_TAG_unspecified_type, name: "decltype(nullptr)")
3781 The ``encoding:`` describes the details of the type. Usually it's one of the
3784 .. code-block:: llvm
3790 DW_ATE_signed_char = 6
3792 DW_ATE_unsigned_char = 8
3794 .. _DISubroutineType:
3799 ``DISubroutineType`` nodes represent subroutine types. Their ``types:`` field
3800 refers to a tuple; the first operand is the return type, while the rest are the
3801 types of the formal arguments in order. If the first operand is ``null``, that
3802 represents a function with no return value (such as ``void foo() {}`` in C++).
3804 .. code-block:: llvm
3806 !0 = !BasicType(name: "int", size: 32, align: 32, DW_ATE_signed)
3807 !1 = !BasicType(name: "char", size: 8, align: 8, DW_ATE_signed_char)
3808 !2 = !DISubroutineType(types: !{null, !0, !1}) ; void (int, char)
3815 ``DIDerivedType`` nodes represent types derived from other types, such as
3818 .. code-block:: llvm
3820 !0 = !DIBasicType(name: "unsigned char", size: 8, align: 8,
3821 encoding: DW_ATE_unsigned_char)
3822 !1 = !DIDerivedType(tag: DW_TAG_pointer_type, baseType: !0, size: 32,
3825 The following ``tag:`` values are valid:
3827 .. code-block:: llvm
3829 DW_TAG_formal_parameter = 5
3831 DW_TAG_pointer_type = 15
3832 DW_TAG_reference_type = 16
3834 DW_TAG_ptr_to_member_type = 31
3835 DW_TAG_const_type = 38
3836 DW_TAG_volatile_type = 53
3837 DW_TAG_restrict_type = 55
3839 ``DW_TAG_member`` is used to define a member of a :ref:`composite type
3840 <DICompositeType>` or :ref:`subprogram <DISubprogram>`. The type of the member
3841 is the ``baseType:``. The ``offset:`` is the member's bit offset.
3842 ``DW_TAG_formal_parameter`` is used to define a member which is a formal
3843 argument of a subprogram.
3845 ``DW_TAG_typedef`` is used to provide a name for the ``baseType:``.
3847 ``DW_TAG_pointer_type``, ``DW_TAG_reference_type``, ``DW_TAG_const_type``,
3848 ``DW_TAG_volatile_type`` and ``DW_TAG_restrict_type`` are used to qualify the
3851 Note that the ``void *`` type is expressed as a type derived from NULL.
3853 .. _DICompositeType:
3858 ``DICompositeType`` nodes represent types composed of other types, like
3859 structures and unions. ``elements:`` points to a tuple of the composed types.
3861 If the source language supports ODR, the ``identifier:`` field gives the unique
3862 identifier used for type merging between modules. When specified, other types
3863 can refer to composite types indirectly via a :ref:`metadata string
3864 <metadata-string>` that matches their identifier.
3866 .. code-block:: llvm
3868 !0 = !DIEnumerator(name: "SixKind", value: 7)
3869 !1 = !DIEnumerator(name: "SevenKind", value: 7)
3870 !2 = !DIEnumerator(name: "NegEightKind", value: -8)
3871 !3 = !DICompositeType(tag: DW_TAG_enumeration_type, name: "Enum", file: !12,
3872 line: 2, size: 32, align: 32, identifier: "_M4Enum",
3873 elements: !{!0, !1, !2})
3875 The following ``tag:`` values are valid:
3877 .. code-block:: llvm
3879 DW_TAG_array_type = 1
3880 DW_TAG_class_type = 2
3881 DW_TAG_enumeration_type = 4
3882 DW_TAG_structure_type = 19
3883 DW_TAG_union_type = 23
3884 DW_TAG_subroutine_type = 21
3885 DW_TAG_inheritance = 28
3888 For ``DW_TAG_array_type``, the ``elements:`` should be :ref:`subrange
3889 descriptors <DISubrange>`, each representing the range of subscripts at that
3890 level of indexing. The ``DIFlagVector`` flag to ``flags:`` indicates that an
3891 array type is a native packed vector.
3893 For ``DW_TAG_enumeration_type``, the ``elements:`` should be :ref:`enumerator
3894 descriptors <DIEnumerator>`, each representing the definition of an enumeration
3895 value for the set. All enumeration type descriptors are collected in the
3896 ``enums:`` field of the :ref:`compile unit <DICompileUnit>`.
3898 For ``DW_TAG_structure_type``, ``DW_TAG_class_type``, and
3899 ``DW_TAG_union_type``, the ``elements:`` should be :ref:`derived types
3900 <DIDerivedType>` with ``tag: DW_TAG_member`` or ``tag: DW_TAG_inheritance``.
3907 ``DISubrange`` nodes are the elements for ``DW_TAG_array_type`` variants of
3908 :ref:`DICompositeType`. ``count: -1`` indicates an empty array.
3910 .. code-block:: llvm
3912 !0 = !DISubrange(count: 5, lowerBound: 0) ; array counting from 0
3913 !1 = !DISubrange(count: 5, lowerBound: 1) ; array counting from 1
3914 !2 = !DISubrange(count: -1) ; empty array.
3921 ``DIEnumerator`` nodes are the elements for ``DW_TAG_enumeration_type``
3922 variants of :ref:`DICompositeType`.
3924 .. code-block:: llvm
3926 !0 = !DIEnumerator(name: "SixKind", value: 7)
3927 !1 = !DIEnumerator(name: "SevenKind", value: 7)
3928 !2 = !DIEnumerator(name: "NegEightKind", value: -8)
3930 DITemplateTypeParameter
3931 """""""""""""""""""""""
3933 ``DITemplateTypeParameter`` nodes represent type parameters to generic source
3934 language constructs. They are used (optionally) in :ref:`DICompositeType` and
3935 :ref:`DISubprogram` ``templateParams:`` fields.
3937 .. code-block:: llvm
3939 !0 = !DITemplateTypeParameter(name: "Ty", type: !1)
3941 DITemplateValueParameter
3942 """"""""""""""""""""""""
3944 ``DITemplateValueParameter`` nodes represent value parameters to generic source
3945 language constructs. ``tag:`` defaults to ``DW_TAG_template_value_parameter``,
3946 but if specified can also be set to ``DW_TAG_GNU_template_template_param`` or
3947 ``DW_TAG_GNU_template_param_pack``. They are used (optionally) in
3948 :ref:`DICompositeType` and :ref:`DISubprogram` ``templateParams:`` fields.
3950 .. code-block:: llvm
3952 !0 = !DITemplateValueParameter(name: "Ty", type: !1, value: i32 7)
3957 ``DINamespace`` nodes represent namespaces in the source language.
3959 .. code-block:: llvm
3961 !0 = !DINamespace(name: "myawesomeproject", scope: !1, file: !2, line: 7)
3966 ``DIGlobalVariable`` nodes represent global variables in the source language.
3968 .. code-block:: llvm
3970 !0 = !DIGlobalVariable(name: "foo", linkageName: "foo", scope: !1,
3971 file: !2, line: 7, type: !3, isLocal: true,
3972 isDefinition: false, variable: i32* @foo,
3975 All global variables should be referenced by the `globals:` field of a
3976 :ref:`compile unit <DICompileUnit>`.
3983 ``DISubprogram`` nodes represent functions from the source language. A
3984 ``DISubprogram`` may be attached to a function definition using ``!dbg``
3985 metadata. The ``variables:`` field points at :ref:`variables <DILocalVariable>`
3986 that must be retained, even if their IR counterparts are optimized out of
3987 the IR. The ``type:`` field must point at an :ref:`DISubroutineType`.
3989 .. code-block:: llvm
3991 define void @_Z3foov() !dbg !0 {
3995 !0 = distinct !DISubprogram(name: "foo", linkageName: "_Zfoov", scope: !1,
3996 file: !2, line: 7, type: !3, isLocal: true,
3997 isDefinition: false, scopeLine: 8,
3999 virtuality: DW_VIRTUALITY_pure_virtual,
4000 virtualIndex: 10, flags: DIFlagPrototyped,
4001 isOptimized: true, templateParams: !5,
4002 declaration: !6, variables: !7)
4009 ``DILexicalBlock`` nodes describe nested blocks within a :ref:`subprogram
4010 <DISubprogram>`. The line number and column numbers are used to distinguish
4011 two lexical blocks at same depth. They are valid targets for ``scope:``
4014 .. code-block:: llvm
4016 !0 = distinct !DILexicalBlock(scope: !1, file: !2, line: 7, column: 35)
4018 Usually lexical blocks are ``distinct`` to prevent node merging based on
4021 .. _DILexicalBlockFile:
4026 ``DILexicalBlockFile`` nodes are used to discriminate between sections of a
4027 :ref:`lexical block <DILexicalBlock>`. The ``file:`` field can be changed to
4028 indicate textual inclusion, or the ``discriminator:`` field can be used to
4029 discriminate between control flow within a single block in the source language.
4031 .. code-block:: llvm
4033 !0 = !DILexicalBlock(scope: !3, file: !4, line: 7, column: 35)
4034 !1 = !DILexicalBlockFile(scope: !0, file: !4, discriminator: 0)
4035 !2 = !DILexicalBlockFile(scope: !0, file: !4, discriminator: 1)
4042 ``DILocation`` nodes represent source debug locations. The ``scope:`` field is
4043 mandatory, and points at an :ref:`DILexicalBlockFile`, an
4044 :ref:`DILexicalBlock`, or an :ref:`DISubprogram`.
4046 .. code-block:: llvm
4048 !0 = !DILocation(line: 2900, column: 42, scope: !1, inlinedAt: !2)
4050 .. _DILocalVariable:
4055 ``DILocalVariable`` nodes represent local variables in the source language. If
4056 the ``arg:`` field is set to non-zero, then this variable is a subprogram
4057 parameter, and it will be included in the ``variables:`` field of its
4058 :ref:`DISubprogram`.
4060 .. code-block:: llvm
4062 !0 = !DILocalVariable(name: "this", arg: 1, scope: !3, file: !2, line: 7,
4063 type: !3, flags: DIFlagArtificial)
4064 !1 = !DILocalVariable(name: "x", arg: 2, scope: !4, file: !2, line: 7,
4066 !2 = !DILocalVariable(name: "y", scope: !5, file: !2, line: 7, type: !3)
4071 ``DIExpression`` nodes represent DWARF expression sequences. They are used in
4072 :ref:`debug intrinsics<dbg_intrinsics>` (such as ``llvm.dbg.declare``) to
4073 describe how the referenced LLVM variable relates to the source language
4076 The current supported vocabulary is limited:
4078 - ``DW_OP_deref`` dereferences the working expression.
4079 - ``DW_OP_plus, 93`` adds ``93`` to the working expression.
4080 - ``DW_OP_bit_piece, 16, 8`` specifies the offset and size (``16`` and ``8``
4081 here, respectively) of the variable piece from the working expression.
4083 .. code-block:: llvm
4085 !0 = !DIExpression(DW_OP_deref)
4086 !1 = !DIExpression(DW_OP_plus, 3)
4087 !2 = !DIExpression(DW_OP_bit_piece, 3, 7)
4088 !3 = !DIExpression(DW_OP_deref, DW_OP_plus, 3, DW_OP_bit_piece, 3, 7)
4093 ``DIObjCProperty`` nodes represent Objective-C property nodes.
4095 .. code-block:: llvm
4097 !3 = !DIObjCProperty(name: "foo", file: !1, line: 7, setter: "setFoo",
4098 getter: "getFoo", attributes: 7, type: !2)
4103 ``DIImportedEntity`` nodes represent entities (such as modules) imported into a
4106 .. code-block:: llvm
4108 !2 = !DIImportedEntity(tag: DW_TAG_imported_module, name: "foo", scope: !0,
4109 entity: !1, line: 7)
4114 In LLVM IR, memory does not have types, so LLVM's own type system is not
4115 suitable for doing TBAA. Instead, metadata is added to the IR to
4116 describe a type system of a higher level language. This can be used to
4117 implement typical C/C++ TBAA, but it can also be used to implement
4118 custom alias analysis behavior for other languages.
4120 The current metadata format is very simple. TBAA metadata nodes have up
4121 to three fields, e.g.:
4123 .. code-block:: llvm
4125 !0 = !{ !"an example type tree" }
4126 !1 = !{ !"int", !0 }
4127 !2 = !{ !"float", !0 }
4128 !3 = !{ !"const float", !2, i64 1 }
4130 The first field is an identity field. It can be any value, usually a
4131 metadata string, which uniquely identifies the type. The most important
4132 name in the tree is the name of the root node. Two trees with different
4133 root node names are entirely disjoint, even if they have leaves with
4136 The second field identifies the type's parent node in the tree, or is
4137 null or omitted for a root node. A type is considered to alias all of
4138 its descendants and all of its ancestors in the tree. Also, a type is
4139 considered to alias all types in other trees, so that bitcode produced
4140 from multiple front-ends is handled conservatively.
4142 If the third field is present, it's an integer which if equal to 1
4143 indicates that the type is "constant" (meaning
4144 ``pointsToConstantMemory`` should return true; see `other useful
4145 AliasAnalysis methods <AliasAnalysis.html#OtherItfs>`_).
4147 '``tbaa.struct``' Metadata
4148 ^^^^^^^^^^^^^^^^^^^^^^^^^^
4150 The :ref:`llvm.memcpy <int_memcpy>` is often used to implement
4151 aggregate assignment operations in C and similar languages, however it
4152 is defined to copy a contiguous region of memory, which is more than
4153 strictly necessary for aggregate types which contain holes due to
4154 padding. Also, it doesn't contain any TBAA information about the fields
4157 ``!tbaa.struct`` metadata can describe which memory subregions in a
4158 memcpy are padding and what the TBAA tags of the struct are.
4160 The current metadata format is very simple. ``!tbaa.struct`` metadata
4161 nodes are a list of operands which are in conceptual groups of three.
4162 For each group of three, the first operand gives the byte offset of a
4163 field in bytes, the second gives its size in bytes, and the third gives
4166 .. code-block:: llvm
4168 !4 = !{ i64 0, i64 4, !1, i64 8, i64 4, !2 }
4170 This describes a struct with two fields. The first is at offset 0 bytes
4171 with size 4 bytes, and has tbaa tag !1. The second is at offset 8 bytes
4172 and has size 4 bytes and has tbaa tag !2.
4174 Note that the fields need not be contiguous. In this example, there is a
4175 4 byte gap between the two fields. This gap represents padding which
4176 does not carry useful data and need not be preserved.
4178 '``noalias``' and '``alias.scope``' Metadata
4179 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4181 ``noalias`` and ``alias.scope`` metadata provide the ability to specify generic
4182 noalias memory-access sets. This means that some collection of memory access
4183 instructions (loads, stores, memory-accessing calls, etc.) that carry
4184 ``noalias`` metadata can specifically be specified not to alias with some other
4185 collection of memory access instructions that carry ``alias.scope`` metadata.
4186 Each type of metadata specifies a list of scopes where each scope has an id and
4187 a domain. When evaluating an aliasing query, if for some domain, the set
4188 of scopes with that domain in one instruction's ``alias.scope`` list is a
4189 subset of (or equal to) the set of scopes for that domain in another
4190 instruction's ``noalias`` list, then the two memory accesses are assumed not to
4193 The metadata identifying each domain is itself a list containing one or two
4194 entries. The first entry is the name of the domain. Note that if the name is a
4195 string then it can be combined across functions and translation units. A
4196 self-reference can be used to create globally unique domain names. A
4197 descriptive string may optionally be provided as a second list entry.
4199 The metadata identifying each scope is also itself a list containing two or
4200 three entries. The first entry is the name of the scope. Note that if the name
4201 is a string then it can be combined across functions and translation units. A
4202 self-reference can be used to create globally unique scope names. A metadata
4203 reference to the scope's domain is the second entry. A descriptive string may
4204 optionally be provided as a third list entry.
4208 .. code-block:: llvm
4210 ; Two scope domains:
4214 ; Some scopes in these domains:
4220 !5 = !{!4} ; A list containing only scope !4
4224 ; These two instructions don't alias:
4225 %0 = load float, float* %c, align 4, !alias.scope !5
4226 store float %0, float* %arrayidx.i, align 4, !noalias !5
4228 ; These two instructions also don't alias (for domain !1, the set of scopes
4229 ; in the !alias.scope equals that in the !noalias list):
4230 %2 = load float, float* %c, align 4, !alias.scope !5
4231 store float %2, float* %arrayidx.i2, align 4, !noalias !6
4233 ; These two instructions may alias (for domain !0, the set of scopes in
4234 ; the !noalias list is not a superset of, or equal to, the scopes in the
4235 ; !alias.scope list):
4236 %2 = load float, float* %c, align 4, !alias.scope !6
4237 store float %0, float* %arrayidx.i, align 4, !noalias !7
4239 '``fpmath``' Metadata
4240 ^^^^^^^^^^^^^^^^^^^^^
4242 ``fpmath`` metadata may be attached to any instruction of floating point
4243 type. It can be used to express the maximum acceptable error in the
4244 result of that instruction, in ULPs, thus potentially allowing the
4245 compiler to use a more efficient but less accurate method of computing
4246 it. ULP is defined as follows:
4248 If ``x`` is a real number that lies between two finite consecutive
4249 floating-point numbers ``a`` and ``b``, without being equal to one
4250 of them, then ``ulp(x) = |b - a|``, otherwise ``ulp(x)`` is the
4251 distance between the two non-equal finite floating-point numbers
4252 nearest ``x``. Moreover, ``ulp(NaN)`` is ``NaN``.
4254 The metadata node shall consist of a single positive floating point
4255 number representing the maximum relative error, for example:
4257 .. code-block:: llvm
4259 !0 = !{ float 2.5 } ; maximum acceptable inaccuracy is 2.5 ULPs
4263 '``range``' Metadata
4264 ^^^^^^^^^^^^^^^^^^^^
4266 ``range`` metadata may be attached only to ``load``, ``call`` and ``invoke`` of
4267 integer types. It expresses the possible ranges the loaded value or the value
4268 returned by the called function at this call site is in. The ranges are
4269 represented with a flattened list of integers. The loaded value or the value
4270 returned is known to be in the union of the ranges defined by each consecutive
4271 pair. Each pair has the following properties:
4273 - The type must match the type loaded by the instruction.
4274 - The pair ``a,b`` represents the range ``[a,b)``.
4275 - Both ``a`` and ``b`` are constants.
4276 - The range is allowed to wrap.
4277 - The range should not represent the full or empty set. That is,
4280 In addition, the pairs must be in signed order of the lower bound and
4281 they must be non-contiguous.
4285 .. code-block:: llvm
4287 %a = load i8, i8* %x, align 1, !range !0 ; Can only be 0 or 1
4288 %b = load i8, i8* %y, align 1, !range !1 ; Can only be 255 (-1), 0 or 1
4289 %c = call i8 @foo(), !range !2 ; Can only be 0, 1, 3, 4 or 5
4290 %d = invoke i8 @bar() to label %cont
4291 unwind label %lpad, !range !3 ; Can only be -2, -1, 3, 4 or 5
4293 !0 = !{ i8 0, i8 2 }
4294 !1 = !{ i8 255, i8 2 }
4295 !2 = !{ i8 0, i8 2, i8 3, i8 6 }
4296 !3 = !{ i8 -2, i8 0, i8 3, i8 6 }
4298 '``unpredictable``' Metadata
4299 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4301 ``unpredictable`` metadata may be attached to any branch or switch
4302 instruction. It can be used to express the unpredictability of control
4303 flow. Similar to the llvm.expect intrinsic, it may be used to alter
4304 optimizations related to compare and branch instructions. The metadata
4305 is treated as a boolean value; if it exists, it signals that the branch
4306 or switch that it is attached to is completely unpredictable.
4311 It is sometimes useful to attach information to loop constructs. Currently,
4312 loop metadata is implemented as metadata attached to the branch instruction
4313 in the loop latch block. This type of metadata refer to a metadata node that is
4314 guaranteed to be separate for each loop. The loop identifier metadata is
4315 specified with the name ``llvm.loop``.
4317 The loop identifier metadata is implemented using a metadata that refers to
4318 itself to avoid merging it with any other identifier metadata, e.g.,
4319 during module linkage or function inlining. That is, each loop should refer
4320 to their own identification metadata even if they reside in separate functions.
4321 The following example contains loop identifier metadata for two separate loop
4324 .. code-block:: llvm
4329 The loop identifier metadata can be used to specify additional
4330 per-loop metadata. Any operands after the first operand can be treated
4331 as user-defined metadata. For example the ``llvm.loop.unroll.count``
4332 suggests an unroll factor to the loop unroller:
4334 .. code-block:: llvm
4336 br i1 %exitcond, label %._crit_edge, label %.lr.ph, !llvm.loop !0
4339 !1 = !{!"llvm.loop.unroll.count", i32 4}
4341 '``llvm.loop.vectorize``' and '``llvm.loop.interleave``'
4342 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4344 Metadata prefixed with ``llvm.loop.vectorize`` or ``llvm.loop.interleave`` are
4345 used to control per-loop vectorization and interleaving parameters such as
4346 vectorization width and interleave count. These metadata should be used in
4347 conjunction with ``llvm.loop`` loop identification metadata. The
4348 ``llvm.loop.vectorize`` and ``llvm.loop.interleave`` metadata are only
4349 optimization hints and the optimizer will only interleave and vectorize loops if
4350 it believes it is safe to do so. The ``llvm.mem.parallel_loop_access`` metadata
4351 which contains information about loop-carried memory dependencies can be helpful
4352 in determining the safety of these transformations.
4354 '``llvm.loop.interleave.count``' Metadata
4355 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4357 This metadata suggests an interleave count to the loop interleaver.
4358 The first operand is the string ``llvm.loop.interleave.count`` and the
4359 second operand is an integer specifying the interleave count. For
4362 .. code-block:: llvm
4364 !0 = !{!"llvm.loop.interleave.count", i32 4}
4366 Note that setting ``llvm.loop.interleave.count`` to 1 disables interleaving
4367 multiple iterations of the loop. If ``llvm.loop.interleave.count`` is set to 0
4368 then the interleave count will be determined automatically.
4370 '``llvm.loop.vectorize.enable``' Metadata
4371 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4373 This metadata selectively enables or disables vectorization for the loop. The
4374 first operand is the string ``llvm.loop.vectorize.enable`` and the second operand
4375 is a bit. If the bit operand value is 1 vectorization is enabled. A value of
4376 0 disables vectorization:
4378 .. code-block:: llvm
4380 !0 = !{!"llvm.loop.vectorize.enable", i1 0}
4381 !1 = !{!"llvm.loop.vectorize.enable", i1 1}
4383 '``llvm.loop.vectorize.width``' Metadata
4384 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4386 This metadata sets the target width of the vectorizer. The first
4387 operand is the string ``llvm.loop.vectorize.width`` and the second
4388 operand is an integer specifying the width. For example:
4390 .. code-block:: llvm
4392 !0 = !{!"llvm.loop.vectorize.width", i32 4}
4394 Note that setting ``llvm.loop.vectorize.width`` to 1 disables
4395 vectorization of the loop. If ``llvm.loop.vectorize.width`` is set to
4396 0 or if the loop does not have this metadata the width will be
4397 determined automatically.
4399 '``llvm.loop.unroll``'
4400 ^^^^^^^^^^^^^^^^^^^^^^
4402 Metadata prefixed with ``llvm.loop.unroll`` are loop unrolling
4403 optimization hints such as the unroll factor. ``llvm.loop.unroll``
4404 metadata should be used in conjunction with ``llvm.loop`` loop
4405 identification metadata. The ``llvm.loop.unroll`` metadata are only
4406 optimization hints and the unrolling will only be performed if the
4407 optimizer believes it is safe to do so.
4409 '``llvm.loop.unroll.count``' Metadata
4410 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4412 This metadata suggests an unroll factor to the loop unroller. The
4413 first operand is the string ``llvm.loop.unroll.count`` and the second
4414 operand is a positive integer specifying the unroll factor. For
4417 .. code-block:: llvm
4419 !0 = !{!"llvm.loop.unroll.count", i32 4}
4421 If the trip count of the loop is less than the unroll count the loop
4422 will be partially unrolled.
4424 '``llvm.loop.unroll.disable``' Metadata
4425 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4427 This metadata disables loop unrolling. The metadata has a single operand
4428 which is the string ``llvm.loop.unroll.disable``. For example:
4430 .. code-block:: llvm
4432 !0 = !{!"llvm.loop.unroll.disable"}
4434 '``llvm.loop.unroll.runtime.disable``' Metadata
4435 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4437 This metadata disables runtime loop unrolling. The metadata has a single
4438 operand which is the string ``llvm.loop.unroll.runtime.disable``. For example:
4440 .. code-block:: llvm
4442 !0 = !{!"llvm.loop.unroll.runtime.disable"}
4444 '``llvm.loop.unroll.enable``' Metadata
4445 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4447 This metadata suggests that the loop should be fully unrolled if the trip count
4448 is known at compile time and partially unrolled if the trip count is not known
4449 at compile time. The metadata has a single operand which is the string
4450 ``llvm.loop.unroll.enable``. For example:
4452 .. code-block:: llvm
4454 !0 = !{!"llvm.loop.unroll.enable"}
4456 '``llvm.loop.unroll.full``' Metadata
4457 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4459 This metadata suggests that the loop should be unrolled fully. The
4460 metadata has a single operand which is the string ``llvm.loop.unroll.full``.
4463 .. code-block:: llvm
4465 !0 = !{!"llvm.loop.unroll.full"}
4470 Metadata types used to annotate memory accesses with information helpful
4471 for optimizations are prefixed with ``llvm.mem``.
4473 '``llvm.mem.parallel_loop_access``' Metadata
4474 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4476 The ``llvm.mem.parallel_loop_access`` metadata refers to a loop identifier,
4477 or metadata containing a list of loop identifiers for nested loops.
4478 The metadata is attached to memory accessing instructions and denotes that
4479 no loop carried memory dependence exist between it and other instructions denoted
4480 with the same loop identifier.
4482 Precisely, given two instructions ``m1`` and ``m2`` that both have the
4483 ``llvm.mem.parallel_loop_access`` metadata, with ``L1`` and ``L2`` being the
4484 set of loops associated with that metadata, respectively, then there is no loop
4485 carried dependence between ``m1`` and ``m2`` for loops in both ``L1`` and
4488 As a special case, if all memory accessing instructions in a loop have
4489 ``llvm.mem.parallel_loop_access`` metadata that refers to that loop, then the
4490 loop has no loop carried memory dependences and is considered to be a parallel
4493 Note that if not all memory access instructions have such metadata referring to
4494 the loop, then the loop is considered not being trivially parallel. Additional
4495 memory dependence analysis is required to make that determination. As a fail
4496 safe mechanism, this causes loops that were originally parallel to be considered
4497 sequential (if optimization passes that are unaware of the parallel semantics
4498 insert new memory instructions into the loop body).
4500 Example of a loop that is considered parallel due to its correct use of
4501 both ``llvm.loop`` and ``llvm.mem.parallel_loop_access``
4502 metadata types that refer to the same loop identifier metadata.
4504 .. code-block:: llvm
4508 %val0 = load i32, i32* %arrayidx, !llvm.mem.parallel_loop_access !0
4510 store i32 %val0, i32* %arrayidx1, !llvm.mem.parallel_loop_access !0
4512 br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !0
4518 It is also possible to have nested parallel loops. In that case the
4519 memory accesses refer to a list of loop identifier metadata nodes instead of
4520 the loop identifier metadata node directly:
4522 .. code-block:: llvm
4526 %val1 = load i32, i32* %arrayidx3, !llvm.mem.parallel_loop_access !2
4528 br label %inner.for.body
4532 %val0 = load i32, i32* %arrayidx1, !llvm.mem.parallel_loop_access !0
4534 store i32 %val0, i32* %arrayidx2, !llvm.mem.parallel_loop_access !0
4536 br i1 %exitcond, label %inner.for.end, label %inner.for.body, !llvm.loop !1
4540 store i32 %val1, i32* %arrayidx4, !llvm.mem.parallel_loop_access !2
4542 br i1 %exitcond, label %outer.for.end, label %outer.for.body, !llvm.loop !2
4544 outer.for.end: ; preds = %for.body
4546 !0 = !{!1, !2} ; a list of loop identifiers
4547 !1 = !{!1} ; an identifier for the inner loop
4548 !2 = !{!2} ; an identifier for the outer loop
4553 The ``llvm.bitsets`` global metadata is used to implement
4554 :doc:`bitsets <BitSets>`.
4556 '``invariant.group``' Metadata
4557 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4559 The ``invariant.group`` metadata may be attached to ``load``/``store`` instructions.
4560 The existence of the ``invariant.group`` metadata on the instruction tells
4561 the optimizer that every ``load`` and ``store`` to the same pointer operand
4562 within the same invariant group can be assumed to load or store the same
4563 value (but see the ``llvm.invariant.group.barrier`` intrinsic which affects
4564 when two pointers are considered the same).
4568 .. code-block:: llvm
4570 @unknownPtr = external global i8
4573 store i8 42, i8* %ptr, !invariant.group !0
4574 call void @foo(i8* %ptr)
4576 %a = load i8, i8* %ptr, !invariant.group !0 ; Can assume that value under %ptr didn't change
4577 call void @foo(i8* %ptr)
4578 %b = load i8, i8* %ptr, !invariant.group !1 ; Can't assume anything, because group changed
4580 %newPtr = call i8* @getPointer(i8* %ptr)
4581 %c = load i8, i8* %newPtr, !invariant.group !0 ; Can't assume anything, because we only have information about %ptr
4583 %unknownValue = load i8, i8* @unknownPtr
4584 store i8 %unknownValue, i8* %ptr, !invariant.group !0 ; Can assume that %unknownValue == 42
4586 call void @foo(i8* %ptr)
4587 %newPtr2 = call i8* @llvm.invariant.group.barrier(i8* %ptr)
4588 %d = load i8, i8* %newPtr2, !invariant.group !0 ; Can't step through invariant.group.barrier to get value of %ptr
4591 declare void @foo(i8*)
4592 declare i8* @getPointer(i8*)
4593 declare i8* @llvm.invariant.group.barrier(i8*)
4595 !0 = !{!"magic ptr"}
4596 !1 = !{!"other ptr"}
4600 Module Flags Metadata
4601 =====================
4603 Information about the module as a whole is difficult to convey to LLVM's
4604 subsystems. The LLVM IR isn't sufficient to transmit this information.
