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 opening curly brace, a list of basic blocks, and a closing curly brace.
645 LLVM function declarations consist of the "``declare``" keyword, an
646 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
647 style <visibility>`, an optional :ref:`DLL storage class <dllstorageclass>`,
648 an optional :ref:`calling convention <callingconv>`,
649 an optional ``unnamed_addr`` attribute, a return type, an optional
650 :ref:`parameter attribute <paramattrs>` for the return type, a function
651 name, a possibly empty list of arguments, an optional alignment, an optional
652 :ref:`garbage collector name <gc>`, an optional :ref:`prefix <prefixdata>`,
653 and an optional :ref:`prologue <prologuedata>`.
655 A function definition contains a list of basic blocks, forming the CFG (Control
656 Flow Graph) for the function. Each basic block may optionally start with a label
657 (giving the basic block a symbol table entry), contains a list of instructions,
658 and ends with a :ref:`terminator <terminators>` instruction (such as a branch or
659 function return). If an explicit label is not provided, a block is assigned an
660 implicit numbered label, using the next value from the same counter as used for
661 unnamed temporaries (:ref:`see above<identifiers>`). For example, if a function
662 entry block does not have an explicit label, it will be assigned label "%0",
663 then the first unnamed temporary in that block will be "%1", etc.
665 The first basic block in a function is special in two ways: it is
666 immediately executed on entrance to the function, and it is not allowed
667 to have predecessor basic blocks (i.e. there can not be any branches to
668 the entry block of a function). Because the block can have no
669 predecessors, it also cannot have any :ref:`PHI nodes <i_phi>`.
671 LLVM allows an explicit section to be specified for functions. If the
672 target supports it, it will emit functions to the section specified.
673 Additionally, the function can be placed in a COMDAT.
675 An explicit alignment may be specified for a function. If not present,
676 or if the alignment is set to zero, the alignment of the function is set
677 by the target to whatever it feels convenient. If an explicit alignment
678 is specified, the function is forced to have at least that much
679 alignment. All alignments must be a power of 2.
681 If the ``unnamed_addr`` attribute is given, the address is known to not
682 be significant and two identical functions can be merged.
686 define [linkage] [visibility] [DLLStorageClass]
688 <ResultType> @<FunctionName> ([argument list])
689 [unnamed_addr] [fn Attrs] [section "name"] [comdat [($name)]]
690 [align N] [gc] [prefix Constant] [prologue Constant]
691 [personality Constant] { ... }
693 The argument list is a comma separated sequence of arguments where each
694 argument is of the following form:
698 <type> [parameter Attrs] [name]
706 Aliases, unlike function or variables, don't create any new data. They
707 are just a new symbol and metadata for an existing position.
709 Aliases have a name and an aliasee that is either a global value or a
712 Aliases may have an optional :ref:`linkage type <linkage>`, an optional
713 :ref:`visibility style <visibility>`, an optional :ref:`DLL storage class
714 <dllstorageclass>` and an optional :ref:`tls model <tls_model>`.
718 @<Name> = [Linkage] [Visibility] [DLLStorageClass] [ThreadLocal] [unnamed_addr] alias <AliaseeTy> @<Aliasee>
720 The linkage must be one of ``private``, ``internal``, ``linkonce``, ``weak``,
721 ``linkonce_odr``, ``weak_odr``, ``external``. Note that some system linkers
722 might not correctly handle dropping a weak symbol that is aliased.
724 Aliases that are not ``unnamed_addr`` are guaranteed to have the same address as
725 the aliasee expression. ``unnamed_addr`` ones are only guaranteed to point
728 Since aliases are only a second name, some restrictions apply, of which
729 some can only be checked when producing an object file:
731 * The expression defining the aliasee must be computable at assembly
732 time. Since it is just a name, no relocations can be used.
734 * No alias in the expression can be weak as the possibility of the
735 intermediate alias being overridden cannot be represented in an
738 * No global value in the expression can be a declaration, since that
739 would require a relocation, which is not possible.
746 Comdat IR provides access to COFF and ELF object file COMDAT functionality.
748 Comdats have a name which represents the COMDAT key. All global objects that
749 specify this key will only end up in the final object file if the linker chooses
750 that key over some other key. Aliases are placed in the same COMDAT that their
751 aliasee computes to, if any.
753 Comdats have a selection kind to provide input on how the linker should
754 choose between keys