4605 The ``llvm.module.flags`` named metadata exists in order to facilitate
4606 this. These flags are in the form of key / value pairs --- much like a
4607 dictionary --- making it easy for any subsystem who cares about a flag to
4610 The ``llvm.module.flags`` metadata contains a list of metadata triplets.
4611 Each triplet has the following form:
4613 - The first element is a *behavior* flag, which specifies the behavior
4614 when two (or more) modules are merged together, and it encounters two
4615 (or more) metadata with the same ID. The supported behaviors are
4617 - The second element is a metadata string that is a unique ID for the
4618 metadata. Each module may only have one flag entry for each unique ID (not
4619 including entries with the **Require** behavior).
4620 - The third element is the value of the flag.
4622 When two (or more) modules are merged together, the resulting
4623 ``llvm.module.flags`` metadata is the union of the modules' flags. That is, for
4624 each unique metadata ID string, there will be exactly one entry in the merged
4625 modules ``llvm.module.flags`` metadata table, and the value for that entry will
4626 be determined by the merge behavior flag, as described below. The only exception
4627 is that entries with the *Require* behavior are always preserved.
4629 The following behaviors are supported:
4640 Emits an error if two values disagree, otherwise the resulting value
4641 is that of the operands.
4645 Emits a warning if two values disagree. The result value will be the
4646 operand for the flag from the first module being linked.
4650 Adds a requirement that another module flag be present and have a
4651 specified value after linking is performed. The value must be a
4652 metadata pair, where the first element of the pair is the ID of the
4653 module flag to be restricted, and the second element of the pair is
4654 the value the module flag should be restricted to. This behavior can
4655 be used to restrict the allowable results (via triggering of an
4656 error) of linking IDs with the **Override** behavior.
4660 Uses the specified value, regardless of the behavior or value of the
4661 other module. If both modules specify **Override**, but the values
4662 differ, an error will be emitted.
4666 Appends the two values, which are required to be metadata nodes.
4670 Appends the two values, which are required to be metadata
4671 nodes. However, duplicate entries in the second list are dropped
4672 during the append operation.
4674 It is an error for a particular unique flag ID to have multiple behaviors,
4675 except in the case of **Require** (which adds restrictions on another metadata
4676 value) or **Override**.
4678 An example of module flags:
4680 .. code-block:: llvm
4682 !0 = !{ i32 1, !"foo", i32 1 }
4683 !1 = !{ i32 4, !"bar", i32 37 }
4684 !2 = !{ i32 2, !"qux", i32 42 }
4685 !3 = !{ i32 3, !"qux",
4690 !llvm.module.flags = !{ !0, !1, !2, !3 }
4692 - Metadata ``!0`` has the ID ``!"foo"`` and the value '1'. The behavior
4693 if two or more ``!"foo"`` flags are seen is to emit an error if their
4694 values are not equal.
4696 - Metadata ``!1`` has the ID ``!"bar"`` and the value '37'. The
4697 behavior if two or more ``!"bar"`` flags are seen is to use the value
4700 - Metadata ``!2`` has the ID ``!"qux"`` and the value '42'. The
4701 behavior if two or more ``!"qux"`` flags are seen is to emit a
4702 warning if their values are not equal.
4704 - Metadata ``!3`` has the ID ``!"qux"`` and the value:
4710 The behavior is to emit an error if the ``llvm.module.flags`` does not
4711 contain a flag with the ID ``!"foo"`` that has the value '1' after linking is
4714 Objective-C Garbage Collection Module Flags Metadata
4715 ----------------------------------------------------
4717 On the Mach-O platform, Objective-C stores metadata about garbage
4718 collection in a special section called "image info". The metadata
4719 consists of a version number and a bitmask specifying what types of
4720 garbage collection are supported (if any) by the file. If two or more
4721 modules are linked together their garbage collection metadata needs to
4722 be merged rather than appended together.
4724 The Objective-C garbage collection module flags metadata consists of the
4725 following key-value pairs:
4734 * - ``Objective-C Version``
4735 - **[Required]** --- The Objective-C ABI version. Valid values are 1 and 2.
4737 * - ``Objective-C Image Info Version``
4738 - **[Required]** --- The version of the image info section. Currently
4741 * - ``Objective-C Image Info Section``
4742 - **[Required]** --- The section to place the metadata. Valid values are
4743 ``"__OBJC, __image_info, regular"`` for Objective-C ABI version 1, and
4744 ``"__DATA,__objc_imageinfo, regular, no_dead_strip"`` for
4745 Objective-C ABI version 2.
4747 * - ``Objective-C Garbage Collection``
4748 - **[Required]** --- Specifies whether garbage collection is supported or
4749 not. Valid values are 0, for no garbage collection, and 2, for garbage
4750 collection supported.
4752 * - ``Objective-C GC Only``
4753 - **[Optional]** --- Specifies that only garbage collection is supported.
4754 If present, its value must be 6. This flag requires that the
4755 ``Objective-C Garbage Collection`` flag have the value 2.
4757 Some important flag interactions:
4759 - If a module with ``Objective-C Garbage Collection`` set to 0 is
4760 merged with a module with ``Objective-C Garbage Collection`` set to
4761 2, then the resulting module has the
4762 ``Objective-C Garbage Collection`` flag set to 0.
4763 - A module with ``Objective-C Garbage Collection`` set to 0 cannot be
4764 merged with a module with ``Objective-C GC Only`` set to 6.
4766 Automatic Linker Flags Module Flags Metadata
4767 --------------------------------------------
4769 Some targets support embedding flags to the linker inside individual object
4770 files. Typically this is used in conjunction with language extensions which
4771 allow source files to explicitly declare the libraries they depend on, and have
4772 these automatically be transmitted to the linker via object files.
4774 These flags are encoded in the IR using metadata in the module flags section,
4775 using the ``Linker Options`` key. The merge behavior for this flag is required
4776 to be ``AppendUnique``, and the value for the key is expected to be a metadata
4777 node which should be a list of other metadata nodes, each of which should be a
4778 list of metadata strings defining linker options.
4780 For example, the following metadata section specifies two separate sets of
4781 linker options, presumably to link against ``libz`` and the ``Cocoa``
4784 !0 = !{ i32 6, !"Linker Options",
4787 !{ !"-framework", !"Cocoa" } } }
4788 !llvm.module.flags = !{ !0 }
4790 The metadata encoding as lists of lists of options, as opposed to a collapsed
4791 list of options, is chosen so that the IR encoding can use multiple option
4792 strings to specify e.g., a single library, while still having that specifier be
4793 preserved as an atomic element that can be recognized by a target specific
4794 assembly writer or object file emitter.
4796 Each individual option is required to be either a valid option for the target's
4797 linker, or an option that is reserved by the target specific assembly writer or
4798 object file emitter. No other aspect of these options is defined by the IR.
4800 C type width Module Flags Metadata
4801 ----------------------------------
4803 The ARM backend emits a section into each generated object file describing the
4804 options that it was compiled with (in a compiler-independent way) to prevent
4805 linking incompatible objects, and to allow automatic library selection. Some
4806 of these options are not visible at the IR level, namely wchar_t width and enum
4809 To pass this information to the backend, these options are encoded in module
4810 flags metadata, using the following key-value pairs:
4820 - * 0 --- sizeof(wchar_t) == 4
4821 * 1 --- sizeof(wchar_t) == 2
4824 - * 0 --- Enums are at least as large as an ``int``.
4825 * 1 --- Enums are stored in the smallest integer type which can
4826 represent all of its values.
4828 For example, the following metadata section specifies that the module was
4829 compiled with a ``wchar_t`` width of 4 bytes, and the underlying type of an
4830 enum is the smallest type which can represent all of its values::
4832 !llvm.module.flags = !{!0, !1}
4833 !0 = !{i32 1, !"short_wchar", i32 1}
4834 !1 = !{i32 1, !"short_enum", i32 0}
4836 .. _intrinsicglobalvariables:
4838 Intrinsic Global Variables
4839 ==========================
4841 LLVM has a number of "magic" global variables that contain data that
4842 affect code generation or other IR semantics. These are documented here.
4843 All globals of this sort should have a section specified as
4844 "``llvm.metadata``". This section and all globals that start with
4845 "``llvm.``" are reserved for use by LLVM.
4849 The '``llvm.used``' Global Variable
4850 -----------------------------------
4852 The ``@llvm.used`` global is an array which has
4853 :ref:`appending linkage <linkage_appending>`. This array contains a list of
4854 pointers to named global variables, functions and aliases which may optionally
4855 have a pointer cast formed of bitcast or getelementptr. For example, a legal
4858 .. code-block:: llvm
4863 @llvm.used = appending global [2 x i8*] [
4865 i8* bitcast (i32* @Y to i8*)
4866 ], section "llvm.metadata"
4868 If a symbol appears in the ``@llvm.used`` list, then the compiler, assembler,
4869 and linker are required to treat the symbol as if there is a reference to the
4870 symbol that it cannot see (which is why they have to be named). For example, if
4871 a variable has internal linkage and no references other than that from the
4872 ``@llvm.used`` list, it cannot be deleted. This is commonly used to represent
4873 references from inline asms and other things the compiler cannot "see", and
4874 corresponds to "``attribute((used))``" in GNU C.
4876 On some targets, the code generator must emit a directive to the
4877 assembler or object file to prevent the assembler and linker from
4878 molesting the symbol.
4880 .. _gv_llvmcompilerused:
4882 The '``llvm.compiler.used``' Global Variable
4883 --------------------------------------------
4885 The ``@llvm.compiler.used`` directive is the same as the ``@llvm.used``
4886 directive, except that it only prevents the compiler from touching the
4887 symbol. On targets that support it, this allows an intelligent linker to
4888 optimize references to the symbol without being impeded as it would be
4891 This is a rare construct that should only be used in rare circumstances,
4892 and should not be exposed to source languages.
4894 .. _gv_llvmglobalctors:
4896 The '``llvm.global_ctors``' Global Variable
4897 -------------------------------------------
4899 .. code-block:: llvm
4901 %0 = type { i32, void ()*, i8* }
4902 @llvm.global_ctors = appending global [1 x %0] [%0 { i32 65535, void ()* @ctor, i8* @data }]
4904 The ``@llvm.global_ctors`` array contains a list of constructor
4905 functions, priorities, and an optional associated global or function.
4906 The functions referenced by this array will be called in ascending order
4907 of priority (i.e. lowest first) when the module is loaded. The order of
4908 functions with the same priority is not defined.
4910 If the third field is present, non-null, and points to a global variable
4911 or function, the initializer function will only run if the associated
4912 data from the current module is not discarded.
4914 .. _llvmglobaldtors:
4916 The '``llvm.global_dtors``' Global Variable
4917 -------------------------------------------
4919 .. code-block:: llvm
4921 %0 = type { i32, void ()*, i8* }
4922 @llvm.global_dtors = appending global [1 x %0] [%0 { i32 65535, void ()* @dtor, i8* @data }]
4924 The ``@llvm.global_dtors`` array contains a list of destructor
4925 functions, priorities, and an optional associated global or function.
4926 The functions referenced by this array will be called in descending
4927 order of priority (i.e. highest first) when the module is unloaded. The
4928 order of functions with the same priority is not defined.
4930 If the third field is present, non-null, and points to a global variable
4931 or function, the destructor function will only run if the associated
4932 data from the current module is not discarded.
4934 Instruction Reference
4935 =====================
4937 The LLVM instruction set consists of several different classifications
4938 of instructions: :ref:`terminator instructions <terminators>`, :ref:`binary
4939 instructions <binaryops>`, :ref:`bitwise binary
4940 instructions <bitwiseops>`, :ref:`memory instructions <memoryops>`, and
4941 :ref:`other instructions <otherops>`.
4945 Terminator Instructions
4946 -----------------------
4948 As mentioned :ref:`previously <functionstructure>`, every basic block in a
4949 program ends with a "Terminator" instruction, which indicates which
4950 block should be executed after the current block is finished. These
4951 terminator instructions typically yield a '``void``' value: they produce
4952 control flow, not values (the one exception being the
4953 ':ref:`invoke <i_invoke>`' instruction).
4955 The terminator instructions are: ':ref:`ret <i_ret>`',
4956 ':ref:`br <i_br>`', ':ref:`switch <i_switch>`',
4957 ':ref:`indirectbr <i_indirectbr>`', ':ref:`invoke <i_invoke>`',
4958 ':ref:`resume <i_resume>`', ':ref:`catchpad <i_catchpad>`',
4959 ':ref:`catchendpad <i_catchendpad>`',
4960 ':ref:`catchret <i_catchret>`',
4961 ':ref:`cleanupendpad <i_cleanupendpad>`',
4962 ':ref:`cleanupret <i_cleanupret>`',
4963 ':ref:`terminatepad <i_terminatepad>`',
4964 and ':ref:`unreachable <i_unreachable>`'.
4968 '``ret``' Instruction
4969 ^^^^^^^^^^^^^^^^^^^^^
4976 ret <type> <value> ; Return a value from a non-void function
4977 ret void ; Return from void function
4982 The '``ret``' instruction is used to return control flow (and optionally
4983 a value) from a function back to the caller.
4985 There are two forms of the '``ret``' instruction: one that returns a
4986 value and then causes control flow, and one that just causes control
4992 The '``ret``' instruction optionally accepts a single argument, the
4993 return value. The type of the return value must be a ':ref:`first
4994 class <t_firstclass>`' type.
4996 A function is not :ref:`well formed <wellformed>` if it it has a non-void
4997 return type and contains a '``ret``' instruction with no return value or
4998 a return value with a type that does not match its type, or if it has a
4999 void return type and contains a '``ret``' instruction with a return
5005 When the '``ret``' instruction is executed, control flow returns back to
5006 the calling function's context. If the caller is a
5007 ":ref:`call <i_call>`" instruction, execution continues at the
5008 instruction after the call. If the caller was an
5009 ":ref:`invoke <i_invoke>`" instruction, execution continues at the
5010 beginning of the "normal" destination block. If the instruction returns
5011 a value, that value shall set the call or invoke instruction's return
5017 .. code-block:: llvm
5019 ret i32 5 ; Return an integer value of 5
5020 ret void ; Return from a void function
5021 ret { i32, i8 } { i32 4, i8 2 } ; Return a struct of values 4 and 2
5025 '``br``' Instruction
5026 ^^^^^^^^^^^^^^^^^^^^
5033 br i1 <cond>, label <iftrue>, label <iffalse>
5034 br label <dest> ; Unconditional branch
5039 The '``br``' instruction is used to cause control flow to transfer to a
5040 different basic block in the current function. There are two forms of
5041 this instruction, corresponding to a conditional branch and an
5042 unconditional branch.
5047 The conditional branch form of the '``br``' instruction takes a single
5048 '``i1``' value and two '``label``' values. The unconditional form of the
5049 '``br``' instruction takes a single '``label``' value as a target.
5054 Upon execution of a conditional '``br``' instruction, the '``i1``'
5055 argument is evaluated. If the value is ``true``, control flows to the
5056 '``iftrue``' ``label`` argument. If "cond" is ``false``, control flows
5057 to the '``iffalse``' ``label`` argument.
5062 .. code-block:: llvm
5065 %cond = icmp eq i32 %a, %b
5066 br i1 %cond, label %IfEqual, label %IfUnequal
5074 '``switch``' Instruction
5075 ^^^^^^^^^^^^^^^^^^^^^^^^
5082 switch <intty> <value>, label <defaultdest> [ <intty> <val>, label <dest> ... ]
5087 The '``switch``' instruction is used to transfer control flow to one of
5088 several different places. It is a generalization of the '``br``'
5089 instruction, allowing a branch to occur to one of many possible
5095 The '``switch``' instruction uses three parameters: an integer
5096 comparison value '``value``', a default '``label``' destination, and an
5097 array of pairs of comparison value constants and '``label``'s. The table
5098 is not allowed to contain duplicate constant entries.
5103 The ``switch`` instruction specifies a table of values and destinations.
5104 When the '``switch``' instruction is executed, this table is searched
5105 for the given value. If the value is found, control flow is transferred
5106 to the corresponding destination; otherwise, control flow is transferred
5107 to the default destination.
5112 Depending on properties of the target machine and the particular
5113 ``switch`` instruction, this instruction may be code generated in
5114 different ways. For example, it could be generated as a series of
5115 chained conditional branches or with a lookup table.
5120 .. code-block:: llvm
5122 ; Emulate a conditional br instruction
5123 %Val = zext i1 %value to i32
5124 switch i32 %Val, label %truedest [ i32 0, label %falsedest ]
5126 ; Emulate an unconditional br instruction
5127 switch i32 0, label %dest [ ]
5129 ; Implement a jump table:
5130 switch i32 %val, label %otherwise [ i32 0, label %onzero
5132 i32 2, label %ontwo ]
5136 '``indirectbr``' Instruction
5137 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5144 indirectbr <somety>* <address>, [ label <dest1>, label <dest2>, ... ]
5149 The '``indirectbr``' instruction implements an indirect branch to a
5150 label within the current function, whose address is specified by
5151 "``address``". Address must be derived from a
5152 :ref:`blockaddress <blockaddress>` constant.
5157 The '``address``' argument is the address of the label to jump to. The
5158 rest of the arguments indicate the full set of possible destinations
5159 that the address may point to. Blocks are allowed to occur multiple
5160 times in the destination list, though this isn't particularly useful.
5162 This destination list is required so that dataflow analysis has an
5163 accurate understanding of the CFG.
5168 Control transfers to the block specified in the address argument. All
5169 possible destination blocks must be listed in the label list, otherwise
5170 this instruction has undefined behavior. This implies that jumps to
5171 labels defined in other functions have undefined behavior as well.
5176 This is typically implemented with a jump through a register.
5181 .. code-block:: llvm
5183 indirectbr i8* %Addr, [ label %bb1, label %bb2, label %bb3 ]
5187 '``invoke``' Instruction
5188 ^^^^^^^^^^^^^^^^^^^^^^^^
5195 <result> = invoke [cconv] [ret attrs] <ptr to function ty> <function ptr val>(<function args>) [fn attrs]
5196 [operand bundles] to label <normal label> unwind label <exception label>
5201 The '``invoke``' instruction causes control to transfer to a specified
5202 function, with the possibility of control flow transfer to either the
5203 '``normal``' label or the '``exception``' label. If the callee function
5204 returns with the "``ret``" instruction, control flow will return to the
5205 "normal" label. If the callee (or any indirect callees) returns via the
5206 ":ref:`resume <i_resume>`" instruction or other exception handling
5207 mechanism, control is interrupted and continued at the dynamically
5208 nearest "exception" label.
5210 The '``exception``' label is a `landing
5211 pad <ExceptionHandling.html#overview>`_ for the exception. As such,
5212 '``exception``' label is required to have the
5213 ":ref:`landingpad <i_landingpad>`" instruction, which contains the
5214 information about the behavior of the program after unwinding happens,
5215 as its first non-PHI instruction. The restrictions on the
5216 "``landingpad``" instruction's tightly couples it to the "``invoke``"
5217 instruction, so that the important information contained within the
5218 "``landingpad``" instruction can't be lost through normal code motion.
5223 This instruction requires several arguments:
5225 #. The optional "cconv" marker indicates which :ref:`calling
5226 convention <callingconv>` the call should use. If none is
5227 specified, the call defaults to using C calling conventions.
5228 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
5229 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
5231 #. '``ptr to function ty``': shall be the signature of the pointer to
5232 function value being invoked. In most cases, this is a direct
5233 function invocation, but indirect ``invoke``'s are just as possible,
5234 branching off an arbitrary pointer to function value.
5235 #. '``function ptr val``': An LLVM value containing a pointer to a
5236 function to be invoked.
5237 #. '``function args``': argument list whose types match the function
5238 signature argument types and parameter attributes. All arguments must
5239 be of :ref:`first class <t_firstclass>` type. If the function signature
5240 indicates the function accepts a variable number of arguments, the
5241 extra arguments can be specified.
5242 #. '``normal label``': the label reached when the called function
5243 executes a '``ret``' instruction.
5244 #. '``exception label``': the label reached when a callee returns via
5245 the :ref:`resume <i_resume>` instruction or other exception handling
5247 #. The optional :ref:`function attributes <fnattrs>` list. Only
5248 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
5249 attributes are valid here.
5250 #. The optional :ref:`operand bundles <opbundles>` list.
5255 This instruction is designed to operate as a standard '``call``'
5256 instruction in most regards. The primary difference is that it
5257 establishes an association with a label, which is used by the runtime
5258 library to unwind the stack.
5260 This instruction is used in languages with destructors to ensure that
5261 proper cleanup is performed in the case of either a ``longjmp`` or a
5262 thrown exception. Additionally, this is important for implementation of
5263 '``catch``' clauses in high-level languages that support them.
5265 For the purposes of the SSA form, the definition of the value returned
5266 by the '``invoke``' instruction is deemed to occur on the edge from the
5267 current block to the "normal" label. If the callee unwinds then no
5268 return value is available.
5273 .. code-block:: llvm
5275 %retval = invoke i32 @Test(i32 15) to label %Continue
5276 unwind label %TestCleanup ; i32:retval set
5277 %retval = invoke coldcc i32 %Testfnptr(i32 15) to label %Continue
5278 unwind label %TestCleanup ; i32:retval set
5282 '``resume``' Instruction
5283 ^^^^^^^^^^^^^^^^^^^^^^^^
5290 resume <type> <value>
5295 The '``resume``' instruction is a terminator instruction that has no
5301 The '``resume``' instruction requires one argument, which must have the
5302 same type as the result of any '``landingpad``' instruction in the same
5308 The '``resume``' instruction resumes propagation of an existing
5309 (in-flight) exception whose unwinding was interrupted with a
5310 :ref:`landingpad <i_landingpad>` instruction.
5315 .. code-block:: llvm
5317 resume { i8*, i32 } %exn
5321 '``catchpad``' Instruction
5322 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5329 <resultval> = catchpad [<args>*]
5330 to label <normal label> unwind label <exception label>
5335 The '``catchpad``' instruction is used by `LLVM's exception handling
5336 system <ExceptionHandling.html#overview>`_ to specify that a basic block
5337 is a catch block --- one where a personality routine attempts to transfer
5338 control to catch an exception.
5339 The ``args`` correspond to whatever information the personality
5340 routine requires to know if this is an appropriate place to catch the
5341 exception. Control is transfered to the ``exception`` label if the
5342 ``catchpad`` is not an appropriate handler for the in-flight exception.
5343 The ``normal`` label should contain the code found in the ``catch``
5344 portion of a ``try``/``catch`` sequence. The ``resultval`` has the type
5345 :ref:`token <t_token>` and is used to match the ``catchpad`` to
5346 corresponding :ref:`catchrets <i_catchret>`.
5351 The instruction takes a list of arbitrary values which are interpreted
5352 by the :ref:`personality function <personalityfn>`.
5354 The ``catchpad`` must be provided a ``normal`` label to transfer control
5355 to if the ``catchpad`` matches the exception and an ``exception``
5356 label to transfer control to if it doesn't.
5361 When the call stack is being unwound due to an exception being thrown,
5362 the exception is compared against the ``args``. If it doesn't match,
5363 then control is transfered to the ``exception`` basic block.
5364 As with calling conventions, how the personality function results are
5365 represented in LLVM IR is target specific.
5367 The ``catchpad`` instruction has several restrictions:
5369 - A catch block is a basic block which is the unwind destination of
5370 an exceptional instruction.
5371 - A catch block must have a '``catchpad``' instruction as its
5372 first non-PHI instruction.
5373 - A catch block's ``exception`` edge must refer to a catch block or a
5375 - There can be only one '``catchpad``' instruction within the
5377 - A basic block that is not a catch block may not include a
5378 '``catchpad``' instruction.
5379 - A catch block which has another catch block as a predecessor may not have
5380 any other predecessors.
5381 - It is undefined behavior for control to transfer from a ``catchpad`` to a
5382 ``ret`` without first executing a ``catchret`` that consumes the
5383 ``catchpad`` or unwinding through its ``catchendpad``.
5384 - It is undefined behavior for control to transfer from a ``catchpad`` to
5385 itself without first executing a ``catchret`` that consumes the
5386 ``catchpad`` or unwinding through its ``catchendpad``.
5391 .. code-block:: llvm
5393 ;; A catch block which can catch an integer.
5394 %tok = catchpad [i8** @_ZTIi]
5395 to label %int.handler unwind label %terminate
5399 '``catchendpad``' Instruction
5400 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5407 catchendpad unwind label <nextaction>
5408 catchendpad unwind to caller
5413 The '``catchendpad``' instruction is used by `LLVM's exception handling
5414 system <ExceptionHandling.html#overview>`_ to communicate to the
5415 :ref:`personality function <personalityfn>` which invokes are associated
5416 with a chain of :ref:`catchpad <i_catchpad>` instructions; propagating an
5417 exception out of a catch handler is represented by unwinding through its
5418 ``catchendpad``. Unwinding to the outer scope when a chain of catch handlers
5419 do not handle an exception is also represented by unwinding through their
5422 The ``nextaction`` label indicates where control should transfer to if
5423 none of the ``catchpad`` instructions are suitable for catching the
5424 in-flight exception.
5426 If a ``nextaction`` label is not present, the instruction unwinds out of
5427 its parent function. The
5428 :ref:`personality function <personalityfn>` will continue processing
5429 exception handling actions in the caller.
5434 The instruction optionally takes a label, ``nextaction``, indicating
5435 where control should transfer to if none of the preceding
5436 ``catchpad`` instructions are suitable for the in-flight exception.
5441 When the call stack is being unwound due to an exception being thrown
5442 and none of the constituent ``catchpad`` instructions match, then
5443 control is transfered to ``nextaction`` if it is present. If it is not
5444 present, control is transfered to the caller.
5446 The ``catchendpad`` instruction has several restrictions:
5448 - A catch-end block is a basic block which is the unwind destination of
5449 an exceptional instruction.
5450 - A catch-end block must have a '``catchendpad``' instruction as its
5451 first non-PHI instruction.
5452 - There can be only one '``catchendpad``' instruction within the
5454 - A basic block that is not a catch-end block may not include a
5455 '``catchendpad``' instruction.
5456 - Exactly one catch block may unwind to a ``catchendpad``.
5457 - It is undefined behavior to execute a ``catchendpad`` if none of the
5458 '``catchpad``'s chained to it have been executed.
5459 - It is undefined behavior to execute a ``catchendpad`` twice without an
5460 intervening execution of one or more of the '``catchpad``'s chained to it.
5461 - It is undefined behavior to execute a ``catchendpad`` if, after the most
5462 recent execution of the normal successor edge of any ``catchpad`` chained
5463 to it, some ``catchret`` consuming that ``catchpad`` has already been
5465 - It is undefined behavior to execute a ``catchendpad`` if, after the most
5466 recent execution of the normal successor edge of any ``catchpad`` chained
5467 to it, any other ``catchpad`` or ``cleanuppad`` has been executed but has
5468 not had a corresponding
5469 ``catchret``/``cleanupret``/``catchendpad``/``cleanupendpad`` executed.
5474 .. code-block:: llvm
5476 catchendpad unwind label %terminate
5477 catchendpad unwind to caller
5481 '``catchret``' Instruction
5482 ^^^^^^^^^^^^^^^^^^^^^^^^^^
5489 catchret <value> to label <normal>
5494 The '``catchret``' instruction is a terminator instruction that has a
5501 The first argument to a '``catchret``' indicates which ``catchpad`` it
5502 exits. It must be a :ref:`catchpad <i_catchpad>`.
5503 The second argument to a '``catchret``' specifies where control will
5509 The '``catchret``' instruction ends the existing (in-flight) exception
5510 whose unwinding was interrupted with a
5511 :ref:`catchpad <i_catchpad>` instruction.
5512 The :ref:`personality function <personalityfn>` gets a chance to execute
5513 arbitrary code to, for example, run a C++ destructor.
5514 Control then transfers to ``normal``.
5515 It may be passed an optional, personality specific, value.
5517 It is undefined behavior to execute a ``catchret`` whose ``catchpad`` has
5520 It is undefined behavior to execute a ``catchret`` if, after the most recent
5521 execution of its ``catchpad``, some ``catchret`` or ``catchendpad`` linked
5522 to the same ``catchpad`` has already been executed.
5524 It is undefined behavior to execute a ``catchret`` if, after the most recent
5525 execution of its ``catchpad``, any other ``catchpad`` or ``cleanuppad`` has
5526 been executed but has not had a corresponding
5527 ``catchret``/``cleanupret``/``catchendpad``/``cleanupendpad`` executed.
5532 .. code-block:: llvm
5534 catchret %catch label %continue
5536 .. _i_cleanupendpad:
5538 '``cleanupendpad``' Instruction
5539 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5546 cleanupendpad <value> unwind label <nextaction>
5547 cleanupendpad <value> unwind to caller
5552 The '``cleanupendpad``' instruction is used by `LLVM's exception handling
5553 system <ExceptionHandling.html#overview>`_ to communicate to the
5554 :ref:`personality function <personalityfn>` which invokes are associated
5555 with a :ref:`cleanuppad <i_cleanuppad>` instructions; propagating an exception
5556 out of a cleanup is represented by unwinding through its ``cleanupendpad``.
5558 The ``nextaction`` label indicates where control should unwind to next, in the
5559 event that a cleanup is exited by means of an(other) exception being raised.
5561 If a ``nextaction`` label is not present, the instruction unwinds out of
5562 its parent function. The
5563 :ref:`personality function <personalityfn>` will continue processing
5564 exception handling actions in the caller.
5569 The '``cleanupendpad``' instruction requires one argument, which indicates
5570 which ``cleanuppad`` it exits, and must be a :ref:`cleanuppad <i_cleanuppad>`.
5571 It also has an optional successor, ``nextaction``, indicating where control
5577 When and exception propagates to a ``cleanupendpad``, control is transfered to
5578 ``nextaction`` if it is present. If it is not present, control is transfered to
5581 The ``cleanupendpad`` instruction has several restrictions:
5583 - A cleanup-end block is a basic block which is the unwind destination of
5584 an exceptional instruction.
5585 - A cleanup-end block must have a '``cleanupendpad``' instruction as its
5586 first non-PHI instruction.
5587 - There can be only one '``cleanupendpad``' instruction within the
5589 - A basic block that is not a cleanup-end block may not include a
5590 '``cleanupendpad``' instruction.
5591 - It is undefined behavior to execute a ``cleanupendpad`` whose ``cleanuppad``
5592 has not been executed.
5593 - It is undefined behavior to execute a ``cleanupendpad`` if, after the most
5594 recent execution of its ``cleanuppad``, some ``cleanupret`` or ``cleanupendpad``
5595 consuming the same ``cleanuppad`` has already been executed.
5596 - It is undefined behavior to execute a ``cleanupendpad`` if, after the most
5597 recent execution of its ``cleanuppad``, any other ``cleanuppad`` or
5598 ``catchpad`` has been executed but has not had a corresponding
5599 ``cleanupret``/``catchret``/``cleanupendpad``/``catchendpad`` executed.
5604 .. code-block:: llvm
5606 cleanupendpad %cleanup unwind label %terminate
5607 cleanupendpad %cleanup unwind to caller
5611 '``cleanupret``' Instruction
5612 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5619 cleanupret <value> unwind label <continue>
5620 cleanupret <value> unwind to caller
5625 The '``cleanupret``' instruction is a terminator instruction that has
5626 an optional successor.
5632 The '``cleanupret``' instruction requires one argument, which indicates
5633 which ``cleanuppad`` it exits, and must be a :ref:`cleanuppad <i_cleanuppad>`.
5634 It also has an optional successor, ``continue``.
5639 The '``cleanupret``' instruction indicates to the
5640 :ref:`personality function <personalityfn>` that one
5641 :ref:`cleanuppad <i_cleanuppad>` it transferred control to has ended.
5642 It transfers control to ``continue`` or unwinds out of the function.
5644 It is undefined behavior to execute a ``cleanupret`` whose ``cleanuppad`` has
5647 It is undefined behavior to execute a ``cleanupret`` if, after the most recent
5648 execution of its ``cleanuppad``, some ``cleanupret`` or ``cleanupendpad``
5649 consuming the same ``cleanuppad`` has already been executed.
5651 It is undefined behavior to execute a ``cleanupret`` if, after the most recent
5652 execution of its ``cleanuppad``, any other ``cleanuppad`` or ``catchpad`` has
5653 been executed but has not had a corresponding
5654 ``cleanupret``/``catchret``/``cleanupendpad``/``catchendpad`` executed.
5659 .. code-block:: llvm
5661 cleanupret %cleanup unwind to caller
5662 cleanupret %cleanup unwind label %continue
5666 '``terminatepad``' Instruction
5667 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5674 terminatepad [<args>*] unwind label <exception label>
5675 terminatepad [<args>*] unwind to caller
5680 The '``terminatepad``' instruction is used by `LLVM's exception handling
5681 system <ExceptionHandling.html#overview>`_ to specify that a basic block
5682 is a terminate block --- one where a personality routine may decide to
5683 terminate the program.
5684 The ``args`` correspond to whatever information the personality
5685 routine requires to know if this is an appropriate place to terminate the
5686 program. Control is transferred to the ``exception`` label if the
5687 personality routine decides not to terminate the program for the
5688 in-flight exception.
5693 The instruction takes a list of arbitrary values which are interpreted
5694 by the :ref:`personality function <personalityfn>`.
5696 The ``terminatepad`` may be given an ``exception`` label to
5697 transfer control to if the in-flight exception matches the ``args``.
5702 When the call stack is being unwound due to an exception being thrown,
5703 the exception is compared against the ``args``. If it matches,
5704 then control is transfered to the ``exception`` basic block. Otherwise,
5705 the program is terminated via personality-specific means. Typically,
5706 the first argument to ``terminatepad`` specifies what function the
5707 personality should defer to in order to terminate the program.
5709 The ``terminatepad`` instruction has several restrictions:
5711 - A terminate block is a basic block which is the unwind destination of
5712 an exceptional instruction.
5713 - A terminate block must have a '``terminatepad``' instruction as its
5714 first non-PHI instruction.
5715 - There can be only one '``terminatepad``' instruction within the
5717 - A basic block that is not a terminate block may not include a
5718 '``terminatepad``' instruction.
5723 .. code-block:: llvm
5725 ;; A terminate block which only permits integers.
5726 terminatepad [i8** @_ZTIi] unwind label %continue
5730 '``unreachable``' Instruction
5731 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5743 The '``unreachable``' instruction has no defined semantics. This
5744 instruction is used to inform the optimizer that a particular portion of
5745 the code is not reachable. This can be used to indicate that the code
5746 after a no-return function cannot be reached, and other facts.
5751 The '``unreachable``' instruction has no defined semantics.
5758 Binary operators are used to do most of the computation in a program.
5759 They require two operands of the same type, execute an operation on
5760 them, and produce a single value. The operands might represent multiple
5761 data, as is the case with the :ref:`vector <t_vector>` data type. The
5762 result value has the same type as its operands.
5764 There are several different binary operators:
5768 '``add``' Instruction
5769 ^^^^^^^^^^^^^^^^^^^^^
5776 <result> = add <ty> <op1>, <op2> ; yields ty:result
5777 <result> = add nuw <ty> <op1>, <op2> ; yields ty:result
5778 <result> = add nsw <ty> <op1>, <op2> ; yields ty:result
5779 <result> = add nuw nsw <ty> <op1>, <op2> ; yields ty:result
5784 The '``add``' instruction returns the sum of its two operands.
5789 The two arguments to the '``add``' instruction must be
5790 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
5791 arguments must have identical types.
5796 The value produced is the integer sum of the two operands.
5798 If the sum has unsigned overflow, the result returned is the
5799 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
5802 Because LLVM integers use a two's complement representation, this
5803 instruction is appropriate for both signed and unsigned integers.
5805 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
5806 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
5807 result value of the ``add`` is a :ref:`poison value <poisonvalues>` if
5808 unsigned and/or signed overflow, respectively, occurs.
5813 .. code-block:: llvm
5815 <result> = add i32 4, %var ; yields i32:result = 4 + %var
5819 '``fadd``' Instruction
5820 ^^^^^^^^^^^^^^^^^^^^^^
5827 <result> = fadd [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
5832 The '``fadd``' instruction returns the sum of its two operands.
5837 The two arguments to the '``fadd``' instruction must be :ref:`floating
5838 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
5839 Both arguments must have identical types.
5844 The value produced is the floating point sum of the two operands. This
5845 instruction can also take any number of :ref:`fast-math flags <fastmath>`,
5846 which are optimization hints to enable otherwise unsafe floating point
5852 .. code-block:: llvm
5854 <result> = fadd float 4.0, %var ; yields float:result = 4.0 + %var
5856 '``sub``' Instruction
5857 ^^^^^^^^^^^^^^^^^^^^^
5864 <result> = sub <ty> <op1>, <op2> ; yields ty:result
5865 <result> = sub nuw <ty> <op1>, <op2> ; yields ty:result
5866 <result> = sub nsw <ty> <op1>, <op2> ; yields ty:result
5867 <result> = sub nuw nsw <ty> <op1>, <op2> ; yields ty:result
5872 The '``sub``' instruction returns the difference of its two operands.
5874 Note that the '``sub``' instruction is used to represent the '``neg``'
5875 instruction present in most other intermediate representations.
5880 The two arguments to the '``sub``' instruction must be
5881 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
5882 arguments must have identical types.
5887 The value produced is the integer difference of the two operands.
5889 If the difference has unsigned overflow, the result returned is the
5890 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
5893 Because LLVM integers use a two's complement representation, this
5894 instruction is appropriate for both signed and unsigned integers.
5896 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
5897 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
5898 result value of the ``sub`` is a :ref:`poison value <poisonvalues>` if
5899 unsigned and/or signed overflow, respectively, occurs.
5904 .. code-block:: llvm
5906 <result> = sub i32 4, %var ; yields i32:result = 4 - %var
5907 <result> = sub i32 0, %val ; yields i32:result = -%var
5911 '``fsub``' Instruction
5912 ^^^^^^^^^^^^^^^^^^^^^^
5919 <result> = fsub [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
5924 The '``fsub``' instruction returns the difference of its two operands.
5926 Note that the '``fsub``' instruction is used to represent the '``fneg``'
5927 instruction present in most other intermediate representations.
5932 The two arguments to the '``fsub``' instruction must be :ref:`floating
5933 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
5934 Both arguments must have identical types.
5939 The value produced is the floating point difference of the two operands.
5940 This instruction can also take any number of :ref:`fast-math
5941 flags <fastmath>`, which are optimization hints to enable otherwise
5942 unsafe floating point optimizations:
5947 .. code-block:: llvm
5949 <result> = fsub float 4.0, %var ; yields float:result = 4.0 - %var
5950 <result> = fsub float -0.0, %val ; yields float:result = -%var
5952 '``mul``' Instruction
5953 ^^^^^^^^^^^^^^^^^^^^^
5960 <result> = mul <ty> <op1>, <op2> ; yields ty:result
5961 <result> = mul nuw <ty> <op1>, <op2> ; yields ty:result
5962 <result> = mul nsw <ty> <op1>, <op2> ; yields ty:result
5963 <result> = mul nuw nsw <ty> <op1>, <op2> ; yields ty:result
5968 The '``mul``' instruction returns the product of its two operands.
5973 The two arguments to the '``mul``' instruction must be
5974 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
5975 arguments must have identical types.
5980 The value produced is the integer product of the two operands.
5982 If the result of the multiplication has unsigned overflow, the result
5983 returned is the mathematical result modulo 2\ :sup:`n`\ , where n is the
5984 bit width of the result.
5986 Because LLVM integers use a two's complement representation, and the
5987 result is the same width as the operands, this instruction returns the
5988 correct result for both signed and unsigned integers. If a full product
5989 (e.g. ``i32`` * ``i32`` -> ``i64``) is needed, the operands should be
5990 sign-extended or zero-extended as appropriate to the width of the full
5993 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
5994 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
5995 result value of the ``mul`` is a :ref:`poison value <poisonvalues>` if
5996 unsigned and/or signed overflow, respectively, occurs.
6001 .. code-block:: llvm
6003 <result> = mul i32 4, %var ; yields i32:result = 4 * %var
6007 '``fmul``' Instruction
6008 ^^^^^^^^^^^^^^^^^^^^^^
6015 <result> = fmul [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
6020 The '``fmul``' instruction returns the product of its two operands.
6025 The two arguments to the '``fmul``' instruction must be :ref:`floating
6026 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
6027 Both arguments must have identical types.
6032 The value produced is the floating point product of the two operands.
6033 This instruction can also take any number of :ref:`fast-math
6034 flags <fastmath>`, which are optimization hints to enable otherwise
6035 unsafe floating point optimizations:
6040 .. code-block:: llvm
6042 <result> = fmul float 4.0, %var ; yields float:result = 4.0 * %var
6044 '``udiv``' Instruction
6045 ^^^^^^^^^^^^^^^^^^^^^^
6052 <result> = udiv <ty> <op1>, <op2> ; yields ty:result
6053 <result> = udiv exact <ty> <op1>, <op2> ; yields ty:result
6058 The '``udiv``' instruction returns the quotient of its two operands.
6063 The two arguments to the '``udiv``' instruction must be
6064 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6065 arguments must have identical types.
6070 The value produced is the unsigned integer quotient of the two operands.
6072 Note that unsigned integer division and signed integer division are
6073 distinct operations; for signed integer division, use '``sdiv``'.
6075 Division by zero leads to undefined behavior.
6077 If the ``exact`` keyword is present, the result value of the ``udiv`` is
6078 a :ref:`poison value <poisonvalues>` if %op1 is not a multiple of %op2 (as
6079 such, "((a udiv exact b) mul b) == a").
6084 .. code-block:: llvm
6086 <result> = udiv i32 4, %var ; yields i32:result = 4 / %var
6088 '``sdiv``' Instruction
6089 ^^^^^^^^^^^^^^^^^^^^^^
6096 <result> = sdiv <ty> <op1>, <op2> ; yields ty:result
6097 <result> = sdiv exact <ty> <op1>, <op2> ; yields ty:result
6102 The '``sdiv``' instruction returns the quotient of its two operands.
6107 The two arguments to the '``sdiv``' instruction must be
6108 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6109 arguments must have identical types.
6114 The value produced is the signed integer quotient of the two operands
6115 rounded towards zero.
6117 Note that signed integer division and unsigned integer division are
6118 distinct operations; for unsigned integer division, use '``udiv``'.
6120 Division by zero leads to undefined behavior. Overflow also leads to
6121 undefined behavior; this is a rare case, but can occur, for example, by
6122 doing a 32-bit division of -2147483648 by -1.
6124 If the ``exact`` keyword is present, the result value of the ``sdiv`` is
6125 a :ref:`poison value <poisonvalues>` if the result would be rounded.
6130 .. code-block:: llvm
6132 <result> = sdiv i32 4, %var ; yields i32:result = 4 / %var
6136 '``fdiv``' Instruction
6137 ^^^^^^^^^^^^^^^^^^^^^^
6144 <result> = fdiv [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
6149 The '``fdiv``' instruction returns the quotient of its two operands.
6154 The two arguments to the '``fdiv``' instruction must be :ref:`floating
6155 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
6156 Both arguments must have identical types.
6161 The value produced is the floating point quotient of the two operands.
6162 This instruction can also take any number of :ref:`fast-math
6163 flags <fastmath>`, which are optimization hints to enable otherwise
6164 unsafe floating point optimizations:
6169 .. code-block:: llvm
6171 <result> = fdiv float 4.0, %var ; yields float:result = 4.0 / %var
6173 '``urem``' Instruction
6174 ^^^^^^^^^^^^^^^^^^^^^^
6181 <result> = urem <ty> <op1>, <op2> ; yields ty:result
6186 The '``urem``' instruction returns the remainder from the unsigned
6187 division of its two arguments.
6192 The two arguments to the '``urem``' instruction must be
6193 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6194 arguments must have identical types.
6199 This instruction returns the unsigned integer *remainder* of a division.
6200 This instruction always performs an unsigned division to get the
6203 Note that unsigned integer remainder and signed integer remainder are
6204 distinct operations; for signed integer remainder, use '``srem``'.
6206 Taking the remainder of a division by zero leads to undefined behavior.
6211 .. code-block:: llvm
6213 <result> = urem i32 4, %var ; yields i32:result = 4 % %var
6215 '``srem``' Instruction
6216 ^^^^^^^^^^^^^^^^^^^^^^
6223 <result> = srem <ty> <op1>, <op2> ; yields ty:result
6228 The '``srem``' instruction returns the remainder from the signed
6229 division of its two operands. This instruction can also take
6230 :ref:`vector <t_vector>` versions of the values in which case the elements
6236 The two arguments to the '``srem``' instruction must be
6237 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6238 arguments must have identical types.
6243 This instruction returns the *remainder* of a division (where the result
6244 is either zero or has the same sign as the dividend, ``op1``), not the
6245 *modulo* operator (where the result is either zero or has the same sign
6246 as the divisor, ``op2``) of a value. For more information about the
6247 difference, see `The Math
6248 Forum <http://mathforum.org/dr.math/problems/anne.4.28.99.html>`_. For a
6249 table of how this is implemented in various languages, please see
6251 operation <http://en.wikipedia.org/wiki/Modulo_operation>`_.
6253 Note that signed integer remainder and unsigned integer remainder are
6254 distinct operations; for unsigned integer remainder, use '``urem``'.
6256 Taking the remainder of a division by zero leads to undefined behavior.
6257 Overflow also leads to undefined behavior; this is a rare case, but can
6258 occur, for example, by taking the remainder of a 32-bit division of
6259 -2147483648 by -1. (The remainder doesn't actually overflow, but this
6260 rule lets srem be implemented using instructions that return both the
6261 result of the division and the remainder.)
6266 .. code-block:: llvm
6268 <result> = srem i32 4, %var ; yields i32:result = 4 % %var
6272 '``frem``' Instruction
6273 ^^^^^^^^^^^^^^^^^^^^^^
6280 <result> = frem [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
6285 The '``frem``' instruction returns the remainder from the division of
6291 The two arguments to the '``frem``' instruction must be :ref:`floating
6292 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
6293 Both arguments must have identical types.
6298 This instruction returns the *remainder* of a division. The remainder
6299 has the same sign as the dividend. This instruction can also take any
6300 number of :ref:`fast-math flags <fastmath>`, which are optimization hints
6301 to enable otherwise unsafe floating point optimizations:
6306 .. code-block:: llvm
6308 <result> = frem float 4.0, %var ; yields float:result = 4.0 % %var
6312 Bitwise Binary Operations
6313 -------------------------
6315 Bitwise binary operators are used to do various forms of bit-twiddling
6316 in a program. They are generally very efficient instructions and can
6317 commonly be strength reduced from other instructions. They require two
6318 operands of the same type, execute an operation on them, and produce a
6319 single value. The resulting value is the same type as its operands.
6321 '``shl``' Instruction
6322 ^^^^^^^^^^^^^^^^^^^^^
6329 <result> = shl <ty> <op1>, <op2> ; yields ty:result
6330 <result> = shl nuw <ty> <op1>, <op2> ; yields ty:result
6331 <result> = shl nsw <ty> <op1>, <op2> ; yields ty:result
6332 <result> = shl nuw nsw <ty> <op1>, <op2> ; yields ty:result
6337 The '``shl``' instruction returns the first operand shifted to the left
6338 a specified number of bits.
6343 Both arguments to the '``shl``' instruction must be the same
6344 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
6345 '``op2``' is treated as an unsigned value.
6350 The value produced is ``op1`` \* 2\ :sup:`op2` mod 2\ :sup:`n`,
6351 where ``n`` is the width of the result. If ``op2`` is (statically or
6352 dynamically) equal to or larger than the number of bits in
6353 ``op1``, the result is undefined. If the arguments are vectors, each
6354 vector element of ``op1`` is shifted by the corresponding shift amount
6357 If the ``nuw`` keyword is present, then the shift produces a :ref:`poison
6358 value <poisonvalues>` if it shifts out any non-zero bits. If the
6359 ``nsw`` keyword is present, then the shift produces a :ref:`poison
6360 value <poisonvalues>` if it shifts out any bits that disagree with the
6361 resultant sign bit. As such, NUW/NSW have the same semantics as they
6362 would if the shift were expressed as a mul instruction with the same
6363 nsw/nuw bits in (mul %op1, (shl 1, %op2)).
6368 .. code-block:: llvm
6370 <result> = shl i32 4, %var ; yields i32: 4 << %var
6371 <result> = shl i32 4, 2 ; yields i32: 16
6372 <result> = shl i32 1, 10 ; yields i32: 1024
6373 <result> = shl i32 1, 32 ; undefined
6374 <result> = shl <2 x i32> < i32 1, i32 1>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 2, i32 4>
6376 '``lshr``' Instruction
6377 ^^^^^^^^^^^^^^^^^^^^^^
6384 <result> = lshr <ty> <op1>, <op2> ; yields ty:result
6385 <result> = lshr exact <ty> <op1>, <op2> ; yields ty:result
6390 The '``lshr``' instruction (logical shift right) returns the first
6391 operand shifted to the right a specified number of bits with zero fill.
6396 Both arguments to the '``lshr``' instruction must be the same
6397 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
6398 '``op2``' is treated as an unsigned value.
6403 This instruction always performs a logical shift right operation. The
6404 most significant bits of the result will be filled with zero bits after
6405 the shift. If ``op2`` is (statically or dynamically) equal to or larger
6406 than the number of bits in ``op1``, the result is undefined. If the
6407 arguments are vectors, each vector element of ``op1`` is shifted by the
6408 corresponding shift amount in ``op2``.
6410 If the ``exact`` keyword is present, the result value of the ``lshr`` is
6411 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
6417 .. code-block:: llvm
6419 <result> = lshr i32 4, 1 ; yields i32:result = 2
6420 <result> = lshr i32 4, 2 ; yields i32:result = 1
6421 <result> = lshr i8 4, 3 ; yields i8:result = 0
6422 <result> = lshr i8 -2, 1 ; yields i8:result = 0x7F
6423 <result> = lshr i32 1, 32 ; undefined
6424 <result> = lshr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 0x7FFFFFFF, i32 1>
6426 '``ashr``' Instruction
6427 ^^^^^^^^^^^^^^^^^^^^^^
6434 <result> = ashr <ty> <op1>, <op2> ; yields ty:result
6435 <result> = ashr exact <ty> <op1>, <op2> ; yields ty:result
6440 The '``ashr``' instruction (arithmetic shift right) returns the first
6441 operand shifted to the right a specified number of bits with sign
6447 Both arguments to the '``ashr``' instruction must be the same
6448 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
6449 '``op2``' is treated as an unsigned value.
6454 This instruction always performs an arithmetic shift right operation,
6455 The most significant bits of the result will be filled with the sign bit
6456 of ``op1``. If ``op2`` is (statically or dynamically) equal to or larger
6457 than the number of bits in ``op1``, the result is undefined. If the
6458 arguments are vectors, each vector element of ``op1`` is shifted by the
6459 corresponding shift amount in ``op2``.
6461 If the ``exact`` keyword is present, the result value of the ``ashr`` is
6462 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
6468 .. code-block:: llvm
6470 <result> = ashr i32 4, 1 ; yields i32:result = 2
6471 <result> = ashr i32 4, 2 ; yields i32:result = 1
6472 <result> = ashr i8 4, 3 ; yields i8:result = 0
6473 <result> = ashr i8 -2, 1 ; yields i8:result = -1
6474 <result> = ashr i32 1, 32 ; undefined
6475 <result> = ashr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 3> ; yields: result=<2 x i32> < i32 -1, i32 0>
6477 '``and``' Instruction
6478 ^^^^^^^^^^^^^^^^^^^^^
6485 <result> = and <ty> <op1>, <op2> ; yields ty:result
6490 The '``and``' instruction returns the bitwise logical and of its two
6496 The two arguments to the '``and``' instruction must be
6497 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6498 arguments must have identical types.
6503 The truth table used for the '``and``' instruction is:
6520 .. code-block:: llvm
6522 <result> = and i32 4, %var ; yields i32:result = 4 & %var
6523 <result> = and i32 15, 40 ; yields i32:result = 8
6524 <result> = and i32 4, 8 ; yields i32:result = 0
6526 '``or``' Instruction
6527 ^^^^^^^^^^^^^^^^^^^^
6534 <result> = or <ty> <op1>, <op2> ; yields ty:result
6539 The '``or``' instruction returns the bitwise logical inclusive or of its
6545 The two arguments to the '``or``' instruction must be
6546 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6547 arguments must have identical types.
6552 The truth table used for the '``or``' instruction is:
6571 <result> = or i32 4, %var ; yields i32:result = 4 | %var
6572 <result> = or i32 15, 40 ; yields i32:result = 47
6573 <result> = or i32 4, 8 ; yields i32:result = 12
6575 '``xor``' Instruction
6576 ^^^^^^^^^^^^^^^^^^^^^
6583 <result> = xor <ty> <op1>, <op2> ; yields ty:result
6588 The '``xor``' instruction returns the bitwise logical exclusive or of
6589 its two operands. The ``xor`` is used to implement the "one's
6590 complement" operation, which is the "~" operator in C.
6595 The two arguments to the '``xor``' instruction must be
6596 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6597 arguments must have identical types.
6602 The truth table used for the '``xor``' instruction is:
6619 .. code-block:: llvm
6621 <result> = xor i32 4, %var ; yields i32:result = 4 ^ %var
6622 <result> = xor i32 15, 40 ; yields i32:result = 39
6623 <result> = xor i32 4, 8 ; yields i32:result = 12
6624 <result> = xor i32 %V, -1 ; yields i32:result = ~%V
6629 LLVM supports several instructions to represent vector operations in a
6630 target-independent manner. These instructions cover the element-access
6631 and vector-specific operations needed to process vectors effectively.
6632 While LLVM does directly support these vector operations, many
6633 sophisticated algorithms will want to use target-specific intrinsics to
6634 take full advantage of a specific target.
6636 .. _i_extractelement:
6638 '``extractelement``' Instruction
6639 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6646 <result> = extractelement <n x <ty>> <val>, <ty2> <idx> ; yields <ty>
6651 The '``extractelement``' instruction extracts a single scalar element
6652 from a vector at a specified index.
6657 The first operand of an '``extractelement``' instruction is a value of
6658 :ref:`vector <t_vector>` type. The second operand is an index indicating
6659 the position from which to extract the element. The index may be a
6660 variable of any integer type.
6665 The result is a scalar of the same type as the element type of ``val``.
6666 Its value is the value at position ``idx`` of ``val``. If ``idx``
6667 exceeds the length of ``val``, the results are undefined.
6672 .. code-block:: llvm
6674 <result> = extractelement <4 x i32> %vec, i32 0 ; yields i32
6676 .. _i_insertelement:
6678 '``insertelement``' Instruction
6679 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6686 <result> = insertelement <n x <ty>> <val>, <ty> <elt>, <ty2> <idx> ; yields <n x <ty>>
6691 The '``insertelement``' instruction inserts a scalar element into a
6692 vector at a specified index.
6697 The first operand of an '``insertelement``' instruction is a value of
6698 :ref:`vector <t_vector>` type. The second operand is a scalar value whose
6699 type must equal the element type of the first operand. The third operand
6700 is an index indicating the position at which to insert the value. The
6701 index may be a variable of any integer type.
6706 The result is a vector of the same type as ``val``. Its element values
6707 are those of ``val`` except at position ``idx``, where it gets the value
6708 ``elt``. If ``idx`` exceeds the length of ``val``, the results are
6714 .. code-block:: llvm
6716 <result> = insertelement <4 x i32> %vec, i32 1, i32 0 ; yields <4 x i32>
6718 .. _i_shufflevector:
6720 '``shufflevector``' Instruction
6721 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6728 <result> = shufflevector <n x <ty>> <v1>, <n x <ty>> <v2>, <m x i32> <mask> ; yields <m x <ty>>
6733 The '``shufflevector``' instruction constructs a permutation of elements
6734 from two input vectors, returning a vector with the same element type as
6735 the input and length that is the same as the shuffle mask.
6740 The first two operands of a '``shufflevector``' instruction are vectors
6741 with the same type. The third argument is a shuffle mask whose element
6742 type is always 'i32'. The result of the instruction is a vector whose
6743 length is the same as the shuffle mask and whose element type is the
6744 same as the element type of the first two operands.
6746 The shuffle mask operand is required to be a constant vector with either
6747 constant integer or undef values.
6752 The elements of the two input vectors are numbered from left to right
6753 across both of the vectors. The shuffle mask operand specifies, for each
6754 element of the result vector, which element of the two input vectors the
6755 result element gets. The element selector may be undef (meaning "don't
6756 care") and the second operand may be undef if performing a shuffle from
6762 .. code-block:: llvm
6764 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
6765 <4 x i32> <i32 0, i32 4, i32 1, i32 5> ; yields <4 x i32>
6766 <result> = shufflevector <4 x i32> %v1, <4 x i32> undef,
6767 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32> - Identity shuffle.
6768 <result> = shufflevector <8 x i32> %v1, <8 x i32> undef,
6769 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32>
6770 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
6771 <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7 > ; yields <8 x i32>
6773 Aggregate Operations
6774 --------------------
6776 LLVM supports several instructions for working with
6777 :ref:`aggregate <t_aggregate>` values.
6781 '``extractvalue``' Instruction
6782 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6789 <result> = extractvalue <aggregate type> <val>, <idx>{, <idx>}*
6794 The '``extractvalue``' instruction extracts the value of a member field
6795 from an :ref:`aggregate <t_aggregate>` value.
6800 The first operand of an '``extractvalue``' instruction is a value of
6801 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The other operands are
6802 constant indices to specify which value to extract in a similar manner
6803 as indices in a '``getelementptr``' instruction.
6805 The major differences to ``getelementptr`` indexing are:
6807 - Since the value being indexed is not a pointer, the first index is
6808 omitted and assumed to be zero.
6809 - At least one index must be specified.
6810 - Not only struct indices but also array indices must be in bounds.
6815 The result is the value at the position in the aggregate specified by
6821 .. code-block:: llvm
6823 <result> = extractvalue {i32, float} %agg, 0 ; yields i32
6827 '``insertvalue``' Instruction
6828 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6835 <result> = insertvalue <aggregate type> <val>, <ty> <elt>, <idx>{, <idx>}* ; yields <aggregate type>
6840 The '``insertvalue``' instruction inserts a value into a member field in
6841 an :ref:`aggregate <t_aggregate>` value.
6846 The first operand of an '``insertvalue``' instruction is a value of
6847 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The second operand is
6848 a first-class value to insert. The following operands are constant
6849 indices indicating the position at which to insert the value in a
6850 similar manner as indices in a '``extractvalue``' instruction. The value
6851 to insert must have the same type as the value identified by the
6857 The result is an aggregate of the same type as ``val``. Its value is
6858 that of ``val`` except that the value at the position specified by the
6859 indices is that of ``elt``.
6864 .. code-block:: llvm
6866 %agg1 = insertvalue {i32, float} undef, i32 1, 0 ; yields {i32 1, float undef}
6867 %agg2 = insertvalue {i32, float} %agg1, float %val, 1 ; yields {i32 1, float %val}
6868 %agg3 = insertvalue {i32, {float}} undef, float %val, 1, 0 ; yields {i32 undef, {float %val}}
6872 Memory Access and Addressing Operations
6873 ---------------------------------------
6875 A key design point of an SSA-based representation is how it represents
6876 memory. In LLVM, no memory locations are in SSA form, which makes things
6877 very simple. This section describes how to read, write, and allocate
6882 '``alloca``' Instruction
6883 ^^^^^^^^^^^^^^^^^^^^^^^^
6890 <result> = alloca [inalloca] <type> [, <ty> <NumElements>] [, align <alignment>] ; yields type*:result
6895 The '``alloca``' instruction allocates memory on the stack frame of the
6896 currently executing function, to be automatically released when this
6897 function returns to its caller. The object is always allocated in the
6898 generic address space (address space zero).
6903 The '``alloca``' instruction allocates ``sizeof(<type>)*NumElements``
6904 bytes of memory on the runtime stack, returning a pointer of the
6905 appropriate type to the program. If "NumElements" is specified, it is
6906 the number of elements allocated, otherwise "NumElements" is defaulted
6907 to be one. If a constant alignment is specified, the value result of the
6908 allocation is guaranteed to be aligned to at least that boundary. The
6909 alignment may not be greater than ``1 << 29``. If not specified, or if
6910 zero, the target can choose to align the allocation on any convenient
6911 boundary compatible with the type.
6913 '``type``' may be any sized type.
6918 Memory is allocated; a pointer is returned. The operation is undefined
6919 if there is insufficient stack space for the allocation. '``alloca``'d
6920 memory is automatically released when the function returns. The
6921 '``alloca``' instruction is commonly used to represent automatic
6922 variables that must have an address available. When the function returns
6923 (either with the ``ret`` or ``resume`` instructions), the memory is
6924 reclaimed. Allocating zero bytes is legal, but the result is undefined.
6925 The order in which memory is allocated (ie., which way the stack grows)
6931 .. code-block:: llvm
6933 %ptr = alloca i32 ; yields i32*:ptr
6934 %ptr = alloca i32, i32 4 ; yields i32*:ptr
6935 %ptr = alloca i32, i32 4, align 1024 ; yields i32*:ptr
6936 %ptr = alloca i32, align 1024 ; yields i32*:ptr
6940 '``load``' Instruction
6941 ^^^^^^^^^^^^^^^^^^^^^^
6948 <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>]
6949 <result> = load atomic [volatile] <ty>* <pointer> [singlethread] <ordering>, align <alignment> [, !invariant.group !<index>]
6950 !<index> = !{ i32 1 }
6951 !<deref_bytes_node> = !{i64 <dereferenceable_bytes>}
6952 !<align_node> = !{ i64 <value_alignment> }
6957 The '``load``' instruction is used to read from memory.
6962 The argument to the ``load`` instruction specifies the memory address
6963 from which to load. The type specified must be a :ref:`first
6964 class <t_firstclass>` type. If the ``load`` is marked as ``volatile``,
6965 then the optimizer is not allowed to modify the number or order of
6966 execution of this ``load`` with other :ref:`volatile
6967 operations <volatile>`.
6969 If the ``load`` is marked as ``atomic``, it takes an extra
6970 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
6971 ``release`` and ``acq_rel`` orderings are not valid on ``load``
6972 instructions. Atomic loads produce :ref:`defined <memmodel>` results
6973 when they may see multiple atomic stores. The type of the pointee must
6974 be an integer type whose bit width is a power of two greater than or
6975 equal to eight and less than or equal to a target-specific size limit.
6976 ``align`` must be explicitly specified on atomic loads, and the load has
6977 undefined behavior if the alignment is not set to a value which is at
6978 least the size in bytes of the pointee. ``!nontemporal`` does not have
6979 any defined semantics for atomic loads.
6981 The optional constant ``align`` argument specifies the alignment of the
6982 operation (that is, the alignment of the memory address). A value of 0
6983 or an omitted ``align`` argument means that the operation has the ABI
6984 alignment for the target. It is the responsibility of the code emitter
6985 to ensure that the alignment information is correct. Overestimating the
6986 alignment results in undefined behavior. Underestimating the alignment
6987 may produce less efficient code. An alignment of 1 is always safe. The
6988 maximum possible alignment is ``1 << 29``.
6990 The optional ``!nontemporal`` metadata must reference a single
6991 metadata name ``<index>`` corresponding to a metadata node with one
6992 ``i32`` entry of value 1. The existence of the ``!nontemporal``
6993 metadata on the instruction tells the optimizer and code generator
6994 that this load is not expected to be reused in the cache. The code
6995 generator may select special instructions to save cache bandwidth, such
6996 as the ``MOVNT`` instruction on x86.
6998 The optional ``!invariant.load`` metadata must reference a single
6999 metadata name ``<index>`` corresponding to a metadata node with no
7000 entries. The existence of the ``!invariant.load`` metadata on the
7001 instruction tells the optimizer and code generator that the address
7002 operand to this load points to memory which can be assumed unchanged.
7003 Being invariant does not imply that a location is dereferenceable,
7004 but it does imply that once the location is known dereferenceable
7005 its value is henceforth unchanging.
7007 The optional ``!invariant.group`` metadata must reference a single metadata name
7008 ``<index>`` corresponding to a metadata node. See ``invariant.group`` metadata.
7010 The optional ``!nonnull`` metadata must reference a single
7011 metadata name ``<index>`` corresponding to a metadata node with no
7012 entries. The existence of the ``!nonnull`` metadata on the
7013 instruction tells the optimizer that the value loaded is known to
7014 never be null. This is analogous to the ``nonnull`` attribute
7015 on parameters and return values. This metadata can only be applied
7016 to loads of a pointer type.
7018 The optional ``!dereferenceable`` metadata must reference a single metadata
7019 name ``<deref_bytes_node>`` corresponding to a metadata node with one ``i64``
7020 entry. The existence of the ``!dereferenceable`` metadata on the instruction
7021 tells the optimizer that the value loaded is known to be dereferenceable.
7022 The number of bytes known to be dereferenceable is specified by the integer
7023 value in the metadata node. This is analogous to the ''dereferenceable''
7024 attribute on parameters and return values. This metadata can only be applied
7025 to loads of a pointer type.
7027 The optional ``!dereferenceable_or_null`` metadata must reference a single
7028 metadata name ``<deref_bytes_node>`` corresponding to a metadata node with one
7029 ``i64`` entry. The existence of the ``!dereferenceable_or_null`` metadata on the
7030 instruction tells the optimizer that the value loaded is known to be either
7031 dereferenceable or null.
7032 The number of bytes known to be dereferenceable is specified by the integer
7033 value in the metadata node. This is analogous to the ''dereferenceable_or_null''
7034 attribute on parameters and return values. This metadata can only be applied
7035 to loads of a pointer type.
7037 The optional ``!align`` metadata must reference a single metadata name
7038 ``<align_node>`` corresponding to a metadata node with one ``i64`` entry.
7039 The existence of the ``!align`` metadata on the instruction tells the
7040 optimizer that the value loaded is known to be aligned to a boundary specified
7041 by the integer value in the metadata node. The alignment must be a power of 2.
7042 This is analogous to the ''align'' attribute on parameters and return values.
7043 This metadata can only be applied to loads of a pointer type.
7048 The location of memory pointed to is loaded. If the value being loaded
7049 is of scalar type then the number of bytes read does not exceed the
7050 minimum number of bytes needed to hold all bits of the type. For
7051 example, loading an ``i24`` reads at most three bytes. When loading a
7052 value of a type like ``i20`` with a size that is not an integral number
7053 of bytes, the result is undefined if the value was not originally
7054 written using a store of the same type.
7059 .. code-block:: llvm
7061 %ptr = alloca i32 ; yields i32*:ptr
7062 store i32 3, i32* %ptr ; yields void
7063 %val = load i32, i32* %ptr ; yields i32:val = i32 3
7067 '``store``' Instruction
7068 ^^^^^^^^^^^^^^^^^^^^^^^
7075 store [volatile] <ty> <value>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>][, !invariant.group !<index>] ; yields void
7076 store atomic [volatile] <ty> <value>, <ty>* <pointer> [singlethread] <ordering>, align <alignment> [, !invariant.group !<index>] ; yields void
7081 The '``store``' instruction is used to write to memory.
7086 There are two arguments to the ``store`` instruction: a value to store
7087 and an address at which to store it. The type of the ``<pointer>``
7088 operand must be a pointer to the :ref:`first class <t_firstclass>` type of
7089 the ``<value>`` operand. If the ``store`` is marked as ``volatile``,
7090 then the optimizer is not allowed to modify the number or order of
7091 execution of this ``store`` with other :ref:`volatile
7092 operations <volatile>`.
7094 If the ``store`` is marked as ``atomic``, it takes an extra
7095 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
7096 ``acquire`` and ``acq_rel`` orderings aren't valid on ``store``
7097 instructions. Atomic loads produce :ref:`defined <memmodel>` results
7098 when they may see multiple atomic stores. The type of the pointee must
7099 be an integer type whose bit width is a power of two greater than or
7100 equal to eight and less than or equal to a target-specific size limit.
7101 ``align`` must be explicitly specified on atomic stores, and the store
7102 has undefined behavior if the alignment is not set to a value which is
7103 at least the size in bytes of the pointee. ``!nontemporal`` does not
7104 have any defined semantics for atomic stores.
7106 The optional constant ``align`` argument specifies the alignment of the
7107 operation (that is, the alignment of the memory address). A value of 0
7108 or an omitted ``align`` argument means that the operation has the ABI
7109 alignment for the target. It is the responsibility of the code emitter
7110 to ensure that the alignment information is correct. Overestimating the
7111 alignment results in undefined behavior. Underestimating the
7112 alignment may produce less efficient code. An alignment of 1 is always
7113 safe. The maximum possible alignment is ``1 << 29``.
7115 The optional ``!nontemporal`` metadata must reference a single metadata
7116 name ``<index>`` corresponding to a metadata node with one ``i32`` entry of
7117 value 1. The existence of the ``!nontemporal`` metadata on the instruction
7118 tells the optimizer and code generator that this load is not expected to
7119 be reused in the cache. The code generator may select special
7120 instructions to save cache bandwidth, such as the MOVNT instruction on
7123 The optional ``!invariant.group`` metadata must reference a
7124 single metadata name ``<index>``. See ``invariant.group`` metadata.
7129 The contents of memory are updated to contain ``<value>`` at the
7130 location specified by the ``<pointer>`` operand. If ``<value>`` is
7131 of scalar type then the number of bytes written does not exceed the
7132 minimum number of bytes needed to hold all bits of the type. For
7133 example, storing an ``i24`` writes at most three bytes. When writing a
7134 value of a type like ``i20`` with a size that is not an integral number
7135 of bytes, it is unspecified what happens to the extra bits that do not
7136 belong to the type, but they will typically be overwritten.
7141 .. code-block:: llvm
7143 %ptr = alloca i32 ; yields i32*:ptr
7144 store i32 3, i32* %ptr ; yields void
7145 %val = load i32, i32* %ptr ; yields i32:val = i32 3
7149 '``fence``' Instruction
7150 ^^^^^^^^^^^^^^^^^^^^^^^
7157 fence [singlethread] <ordering> ; yields void
7162 The '``fence``' instruction is used to introduce happens-before edges
7168 '``fence``' instructions take an :ref:`ordering <ordering>` argument which
7169 defines what *synchronizes-with* edges they add. They can only be given
7170 ``acquire``, ``release``, ``acq_rel``, and ``seq_cst`` orderings.
7175 A fence A which has (at least) ``release`` ordering semantics
7176 *synchronizes with* a fence B with (at least) ``acquire`` ordering
7177 semantics if and only if there exist atomic operations X and Y, both
7178 operating on some atomic object M, such that A is sequenced before X, X
7179 modifies M (either directly or through some side effect of a sequence
7180 headed by X), Y is sequenced before B, and Y observes M. This provides a
7181 *happens-before* dependency between A and B. Rather than an explicit
7182 ``fence``, one (but not both) of the atomic operations X or Y might
7183 provide a ``release`` or ``acquire`` (resp.) ordering constraint and
7184 still *synchronize-with* the explicit ``fence`` and establish the
7185 *happens-before* edge.
7187 A ``fence`` which has ``seq_cst`` ordering, in addition to having both
7188 ``acquire`` and ``release`` semantics specified above, participates in
7189 the global program order of other ``seq_cst`` operations and/or fences.
7191 The optional ":ref:`singlethread <singlethread>`" argument specifies
7192 that the fence only synchronizes with other fences in the same thread.
7193 (This is useful for interacting with signal handlers.)
7198 .. code-block:: llvm
7200 fence acquire ; yields void
7201 fence singlethread seq_cst ; yields void
7205 '``cmpxchg``' Instruction
7206 ^^^^^^^^^^^^^^^^^^^^^^^^^
7213 cmpxchg [weak] [volatile] <ty>* <pointer>, <ty> <cmp>, <ty> <new> [singlethread] <success ordering> <failure ordering> ; yields { ty, i1 }
7218 The '``cmpxchg``' instruction is used to atomically modify memory. It
7219 loads a value in memory and compares it to a given value. If they are
7220 equal, it tries to store a new value into the memory.
7225 There are three arguments to the '``cmpxchg``' instruction: an address
7226 to operate on, a value to compare to the value currently be at that
7227 address, and a new value to place at that address if the compared values
7228 are equal. The type of '<cmp>' must be an integer type whose bit width
7229 is a power of two greater than or equal to eight and less than or equal
7230 to a target-specific size limit. '<cmp>' and '<new>' must have the same
7231 type, and the type of '<pointer>' must be a pointer to that type. If the
7232 ``cmpxchg`` is marked as ``volatile``, then the optimizer is not allowed
7233 to modify the number or order of execution of this ``cmpxchg`` with
7234 other :ref:`volatile operations <volatile>`.
7236 The success and failure :ref:`ordering <ordering>` arguments specify how this
7237 ``cmpxchg`` synchronizes with other atomic operations. Both ordering parameters
7238 must be at least ``monotonic``, the ordering constraint on failure must be no
7239 stronger than that on success, and the failure ordering cannot be either
7240 ``release`` or ``acq_rel``.
7242 The optional "``singlethread``" argument declares that the ``cmpxchg``
7243 is only atomic with respect to code (usually signal handlers) running in
7244 the same thread as the ``cmpxchg``. Otherwise the cmpxchg is atomic with
7245 respect to all other code in the system.
7247 The pointer passed into cmpxchg must have alignment greater than or
7248 equal to the size in memory of the operand.
7253 The contents of memory at the location specified by the '``<pointer>``' operand
7254 is read and compared to '``<cmp>``'; if the read value is the equal, the
7255 '``<new>``' is written. The original value at the location is returned, together
7256 with a flag indicating success (true) or failure (false).
7258 If the cmpxchg operation is marked as ``weak`` then a spurious failure is
7259 permitted: the operation may not write ``<new>`` even if the comparison
7262 If the cmpxchg operation is strong (the default), the i1 value is 1 if and only
7263 if the value loaded equals ``cmp``.
7265 A successful ``cmpxchg`` is a read-modify-write instruction for the purpose of
7266 identifying release sequences. A failed ``cmpxchg`` is equivalent to an atomic
7267 load with an ordering parameter determined the second ordering parameter.
7272 .. code-block:: llvm
7275 %orig = atomic load i32, i32* %ptr unordered ; yields i32
7279 %cmp = phi i32 [ %orig, %entry ], [%old, %loop]
7280 %squared = mul i32 %cmp, %cmp
7281 %val_success = cmpxchg i32* %ptr, i32 %cmp, i32 %squared acq_rel monotonic ; yields { i32, i1 }
7282 %value_loaded = extractvalue { i32, i1 } %val_success, 0
7283 %success = extractvalue { i32, i1 } %val_success, 1
7284 br i1 %success, label %done, label %loop
7291 '``atomicrmw``' Instruction
7292 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7299 atomicrmw [volatile] <operation> <ty>* <pointer>, <ty> <value> [singlethread] <ordering> ; yields ty
7304 The '``atomicrmw``' instruction is used to atomically modify memory.
7309 There are three arguments to the '``atomicrmw``' instruction: an
7310 operation to apply, an address whose value to modify, an argument to the
7311 operation. The operation must be one of the following keywords:
7325 The type of '<value>' must be an integer type whose bit width is a power
7326 of two greater than or equal to eight and less than or equal to a
7327 target-specific size limit. The type of the '``<pointer>``' operand must
7328 be a pointer to that type. If the ``atomicrmw`` is marked as
7329 ``volatile``, then the optimizer is not allowed to modify the number or
7330 order of execution of this ``atomicrmw`` with other :ref:`volatile
7331 operations <volatile>`.
7336 The contents of memory at the location specified by the '``<pointer>``'
7337 operand are atomically read, modified, and written back. The original
7338 value at the location is returned. The modification is specified by the
7341 - xchg: ``*ptr = val``
7342 - add: ``*ptr = *ptr + val``
7343 - sub: ``*ptr = *ptr - val``
7344 - and: ``*ptr = *ptr & val``
7345 - nand: ``*ptr = ~(*ptr & val)``
7346 - or: ``*ptr = *ptr | val``
7347 - xor: ``*ptr = *ptr ^ val``
7348 - max: ``*ptr = *ptr > val ? *ptr : val`` (using a signed comparison)
7349 - min: ``*ptr = *ptr < val ? *ptr : val`` (using a signed comparison)
7350 - umax: ``*ptr = *ptr > val ? *ptr : val`` (using an unsigned
7352 - umin: ``*ptr = *ptr < val ? *ptr : val`` (using an unsigned
7358 .. code-block:: llvm
7360 %old = atomicrmw add i32* %ptr, i32 1 acquire ; yields i32
7362 .. _i_getelementptr:
7364 '``getelementptr``' Instruction
7365 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7372 <result> = getelementptr <ty>, <ty>* <ptrval>{, <ty> <idx>}*
7373 <result> = getelementptr inbounds <ty>, <ty>* <ptrval>{, <ty> <idx>}*
7374 <result> = getelementptr <ty>, <ptr vector> <ptrval>, <vector index type> <idx>
7379 The '``getelementptr``' instruction is used to get the address of a
7380 subelement of an :ref:`aggregate <t_aggregate>` data structure. It performs
7381 address calculation only and does not access memory. The instruction can also
7382 be used to calculate a vector of such addresses.
7387 The first argument is always a type used as the basis for the calculations.
7388 The second argument is always a pointer or a vector of pointers, and is the
7389 base address to start from. The remaining arguments are indices
7390 that indicate which of the elements of the aggregate object are indexed.
7391 The interpretation of each index is dependent on the type being indexed
7392 into. The first index always indexes the pointer value given as the
7393 first argument, the second index indexes a value of the type pointed to
7394 (not necessarily the value directly pointed to, since the first index
7395 can be non-zero), etc. The first type indexed into must be a pointer
7396 value, subsequent types can be arrays, vectors, and structs. Note that
7397 subsequent types being indexed into can never be pointers, since that
7398 would require loading the pointer before continuing calculation.
7400 The type of each index argument depends on the type it is indexing into.
7401 When indexing into a (optionally packed) structure, only ``i32`` integer
7402 **constants** are allowed (when using a vector of indices they must all
7403 be the **same** ``i32`` integer constant). When indexing into an array,
7404 pointer or vector, integers of any width are allowed, and they are not
7405 required to be constant. These integers are treated as signed values
7408 For example, let's consider a C code fragment and how it gets compiled
7424 int *foo(struct ST *s) {
7425 return &s[1].Z.B[5][13];
7428 The LLVM code generated by Clang is:
7430 .. code-block:: llvm
7432 %struct.RT = type { i8, [10 x [20 x i32]], i8 }
7433 %struct.ST = type { i32, double, %struct.RT }
7435 define i32* @foo(%struct.ST* %s) nounwind uwtable readnone optsize ssp {
7437 %arrayidx = getelementptr inbounds %struct.ST, %struct.ST* %s, i64 1, i32 2, i32 1, i64 5, i64 13
7444 In the example above, the first index is indexing into the
7445 '``%struct.ST*``' type, which is a pointer, yielding a '``%struct.ST``'
7446 = '``{ i32, double, %struct.RT }``' type, a structure. The second index
7447 indexes into the third element of the structure, yielding a
7448 '``%struct.RT``' = '``{ i8 , [10 x [20 x i32]], i8 }``' type, another
7449 structure. The third index indexes into the second element of the
7450 structure, yielding a '``[10 x [20 x i32]]``' type, an array. The two
7451 dimensions of the array are subscripted into, yielding an '``i32``'
7452 type. The '``getelementptr``' instruction returns a pointer to this
7453 element, thus computing a value of '``i32*``' type.
7455 Note that it is perfectly legal to index partially through a structure,
7456 returning a pointer to an inner element. Because of this, the LLVM code
7457 for the given testcase is equivalent to:
7459 .. code-block:: llvm
7461 define i32* @foo(%struct.ST* %s) {
7462 %t1 = getelementptr %struct.ST, %struct.ST* %s, i32 1 ; yields %struct.ST*:%t1
7463 %t2 = getelementptr %struct.ST, %struct.ST* %t1, i32 0, i32 2 ; yields %struct.RT*:%t2
7464 %t3 = getelementptr %struct.RT, %struct.RT* %t2, i32 0, i32 1 ; yields [10 x [20 x i32]]*:%t3
7465 %t4 = getelementptr [10 x [20 x i32]], [10 x [20 x i32]]* %t3, i32 0, i32 5 ; yields [20 x i32]*:%t4
7466 %t5 = getelementptr [20 x i32], [20 x i32]* %t4, i32 0, i32 13 ; yields i32*:%t5
7470 If the ``inbounds`` keyword is present, the result value of the
7471 ``getelementptr`` is a :ref:`poison value <poisonvalues>` if the base
7472 pointer is not an *in bounds* address of an allocated object, or if any
7473 of the addresses that would be formed by successive addition of the
7474 offsets implied by the indices to the base address with infinitely
7475 precise signed arithmetic are not an *in bounds* address of that
7476 allocated object. The *in bounds* addresses for an allocated object are
7477 all the addresses that point into the object, plus the address one byte
7478 past the end. In cases where the base is a vector of pointers the
7479 ``inbounds`` keyword applies to each of the computations element-wise.
7481 If the ``inbounds`` keyword is not present, the offsets are added to the
7482 base address with silently-wrapping two's complement arithmetic. If the
7483 offsets have a different width from the pointer, they are sign-extended
7484 or truncated to the width of the pointer. The result value of the
7485 ``getelementptr`` may be outside the object pointed to by the base
7486 pointer. The result value may not necessarily be used to access memory
7487 though, even if it happens to point into allocated storage. See the
7488 :ref:`Pointer Aliasing Rules <pointeraliasing>` section for more
7491 The getelementptr instruction is often confusing. For some more insight
7492 into how it works, see :doc:`the getelementptr FAQ <GetElementPtr>`.
7497 .. code-block:: llvm
7499 ; yields [12 x i8]*:aptr
7500 %aptr = getelementptr {i32, [12 x i8]}, {i32, [12 x i8]}* %saptr, i64 0, i32 1
7502 %vptr = getelementptr {i32, <2 x i8>}, {i32, <2 x i8>}* %svptr, i64 0, i32 1, i32 1
7504 %eptr = getelementptr [12 x i8], [12 x i8]* %aptr, i64 0, i32 1
7506 %iptr = getelementptr [10 x i32], [10 x i32]* @arr, i16 0, i16 0
7511 The ``getelementptr`` returns a vector of pointers, instead of a single address,
7512 when one or more of its arguments is a vector. In such cases, all vector
7513 arguments should have the same number of elements, and every scalar argument
7514 will be effectively broadcast into a vector during address calculation.
7516 .. code-block:: llvm
7518 ; All arguments are vectors:
7519 ; A[i] = ptrs[i] + offsets[i]*sizeof(i8)
7520 %A = getelementptr i8, <4 x i8*> %ptrs, <4 x i64> %offsets
7522 ; Add the same scalar offset to each pointer of a vector:
7523 ; A[i] = ptrs[i] + offset*sizeof(i8)
7524 %A = getelementptr i8, <4 x i8*> %ptrs, i64 %offset
7526 ; Add distinct offsets to the same pointer:
7527 ; A[i] = ptr + offsets[i]*sizeof(i8)
7528 %A = getelementptr i8, i8* %ptr, <4 x i64> %offsets
7530 ; In all cases described above the type of the result is <4 x i8*>
7532 The two following instructions are equivalent:
7534 .. code-block:: llvm
7536 getelementptr %struct.ST, <4 x %struct.ST*> %s, <4 x i64> %ind1,
7537 <4 x i32> <i32 2, i32 2, i32 2, i32 2>,
7538 <4 x i32> <i32 1, i32 1, i32 1, i32 1>,
7540 <4 x i64> <i64 13, i64 13, i64 13, i64 13>
7542 getelementptr %struct.ST, <4 x %struct.ST*> %s, <4 x i64> %ind1,
7543 i32 2, i32 1, <4 x i32> %ind4, i64 13
7545 Let's look at the C code, where the vector version of ``getelementptr``
7550 // Let's assume that we vectorize the following loop:
7551 double *A, B; int *C;
7552 for (int i = 0; i < size; ++i) {
7556 .. code-block:: llvm
7558 ; get pointers for 8 elements from array B
7559 %ptrs = getelementptr double, double* %B, <8 x i32> %C
7560 ; load 8 elements from array B into A
7561 %A = call <8 x double> @llvm.masked.gather.v8f64(<8 x double*> %ptrs,
7562 i32 8, <8 x i1> %mask, <8 x double> %passthru)
7564 Conversion Operations
7565 ---------------------
7567 The instructions in this category are the conversion instructions
7568 (casting) which all take a single operand and a type. They perform
7569 various bit conversions on the operand.
7571 '``trunc .. to``' Instruction
7572 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7579 <result> = trunc <ty> <value> to <ty2> ; yields ty2
7584 The '``trunc``' instruction truncates its operand to the type ``ty2``.
7589 The '``trunc``' instruction takes a value to trunc, and a type to trunc
7590 it to. Both types must be of :ref:`integer <t_integer>` types, or vectors
7591 of the same number of integers. The bit size of the ``value`` must be
7592 larger than the bit size of the destination type, ``ty2``. Equal sized
7593 types are not allowed.
7598 The '``trunc``' instruction truncates the high order bits in ``value``
7599 and converts the remaining bits to ``ty2``. Since the source size must
7600 be larger than the destination size, ``trunc`` cannot be a *no-op cast*.
7601 It will always truncate bits.
7606 .. code-block:: llvm
7608 %X = trunc i32 257 to i8 ; yields i8:1
7609 %Y = trunc i32 123 to i1 ; yields i1:true
7610 %Z = trunc i32 122 to i1 ; yields i1:false
7611 %W = trunc <2 x i16> <i16 8, i16 7> to <2 x i8> ; yields <i8 8, i8 7>
7613 '``zext .. to``' Instruction
7614 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7621 <result> = zext <ty> <value> to <ty2> ; yields ty2
7626 The '``zext``' instruction zero extends its operand to type ``ty2``.
7631 The '``zext``' instruction takes a value to cast, and a type to cast it
7632 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
7633 the same number of integers. The bit size of the ``value`` must be
7634 smaller than the bit size of the destination type, ``ty2``.
7639 The ``zext`` fills the high order bits of the ``value`` with zero bits
7640 until it reaches the size of the destination type, ``ty2``.
7642 When zero extending from i1, the result will always be either 0 or 1.
7647 .. code-block:: llvm
7649 %X = zext i32 257 to i64 ; yields i64:257
7650 %Y = zext i1 true to i32 ; yields i32:1
7651 %Z = zext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
7653 '``sext .. to``' Instruction
7654 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7661 <result> = sext <ty> <value> to <ty2> ; yields ty2
7666 The '``sext``' sign extends ``value`` to the type ``ty2``.
7671 The '``sext``' instruction takes a value to cast, and a type to cast it
7672 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
7673 the same number of integers. The bit size of the ``value`` must be
7674 smaller than the bit size of the destination type, ``ty2``.
7679 The '``sext``' instruction performs a sign extension by copying the sign
7680 bit (highest order bit) of the ``value`` until it reaches the bit size
7681 of the type ``ty2``.
7683 When sign extending from i1, the extension always results in -1 or 0.
7688 .. code-block:: llvm
7690 %X = sext i8 -1 to i16 ; yields i16 :65535
7691 %Y = sext i1 true to i32 ; yields i32:-1
7692 %Z = sext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
7694 '``fptrunc .. to``' Instruction
7695 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7702 <result> = fptrunc <ty> <value> to <ty2> ; yields ty2
7707 The '``fptrunc``' instruction truncates ``value`` to type ``ty2``.
7712 The '``fptrunc``' instruction takes a :ref:`floating point <t_floating>`
7713 value to cast and a :ref:`floating point <t_floating>` type to cast it to.
7714 The size of ``value`` must be larger than the size of ``ty2``. This
7715 implies that ``fptrunc`` cannot be used to make a *no-op cast*.
7720 The '``fptrunc``' instruction casts a ``value`` from a larger
7721 :ref:`floating point <t_floating>` type to a smaller :ref:`floating
7722 point <t_floating>` type. If the value cannot fit (i.e. overflows) within the
7723 destination type, ``ty2``, then the results are undefined. If the cast produces
7724 an inexact result, how rounding is performed (e.g. truncation, also known as
7725 round to zero) is undefined.
7730 .. code-block:: llvm
7732 %X = fptrunc double 123.0 to float ; yields float:123.0
7733 %Y = fptrunc double 1.0E+300 to float ; yields undefined
7735 '``fpext .. to``' Instruction
7736 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7743 <result> = fpext <ty> <value> to <ty2> ; yields ty2
7748 The '``fpext``' extends a floating point ``value`` to a larger floating
7754 The '``fpext``' instruction takes a :ref:`floating point <t_floating>`
7755 ``value`` to cast, and a :ref:`floating point <t_floating>` type to cast it
7756 to. The source type must be smaller than the destination type.
7761 The '``fpext``' instruction extends the ``value`` from a smaller
7762 :ref:`floating point <t_floating>` type to a larger :ref:`floating
7763 point <t_floating>` type. The ``fpext`` cannot be used to make a
7764 *no-op cast* because it always changes bits. Use ``bitcast`` to make a
7765 *no-op cast* for a floating point cast.
7770 .. code-block:: llvm
7772 %X = fpext float 3.125 to double ; yields double:3.125000e+00
7773 %Y = fpext double %X to fp128 ; yields fp128:0xL00000000000000004000900000000000
7775 '``fptoui .. to``' Instruction
7776 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7783 <result> = fptoui <ty> <value> to <ty2> ; yields ty2
7788 The '``fptoui``' converts a floating point ``value`` to its unsigned
7789 integer equivalent of type ``ty2``.
7794 The '``fptoui``' instruction takes a value to cast, which must be a
7795 scalar or vector :ref:`floating point <t_floating>` value, and a type to
7796 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
7797 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
7798 type with the same number of elements as ``ty``
7803 The '``fptoui``' instruction converts its :ref:`floating
7804 point <t_floating>` operand into the nearest (rounding towards zero)
7805 unsigned integer value. If the value cannot fit in ``ty2``, the results
7811 .. code-block:: llvm
7813 %X = fptoui double 123.0 to i32 ; yields i32:123
7814 %Y = fptoui float 1.0E+300 to i1 ; yields undefined:1
7815 %Z = fptoui float 1.04E+17 to i8 ; yields undefined:1
7817 '``fptosi .. to``' Instruction
7818 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7825 <result> = fptosi <ty> <value> to <ty2> ; yields ty2
7830 The '``fptosi``' instruction converts :ref:`floating point <t_floating>`
7831 ``value`` to type ``ty2``.
7836 The '``fptosi``' instruction takes a value to cast, which must be a
7837 scalar or vector :ref:`floating point <t_floating>` value, and a type to
7838 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
7839 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
7840 type with the same number of elements as ``ty``
7845 The '``fptosi``' instruction converts its :ref:`floating
7846 point <t_floating>` operand into the nearest (rounding towards zero)
7847 signed integer value. If the value cannot fit in ``ty2``, the results
7853 .. code-block:: llvm
7855 %X = fptosi double -123.0 to i32 ; yields i32:-123
7856 %Y = fptosi float 1.0E-247 to i1 ; yields undefined:1
7857 %Z = fptosi float 1.04E+17 to i8 ; yields undefined:1
7859 '``uitofp .. to``' Instruction
7860 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7867 <result> = uitofp <ty> <value> to <ty2> ; yields ty2
7872 The '``uitofp``' instruction regards ``value`` as an unsigned integer
7873 and converts that value to the ``ty2`` type.
7878 The '``uitofp``' instruction takes a value to cast, which must be a
7879 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
7880 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
7881 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
7882 type with the same number of elements as ``ty``
7887 The '``uitofp``' instruction interprets its operand as an unsigned
7888 integer quantity and converts it to the corresponding floating point
7889 value. If the value cannot fit in the floating point value, the results
7895 .. code-block:: llvm
7897 %X = uitofp i32 257 to float ; yields float:257.0
7898 %Y = uitofp i8 -1 to double ; yields double:255.0
7900 '``sitofp .. to``' Instruction
7901 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7908 <result> = sitofp <ty> <value> to <ty2> ; yields ty2
7913 The '``sitofp``' instruction regards ``value`` as a signed integer and
7914 converts that value to the ``ty2`` type.
7919 The '``sitofp``' instruction takes a value to cast, which must be a
7920 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
7921 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
7922 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
7923 type with the same number of elements as ``ty``
7928 The '``sitofp``' instruction interprets its operand as a signed integer
7929 quantity and converts it to the corresponding floating point value. If
7930 the value cannot fit in the floating point value, the results are
7936 .. code-block:: llvm
7938 %X = sitofp i32 257 to float ; yields float:257.0
7939 %Y = sitofp i8 -1 to double ; yields double:-1.0
7943 '``ptrtoint .. to``' Instruction
7944 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7951 <result> = ptrtoint <ty> <value> to <ty2> ; yields ty2
7956 The '``ptrtoint``' instruction converts the pointer or a vector of
7957 pointers ``value`` to the integer (or vector of integers) type ``ty2``.
7962 The '``ptrtoint``' instruction takes a ``value`` to cast, which must be
7963 a value of type :ref:`pointer <t_pointer>` or a vector of pointers, and a
7964 type to cast it to ``ty2``, which must be an :ref:`integer <t_integer>` or
7965 a vector of integers type.
7970 The '``ptrtoint``' instruction converts ``value`` to integer type
7971 ``ty2`` by interpreting the pointer value as an integer and either
7972 truncating or zero extending that value to the size of the integer type.
7973 If ``value`` is smaller than ``ty2`` then a zero extension is done. If
7974 ``value`` is larger than ``ty2`` then a truncation is done. If they are
7975 the same size, then nothing is done (*no-op cast*) other than a type
7981 .. code-block:: llvm
7983 %X = ptrtoint i32* %P to i8 ; yields truncation on 32-bit architecture
7984 %Y = ptrtoint i32* %P to i64 ; yields zero extension on 32-bit architecture
7985 %Z = ptrtoint <4 x i32*> %P to <4 x i64>; yields vector zero extension for a vector of addresses on 32-bit architecture
7989 '``inttoptr .. to``' Instruction
7990 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7997 <result> = inttoptr <ty> <value> to <ty2> ; yields ty2
8002 The '``inttoptr``' instruction converts an integer ``value`` to a
8003 pointer type, ``ty2``.
8008 The '``inttoptr``' instruction takes an :ref:`integer <t_integer>` value to
8009 cast, and a type to cast it to, which must be a :ref:`pointer <t_pointer>`
8015 The '``inttoptr``' instruction converts ``value`` to type ``ty2`` by
8016 applying either a zero extension or a truncation depending on the size
8017 of the integer ``value``. If ``value`` is larger than the size of a
8018 pointer then a truncation is done. If ``value`` is smaller than the size
8019 of a pointer then a zero extension is done. If they are the same size,
8020 nothing is done (*no-op cast*).
8025 .. code-block:: llvm
8027 %X = inttoptr i32 255 to i32* ; yields zero extension on 64-bit architecture
8028 %Y = inttoptr i32 255 to i32* ; yields no-op on 32-bit architecture
8029 %Z = inttoptr i64 0 to i32* ; yields truncation on 32-bit architecture
8030 %Z = inttoptr <4 x i32> %G to <4 x i8*>; yields truncation of vector G to four pointers
8034 '``bitcast .. to``' Instruction
8035 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8042 <result> = bitcast <ty> <value> to <ty2> ; yields ty2
8047 The '``bitcast``' instruction converts ``value`` to type ``ty2`` without
8053 The '``bitcast``' instruction takes a value to cast, which must be a
8054 non-aggregate first class value, and a type to cast it to, which must
8055 also be a non-aggregate :ref:`first class <t_firstclass>` type. The
8056 bit sizes of ``value`` and the destination type, ``ty2``, must be
8057 identical. If the source type is a pointer, the destination type must
8058 also be a pointer of the same size. This instruction supports bitwise
8059 conversion of vectors to integers and to vectors of other types (as
8060 long as they have the same size).
8065 The '``bitcast``' instruction converts ``value`` to type ``ty2``. It
8066 is always a *no-op cast* because no bits change with this
8067 conversion. The conversion is done as if the ``value`` had been stored
8068 to memory and read back as type ``ty2``. Pointer (or vector of
8069 pointers) types may only be converted to other pointer (or vector of
8070 pointers) types with the same address space through this instruction.
8071 To convert pointers to other types, use the :ref:`inttoptr <i_inttoptr>`
8072 or :ref:`ptrtoint <i_ptrtoint>` instructions first.
8077 .. code-block:: llvm
8079 %X = bitcast i8 255 to i8 ; yields i8 :-1
8080 %Y = bitcast i32* %x to sint* ; yields sint*:%x
8081 %Z = bitcast <2 x int> %V to i64; ; yields i64: %V
8082 %Z = bitcast <2 x i32*> %V to <2 x i64*> ; yields <2 x i64*>
8084 .. _i_addrspacecast:
8086 '``addrspacecast .. to``' Instruction
8087 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8094 <result> = addrspacecast <pty> <ptrval> to <pty2> ; yields pty2
8099 The '``addrspacecast``' instruction converts ``ptrval`` from ``pty`` in
8100 address space ``n`` to type ``pty2`` in address space ``m``.
8105 The '``addrspacecast``' instruction takes a pointer or vector of pointer value
8106 to cast and a pointer type to cast it to, which must have a different
8112 The '``addrspacecast``' instruction converts the pointer value
8113 ``ptrval`` to type ``pty2``. It can be a *no-op cast* or a complex
8114 value modification, depending on the target and the address space
8115 pair. Pointer conversions within the same address space must be
8116 performed with the ``bitcast`` instruction. Note that if the address space
8117 conversion is legal then both result and operand refer to the same memory
8123 .. code-block:: llvm
8125 %X = addrspacecast i32* %x to i32 addrspace(1)* ; yields i32 addrspace(1)*:%x
8126 %Y = addrspacecast i32 addrspace(1)* %y to i64 addrspace(2)* ; yields i64 addrspace(2)*:%y
8127 %Z = addrspacecast <4 x i32*> %z to <4 x float addrspace(3)*> ; yields <4 x float addrspace(3)*>:%z
8134 The instructions in this category are the "miscellaneous" instructions,
8135 which defy better classification.
8139 '``icmp``' Instruction
8140 ^^^^^^^^^^^^^^^^^^^^^^
8147 <result> = icmp <cond> <ty> <op1>, <op2> ; yields i1 or <N x i1>:result
8152 The '``icmp``' instruction returns a boolean value or a vector of
8153 boolean values based on comparison of its two integer, integer vector,
8154 pointer, or pointer vector operands.
8159 The '``icmp``' instruction takes three operands. The first operand is
8160 the condition code indicating the kind of comparison to perform. It is
8161 not a value, just a keyword. The possible condition code are:
8164 #. ``ne``: not equal
8165 #. ``ugt``: unsigned greater than
8166 #. ``uge``: unsigned greater or equal
8167 #. ``ult``: unsigned less than
8168 #. ``ule``: unsigned less or equal
8169 #. ``sgt``: signed greater than
8170 #. ``sge``: signed greater or equal
8171 #. ``slt``: signed less than
8172 #. ``sle``: signed less or equal
8174 The remaining two arguments must be :ref:`integer <t_integer>` or
8175 :ref:`pointer <t_pointer>` or integer :ref:`vector <t_vector>` typed. They
8176 must also be identical types.
8181 The '``icmp``' compares ``op1`` and ``op2`` according to the condition
8182 code given as ``cond``. The comparison performed always yields either an
8183 :ref:`i1 <t_integer>` or vector of ``i1`` result, as follows:
8185 #. ``eq``: yields ``true`` if the operands are equal, ``false``
8186 otherwise. No sign interpretation is necessary or performed.
8187 #. ``ne``: yields ``true`` if the operands are unequal, ``false``
8188 otherwise. No sign interpretation is necessary or performed.
8189 #. ``ugt``: interprets the operands as unsigned values and yields
8190 ``true`` if ``op1`` is greater than ``op2``.
8191 #. ``uge``: interprets the operands as unsigned values and yields
8192 ``true`` if ``op1`` is greater than or equal to ``op2``.
8193 #. ``ult``: interprets the operands as unsigned values and yields
8194 ``true`` if ``op1`` is less than ``op2``.
8195 #. ``ule``: interprets the operands as unsigned values and yields
8196 ``true`` if ``op1`` is less than or equal to ``op2``.
8197 #. ``sgt``: interprets the operands as signed values and yields ``true``
8198 if ``op1`` is greater than ``op2``.
8199 #. ``sge``: interprets the operands as signed values and yields ``true``
8200 if ``op1`` is greater than or equal to ``op2``.
8201 #. ``slt``: interprets the operands as signed values and yields ``true``
8202 if ``op1`` is less than ``op2``.
8203 #. ``sle``: interprets the operands as signed values and yields ``true``
8204 if ``op1`` is less than or equal to ``op2``.
8206 If the operands are :ref:`pointer <t_pointer>` typed, the pointer values
8207 are compared as if they were integers.
8209 If the operands are integer vectors, then they are compared element by
8210 element. The result is an ``i1`` vector with the same number of elements
8211 as the values being compared. Otherwise, the result is an ``i1``.
8216 .. code-block:: llvm
8218 <result> = icmp eq i32 4, 5 ; yields: result=false
8219 <result> = icmp ne float* %X, %X ; yields: result=false
8220 <result> = icmp ult i16 4, 5 ; yields: result=true
8221 <result> = icmp sgt i16 4, 5 ; yields: result=false
8222 <result> = icmp ule i16 -4, 5 ; yields: result=false
8223 <result> = icmp sge i16 4, 5 ; yields: result=false
8225 Note that the code generator does not yet support vector types with the
8226 ``icmp`` instruction.
8230 '``fcmp``' Instruction
8231 ^^^^^^^^^^^^^^^^^^^^^^
8238 <result> = fcmp [fast-math flags]* <cond> <ty> <op1>, <op2> ; yields i1 or <N x i1>:result
8243 The '``fcmp``' instruction returns a boolean value or vector of boolean
8244 values based on comparison of its operands.
8246 If the operands are floating point scalars, then the result type is a
8247 boolean (:ref:`i1 <t_integer>`).
8249 If the operands are floating point vectors, then the result type is a
8250 vector of boolean with the same number of elements as the operands being
8256 The '``fcmp``' instruction takes three operands. The first operand is
8257 the condition code indicating the kind of comparison to perform. It is
8258 not a value, just a keyword. The possible condition code are:
8260 #. ``false``: no comparison, always returns false
8261 #. ``oeq``: ordered and equal
8262 #. ``ogt``: ordered and greater than
8263 #. ``oge``: ordered and greater than or equal
8264 #. ``olt``: ordered and less than
8265 #. ``ole``: ordered and less than or equal
8266 #. ``one``: ordered and not equal
8267 #. ``ord``: ordered (no nans)
8268 #. ``ueq``: unordered or equal
8269 #. ``ugt``: unordered or greater than
8270 #. ``uge``: unordered or greater than or equal
8271 #. ``ult``: unordered or less than
8272 #. ``ule``: unordered or less than or equal
8273 #. ``une``: unordered or not equal
8274 #. ``uno``: unordered (either nans)
8275 #. ``true``: no comparison, always returns true
8277 *Ordered* means that neither operand is a QNAN while *unordered* means
8278 that either operand may be a QNAN.
8280 Each of ``val1`` and ``val2`` arguments must be either a :ref:`floating
8281 point <t_floating>` type or a :ref:`vector <t_vector>` of floating point
8282 type. They must have identical types.
8287 The '``fcmp``' instruction compares ``op1`` and ``op2`` according to the
8288 condition code given as ``cond``. If the operands are vectors, then the
8289 vectors are compared element by element. Each comparison performed
8290 always yields an :ref:`i1 <t_integer>` result, as follows:
8292 #. ``false``: always yields ``false``, regardless of operands.
8293 #. ``oeq``: yields ``true`` if both operands are not a QNAN and ``op1``
8294 is equal to ``op2``.
8295 #. ``ogt``: yields ``true`` if both operands are not a QNAN and ``op1``
8296 is greater than ``op2``.
8297 #. ``oge``: yields ``true`` if both operands are not a QNAN and ``op1``
8298 is greater than or equal to ``op2``.
8299 #. ``olt``: yields ``true`` if both operands are not a QNAN and ``op1``
8300 is less than ``op2``.
8301 #. ``ole``: yields ``true`` if both operands are not a QNAN and ``op1``
8302 is less than or equal to ``op2``.
8303 #. ``one``: yields ``true`` if both operands are not a QNAN and ``op1``
8304 is not equal to ``op2``.
8305 #. ``ord``: yields ``true`` if both operands are not a QNAN.
8306 #. ``ueq``: yields ``true`` if either operand is a QNAN or ``op1`` is
8308 #. ``ugt``: yields ``true`` if either operand is a QNAN or ``op1`` is
8309 greater than ``op2``.
8310 #. ``uge``: yields ``true`` if either operand is a QNAN or ``op1`` is
8311 greater than or equal to ``op2``.
8312 #. ``ult``: yields ``true`` if either operand is a QNAN or ``op1`` is
8314 #. ``ule``: yields ``true`` if either operand is a QNAN or ``op1`` is
8315 less than or equal to ``op2``.
8316 #. ``une``: yields ``true`` if either operand is a QNAN or ``op1`` is
8317 not equal to ``op2``.
8318 #. ``uno``: yields ``true`` if either operand is a QNAN.
8319 #. ``true``: always yields ``true``, regardless of operands.
8321 The ``fcmp`` instruction can also optionally take any number of
8322 :ref:`fast-math flags <fastmath>`, which are optimization hints to enable
8323 otherwise unsafe floating point optimizations.
8325 Any set of fast-math flags are legal on an ``fcmp`` instruction, but the
8326 only flags that have any effect on its semantics are those that allow
8327 assumptions to be made about the values of input arguments; namely
8328 ``nnan``, ``ninf``, and ``nsz``. See :ref:`fastmath` for more information.
8333 .. code-block:: llvm
8335 <result> = fcmp oeq float 4.0, 5.0 ; yields: result=false
8336 <result> = fcmp one float 4.0, 5.0 ; yields: result=true
8337 <result> = fcmp olt float 4.0, 5.0 ; yields: result=true
8338 <result> = fcmp ueq double 1.0, 2.0 ; yields: result=false
8340 Note that the code generator does not yet support vector types with the
8341 ``fcmp`` instruction.
8345 '``phi``' Instruction
8346 ^^^^^^^^^^^^^^^^^^^^^
8353 <result> = phi <ty> [ <val0>, <label0>], ...
8358 The '``phi``' instruction is used to implement the φ node in the SSA
8359 graph representing the function.
8364 The type of the incoming values is specified with the first type field.
8365 After this, the '``phi``' instruction takes a list of pairs as
8366 arguments, with one pair for each predecessor basic block of the current
8367 block. Only values of :ref:`first class <t_firstclass>` type may be used as
8368 the value arguments to the PHI node. Only labels may be used as the
8371 There must be no non-phi instructions between the start of a basic block
8372 and the PHI instructions: i.e. PHI instructions must be first in a basic
8375 For the purposes of the SSA form, the use of each incoming value is
8376 deemed to occur on the edge from the corresponding predecessor block to
8377 the current block (but after any definition of an '``invoke``'
8378 instruction's return value on the same edge).
8383 At runtime, the '``phi``' instruction logically takes on the value
8384 specified by the pair corresponding to the predecessor basic block that
8385 executed just prior to the current block.
8390 .. code-block:: llvm
8392 Loop: ; Infinite loop that counts from 0 on up...
8393 %indvar = phi i32 [ 0, %LoopHeader ], [ %nextindvar, %Loop ]
8394 %nextindvar = add i32 %indvar, 1
8399 '``select``' Instruction
8400 ^^^^^^^^^^^^^^^^^^^^^^^^
8407 <result> = select selty <cond>, <ty> <val1>, <ty> <val2> ; yields ty
8409 selty is either i1 or {<N x i1>}
8414 The '``select``' instruction is used to choose one value based on a
8415 condition, without IR-level branching.
8420 The '``select``' instruction requires an 'i1' value or a vector of 'i1'
8421 values indicating the condition, and two values of the same :ref:`first
8422 class <t_firstclass>` type.
8427 If the condition is an i1 and it evaluates to 1, the instruction returns
8428 the first value argument; otherwise, it returns the second value
8431 If the condition is a vector of i1, then the value arguments must be
8432 vectors of the same size, and the selection is done element by element.
8434 If the condition is an i1 and the value arguments are vectors of the
8435 same size, then an entire vector is selected.
8440 .. code-block:: llvm
8442 %X = select i1 true, i8 17, i8 42 ; yields i8:17
8446 '``call``' Instruction
8447 ^^^^^^^^^^^^^^^^^^^^^^
8454 <result> = [tail | musttail | notail ] call [cconv] [ret attrs] <ty> [<fnty>*] <fnptrval>(<function args>) [fn attrs]
8460 The '``call``' instruction represents a simple function call.
8465 This instruction requires several arguments:
8467 #. The optional ``tail`` and ``musttail`` markers indicate that the optimizers
8468 should perform tail call optimization. The ``tail`` marker is a hint that
8469 `can be ignored <CodeGenerator.html#sibcallopt>`_. The ``musttail`` marker
8470 means that the call must be tail call optimized in order for the program to
8471 be correct. The ``musttail`` marker provides these guarantees:
8473 #. The call will not cause unbounded stack growth if it is part of a
8474 recursive cycle in the call graph.
8475 #. Arguments with the :ref:`inalloca <attr_inalloca>` attribute are
8478 Both markers imply that the callee does not access allocas or varargs from
8479 the caller. Calls marked ``musttail`` must obey the following additional
8482 - The call must immediately precede a :ref:`ret <i_ret>` instruction,
8483 or a pointer bitcast followed by a ret instruction.
8484 - The ret instruction must return the (possibly bitcasted) value
8485 produced by the call or void.
8486 - The caller and callee prototypes must match. Pointer types of
8487 parameters or return types may differ in pointee type, but not
8489 - The calling conventions of the caller and callee must match.
8490 - All ABI-impacting function attributes, such as sret, byval, inreg,
8491 returned, and inalloca, must match.
8492 - The callee must be varargs iff the caller is varargs. Bitcasting a
8493 non-varargs function to the appropriate varargs type is legal so
8494 long as the non-varargs prefixes obey the other rules.
8496 Tail call optimization for calls marked ``tail`` is guaranteed to occur if
8497 the following conditions are met:
8499 - Caller and callee both have the calling convention ``fastcc``.
8500 - The call is in tail position (ret immediately follows call and ret
8501 uses value of call or is void).
8502 - Option ``-tailcallopt`` is enabled, or
8503 ``llvm::GuaranteedTailCallOpt`` is ``true``.
8504 - `Platform-specific constraints are
8505 met. <CodeGenerator.html#tailcallopt>`_
8507 #. The optional ``notail`` marker indicates that the optimizers should not add
8508 ``tail`` or ``musttail`` markers to the call. It is used to prevent tail
8509 call optimization from being performed on the call.
8511 #. The optional "cconv" marker indicates which :ref:`calling
8512 convention <callingconv>` the call should use. If none is
8513 specified, the call defaults to using C calling conventions. The
8514 calling convention of the call must match the calling convention of
8515 the target function, or else the behavior is undefined.
8516 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
8517 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
8519 #. '``ty``': the type of the call instruction itself which is also the
8520 type of the return value. Functions that return no value are marked
8522 #. '``fnty``': shall be the signature of the pointer to function value
8523 being invoked. The argument types must match the types implied by
8524 this signature. This type can be omitted if the function is not
8525 varargs and if the function type does not return a pointer to a
8527 #. '``fnptrval``': An LLVM value containing a pointer to a function to
8528 be invoked. In most cases, this is a direct function invocation, but
8529 indirect ``call``'s are just as possible, calling an arbitrary pointer
8531 #. '``function args``': argument list whose types match the function
8532 signature argument types and parameter attributes. All arguments must
8533 be of :ref:`first class <t_firstclass>` type. If the function signature
8534 indicates the function accepts a variable number of arguments, the
8535 extra arguments can be specified.
8536 #. The optional :ref:`function attributes <fnattrs>` list. Only
8537 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
8538 attributes are valid here.
8539 #. The optional :ref:`operand bundles <opbundles>` list.
8544 The '``call``' instruction is used to cause control flow to transfer to
8545 a specified function, with its incoming arguments bound to the specified
8546 values. Upon a '``ret``' instruction in the called function, control
8547 flow continues with the instruction after the function call, and the
8548 return value of the function is bound to the result argument.
8553 .. code-block:: llvm
8555 %retval = call i32 @test(i32 %argc)
8556 call i32 (i8*, ...)* @printf(i8* %msg, i32 12, i8 42) ; yields i32
8557 %X = tail call i32 @foo() ; yields i32
8558 %Y = tail call fastcc i32 @foo() ; yields i32
8559 call void %foo(i8 97 signext)
8561 %struct.A = type { i32, i8 }
8562 %r = call %struct.A @foo() ; yields { i32, i8 }
8563 %gr = extractvalue %struct.A %r, 0 ; yields i32
8564 %gr1 = extractvalue %struct.A %r, 1 ; yields i8
8565 %Z = call void @foo() noreturn ; indicates that %foo never returns normally
8566 %ZZ = call zeroext i32 @bar() ; Return value is %zero extended
8568 llvm treats calls to some functions with names and arguments that match
8569 the standard C99 library as being the C99 library functions, and may
8570 perform optimizations or generate code for them under that assumption.
8571 This is something we'd like to change in the future to provide better
8572 support for freestanding environments and non-C-based languages.
8576 '``va_arg``' Instruction
8577 ^^^^^^^^^^^^^^^^^^^^^^^^
8584 <resultval> = va_arg <va_list*> <arglist>, <argty>
8589 The '``va_arg``' instruction is used to access arguments passed through
8590 the "variable argument" area of a function call. It is used to implement
8591 the ``va_arg`` macro in C.
8596 This instruction takes a ``va_list*`` value and the type of the
8597 argument. It returns a value of the specified argument type and
8598 increments the ``va_list`` to point to the next argument. The actual
8599 type of ``va_list`` is target specific.
8604 The '``va_arg``' instruction loads an argument of the specified type
8605 from the specified ``va_list`` and causes the ``va_list`` to point to
8606 the next argument. For more information, see the variable argument
8607 handling :ref:`Intrinsic Functions <int_varargs>`.
8609 It is legal for this instruction to be called in a function which does
8610 not take a variable number of arguments, for example, the ``vfprintf``
8613 ``va_arg`` is an LLVM instruction instead of an :ref:`intrinsic
8614 function <intrinsics>` because it takes a type as an argument.
8619 See the :ref:`variable argument processing <int_varargs>` section.
8621 Note that the code generator does not yet fully support va\_arg on many
8622 targets. Also, it does not currently support va\_arg with aggregate
8623 types on any target.
8627 '``landingpad``' Instruction
8628 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8635 <resultval> = landingpad <resultty> <clause>+
8636 <resultval> = landingpad <resultty> cleanup <clause>*
8638 <clause> := catch <type> <value>
8639 <clause> := filter <array constant type> <array constant>
8644 The '``landingpad``' instruction is used by `LLVM's exception handling
8645 system <ExceptionHandling.html#overview>`_ to specify that a basic block
8646 is a landing pad --- one where the exception lands, and corresponds to the
8647 code found in the ``catch`` portion of a ``try``/``catch`` sequence. It
8648 defines values supplied by the :ref:`personality function <personalityfn>` upon
8649 re-entry to the function. The ``resultval`` has the type ``resultty``.
8655 ``cleanup`` flag indicates that the landing pad block is a cleanup.
8657 A ``clause`` begins with the clause type --- ``catch`` or ``filter`` --- and
8658 contains the global variable representing the "type" that may be caught
8659 or filtered respectively. Unlike the ``catch`` clause, the ``filter``
8660 clause takes an array constant as its argument. Use
8661 "``[0 x i8**] undef``" for a filter which cannot throw. The
8662 '``landingpad``' instruction must contain *at least* one ``clause`` or
8663 the ``cleanup`` flag.
8668 The '``landingpad``' instruction defines the values which are set by the
8669 :ref:`personality function <personalityfn>` upon re-entry to the function, and
8670 therefore the "result type" of the ``landingpad`` instruction. As with
8671 calling conventions, how the personality function results are
8672 represented in LLVM IR is target specific.
8674 The clauses are applied in order from top to bottom. If two
8675 ``landingpad`` instructions are merged together through inlining, the
8676 clauses from the calling function are appended to the list of clauses.
8677 When the call stack is being unwound due to an exception being thrown,
8678 the exception is compared against each ``clause`` in turn. If it doesn't
8679 match any of the clauses, and the ``cleanup`` flag is not set, then
8680 unwinding continues further up the call stack.
8682 The ``landingpad`` instruction has several restrictions:
8684 - A landing pad block is a basic block which is the unwind destination
8685 of an '``invoke``' instruction.
8686 - A landing pad block must have a '``landingpad``' instruction as its
8687 first non-PHI instruction.
8688 - There can be only one '``landingpad``' instruction within the landing
8690 - A basic block that is not a landing pad block may not include a
8691 '``landingpad``' instruction.
8696 .. code-block:: llvm
8698 ;; A landing pad which can catch an integer.
8699 %res = landingpad { i8*, i32 }
8701 ;; A landing pad that is a cleanup.
8702 %res = landingpad { i8*, i32 }
8704 ;; A landing pad which can catch an integer and can only throw a double.
8705 %res = landingpad { i8*, i32 }
8707 filter [1 x i8**] [@_ZTId]
8711 '``cleanuppad``' Instruction
8712 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8719 <resultval> = cleanuppad [<args>*]
8724 The '``cleanuppad``' instruction is used by `LLVM's exception handling
8725 system <ExceptionHandling.html#overview>`_ to specify that a basic block
8726 is a cleanup block --- one where a personality routine attempts to
8727 transfer control to run cleanup actions.
8728 The ``args`` correspond to whatever additional
8729 information the :ref:`personality function <personalityfn>` requires to
8730 execute the cleanup.
8731 The ``resultval`` has the type :ref:`token <t_token>` and is used to
8732 match the ``cleanuppad`` to corresponding :ref:`cleanuprets <i_cleanupret>`
8733 and :ref:`cleanupendpads <i_cleanupendpad>`.
8738 The instruction takes a list of arbitrary values which are interpreted
8739 by the :ref:`personality function <personalityfn>`.
8744 When the call stack is being unwound due to an exception being thrown,
8745 the :ref:`personality function <personalityfn>` transfers control to the
8746 ``cleanuppad`` with the aid of the personality-specific arguments.
8747 As with calling conventions, how the personality function results are
8748 represented in LLVM IR is target specific.
8750 The ``cleanuppad`` instruction has several restrictions:
8752 - A cleanup block is a basic block which is the unwind destination of
8753 an exceptional instruction.
8754 - A cleanup block must have a '``cleanuppad``' instruction as its
8755 first non-PHI instruction.
8756 - There can be only one '``cleanuppad``' instruction within the
8758 - A basic block that is not a cleanup block may not include a
8759 '``cleanuppad``' instruction.
8760 - All '``cleanupret``'s and '``cleanupendpad``'s which consume a ``cleanuppad``
8761 must have the same exceptional successor.
8762 - It is undefined behavior for control to transfer from a ``cleanuppad`` to a
8763 ``ret`` without first executing a ``cleanupret`` or ``cleanupendpad`` that
8764 consumes the ``cleanuppad``.
8765 - It is undefined behavior for control to transfer from a ``cleanuppad`` to
8766 itself without first executing a ``cleanupret`` or ``cleanupendpad`` that
8767 consumes the ``cleanuppad``.
8772 .. code-block:: llvm
8774 %tok = cleanuppad []
8781 LLVM supports the notion of an "intrinsic function". These functions
8782 have well known names and semantics and are required to follow certain
8783 restrictions. Overall, these intrinsics represent an extension mechanism
8784 for the LLVM language that does not require changing all of the
8785 transformations in LLVM when adding to the language (or the bitcode
8786 reader/writer, the parser, etc...).
8788 Intrinsic function names must all start with an "``llvm.``" prefix. This
8789 prefix is reserved in LLVM for intrinsic names; thus, function names may
8790 not begin with this prefix. Intrinsic functions must always be external
8791 functions: you cannot define the body of intrinsic functions. Intrinsic
8792 functions may only be used in call or invoke instructions: it is illegal
8793 to take the address of an intrinsic function. Additionally, because
8794 intrinsic functions are part of the LLVM language, it is required if any
8795 are added that they be documented here.
8797 Some intrinsic functions can be overloaded, i.e., the intrinsic
8798 represents a family of functions that perform the same operation but on
8799 different data types. Because LLVM can represent over 8 million
8800 different integer types, overloading is used commonly to allow an
8801 intrinsic function to operate on any integer type. One or more of the
8802 argument types or the result type can be overloaded to accept any
8803 integer type. Argument types may also be defined as exactly matching a
8804 previous argument's type or the result type. This allows an intrinsic
8805 function which accepts multiple arguments, but needs all of them to be
8806 of the same type, to only be overloaded with respect to a single
8807 argument or the result.
8809 Overloaded intrinsics will have the names of its overloaded argument
8810 types encoded into its function name, each preceded by a period. Only
8811 those types which are overloaded result in a name suffix. Arguments
8812 whose type is matched against another type do not. For example, the
8813 ``llvm.ctpop`` function can take an integer of any width and returns an
8814 integer of exactly the same integer width. This leads to a family of
8815 functions such as ``i8 @llvm.ctpop.i8(i8 %val)`` and
8816 ``i29 @llvm.ctpop.i29(i29 %val)``. Only one type, the return type, is
8817 overloaded, and only one type suffix is required. Because the argument's
8818 type is matched against the return type, it does not require its own
8821 To learn how to add an intrinsic function, please see the `Extending
8822 LLVM Guide <ExtendingLLVM.html>`_.
8826 Variable Argument Handling Intrinsics
8827 -------------------------------------
8829 Variable argument support is defined in LLVM with the
8830 :ref:`va_arg <i_va_arg>` instruction and these three intrinsic
8831 functions. These functions are related to the similarly named macros
8832 defined in the ``<stdarg.h>`` header file.
8834 All of these functions operate on arguments that use a target-specific
8835 value type "``va_list``". The LLVM assembly language reference manual
8836 does not define what this type is, so all transformations should be
8837 prepared to handle these functions regardless of the type used.
8839 This example shows how the :ref:`va_arg <i_va_arg>` instruction and the
8840 variable argument handling intrinsic functions are used.
8842 .. code-block:: llvm
8844 ; This struct is different for every platform. For most platforms,
8845 ; it is merely an i8*.
8846 %struct.va_list = type { i8* }
8848 ; For Unix x86_64 platforms, va_list is the following struct:
8849 ; %struct.va_list = type { i32, i32, i8*, i8* }
8851 define i32 @test(i32 %X, ...) {
8852 ; Initialize variable argument processing
8853 %ap = alloca %struct.va_list
8854 %ap2 = bitcast %struct.va_list* %ap to i8*
8855 call void @llvm.va_start(i8* %ap2)
8857 ; Read a single integer argument
8858 %tmp = va_arg i8* %ap2, i32
8860 ; Demonstrate usage of llvm.va_copy and llvm.va_end
8862 %aq2 = bitcast i8** %aq to i8*
8863 call void @llvm.va_copy(i8* %aq2, i8* %ap2)
8864 call void @llvm.va_end(i8* %aq2)
8866 ; Stop processing of arguments.
8867 call void @llvm.va_end(i8* %ap2)
8871 declare void @llvm.va_start(i8*)
8872 declare void @llvm.va_copy(i8*, i8*)
8873 declare void @llvm.va_end(i8*)
8877 '``llvm.va_start``' Intrinsic
8878 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8885 declare void @llvm.va_start(i8* <arglist>)
8890 The '``llvm.va_start``' intrinsic initializes ``*<arglist>`` for
8891 subsequent use by ``va_arg``.
8896 The argument is a pointer to a ``va_list`` element to initialize.
8901 The '``llvm.va_start``' intrinsic works just like the ``va_start`` macro
8902 available in C. In a target-dependent way, it initializes the
8903 ``va_list`` element to which the argument points, so that the next call
8904 to ``va_arg`` will produce the first variable argument passed to the
8905 function. Unlike the C ``va_start`` macro, this intrinsic does not need
8906 to know the last argument of the function as the compiler can figure
8909 '``llvm.va_end``' Intrinsic
8910 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8917 declare void @llvm.va_end(i8* <arglist>)
8922 The '``llvm.va_end``' intrinsic destroys ``*<arglist>``, which has been
8923 initialized previously with ``llvm.va_start`` or ``llvm.va_copy``.
8928 The argument is a pointer to a ``va_list`` to destroy.
8933 The '``llvm.va_end``' intrinsic works just like the ``va_end`` macro
8934 available in C. In a target-dependent way, it destroys the ``va_list``
8935 element to which the argument points. Calls to
8936 :ref:`llvm.va_start <int_va_start>` and
8937 :ref:`llvm.va_copy <int_va_copy>` must be matched exactly with calls to
8942 '``llvm.va_copy``' Intrinsic
8943 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8950 declare void @llvm.va_copy(i8* <destarglist>, i8* <srcarglist>)
8955 The '``llvm.va_copy``' intrinsic copies the current argument position
8956 from the source argument list to the destination argument list.
8961 The first argument is a pointer to a ``va_list`` element to initialize.
8962 The second argument is a pointer to a ``va_list`` element to copy from.
8967 The '``llvm.va_copy``' intrinsic works just like the ``va_copy`` macro
8968 available in C. In a target-dependent way, it copies the source
8969 ``va_list`` element into the destination ``va_list`` element. This
8970 intrinsic is necessary because the `` llvm.va_start`` intrinsic may be
8971 arbitrarily complex and require, for example, memory allocation.
8973 Accurate Garbage Collection Intrinsics
8974 --------------------------------------
8976 LLVM's support for `Accurate Garbage Collection <GarbageCollection.html>`_
8977 (GC) requires the frontend to generate code containing appropriate intrinsic
8978 calls and select an appropriate GC strategy which knows how to lower these
8979 intrinsics in a manner which is appropriate for the target collector.
8981 These intrinsics allow identification of :ref:`GC roots on the
8982 stack <int_gcroot>`, as well as garbage collector implementations that
8983 require :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers.
8984 Frontends for type-safe garbage collected languages should generate
8985 these intrinsics to make use of the LLVM garbage collectors. For more
8986 details, see `Garbage Collection with LLVM <GarbageCollection.html>`_.
8988 Experimental Statepoint Intrinsics
8989 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8991 LLVM provides an second experimental set of intrinsics for describing garbage
8992 collection safepoints in compiled code. These intrinsics are an alternative
8993 to the ``llvm.gcroot`` intrinsics, but are compatible with the ones for
8994 :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers. The
8995 differences in approach are covered in the `Garbage Collection with LLVM
8996 <GarbageCollection.html>`_ documentation. The intrinsics themselves are
8997 described in :doc:`Statepoints`.
9001 '``llvm.gcroot``' Intrinsic
9002 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9009 declare void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
9014 The '``llvm.gcroot``' intrinsic declares the existence of a GC root to
9015 the code generator, and allows some metadata to be associated with it.
9020 The first argument specifies the address of a stack object that contains
9021 the root pointer. The second pointer (which must be either a constant or
9022 a global value address) contains the meta-data to be associated with the
9028 At runtime, a call to this intrinsic stores a null pointer into the
9029 "ptrloc" location. At compile-time, the code generator generates
9030 information to allow the runtime to find the pointer at GC safe points.
9031 The '``llvm.gcroot``' intrinsic may only be used in a function which
9032 :ref:`specifies a GC algorithm <gc>`.
9036 '``llvm.gcread``' Intrinsic
9037 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9044 declare i8* @llvm.gcread(i8* %ObjPtr, i8** %Ptr)
9049 The '``llvm.gcread``' intrinsic identifies reads of references from heap
9050 locations, allowing garbage collector implementations that require read
9056 The second argument is the address to read from, which should be an
9057 address allocated from the garbage collector. The first object is a
9058 pointer to the start of the referenced object, if needed by the language
9059 runtime (otherwise null).
9064 The '``llvm.gcread``' intrinsic has the same semantics as a load
9065 instruction, but may be replaced with substantially more complex code by
9066 the garbage collector runtime, as needed. The '``llvm.gcread``'
9067 intrinsic may only be used in a function which :ref:`specifies a GC
9072 '``llvm.gcwrite``' Intrinsic
9073 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9080 declare void @llvm.gcwrite(i8* %P1, i8* %Obj, i8** %P2)
9085 The '``llvm.gcwrite``' intrinsic identifies writes of references to heap
9086 locations, allowing garbage collector implementations that require write
9087 barriers (such as generational or reference counting collectors).
9092 The first argument is the reference to store, the second is the start of
9093 the object to store it to, and the third is the address of the field of
9094 Obj to store to. If the runtime does not require a pointer to the
9095 object, Obj may be null.
9100 The '``llvm.gcwrite``' intrinsic has the same semantics as a store
9101 instruction, but may be replaced with substantially more complex code by
9102 the garbage collector runtime, as needed. The '``llvm.gcwrite``'
9103 intrinsic may only be used in a function which :ref:`specifies a GC
9106 Code Generator Intrinsics
9107 -------------------------
9109 These intrinsics are provided by LLVM to expose special features that
9110 may only be implemented with code generator support.
9112 '``llvm.returnaddress``' Intrinsic
9113 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9120 declare i8 *@llvm.returnaddress(i32 <level>)
9125 The '``llvm.returnaddress``' intrinsic attempts to compute a
9126 target-specific value indicating the return address of the current
9127 function or one of its callers.
9132 The argument to this intrinsic indicates which function to return the
9133 address for. Zero indicates the calling function, one indicates its
9134 caller, etc. The argument is **required** to be a constant integer
9140 The '``llvm.returnaddress``' intrinsic either returns a pointer
9141 indicating the return address of the specified call frame, or zero if it
9142 cannot be identified. The value returned by this intrinsic is likely to
9143 be incorrect or 0 for arguments other than zero, so it should only be
9144 used for debugging purposes.
9146 Note that calling this intrinsic does not prevent function inlining or
9147 other aggressive transformations, so the value returned may not be that
9148 of the obvious source-language caller.
9150 '``llvm.frameaddress``' Intrinsic
9151 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9158 declare i8* @llvm.frameaddress(i32 <level>)
9163 The '``llvm.frameaddress``' intrinsic attempts to return the
9164 target-specific frame pointer value for the specified stack frame.
9169 The argument to this intrinsic indicates which function to return the
9170 frame pointer for. Zero indicates the calling function, one indicates
9171 its caller, etc. The argument is **required** to be a constant integer
9177 The '``llvm.frameaddress``' intrinsic either returns a pointer
9178 indicating the frame address of the specified call frame, or zero if it
9179 cannot be identified. The value returned by this intrinsic is likely to
9180 be incorrect or 0 for arguments other than zero, so it should only be
9181 used for debugging purposes.
9183 Note that calling this intrinsic does not prevent function inlining or
9184 other aggressive transformations, so the value returned may not be that
9185 of the obvious source-language caller.
9187 '``llvm.localescape``' and '``llvm.localrecover``' Intrinsics
9188 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9195 declare void @llvm.localescape(...)
9196 declare i8* @llvm.localrecover(i8* %func, i8* %fp, i32 %idx)
9201 The '``llvm.localescape``' intrinsic escapes offsets of a collection of static
9202 allocas, and the '``llvm.localrecover``' intrinsic applies those offsets to a
9203 live frame pointer to recover the address of the allocation. The offset is
9204 computed during frame layout of the caller of ``llvm.localescape``.
9209 All arguments to '``llvm.localescape``' must be pointers to static allocas or
9210 casts of static allocas. Each function can only call '``llvm.localescape``'
9211 once, and it can only do so from the entry block.
9213 The ``func`` argument to '``llvm.localrecover``' must be a constant
9214 bitcasted pointer to a function defined in the current module. The code
9215 generator cannot determine the frame allocation offset of functions defined in
9218 The ``fp`` argument to '``llvm.localrecover``' must be a frame pointer of a
9219 call frame that is currently live. The return value of '``llvm.localaddress``'
9220 is one way to produce such a value, but various runtimes also expose a suitable
9221 pointer in platform-specific ways.
9223 The ``idx`` argument to '``llvm.localrecover``' indicates which alloca passed to
9224 '``llvm.localescape``' to recover. It is zero-indexed.
9229 These intrinsics allow a group of functions to share access to a set of local
9230 stack allocations of a one parent function. The parent function may call the
9231 '``llvm.localescape``' intrinsic once from the function entry block, and the
9232 child functions can use '``llvm.localrecover``' to access the escaped allocas.
9233 The '``llvm.localescape``' intrinsic blocks inlining, as inlining changes where
9234 the escaped allocas are allocated, which would break attempts to use
9235 '``llvm.localrecover``'.
9237 .. _int_read_register:
9238 .. _int_write_register:
9240 '``llvm.read_register``' and '``llvm.write_register``' Intrinsics
9241 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9248 declare i32 @llvm.read_register.i32(metadata)
9249 declare i64 @llvm.read_register.i64(metadata)
9250 declare void @llvm.write_register.i32(metadata, i32 @value)
9251 declare void @llvm.write_register.i64(metadata, i64 @value)
9257 The '``llvm.read_register``' and '``llvm.write_register``' intrinsics
9258 provides access to the named register. The register must be valid on
9259 the architecture being compiled to. The type needs to be compatible
9260 with the register being read.
9265 The '``llvm.read_register``' intrinsic returns the current value of the
9266 register, where possible. The '``llvm.write_register``' intrinsic sets
9267 the current value of the register, where possible.
9269 This is useful to implement named register global variables that need
9270 to always be mapped to a specific register, as is common practice on
9271 bare-metal programs including OS kernels.
9273 The compiler doesn't check for register availability or use of the used
9274 register in surrounding code, including inline assembly. Because of that,
9275 allocatable registers are not supported.
9277 Warning: So far it only works with the stack pointer on selected
9278 architectures (ARM, AArch64, PowerPC and x86_64). Significant amount of
9279 work is needed to support other registers and even more so, allocatable
9284 '``llvm.stacksave``' Intrinsic
9285 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9292 declare i8* @llvm.stacksave()
9297 The '``llvm.stacksave``' intrinsic is used to remember the current state
9298 of the function stack, for use with
9299 :ref:`llvm.stackrestore <int_stackrestore>`. This is useful for
9300 implementing language features like scoped automatic variable sized
9306 This intrinsic returns a opaque pointer value that can be passed to
9307 :ref:`llvm.stackrestore <int_stackrestore>`. When an
9308 ``llvm.stackrestore`` intrinsic is executed with a value saved from
9309 ``llvm.stacksave``, it effectively restores the state of the stack to
9310 the state it was in when the ``llvm.stacksave`` intrinsic executed. In
9311 practice, this pops any :ref:`alloca <i_alloca>` blocks from the stack that
9312 were allocated after the ``llvm.stacksave`` was executed.
9314 .. _int_stackrestore:
9316 '``llvm.stackrestore``' Intrinsic
9317 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9324 declare void @llvm.stackrestore(i8* %ptr)
9329 The '``llvm.stackrestore``' intrinsic is used to restore the state of
9330 the function stack to the state it was in when the corresponding
9331 :ref:`llvm.stacksave <int_stacksave>` intrinsic executed. This is
9332 useful for implementing language features like scoped automatic variable
9333 sized arrays in C99.
9338 See the description for :ref:`llvm.stacksave <int_stacksave>`.
9340 '``llvm.prefetch``' Intrinsic
9341 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9348 declare void @llvm.prefetch(i8* <address>, i32 <rw>, i32 <locality>, i32 <cache type>)
9353 The '``llvm.prefetch``' intrinsic is a hint to the code generator to
9354 insert a prefetch instruction if supported; otherwise, it is a noop.
9355 Prefetches have no effect on the behavior of the program but can change
9356 its performance characteristics.
9361 ``address`` is the address to be prefetched, ``rw`` is the specifier
9362 determining if the fetch should be for a read (0) or write (1), and
9363 ``locality`` is a temporal locality specifier ranging from (0) - no
9364 locality, to (3) - extremely local keep in cache. The ``cache type``
9365 specifies whether the prefetch is performed on the data (1) or
9366 instruction (0) cache. The ``rw``, ``locality`` and ``cache type``
9367 arguments must be constant integers.
9372 This intrinsic does not modify the behavior of the program. In
9373 particular, prefetches cannot trap and do not produce a value. On
9374 targets that support this intrinsic, the prefetch can provide hints to
9375 the processor cache for better performance.
9377 '``llvm.pcmarker``' Intrinsic
9378 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9385 declare void @llvm.pcmarker(i32 <id>)
9390 The '``llvm.pcmarker``' intrinsic is a method to export a Program
9391 Counter (PC) in a region of code to simulators and other tools. The
9392 method is target specific, but it is expected that the marker will use
9393 exported symbols to transmit the PC of the marker. The marker makes no
9394 guarantees that it will remain with any specific instruction after
9395 optimizations. It is possible that the presence of a marker will inhibit
9396 optimizations. The intended use is to be inserted after optimizations to
9397 allow correlations of simulation runs.
9402 ``id`` is a numerical id identifying the marker.
9407 This intrinsic does not modify the behavior of the program. Backends
9408 that do not support this intrinsic may ignore it.
9410 '``llvm.readcyclecounter``' Intrinsic
9411 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9418 declare i64 @llvm.readcyclecounter()
9423 The '``llvm.readcyclecounter``' intrinsic provides access to the cycle
9424 counter register (or similar low latency, high accuracy clocks) on those
9425 targets that support it. On X86, it should map to RDTSC. On Alpha, it
9426 should map to RPCC. As the backing counters overflow quickly (on the
9427 order of 9 seconds on alpha), this should only be used for small
9433 When directly supported, reading the cycle counter should not modify any
9434 memory. Implementations are allowed to either return a application
9435 specific value or a system wide value. On backends without support, this
9436 is lowered to a constant 0.
9438 Note that runtime support may be conditional on the privilege-level code is
9439 running at and the host platform.
9441 '``llvm.clear_cache``' Intrinsic
9442 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9449 declare void @llvm.clear_cache(i8*, i8*)
9454 The '``llvm.clear_cache``' intrinsic ensures visibility of modifications
9455 in the specified range to the execution unit of the processor. On
9456 targets with non-unified instruction and data cache, the implementation
9457 flushes the instruction cache.
9462 On platforms with coherent instruction and data caches (e.g. x86), this
9463 intrinsic is a nop. On platforms with non-coherent instruction and data
9464 cache (e.g. ARM, MIPS), the intrinsic is lowered either to appropriate
9465 instructions or a system call, if cache flushing requires special
9468 The default behavior is to emit a call to ``__clear_cache`` from the run
9471 This instrinsic does *not* empty the instruction pipeline. Modifications
9472 of the current function are outside the scope of the intrinsic.
9474 '``llvm.instrprof_increment``' Intrinsic
9475 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9482 declare void @llvm.instrprof_increment(i8* <name>, i64 <hash>,
9483 i32 <num-counters>, i32 <index>)
9488 The '``llvm.instrprof_increment``' intrinsic can be emitted by a
9489 frontend for use with instrumentation based profiling. These will be
9490 lowered by the ``-instrprof`` pass to generate execution counts of a
9496 The first argument is a pointer to a global variable containing the
9497 name of the entity being instrumented. This should generally be the
9498 (mangled) function name for a set of counters.
9500 The second argument is a hash value that can be used by the consumer
9501 of the profile data to detect changes to the instrumented source, and
9502 the third is the number of counters associated with ``name``. It is an
9503 error if ``hash`` or ``num-counters`` differ between two instances of
9504 ``instrprof_increment`` that refer to the same name.
9506 The last argument refers to which of the counters for ``name`` should
9507 be incremented. It should be a value between 0 and ``num-counters``.
9512 This intrinsic represents an increment of a profiling counter. It will
9513 cause the ``-instrprof`` pass to generate the appropriate data
9514 structures and the code to increment the appropriate value, in a
9515 format that can be written out by a compiler runtime and consumed via
9516 the ``llvm-profdata`` tool.
9518 '``llvm.instrprof_value_profile``' Intrinsic
9519 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9526 declare void @llvm.instrprof_value_profile(i8* <name>, i64 <hash>,
9527 i64 <value>, i32 <value_kind>,
9533 The '``llvm.instrprof_value_profile``' intrinsic can be emitted by a
9534 frontend for use with instrumentation based profiling. This will be
9535 lowered by the ``-instrprof`` pass to find out the target values,
9536 instrumented expressions take in a program at runtime.
9541 The first argument is a pointer to a global variable containing the
9542 name of the entity being instrumented. ``name`` should generally be the
9543 (mangled) function name for a set of counters.
9545 The second argument is a hash value that can be used by the consumer
9546 of the profile data to detect changes to the instrumented source. It
9547 is an error if ``hash`` differs between two instances of
9548 ``llvm.instrprof_*`` that refer to the same name.
9550 The third argument is the value of the expression being profiled. The profiled
9551 expression's value should be representable as an unsigned 64-bit value. The
9552 fourth argument represents the kind of value profiling that is being done. The
9553 supported value profiling kinds are enumerated through the
9554 ``InstrProfValueKind`` type declared in the
9555 ``<include/llvm/ProfileData/InstrProf.h>`` header file. The last argument is the
9556 index of the instrumented expression within ``name``. It should be >= 0.
9561 This intrinsic represents the point where a call to a runtime routine
9562 should be inserted for value profiling of target expressions. ``-instrprof``
9563 pass will generate the appropriate data structures and replace the
9564 ``llvm.instrprof_value_profile`` intrinsic with the call to the profile
9565 runtime library with proper arguments.
9567 Standard C Library Intrinsics
9568 -----------------------------
9570 LLVM provides intrinsics for a few important standard C library
9571 functions. These intrinsics allow source-language front-ends to pass
9572 information about the alignment of the pointer arguments to the code
9573 generator, providing opportunity for more efficient code generation.
9577 '``llvm.memcpy``' Intrinsic
9578 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9583 This is an overloaded intrinsic. You can use ``llvm.memcpy`` on any
9584 integer bit width and for different address spaces. Not all targets
9585 support all bit widths however.
9589 declare void @llvm.memcpy.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
9590 i32 <len>, i32 <align>, i1 <isvolatile>)
9591 declare void @llvm.memcpy.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
9592 i64 <len>, i32 <align>, i1 <isvolatile>)
9597 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
9598 source location to the destination location.
9600 Note that, unlike the standard libc function, the ``llvm.memcpy.*``
9601 intrinsics do not return a value, takes extra alignment/isvolatile
9602 arguments and the pointers can be in specified address spaces.
9607 The first argument is a pointer to the destination, the second is a
9608 pointer to the source. The third argument is an integer argument
9609 specifying the number of bytes to copy, the fourth argument is the
9610 alignment of the source and destination locations, and the fifth is a
9611 boolean indicating a volatile access.
9613 If the call to this intrinsic has an alignment value that is not 0 or 1,
9614 then the caller guarantees that both the source and destination pointers
9615 are aligned to that boundary.
9617 If the ``isvolatile`` parameter is ``true``, the ``llvm.memcpy`` call is
9618 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
9619 very cleanly specified and it is unwise to depend on it.
9624 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
9625 source location to the destination location, which are not allowed to
9626 overlap. It copies "len" bytes of memory over. If the argument is known
9627 to be aligned to some boundary, this can be specified as the fourth
9628 argument, otherwise it should be set to 0 or 1 (both meaning no alignment).
9630 '``llvm.memmove``' Intrinsic
9631 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9636 This is an overloaded intrinsic. You can use llvm.memmove on any integer
9637 bit width and for different address space. Not all targets support all
9642 declare void @llvm.memmove.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
9643 i32 <len>, i32 <align>, i1 <isvolatile>)
9644 declare void @llvm.memmove.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
9645 i64 <len>, i32 <align>, i1 <isvolatile>)
9650 The '``llvm.memmove.*``' intrinsics move a block of memory from the
9651 source location to the destination location. It is similar to the
9652 '``llvm.memcpy``' intrinsic but allows the two memory locations to
9655 Note that, unlike the standard libc function, the ``llvm.memmove.*``
9656 intrinsics do not return a value, takes extra alignment/isvolatile
9657 arguments and the pointers can be in specified address spaces.
9662 The first argument is a pointer to the destination, the second is a
9663 pointer to the source. The third argument is an integer argument
9664 specifying the number of bytes to copy, the fourth argument is the
9665 alignment of the source and destination locations, and the fifth is a
9666 boolean indicating a volatile access.
9668 If the call to this intrinsic has an alignment value that is not 0 or 1,
9669 then the caller guarantees that the source and destination pointers are
9670 aligned to that boundary.
9672 If the ``isvolatile`` parameter is ``true``, the ``llvm.memmove`` call
9673 is a :ref:`volatile operation <volatile>`. The detailed access behavior is
9674 not very cleanly specified and it is unwise to depend on it.
9679 The '``llvm.memmove.*``' intrinsics copy a block of memory from the
9680 source location to the destination location, which may overlap. It
9681 copies "len" bytes of memory over. If the argument is known to be
9682 aligned to some boundary, this can be specified as the fourth argument,
9683 otherwise it should be set to 0 or 1 (both meaning no alignment).
9685 '``llvm.memset.*``' Intrinsics
9686 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9691 This is an overloaded intrinsic. You can use llvm.memset on any integer
9692 bit width and for different address spaces. However, not all targets
9693 support all bit widths.
9697 declare void @llvm.memset.p0i8.i32(i8* <dest>, i8 <val>,
9698 i32 <len>, i32 <align>, i1 <isvolatile>)
9699 declare void @llvm.memset.p0i8.i64(i8* <dest>, i8 <val>,
9700 i64 <len>, i32 <align>, i1 <isvolatile>)
9705 The '``llvm.memset.*``' intrinsics fill a block of memory with a
9706 particular byte value.
9708 Note that, unlike the standard libc function, the ``llvm.memset``
9709 intrinsic does not return a value and takes extra alignment/volatile
9710 arguments. Also, the destination can be in an arbitrary address space.
9715 The first argument is a pointer to the destination to fill, the second
9716 is the byte value with which to fill it, the third argument is an
9717 integer argument specifying the number of bytes to fill, and the fourth
9718 argument is the known alignment of the destination location.
9720 If the call to this intrinsic has an alignment value that is not 0 or 1,
9721 then the caller guarantees that the destination pointer is aligned to
9724 If the ``isvolatile`` parameter is ``true``, the ``llvm.memset`` call is
9725 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
9726 very cleanly specified and it is unwise to depend on it.
9731 The '``llvm.memset.*``' intrinsics fill "len" bytes of memory starting
9732 at the destination location. If the argument is known to be aligned to
9733 some boundary, this can be specified as the fourth argument, otherwise
9734 it should be set to 0 or 1 (both meaning no alignment).
9736 '``llvm.sqrt.*``' Intrinsic
9737 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9742 This is an overloaded intrinsic. You can use ``llvm.sqrt`` on any
9743 floating point or vector of floating point type. Not all targets support
9748 declare float @llvm.sqrt.f32(float %Val)
9749 declare double @llvm.sqrt.f64(double %Val)
9750 declare x86_fp80 @llvm.sqrt.f80(x86_fp80 %Val)
9751 declare fp128 @llvm.sqrt.f128(fp128 %Val)
9752 declare ppc_fp128 @llvm.sqrt.ppcf128(ppc_fp128 %Val)
9757 The '``llvm.sqrt``' intrinsics return the sqrt of the specified operand,
9758 returning the same value as the libm '``sqrt``' functions would. Unlike
9759 ``sqrt`` in libm, however, ``llvm.sqrt`` has undefined behavior for
9760 negative numbers other than -0.0 (which allows for better optimization,
9761 because there is no need to worry about errno being set).
9762 ``llvm.sqrt(-0.0)`` is defined to return -0.0 like IEEE sqrt.
9767 The argument and return value are floating point numbers of the same
9773 This function returns the sqrt of the specified operand if it is a
9774 nonnegative floating point number.
9776 '``llvm.powi.*``' Intrinsic
9777 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9782 This is an overloaded intrinsic. You can use ``llvm.powi`` on any
9783 floating point or vector of floating point type. Not all targets support
9788 declare float @llvm.powi.f32(float %Val, i32 %power)
9789 declare double @llvm.powi.f64(double %Val, i32 %power)
9790 declare x86_fp80 @llvm.powi.f80(x86_fp80 %Val, i32 %power)
9791 declare fp128 @llvm.powi.f128(fp128 %Val, i32 %power)
9792 declare ppc_fp128 @llvm.powi.ppcf128(ppc_fp128 %Val, i32 %power)
9797 The '``llvm.powi.*``' intrinsics return the first operand raised to the
9798 specified (positive or negative) power. The order of evaluation of
9799 multiplications is not defined. When a vector of floating point type is
9800 used, the second argument remains a scalar integer value.
9805 The second argument is an integer power, and the first is a value to
9806 raise to that power.
9811 This function returns the first value raised to the second power with an
9812 unspecified sequence of rounding operations.
9814 '``llvm.sin.*``' Intrinsic
9815 ^^^^^^^^^^^^^^^^^^^^^^^^^^
9820 This is an overloaded intrinsic. You can use ``llvm.sin`` on any
9821 floating point or vector of floating point type. Not all targets support
9826 declare float @llvm.sin.f32(float %Val)
9827 declare double @llvm.sin.f64(double %Val)
9828 declare x86_fp80 @llvm.sin.f80(x86_fp80 %Val)
9829 declare fp128 @llvm.sin.f128(fp128 %Val)
9830 declare ppc_fp128 @llvm.sin.ppcf128(ppc_fp128 %Val)
9835 The '``llvm.sin.*``' intrinsics return the sine of the operand.
9840 The argument and return value are floating point numbers of the same
9846 This function returns the sine of the specified operand, returning the
9847 same values as the libm ``sin`` functions would, and handles error
9848 conditions in the same way.
9850 '``llvm.cos.*``' Intrinsic
9851 ^^^^^^^^^^^^^^^^^^^^^^^^^^
9856 This is an overloaded intrinsic. You can use ``llvm.cos`` on any
9857 floating point or vector of floating point type. Not all targets support
9862 declare float @llvm.cos.f32(float %Val)
9863 declare double @llvm.cos.f64(double %Val)
9864 declare x86_fp80 @llvm.cos.f80(x86_fp80 %Val)
9865 declare fp128 @llvm.cos.f128(fp128 %Val)
9866 declare ppc_fp128 @llvm.cos.ppcf128(ppc_fp128 %Val)
9871 The '``llvm.cos.*``' intrinsics return the cosine of the operand.
9876 The argument and return value are floating point numbers of the same
9882 This function returns the cosine of the specified operand, returning the
9883 same values as the libm ``cos`` functions would, and handles error
9884 conditions in the same way.
9886 '``llvm.pow.*``' Intrinsic
9887 ^^^^^^^^^^^^^^^^^^^^^^^^^^
9892 This is an overloaded intrinsic. You can use ``llvm.pow`` on any
9893 floating point or vector of floating point type. Not all targets support
9898 declare float @llvm.pow.f32(float %Val, float %Power)
9899 declare double @llvm.pow.f64(double %Val, double %Power)
9900 declare x86_fp80 @llvm.pow.f80(x86_fp80 %Val, x86_fp80 %Power)
9901 declare fp128 @llvm.pow.f128(fp128 %Val, fp128 %Power)
9902 declare ppc_fp128 @llvm.pow.ppcf128(ppc_fp128 %Val, ppc_fp128 Power)
9907 The '``llvm.pow.*``' intrinsics return the first operand raised to the
9908 specified (positive or negative) power.
9913 The second argument is a floating point power, and the first is a value
9914 to raise to that power.
9919 This function returns the first value raised to the second power,
9920 returning the same values as the libm ``pow`` functions would, and
9921 handles error conditions in the same way.
9923 '``llvm.exp.*``' Intrinsic
9924 ^^^^^^^^^^^^^^^^^^^^^^^^^^
9929 This is an overloaded intrinsic. You can use ``llvm.exp`` on any
9930 floating point or vector of floating point type. Not all targets support
9935 declare float @llvm.exp.f32(float %Val)
9936 declare double @llvm.exp.f64(double %Val)
9937 declare x86_fp80 @llvm.exp.f80(x86_fp80 %Val)
9938 declare fp128 @llvm.exp.f128(fp128 %Val)
9939 declare ppc_fp128 @llvm.exp.ppcf128(ppc_fp128 %Val)
9944 The '``llvm.exp.*``' intrinsics perform the exp function.
9949 The argument and return value are floating point numbers of the same
9955 This function returns the same values as the libm ``exp`` functions
9956 would, and handles error conditions in the same way.
9958 '``llvm.exp2.*``' Intrinsic
9959 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9964 This is an overloaded intrinsic. You can use ``llvm.exp2`` on any
9965 floating point or vector of floating point type. Not all targets support
9970 declare float @llvm.exp2.f32(float %Val)
9971 declare double @llvm.exp2.f64(double %Val)
9972 declare x86_fp80 @llvm.exp2.f80(x86_fp80 %Val)
9973 declare fp128 @llvm.exp2.f128(fp128 %Val)
9974 declare ppc_fp128 @llvm.exp2.ppcf128(ppc_fp128 %Val)
9979 The '``llvm.exp2.*``' intrinsics perform the exp2 function.
9984 The argument and return value are floating point numbers of the same
9990 This function returns the same values as the libm ``exp2`` functions
9991 would, and handles error conditions in the same way.
9993 '``llvm.log.*``' Intrinsic
9994 ^^^^^^^^^^^^^^^^^^^^^^^^^^
9999 This is an overloaded intrinsic. You can use ``llvm.log`` on any
10000 floating point or vector of floating point type. Not all targets support
10005 declare float @llvm.log.f32(float %Val)
10006 declare double @llvm.log.f64(double %Val)
10007 declare x86_fp80 @llvm.log.f80(x86_fp80 %Val)
10008 declare fp128 @llvm.log.f128(fp128 %Val)
10009 declare ppc_fp128 @llvm.log.ppcf128(ppc_fp128 %Val)
10014 The '``llvm.log.*``' intrinsics perform the log function.
10019 The argument and return value are floating point numbers of the same
10025 This function returns the same values as the libm ``log`` functions
10026 would, and handles error conditions in the same way.
10028 '``llvm.log10.*``' Intrinsic
10029 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10034 This is an overloaded intrinsic. You can use ``llvm.log10`` on any
10035 floating point or vector of floating point type. Not all targets support
10040 declare float @llvm.log10.f32(float %Val)
10041 declare double @llvm.log10.f64(double %Val)
10042 declare x86_fp80 @llvm.log10.f80(x86_fp80 %Val)
10043 declare fp128 @llvm.log10.f128(fp128 %Val)
10044 declare ppc_fp128 @llvm.log10.ppcf128(ppc_fp128 %Val)
10049 The '``llvm.log10.*``' intrinsics perform the log10 function.
10054 The argument and return value are floating point numbers of the same
10060 This function returns the same values as the libm ``log10`` functions
10061 would, and handles error conditions in the same way.
10063 '``llvm.log2.*``' Intrinsic
10064 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10069 This is an overloaded intrinsic. You can use ``llvm.log2`` on any
10070 floating point or vector of floating point type. Not all targets support
10075 declare float @llvm.log2.f32(float %Val)
10076 declare double @llvm.log2.f64(double %Val)
10077 declare x86_fp80 @llvm.log2.f80(x86_fp80 %Val)
10078 declare fp128 @llvm.log2.f128(fp128 %Val)
10079 declare ppc_fp128 @llvm.log2.ppcf128(ppc_fp128 %Val)
10084 The '``llvm.log2.*``' intrinsics perform the log2 function.
10089 The argument and return value are floating point numbers of the same
10095 This function returns the same values as the libm ``log2`` functions
10096 would, and handles error conditions in the same way.
10098 '``llvm.fma.*``' Intrinsic
10099 ^^^^^^^^^^^^^^^^^^^^^^^^^^
10104 This is an overloaded intrinsic. You can use ``llvm.fma`` on any
10105 floating point or vector of floating point type. Not all targets support
10110 declare float @llvm.fma.f32(float %a, float %b, float %c)
10111 declare double @llvm.fma.f64(double %a, double %b, double %c)
10112 declare x86_fp80 @llvm.fma.f80(x86_fp80 %a, x86_fp80 %b, x86_fp80 %c)
10113 declare fp128 @llvm.fma.f128(fp128 %a, fp128 %b, fp128 %c)
10114 declare ppc_fp128 @llvm.fma.ppcf128(ppc_fp128 %a, ppc_fp128 %b, ppc_fp128 %c)
10119 The '``llvm.fma.*``' intrinsics perform the fused multiply-add
10125 The argument and return value are floating point numbers of the same
10131 This function returns the same values as the libm ``fma`` functions
10132 would, and does not set errno.
10134 '``llvm.fabs.*``' Intrinsic
10135 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10140 This is an overloaded intrinsic. You can use ``llvm.fabs`` on any
10141 floating point or vector of floating point type. Not all targets support
10146 declare float @llvm.fabs.f32(float %Val)
10147 declare double @llvm.fabs.f64(double %Val)
10148 declare x86_fp80 @llvm.fabs.f80(x86_fp80 %Val)
10149 declare fp128 @llvm.fabs.f128(fp128 %Val)
10150 declare ppc_fp128 @llvm.fabs.ppcf128(ppc_fp128 %Val)
10155 The '``llvm.fabs.*``' intrinsics return the absolute value of the
10161 The argument and return value are floating point numbers of the same
10167 This function returns the same values as the libm ``fabs`` functions
10168 would, and handles error conditions in the same way.
10170 '``llvm.minnum.*``' Intrinsic
10171 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10176 This is an overloaded intrinsic. You can use ``llvm.minnum`` on any
10177 floating point or vector of floating point type. Not all targets support
10182 declare float @llvm.minnum.f32(float %Val0, float %Val1)
10183 declare double @llvm.minnum.f64(double %Val0, double %Val1)
10184 declare x86_fp80 @llvm.minnum.f80(x86_fp80 %Val0, x86_fp80 %Val1)
10185 declare fp128 @llvm.minnum.f128(fp128 %Val0, fp128 %Val1)
10186 declare ppc_fp128 @llvm.minnum.ppcf128(ppc_fp128 %Val0, ppc_fp128 %Val1)
10191 The '``llvm.minnum.*``' intrinsics return the minimum of the two
10198 The arguments and return value are floating point numbers of the same
10204 Follows the IEEE-754 semantics for minNum, which also match for libm's
10207 If either operand is a NaN, returns the other non-NaN operand. Returns
10208 NaN only if both operands are NaN. If the operands compare equal,
10209 returns a value that compares equal to both operands. This means that
10210 fmin(+/-0.0, +/-0.0) could return either -0.0 or 0.0.
10212 '``llvm.maxnum.*``' Intrinsic
10213 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10218 This is an overloaded intrinsic. You can use ``llvm.maxnum`` on any
10219 floating point or vector of floating point type. Not all targets support
10224 declare float @llvm.maxnum.f32(float %Val0, float %Val1l)
10225 declare double @llvm.maxnum.f64(double %Val0, double %Val1)
10226 declare x86_fp80 @llvm.maxnum.f80(x86_fp80 %Val0, x86_fp80 %Val1)
10227 declare fp128 @llvm.maxnum.f128(fp128 %Val0, fp128 %Val1)
10228 declare ppc_fp128 @llvm.maxnum.ppcf128(ppc_fp128 %Val0, ppc_fp128 %Val1)
10233 The '``llvm.maxnum.*``' intrinsics return the maximum of the two
10240 The arguments and return value are floating point numbers of the same
10245 Follows the IEEE-754 semantics for maxNum, which also match for libm's
10248 If either operand is a NaN, returns the other non-NaN operand. Returns
10249 NaN only if both operands are NaN. If the operands compare equal,
10250 returns a value that compares equal to both operands. This means that
10251 fmax(+/-0.0, +/-0.0) could return either -0.0 or 0.0.
10253 '``llvm.copysign.*``' Intrinsic
10254 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10259 This is an overloaded intrinsic. You can use ``llvm.copysign`` on any
10260 floating point or vector of floating point type. Not all targets support
10265 declare float @llvm.copysign.f32(float %Mag, float %Sgn)
10266 declare double @llvm.copysign.f64(double %Mag, double %Sgn)
10267 declare x86_fp80 @llvm.copysign.f80(x86_fp80 %Mag, x86_fp80 %Sgn)
10268 declare fp128 @llvm.copysign.f128(fp128 %Mag, fp128 %Sgn)
10269 declare ppc_fp128 @llvm.copysign.ppcf128(ppc_fp128 %Mag, ppc_fp128 %Sgn)
10274 The '``llvm.copysign.*``' intrinsics return a value with the magnitude of the
10275 first operand and the sign of the second operand.
10280 The arguments and return value are floating point numbers of the same
10286 This function returns the same values as the libm ``copysign``
10287 functions would, and handles error conditions in the same way.
10289 '``llvm.floor.*``' Intrinsic
10290 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10295 This is an overloaded intrinsic. You can use ``llvm.floor`` on any
10296 floating point or vector of floating point type. Not all targets support
10301 declare float @llvm.floor.f32(float %Val)
10302 declare double @llvm.floor.f64(double %Val)
10303 declare x86_fp80 @llvm.floor.f80(x86_fp80 %Val)
10304 declare fp128 @llvm.floor.f128(fp128 %Val)
10305 declare ppc_fp128 @llvm.floor.ppcf128(ppc_fp128 %Val)
10310 The '``llvm.floor.*``' intrinsics return the floor of the operand.
10315 The argument and return value are floating point numbers of the same
10321 This function returns the same values as the libm ``floor`` functions
10322 would, and handles error conditions in the same way.
10324 '``llvm.ceil.*``' Intrinsic
10325 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10330 This is an overloaded intrinsic. You can use ``llvm.ceil`` on any
10331 floating point or vector of floating point type. Not all targets support
10336 declare float @llvm.ceil.f32(float %Val)
10337 declare double @llvm.ceil.f64(double %Val)
10338 declare x86_fp80 @llvm.ceil.f80(x86_fp80 %Val)
10339 declare fp128 @llvm.ceil.f128(fp128 %Val)
10340 declare ppc_fp128 @llvm.ceil.ppcf128(ppc_fp128 %Val)
10345 The '``llvm.ceil.*``' intrinsics return the ceiling of the operand.
10350 The argument and return value are floating point numbers of the same
10356 This function returns the same values as the libm ``ceil`` functions
10357 would, and handles error conditions in the same way.
10359 '``llvm.trunc.*``' Intrinsic
10360 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10365 This is an overloaded intrinsic. You can use ``llvm.trunc`` on any
10366 floating point or vector of floating point type. Not all targets support
10371 declare float @llvm.trunc.f32(float %Val)
10372 declare double @llvm.trunc.f64(double %Val)
10373 declare x86_fp80 @llvm.trunc.f80(x86_fp80 %Val)
10374 declare fp128 @llvm.trunc.f128(fp128 %Val)
10375 declare ppc_fp128 @llvm.trunc.ppcf128(ppc_fp128 %Val)
10380 The '``llvm.trunc.*``' intrinsics returns the operand rounded to the
10381 nearest integer not larger in magnitude than the operand.
10386 The argument and return value are floating point numbers of the same
10392 This function returns the same values as the libm ``trunc`` functions
10393 would, and handles error conditions in the same way.
10395 '``llvm.rint.*``' Intrinsic
10396 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10401 This is an overloaded intrinsic. You can use ``llvm.rint`` on any
10402 floating point or vector of floating point type. Not all targets support
10407 declare float @llvm.rint.f32(float %Val)
10408 declare double @llvm.rint.f64(double %Val)
10409 declare x86_fp80 @llvm.rint.f80(x86_fp80 %Val)
10410 declare fp128 @llvm.rint.f128(fp128 %Val)
10411 declare ppc_fp128 @llvm.rint.ppcf128(ppc_fp128 %Val)
10416 The '``llvm.rint.*``' intrinsics returns the operand rounded to the
10417 nearest integer. It may raise an inexact floating-point exception if the
10418 operand isn't an integer.
10423 The argument and return value are floating point numbers of the same
10429 This function returns the same values as the libm ``rint`` functions
10430 would, and handles error conditions in the same way.
10432 '``llvm.nearbyint.*``' Intrinsic
10433 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10438 This is an overloaded intrinsic. You can use ``llvm.nearbyint`` on any
10439 floating point or vector of floating point type. Not all targets support
10444 declare float @llvm.nearbyint.f32(float %Val)
10445 declare double @llvm.nearbyint.f64(double %Val)
10446 declare x86_fp80 @llvm.nearbyint.f80(x86_fp80 %Val)
10447 declare fp128 @llvm.nearbyint.f128(fp128 %Val)
10448 declare ppc_fp128 @llvm.nearbyint.ppcf128(ppc_fp128 %Val)
10453 The '``llvm.nearbyint.*``' intrinsics returns the operand rounded to the
10459 The argument and return value are floating point numbers of the same
10465 This function returns the same values as the libm ``nearbyint``
10466 functions would, and handles error conditions in the same way.
10468 '``llvm.round.*``' Intrinsic
10469 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10474 This is an overloaded intrinsic. You can use ``llvm.round`` on any
10475 floating point or vector of floating point type. Not all targets support
10480 declare float @llvm.round.f32(float %Val)
10481 declare double @llvm.round.f64(double %Val)
10482 declare x86_fp80 @llvm.round.f80(x86_fp80 %Val)
10483 declare fp128 @llvm.round.f128(fp128 %Val)
10484 declare ppc_fp128 @llvm.round.ppcf128(ppc_fp128 %Val)
10489 The '``llvm.round.*``' intrinsics returns the operand rounded to the
10495 The argument and return value are floating point numbers of the same
10501 This function returns the same values as the libm ``round``
10502 functions would, and handles error conditions in the same way.
10504 Bit Manipulation Intrinsics
10505 ---------------------------
10507 LLVM provides intrinsics for a few important bit manipulation
10508 operations. These allow efficient code generation for some algorithms.
10510 '``llvm.bitreverse.*``' Intrinsics
10511 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10516 This is an overloaded intrinsic function. You can use bitreverse on any
10521 declare i16 @llvm.bitreverse.i16(i16 <id>)
10522 declare i32 @llvm.bitreverse.i32(i32 <id>)
10523 declare i64 @llvm.bitreverse.i64(i64 <id>)
10528 The '``llvm.bitreverse``' family of intrinsics is used to reverse the
10529 bitpattern of an integer value; for example ``0b1234567`` becomes
10535 The ``llvm.bitreverse.iN`` intrinsic returns an i16 value that has bit
10536 ``M`` in the input moved to bit ``N-M`` in the output.
10538 '``llvm.bswap.*``' Intrinsics
10539 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10544 This is an overloaded intrinsic function. You can use bswap on any
10545 integer type that is an even number of bytes (i.e. BitWidth % 16 == 0).
10549 declare i16 @llvm.bswap.i16(i16 <id>)
10550 declare i32 @llvm.bswap.i32(i32 <id>)
10551 declare i64 @llvm.bswap.i64(i64 <id>)
10556 The '``llvm.bswap``' family of intrinsics is used to byte swap integer
10557 values with an even number of bytes (positive multiple of 16 bits).
10558 These are useful for performing operations on data that is not in the
10559 target's native byte order.
10564 The ``llvm.bswap.i16`` intrinsic returns an i16 value that has the high
10565 and low byte of the input i16 swapped. Similarly, the ``llvm.bswap.i32``
10566 intrinsic returns an i32 value that has the four bytes of the input i32
10567 swapped, so that if the input bytes are numbered 0, 1, 2, 3 then the
10568 returned i32 will have its bytes in 3, 2, 1, 0 order. The
10569 ``llvm.bswap.i48``, ``llvm.bswap.i64`` and other intrinsics extend this
10570 concept to additional even-byte lengths (6 bytes, 8 bytes and more,
10573 '``llvm.ctpop.*``' Intrinsic
10574 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10579 This is an overloaded intrinsic. You can use llvm.ctpop on any integer
10580 bit width, or on any vector with integer elements. Not all targets
10581 support all bit widths or vector types, however.
10585 declare i8 @llvm.ctpop.i8(i8 <src>)
10586 declare i16 @llvm.ctpop.i16(i16 <src>)
10587 declare i32 @llvm.ctpop.i32(i32 <src>)
10588 declare i64 @llvm.ctpop.i64(i64 <src>)
10589 declare i256 @llvm.ctpop.i256(i256 <src>)
10590 declare <2 x i32> @llvm.ctpop.v2i32(<2 x i32> <src>)
10595 The '``llvm.ctpop``' family of intrinsics counts the number of bits set
10601 The only argument is the value to be counted. The argument may be of any
10602 integer type, or a vector with integer elements. The return type must
10603 match the argument type.
10608 The '``llvm.ctpop``' intrinsic counts the 1's in a variable, or within
10609 each element of a vector.
10611 '``llvm.ctlz.*``' Intrinsic
10612 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10617 This is an overloaded intrinsic. You can use ``llvm.ctlz`` on any
10618 integer bit width, or any vector whose elements are integers. Not all
10619 targets support all bit widths or vector types, however.
10623 declare i8 @llvm.ctlz.i8 (i8 <src>, i1 <is_zero_undef>)
10624 declare i16 @llvm.ctlz.i16 (i16 <src>, i1 <is_zero_undef>)
10625 declare i32 @llvm.ctlz.i32 (i32 <src>, i1 <is_zero_undef>)
10626 declare i64 @llvm.ctlz.i64 (i64 <src>, i1 <is_zero_undef>)
10627 declare i256 @llvm.ctlz.i256(i256 <src>, i1 <is_zero_undef>)
10628 declase <2 x i32> @llvm.ctlz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
10633 The '``llvm.ctlz``' family of intrinsic functions counts the number of
10634 leading zeros in a variable.
10639 The first argument is the value to be counted. This argument may be of
10640 any integer type, or a vector with integer element type. The return
10641 type must match the first argument type.
10643 The second argument must be a constant and is a flag to indicate whether
10644 the intrinsic should ensure that a zero as the first argument produces a
10645 defined result. Historically some architectures did not provide a
10646 defined result for zero values as efficiently, and many algorithms are
10647 now predicated on avoiding zero-value inputs.
10652 The '``llvm.ctlz``' intrinsic counts the leading (most significant)
10653 zeros in a variable, or within each element of the vector. If
10654 ``src == 0`` then the result is the size in bits of the type of ``src``
10655 if ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
10656 ``llvm.ctlz(i32 2) = 30``.
10658 '``llvm.cttz.*``' Intrinsic
10659 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10664 This is an overloaded intrinsic. You can use ``llvm.cttz`` on any
10665 integer bit width, or any vector of integer elements. Not all targets
10666 support all bit widths or vector types, however.
10670 declare i8 @llvm.cttz.i8 (i8 <src>, i1 <is_zero_undef>)
10671 declare i16 @llvm.cttz.i16 (i16 <src>, i1 <is_zero_undef>)
10672 declare i32 @llvm.cttz.i32 (i32 <src>, i1 <is_zero_undef>)
10673 declare i64 @llvm.cttz.i64 (i64 <src>, i1 <is_zero_undef>)
10674 declare i256 @llvm.cttz.i256(i256 <src>, i1 <is_zero_undef>)
10675 declase <2 x i32> @llvm.cttz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
10680 The '``llvm.cttz``' family of intrinsic functions counts the number of
10686 The first argument is the value to be counted. This argument may be of
10687 any integer type, or a vector with integer element type. The return
10688 type must match the first argument type.
10690 The second argument must be a constant and is a flag to indicate whether
10691 the intrinsic should ensure that a zero as the first argument produces a
10692 defined result. Historically some architectures did not provide a
10693 defined result for zero values as efficiently, and many algorithms are
10694 now predicated on avoiding zero-value inputs.
10699 The '``llvm.cttz``' intrinsic counts the trailing (least significant)
10700 zeros in a variable, or within each element of a vector. If ``src == 0``
10701 then the result is the size in bits of the type of ``src`` if
10702 ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
10703 ``llvm.cttz(2) = 1``.
10707 Arithmetic with Overflow Intrinsics
10708 -----------------------------------
10710 LLVM provides intrinsics for some arithmetic with overflow operations.
10712 '``llvm.sadd.with.overflow.*``' Intrinsics
10713 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10718 This is an overloaded intrinsic. You can use ``llvm.sadd.with.overflow``
10719 on any integer bit width.
10723 declare {i16, i1} @llvm.sadd.with.overflow.i16(i16 %a, i16 %b)
10724 declare {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
10725 declare {i64, i1} @llvm.sadd.with.overflow.i64(i64 %a, i64 %b)
10730 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
10731 a signed addition of the two arguments, and indicate whether an overflow
10732 occurred during the signed summation.
10737 The arguments (%a and %b) and the first element of the result structure
10738 may be of integer types of any bit width, but they must have the same
10739 bit width. The second element of the result structure must be of type
10740 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
10746 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
10747 a signed addition of the two variables. They return a structure --- the
10748 first element of which is the signed summation, and the second element
10749 of which is a bit specifying if the signed summation resulted in an
10755 .. code-block:: llvm
10757 %res = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
10758 %sum = extractvalue {i32, i1} %res, 0
10759 %obit = extractvalue {i32, i1} %res, 1
10760 br i1 %obit, label %overflow, label %normal
10762 '``llvm.uadd.with.overflow.*``' Intrinsics
10763 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10768 This is an overloaded intrinsic. You can use ``llvm.uadd.with.overflow``
10769 on any integer bit width.
10773 declare {i16, i1} @llvm.uadd.with.overflow.i16(i16 %a, i16 %b)
10774 declare {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
10775 declare {i64, i1} @llvm.uadd.with.overflow.i64(i64 %a, i64 %b)
10780 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
10781 an unsigned addition of the two arguments, and indicate whether a carry
10782 occurred during the unsigned summation.
10787 The arguments (%a and %b) and the first element of the result structure
10788 may be of integer types of any bit width, but they must have the same
10789 bit width. The second element of the result structure must be of type
10790 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
10796 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
10797 an unsigned addition of the two arguments. They return a structure --- the
10798 first element of which is the sum, and the second element of which is a
10799 bit specifying if the unsigned summation resulted in a carry.
10804 .. code-block:: llvm
10806 %res = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
10807 %sum = extractvalue {i32, i1} %res, 0
10808 %obit = extractvalue {i32, i1} %res, 1
10809 br i1 %obit, label %carry, label %normal
10811 '``llvm.ssub.with.overflow.*``' Intrinsics
10812 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10817 This is an overloaded intrinsic. You can use ``llvm.ssub.with.overflow``
10818 on any integer bit width.
10822 declare {i16, i1} @llvm.ssub.with.overflow.i16(i16 %a, i16 %b)
10823 declare {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
10824 declare {i64, i1} @llvm.ssub.with.overflow.i64(i64 %a, i64 %b)
10829 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
10830 a signed subtraction of the two arguments, and indicate whether an
10831 overflow occurred during the signed subtraction.
10836 The arguments (%a and %b) and the first element of the result structure
10837 may be of integer types of any bit width, but they must have the same
10838 bit width. The second element of the result structure must be of type
10839 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
10845 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
10846 a signed subtraction of the two arguments. They return a structure --- the
10847 first element of which is the subtraction, and the second element of
10848 which is a bit specifying if the signed subtraction resulted in an
10854 .. code-block:: llvm
10856 %res = call {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
10857 %sum = extractvalue {i32, i1} %res, 0
10858 %obit = extractvalue {i32, i1} %res, 1
10859 br i1 %obit, label %overflow, label %normal
10861 '``llvm.usub.with.overflow.*``' Intrinsics
10862 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10867 This is an overloaded intrinsic. You can use ``llvm.usub.with.overflow``
10868 on any integer bit width.
10872 declare {i16, i1} @llvm.usub.with.overflow.i16(i16 %a, i16 %b)
10873 declare {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
10874 declare {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
10879 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
10880 an unsigned subtraction of the two arguments, and indicate whether an
10881 overflow occurred during the unsigned subtraction.
10886 The arguments (%a and %b) and the first element of the result structure
10887 may be of integer types of any bit width, but they must have the same
10888 bit width. The second element of the result structure must be of type
10889 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
10895 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
10896 an unsigned subtraction of the two arguments. They return a structure ---
10897 the first element of which is the subtraction, and the second element of
10898 which is a bit specifying if the unsigned subtraction resulted in an
10904 .. code-block:: llvm
10906 %res = call {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
10907 %sum = extractvalue {i32, i1} %res, 0
10908 %obit = extractvalue {i32, i1} %res, 1
10909 br i1 %obit, label %overflow, label %normal
10911 '``llvm.smul.with.overflow.*``' Intrinsics
10912 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10917 This is an overloaded intrinsic. You can use ``llvm.smul.with.overflow``
10918 on any integer bit width.
10922 declare {i16, i1} @llvm.smul.with.overflow.i16(i16 %a, i16 %b)
10923 declare {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
10924 declare {i64, i1} @llvm.smul.with.overflow.i64(i64 %a, i64 %b)
10929 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
10930 a signed multiplication of the two arguments, and indicate whether an
10931 overflow occurred during the signed multiplication.
10936 The arguments (%a and %b) and the first element of the result structure
10937 may be of integer types of any bit width, but they must have the same
10938 bit width. The second element of the result structure must be of type
10939 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
10945 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
10946 a signed multiplication of the two arguments. They return a structure ---
10947 the first element of which is the multiplication, and the second element
10948 of which is a bit specifying if the signed multiplication resulted in an
10954 .. code-block:: llvm
10956 %res = call {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
10957 %sum = extractvalue {i32, i1} %res, 0
10958 %obit = extractvalue {i32, i1} %res, 1
10959 br i1 %obit, label %overflow, label %normal
10961 '``llvm.umul.with.overflow.*``' Intrinsics
10962 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10967 This is an overloaded intrinsic. You can use ``llvm.umul.with.overflow``
10968 on any integer bit width.
10972 declare {i16, i1} @llvm.umul.with.overflow.i16(i16 %a, i16 %b)
10973 declare {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
10974 declare {i64, i1} @llvm.umul.with.overflow.i64(i64 %a, i64 %b)
10979 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
10980 a unsigned multiplication of the two arguments, and indicate whether an
10981 overflow occurred during the unsigned multiplication.
10986 The arguments (%a and %b) and the first element of the result structure
10987 may be of integer types of any bit width, but they must have the same
10988 bit width. The second element of the result structure must be of type
10989 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
10995 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
10996 an unsigned multiplication of the two arguments. They return a structure ---
10997 the first element of which is the multiplication, and the second
10998 element of which is a bit specifying if the unsigned multiplication
10999 resulted in an overflow.
11004 .. code-block:: llvm
11006 %res = call {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
11007 %sum = extractvalue {i32, i1} %res, 0
11008 %obit = extractvalue {i32, i1} %res, 1
11009 br i1 %obit, label %overflow, label %normal
11011 Specialised Arithmetic Intrinsics
11012 ---------------------------------
11014 '``llvm.canonicalize.*``' Intrinsic
11015 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11022 declare float @llvm.canonicalize.f32(float %a)
11023 declare double @llvm.canonicalize.f64(double %b)
11028 The '``llvm.canonicalize.*``' intrinsic returns the platform specific canonical
11029 encoding of a floating point number. This canonicalization is useful for
11030 implementing certain numeric primitives such as frexp. The canonical encoding is
11031 defined by IEEE-754-2008 to be:
11035 2.1.8 canonical encoding: The preferred encoding of a floating-point
11036 representation in a format. Applied to declets, significands of finite
11037 numbers, infinities, and NaNs, especially in decimal formats.
11039 This operation can also be considered equivalent to the IEEE-754-2008
11040 conversion of a floating-point value to the same format. NaNs are handled
11041 according to section 6.2.
11043 Examples of non-canonical encodings:
11045 - x87 pseudo denormals, pseudo NaNs, pseudo Infinity, Unnormals. These are
11046 converted to a canonical representation per hardware-specific protocol.
11047 - Many normal decimal floating point numbers have non-canonical alternative
11049 - Some machines, like GPUs or ARMv7 NEON, do not support subnormal values.
11050 These are treated as non-canonical encodings of zero and with be flushed to
11051 a zero of the same sign by this operation.
11053 Note that per IEEE-754-2008 6.2, systems that support signaling NaNs with
11054 default exception handling must signal an invalid exception, and produce a
11057 This function should always be implementable as multiplication by 1.0, provided
11058 that the compiler does not constant fold the operation. Likewise, division by
11059 1.0 and ``llvm.minnum(x, x)`` are possible implementations. Addition with
11060 -0.0 is also sufficient provided that the rounding mode is not -Infinity.
11062 ``@llvm.canonicalize`` must preserve the equality relation. That is:
11064 - ``(@llvm.canonicalize(x) == x)`` is equivalent to ``(x == x)``
11065 - ``(@llvm.canonicalize(x) == @llvm.canonicalize(y))`` is equivalent to
11068 Additionally, the sign of zero must be conserved:
11069 ``@llvm.canonicalize(-0.0) = -0.0`` and ``@llvm.canonicalize(+0.0) = +0.0``
11071 The payload bits of a NaN must be conserved, with two exceptions.
11072 First, environments which use only a single canonical representation of NaN
11073 must perform said canonicalization. Second, SNaNs must be quieted per the
11076 The canonicalization operation may be optimized away if:
11078 - The input is known to be canonical. For example, it was produced by a
11079 floating-point operation that is required by the standard to be canonical.
11080 - The result is consumed only by (or fused with) other floating-point
11081 operations. That is, the bits of the floating point value are not examined.
11083 '``llvm.fmuladd.*``' Intrinsic
11084 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11091 declare float @llvm.fmuladd.f32(float %a, float %b, float %c)
11092 declare double @llvm.fmuladd.f64(double %a, double %b, double %c)
11097 The '``llvm.fmuladd.*``' intrinsic functions represent multiply-add
11098 expressions that can be fused if the code generator determines that (a) the
11099 target instruction set has support for a fused operation, and (b) that the
11100 fused operation is more efficient than the equivalent, separate pair of mul
11101 and add instructions.
11106 The '``llvm.fmuladd.*``' intrinsics each take three arguments: two
11107 multiplicands, a and b, and an addend c.
11116 %0 = call float @llvm.fmuladd.f32(%a, %b, %c)
11118 is equivalent to the expression a \* b + c, except that rounding will
11119 not be performed between the multiplication and addition steps if the
11120 code generator fuses the operations. Fusion is not guaranteed, even if
11121 the target platform supports it. If a fused multiply-add is required the
11122 corresponding llvm.fma.\* intrinsic function should be used
11123 instead. This never sets errno, just as '``llvm.fma.*``'.
11128 .. code-block:: llvm
11130 %r2 = call float @llvm.fmuladd.f32(float %a, float %b, float %c) ; yields float:r2 = (a * b) + c
11133 '``llvm.uabsdiff.*``' and '``llvm.sabsdiff.*``' Intrinsics
11134 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11138 This is an overloaded intrinsic. The loaded data is a vector of any integer bit width.
11140 .. code-block:: llvm
11142 declare <4 x integer> @llvm.uabsdiff.v4i32(<4 x integer> %a, <4 x integer> %b)
11148 The ``llvm.uabsdiff`` intrinsic returns a vector result of the absolute difference
11149 of the two operands, treating them both as unsigned integers. The intermediate
11150 calculations are computed using infinitely precise unsigned arithmetic. The final
11151 result will be truncated to the given type.
11153 The ``llvm.sabsdiff`` intrinsic returns a vector result of the absolute difference of
11154 the two operands, treating them both as signed integers. If the result overflows, the
11155 behavior is undefined.
11159 These intrinsics are primarily used during the code generation stage of compilation.
11160 They are generated by compiler passes such as the Loop and SLP vectorizers. It is not
11161 recommended for users to create them manually.
11166 Both intrinsics take two integer of the same bitwidth.
11173 call <4 x i32> @llvm.uabsdiff.v4i32(<4 x i32> %a, <4 x i32> %b)
11177 %1 = zext <4 x i32> %a to <4 x i64>
11178 %2 = zext <4 x i32> %b to <4 x i64>
11179 %sub = sub <4 x i64> %1, %2
11180 %trunc = trunc <4 x i64> to <4 x i32>
11182 and the expression::
11184 call <4 x i32> @llvm.sabsdiff.v4i32(<4 x i32> %a, <4 x i32> %b)
11188 %sub = sub nsw <4 x i32> %a, %b
11189 %ispos = icmp sge <4 x i32> %sub, zeroinitializer
11190 %neg = sub nsw <4 x i32> zeroinitializer, %sub
11191 %1 = select <4 x i1> %ispos, <4 x i32> %sub, <4 x i32> %neg
11194 Half Precision Floating Point Intrinsics
11195 ----------------------------------------
11197 For most target platforms, half precision floating point is a
11198 storage-only format. This means that it is a dense encoding (in memory)
11199 but does not support computation in the format.
11201 This means that code must first load the half-precision floating point
11202 value as an i16, then convert it to float with
11203 :ref:`llvm.convert.from.fp16 <int_convert_from_fp16>`. Computation can
11204 then be performed on the float value (including extending to double
11205 etc). To store the value back to memory, it is first converted to float
11206 if needed, then converted to i16 with
11207 :ref:`llvm.convert.to.fp16 <int_convert_to_fp16>`, then storing as an
11210 .. _int_convert_to_fp16:
11212 '``llvm.convert.to.fp16``' Intrinsic
11213 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11220 declare i16 @llvm.convert.to.fp16.f32(float %a)
11221 declare i16 @llvm.convert.to.fp16.f64(double %a)
11226 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion from a
11227 conventional floating point type to half precision floating point format.
11232 The intrinsic function contains single argument - the value to be
11238 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion from a
11239 conventional floating point format to half precision floating point format. The
11240 return value is an ``i16`` which contains the converted number.
11245 .. code-block:: llvm
11247 %res = call i16 @llvm.convert.to.fp16.f32(float %a)
11248 store i16 %res, i16* @x, align 2
11250 .. _int_convert_from_fp16:
11252 '``llvm.convert.from.fp16``' Intrinsic
11253 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11260 declare float @llvm.convert.from.fp16.f32(i16 %a)
11261 declare double @llvm.convert.from.fp16.f64(i16 %a)
11266 The '``llvm.convert.from.fp16``' intrinsic function performs a
11267 conversion from half precision floating point format to single precision
11268 floating point format.
11273 The intrinsic function contains single argument - the value to be
11279 The '``llvm.convert.from.fp16``' intrinsic function performs a
11280 conversion from half single precision floating point format to single
11281 precision floating point format. The input half-float value is
11282 represented by an ``i16`` value.
11287 .. code-block:: llvm
11289 %a = load i16, i16* @x, align 2
11290 %res = call float @llvm.convert.from.fp16(i16 %a)
11292 .. _dbg_intrinsics:
11294 Debugger Intrinsics
11295 -------------------
11297 The LLVM debugger intrinsics (which all start with ``llvm.dbg.``
11298 prefix), are described in the `LLVM Source Level
11299 Debugging <SourceLevelDebugging.html#format_common_intrinsics>`_
11302 Exception Handling Intrinsics
11303 -----------------------------
11305 The LLVM exception handling intrinsics (which all start with
11306 ``llvm.eh.`` prefix), are described in the `LLVM Exception
11307 Handling <ExceptionHandling.html#format_common_intrinsics>`_ document.
11309 .. _int_trampoline:
11311 Trampoline Intrinsics
11312 ---------------------
11314 These intrinsics make it possible to excise one parameter, marked with
11315 the :ref:`nest <nest>` attribute, from a function. The result is a
11316 callable function pointer lacking the nest parameter - the caller does
11317 not need to provide a value for it. Instead, the value to use is stored
11318 in advance in a "trampoline", a block of memory usually allocated on the
11319 stack, which also contains code to splice the nest value into the
11320 argument list. This is used to implement the GCC nested function address
11323 For example, if the function is ``i32 f(i8* nest %c, i32 %x, i32 %y)``
11324 then the resulting function pointer has signature ``i32 (i32, i32)*``.
11325 It can be created as follows:
11327 .. code-block:: llvm
11329 %tramp = alloca [10 x i8], align 4 ; size and alignment only correct for X86
11330 %tramp1 = getelementptr [10 x i8], [10 x i8]* %tramp, i32 0, i32 0
11331 call i8* @llvm.init.trampoline(i8* %tramp1, i8* bitcast (i32 (i8*, i32, i32)* @f to i8*), i8* %nval)
11332 %p = call i8* @llvm.adjust.trampoline(i8* %tramp1)
11333 %fp = bitcast i8* %p to i32 (i32, i32)*
11335 The call ``%val = call i32 %fp(i32 %x, i32 %y)`` is then equivalent to
11336 ``%val = call i32 %f(i8* %nval, i32 %x, i32 %y)``.
11340 '``llvm.init.trampoline``' Intrinsic
11341 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11348 declare void @llvm.init.trampoline(i8* <tramp>, i8* <func>, i8* <nval>)
11353 This fills the memory pointed to by ``tramp`` with executable code,
11354 turning it into a trampoline.
11359 The ``llvm.init.trampoline`` intrinsic takes three arguments, all
11360 pointers. The ``tramp`` argument must point to a sufficiently large and
11361 sufficiently aligned block of memory; this memory is written to by the
11362 intrinsic. Note that the size and the alignment are target-specific -
11363 LLVM currently provides no portable way of determining them, so a
11364 front-end that generates this intrinsic needs to have some
11365 target-specific knowledge. The ``func`` argument must hold a function
11366 bitcast to an ``i8*``.
11371 The block of memory pointed to by ``tramp`` is filled with target
11372 dependent code, turning it into a function. Then ``tramp`` needs to be
11373 passed to :ref:`llvm.adjust.trampoline <int_at>` to get a pointer which can
11374 be :ref:`bitcast (to a new function) and called <int_trampoline>`. The new
11375 function's signature is the same as that of ``func`` with any arguments
11376 marked with the ``nest`` attribute removed. At most one such ``nest``
11377 argument is allowed, and it must be of pointer type. Calling the new
11378 function is equivalent to calling ``func`` with the same argument list,
11379 but with ``nval`` used for the missing ``nest`` argument. If, after
11380 calling ``llvm.init.trampoline``, the memory pointed to by ``tramp`` is
11381 modified, then the effect of any later call to the returned function
11382 pointer is undefined.
11386 '``llvm.adjust.trampoline``' Intrinsic
11387 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11394 declare i8* @llvm.adjust.trampoline(i8* <tramp>)
11399 This performs any required machine-specific adjustment to the address of
11400 a trampoline (passed as ``tramp``).
11405 ``tramp`` must point to a block of memory which already has trampoline
11406 code filled in by a previous call to
11407 :ref:`llvm.init.trampoline <int_it>`.
11412 On some architectures the address of the code to be executed needs to be
11413 different than the address where the trampoline is actually stored. This
11414 intrinsic returns the executable address corresponding to ``tramp``
11415 after performing the required machine specific adjustments. The pointer
11416 returned can then be :ref:`bitcast and executed <int_trampoline>`.
11418 .. _int_mload_mstore:
11420 Masked Vector Load and Store Intrinsics
11421 ---------------------------------------
11423 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.
11427 '``llvm.masked.load.*``' Intrinsics
11428 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11432 This is an overloaded intrinsic. The loaded data is a vector of any integer, floating point or pointer data type.
11436 declare <16 x float> @llvm.masked.load.v16f32 (<16 x float>* <ptr>, i32 <alignment>, <16 x i1> <mask>, <16 x float> <passthru>)
11437 declare <2 x double> @llvm.masked.load.v2f64 (<2 x double>* <ptr>, i32 <alignment>, <2 x i1> <mask>, <2 x double> <passthru>)
11438 ;; The data is a vector of pointers to double
11439 declare <8 x double*> @llvm.masked.load.v8p0f64 (<8 x double*>* <ptr>, i32 <alignment>, <8 x i1> <mask>, <8 x double*> <passthru>)
11440 ;; The data is a vector of function pointers
11441 declare <8 x i32 ()*> @llvm.masked.load.v8p0f_i32f (<8 x i32 ()*>* <ptr>, i32 <alignment>, <8 x i1> <mask>, <8 x i32 ()*> <passthru>)
11446 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.
11452 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.
11458 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.
11459 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.
11464 %res = call <16 x float> @llvm.masked.load.v16f32 (<16 x float>* %ptr, i32 4, <16 x i1>%mask, <16 x float> %passthru)
11466 ;; The result of the two following instructions is identical aside from potential memory access exception
11467 %loadlal = load <16 x float>, <16 x float>* %ptr, align 4
11468 %res = select <16 x i1> %mask, <16 x float> %loadlal, <16 x float> %passthru
11472 '``llvm.masked.store.*``' Intrinsics
11473 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11477 This is an overloaded intrinsic. The data stored in memory is a vector of any integer, floating point or pointer data type.
11481 declare void @llvm.masked.store.v8i32 (<8 x i32> <value>, <8 x i32>* <ptr>, i32 <alignment>, <8 x i1> <mask>)
11482 declare void @llvm.masked.store.v16f32 (<16 x float> <value>, <16 x float>* <ptr>, i32 <alignment>, <16 x i1> <mask>)
11483 ;; The data is a vector of pointers to double
11484 declare void @llvm.masked.store.v8p0f64 (<8 x double*> <value>, <8 x double*>* <ptr>, i32 <alignment>, <8 x i1> <mask>)
11485 ;; The data is a vector of function pointers
11486 declare void @llvm.masked.store.v4p0f_i32f (<4 x i32 ()*> <value>, <4 x i32 ()*>* <ptr>, i32 <alignment>, <4 x i1> <mask>)
11491 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.
11496 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.
11502 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.
11503 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.
11507 call void @llvm.masked.store.v16f32(<16 x float> %value, <16 x float>* %ptr, i32 4, <16 x i1> %mask)
11509 ;; The result of the following instructions is identical aside from potential data races and memory access exceptions
11510 %oldval = load <16 x float>, <16 x float>* %ptr, align 4
11511 %res = select <16 x i1> %mask, <16 x float> %value, <16 x float> %oldval
11512 store <16 x float> %res, <16 x float>* %ptr, align 4
11515 Masked Vector Gather and Scatter Intrinsics
11516 -------------------------------------------
11518 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.
11522 '``llvm.masked.gather.*``' Intrinsics
11523 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11527 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.
11531 declare <16 x float> @llvm.masked.gather.v16f32 (<16 x float*> <ptrs>, i32 <alignment>, <16 x i1> <mask>, <16 x float> <passthru>)
11532 declare <2 x double> @llvm.masked.gather.v2f64 (<2 x double*> <ptrs>, i32 <alignment>, <2 x i1> <mask>, <2 x double> <passthru>)
11533 declare <8 x float*> @llvm.masked.gather.v8p0f32 (<8 x float**> <ptrs>, i32 <alignment>, <8 x i1> <mask>, <8 x float*> <passthru>)
11538 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.
11544 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.
11550 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.
11551 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.
11556 %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>)
11558 ;; The gather with all-true mask is equivalent to the following instruction sequence
11559 %ptr0 = extractelement <4 x double*> %ptrs, i32 0
11560 %ptr1 = extractelement <4 x double*> %ptrs, i32 1
11561 %ptr2 = extractelement <4 x double*> %ptrs, i32 2
11562 %ptr3 = extractelement <4 x double*> %ptrs, i32 3
11564 %val0 = load double, double* %ptr0, align 8
11565 %val1 = load double, double* %ptr1, align 8
11566 %val2 = load double, double* %ptr2, align 8
11567 %val3 = load double, double* %ptr3, align 8
11569 %vec0 = insertelement <4 x double>undef, %val0, 0
11570 %vec01 = insertelement <4 x double>%vec0, %val1, 1
11571 %vec012 = insertelement <4 x double>%vec01, %val2, 2
11572 %vec0123 = insertelement <4 x double>%vec012, %val3, 3
11576 '``llvm.masked.scatter.*``' Intrinsics
11577 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11581 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.
11585 declare void @llvm.masked.scatter.v8i32 (<8 x i32> <value>, <8 x i32*> <ptrs>, i32 <alignment>, <8 x i1> <mask>)
11586 declare void @llvm.masked.scatter.v16f32 (<16 x float> <value>, <16 x float*> <ptrs>, i32 <alignment>, <16 x i1> <mask>)
11587 declare void @llvm.masked.scatter.v4p0f64 (<4 x double*> <value>, <4 x double**> <ptrs>, i32 <alignment>, <4 x i1> <mask>)
11592 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.
11597 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.
11603 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.
11607 ;; This instruction unconditionaly stores data vector in multiple addresses
11608 call @llvm.masked.scatter.v8i32 (<8 x i32> %value, <8 x i32*> %ptrs, i32 4, <8 x i1> <true, true, .. true>)
11610 ;; It is equivalent to a list of scalar stores
11611 %val0 = extractelement <8 x i32> %value, i32 0
11612 %val1 = extractelement <8 x i32> %value, i32 1
11614 %val7 = extractelement <8 x i32> %value, i32 7
11615 %ptr0 = extractelement <8 x i32*> %ptrs, i32 0
11616 %ptr1 = extractelement <8 x i32*> %ptrs, i32 1
11618 %ptr7 = extractelement <8 x i32*> %ptrs, i32 7
11619 ;; Note: the order of the following stores is important when they overlap:
11620 store i32 %val0, i32* %ptr0, align 4
11621 store i32 %val1, i32* %ptr1, align 4
11623 store i32 %val7, i32* %ptr7, align 4
11629 This class of intrinsics provides information about the lifetime of
11630 memory objects and ranges where variables are immutable.
11634 '``llvm.lifetime.start``' Intrinsic
11635 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11642 declare void @llvm.lifetime.start(i64 <size>, i8* nocapture <ptr>)
11647 The '``llvm.lifetime.start``' intrinsic specifies the start of a memory
11653 The first argument is a constant integer representing the size of the
11654 object, or -1 if it is variable sized. The second argument is a pointer
11660 This intrinsic indicates that before this point in the code, the value
11661 of the memory pointed to by ``ptr`` is dead. This means that it is known
11662 to never be used and has an undefined value. A load from the pointer
11663 that precedes this intrinsic can be replaced with ``'undef'``.
11667 '``llvm.lifetime.end``' Intrinsic
11668 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11675 declare void @llvm.lifetime.end(i64 <size>, i8* nocapture <ptr>)
11680 The '``llvm.lifetime.end``' intrinsic specifies the end of a memory
11686 The first argument is a constant integer representing the size of the
11687 object, or -1 if it is variable sized. The second argument is a pointer
11693 This intrinsic indicates that after this point in the code, the value of
11694 the memory pointed to by ``ptr`` is dead. This means that it is known to
11695 never be used and has an undefined value. Any stores into the memory
11696 object following this intrinsic may be removed as dead.
11698 '``llvm.invariant.start``' Intrinsic
11699 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11706 declare {}* @llvm.invariant.start(i64 <size>, i8* nocapture <ptr>)
11711 The '``llvm.invariant.start``' intrinsic specifies that the contents of
11712 a memory object will not change.
11717 The first argument is a constant integer representing the size of the
11718 object, or -1 if it is variable sized. The second argument is a pointer
11724 This intrinsic indicates that until an ``llvm.invariant.end`` that uses
11725 the return value, the referenced memory location is constant and
11728 '``llvm.invariant.end``' Intrinsic
11729 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11736 declare void @llvm.invariant.end({}* <start>, i64 <size>, i8* nocapture <ptr>)
11741 The '``llvm.invariant.end``' intrinsic specifies that the contents of a
11742 memory object are mutable.
11747 The first argument is the matching ``llvm.invariant.start`` intrinsic.
11748 The second argument is a constant integer representing the size of the
11749 object, or -1 if it is variable sized and the third argument is a
11750 pointer to the object.
11755 This intrinsic indicates that the memory is mutable again.
11757 '``llvm.invariant.group.barrier``' Intrinsic
11758 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11765 declare i8* @llvm.invariant.group.barrier(i8* <ptr>)
11770 The '``llvm.invariant.group.barrier``' intrinsic can be used when an invariant
11771 established by invariant.group metadata no longer holds, to obtain a new pointer
11772 value that does not carry the invariant information.
11778 The ``llvm.invariant.group.barrier`` takes only one argument, which is
11779 the pointer to the memory for which the ``invariant.group`` no longer holds.
11784 Returns another pointer that aliases its argument but which is considered different
11785 for the purposes of ``load``/``store`` ``invariant.group`` metadata.
11790 This class of intrinsics is designed to be generic and has no specific
11793 '``llvm.var.annotation``' Intrinsic
11794 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11801 declare void @llvm.var.annotation(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
11806 The '``llvm.var.annotation``' intrinsic.
11811 The first argument is a pointer to a value, the second is a pointer to a
11812 global string, the third is a pointer to a global string which is the
11813 source file name, and the last argument is the line number.
11818 This intrinsic allows annotation of local variables with arbitrary
11819 strings. This can be useful for special purpose optimizations that want
11820 to look for these annotations. These have no other defined use; they are
11821 ignored by code generation and optimization.
11823 '``llvm.ptr.annotation.*``' Intrinsic
11824 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11829 This is an overloaded intrinsic. You can use '``llvm.ptr.annotation``' on a
11830 pointer to an integer of any width. *NOTE* you must specify an address space for
11831 the pointer. The identifier for the default address space is the integer
11836 declare i8* @llvm.ptr.annotation.p<address space>i8(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
11837 declare i16* @llvm.ptr.annotation.p<address space>i16(i16* <val>, i8* <str>, i8* <str>, i32 <int>)
11838 declare i32* @llvm.ptr.annotation.p<address space>i32(i32* <val>, i8* <str>, i8* <str>, i32 <int>)
11839 declare i64* @llvm.ptr.annotation.p<address space>i64(i64* <val>, i8* <str>, i8* <str>, i32 <int>)
11840 declare i256* @llvm.ptr.annotation.p<address space>i256(i256* <val>, i8* <str>, i8* <str>, i32 <int>)
11845 The '``llvm.ptr.annotation``' intrinsic.
11850 The first argument is a pointer to an integer value of arbitrary bitwidth
11851 (result of some expression), the second is a pointer to a global string, the
11852 third is a pointer to a global string which is the source file name, and the
11853 last argument is the line number. It returns the value of the first argument.
11858 This intrinsic allows annotation of a pointer to an integer with arbitrary
11859 strings. This can be useful for special purpose optimizations that want to look
11860 for these annotations. These have no other defined use; they are ignored by code
11861 generation and optimization.
11863 '``llvm.annotation.*``' Intrinsic
11864 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11869 This is an overloaded intrinsic. You can use '``llvm.annotation``' on
11870 any integer bit width.
11874 declare i8 @llvm.annotation.i8(i8 <val>, i8* <str>, i8* <str>, i32 <int>)
11875 declare i16 @llvm.annotation.i16(i16 <val>, i8* <str>, i8* <str>, i32 <int>)
11876 declare i32 @llvm.annotation.i32(i32 <val>, i8* <str>, i8* <str>, i32 <int>)
11877 declare i64 @llvm.annotation.i64(i64 <val>, i8* <str>, i8* <str>, i32 <int>)
11878 declare i256 @llvm.annotation.i256(i256 <val>, i8* <str>, i8* <str>, i32 <int>)
11883 The '``llvm.annotation``' intrinsic.
11888 The first argument is an integer value (result of some expression), the
11889 second is a pointer to a global string, the third is a pointer to a
11890 global string which is the source file name, and the last argument is
11891 the line number. It returns the value of the first argument.
11896 This intrinsic allows annotations to be put on arbitrary expressions
11897 with arbitrary strings. This can be useful for special purpose
11898 optimizations that want to look for these annotations. These have no
11899 other defined use; they are ignored by code generation and optimization.
11901 '``llvm.trap``' Intrinsic
11902 ^^^^^^^^^^^^^^^^^^^^^^^^^
11909 declare void @llvm.trap() noreturn nounwind
11914 The '``llvm.trap``' intrinsic.
11924 This intrinsic is lowered to the target dependent trap instruction. If
11925 the target does not have a trap instruction, this intrinsic will be
11926 lowered to a call of the ``abort()`` function.
11928 '``llvm.debugtrap``' Intrinsic
11929 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11936 declare void @llvm.debugtrap() nounwind
11941 The '``llvm.debugtrap``' intrinsic.
11951 This intrinsic is lowered to code which is intended to cause an
11952 execution trap with the intention of requesting the attention of a
11955 '``llvm.stackprotector``' Intrinsic
11956 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11963 declare void @llvm.stackprotector(i8* <guard>, i8** <slot>)
11968 The ``llvm.stackprotector`` intrinsic takes the ``guard`` and stores it
11969 onto the stack at ``slot``. The stack slot is adjusted to ensure that it
11970 is placed on the stack before local variables.
11975 The ``llvm.stackprotector`` intrinsic requires two pointer arguments.
11976 The first argument is the value loaded from the stack guard
11977 ``@__stack_chk_guard``. The second variable is an ``alloca`` that has
11978 enough space to hold the value of the guard.
11983 This intrinsic causes the prologue/epilogue inserter to force the position of
11984 the ``AllocaInst`` stack slot to be before local variables on the stack. This is
11985 to ensure that if a local variable on the stack is overwritten, it will destroy
11986 the value of the guard. When the function exits, the guard on the stack is
11987 checked against the original guard by ``llvm.stackprotectorcheck``. If they are
11988 different, then ``llvm.stackprotectorcheck`` causes the program to abort by
11989 calling the ``__stack_chk_fail()`` function.
11991 '``llvm.stackprotectorcheck``' Intrinsic
11992 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11999 declare void @llvm.stackprotectorcheck(i8** <guard>)
12004 The ``llvm.stackprotectorcheck`` intrinsic compares ``guard`` against an already
12005 created stack protector and if they are not equal calls the
12006 ``__stack_chk_fail()`` function.
12011 The ``llvm.stackprotectorcheck`` intrinsic requires one pointer argument, the
12012 the variable ``@__stack_chk_guard``.
12017 This intrinsic is provided to perform the stack protector check by comparing
12018 ``guard`` with the stack slot created by ``llvm.stackprotector`` and if the
12019 values do not match call the ``__stack_chk_fail()`` function.
12021 The reason to provide this as an IR level intrinsic instead of implementing it
12022 via other IR operations is that in order to perform this operation at the IR
12023 level without an intrinsic, one would need to create additional basic blocks to
12024 handle the success/failure cases. This makes it difficult to stop the stack
12025 protector check from disrupting sibling tail calls in Codegen. With this
12026 intrinsic, we are able to generate the stack protector basic blocks late in
12027 codegen after the tail call decision has occurred.
12029 '``llvm.objectsize``' Intrinsic
12030 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12037 declare i32 @llvm.objectsize.i32(i8* <object>, i1 <min>)
12038 declare i64 @llvm.objectsize.i64(i8* <object>, i1 <min>)
12043 The ``llvm.objectsize`` intrinsic is designed to provide information to
12044 the optimizers to determine at compile time whether a) an operation
12045 (like memcpy) will overflow a buffer that corresponds to an object, or
12046 b) that a runtime check for overflow isn't necessary. An object in this
12047 context means an allocation of a specific class, structure, array, or
12053 The ``llvm.objectsize`` intrinsic takes two arguments. The first
12054 argument is a pointer to or into the ``object``. The second argument is
12055 a boolean and determines whether ``llvm.objectsize`` returns 0 (if true)
12056 or -1 (if false) when the object size is unknown. The second argument
12057 only accepts constants.
12062 The ``llvm.objectsize`` intrinsic is lowered to a constant representing
12063 the size of the object concerned. If the size cannot be determined at
12064 compile time, ``llvm.objectsize`` returns ``i32/i64 -1 or 0`` (depending
12065 on the ``min`` argument).
12067 '``llvm.expect``' Intrinsic
12068 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
12073 This is an overloaded intrinsic. You can use ``llvm.expect`` on any
12078 declare i1 @llvm.expect.i1(i1 <val>, i1 <expected_val>)
12079 declare i32 @llvm.expect.i32(i32 <val>, i32 <expected_val>)
12080 declare i64 @llvm.expect.i64(i64 <val>, i64 <expected_val>)
12085 The ``llvm.expect`` intrinsic provides information about expected (the
12086 most probable) value of ``val``, which can be used by optimizers.
12091 The ``llvm.expect`` intrinsic takes two arguments. The first argument is
12092 a value. The second argument is an expected value, this needs to be a
12093 constant value, variables are not allowed.
12098 This intrinsic is lowered to the ``val``.
12102 '``llvm.assume``' Intrinsic
12103 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12110 declare void @llvm.assume(i1 %cond)
12115 The ``llvm.assume`` allows the optimizer to assume that the provided
12116 condition is true. This information can then be used in simplifying other parts
12122 The condition which the optimizer may assume is always true.
12127 The intrinsic allows the optimizer to assume that the provided condition is
12128 always true whenever the control flow reaches the intrinsic call. No code is
12129 generated for this intrinsic, and instructions that contribute only to the
12130 provided condition are not used for code generation. If the condition is
12131 violated during execution, the behavior is undefined.
12133 Note that the optimizer might limit the transformations performed on values
12134 used by the ``llvm.assume`` intrinsic in order to preserve the instructions
12135 only used to form the intrinsic's input argument. This might prove undesirable
12136 if the extra information provided by the ``llvm.assume`` intrinsic does not cause
12137 sufficient overall improvement in code quality. For this reason,
12138 ``llvm.assume`` should not be used to document basic mathematical invariants
12139 that the optimizer can otherwise deduce or facts that are of little use to the
12144 '``llvm.bitset.test``' Intrinsic
12145 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12152 declare i1 @llvm.bitset.test(i8* %ptr, metadata %bitset) nounwind readnone
12158 The first argument is a pointer to be tested. The second argument is a
12159 metadata object representing an identifier for a :doc:`bitset <BitSets>`.
12164 The ``llvm.bitset.test`` intrinsic tests whether the given pointer is a
12165 member of the given bitset.
12167 '``llvm.donothing``' Intrinsic
12168 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
12175 declare void @llvm.donothing() nounwind readnone
12180 The ``llvm.donothing`` intrinsic doesn't perform any operation. It's one of only
12181 two intrinsics (besides ``llvm.experimental.patchpoint``) that can be called
12182 with an invoke instruction.
12192 This intrinsic does nothing, and it's removed by optimizers and ignored
12195 Stack Map Intrinsics
12196 --------------------
12198 LLVM provides experimental intrinsics to support runtime patching
12199 mechanisms commonly desired in dynamic language JITs. These intrinsics
12200 are described in :doc:`StackMaps`.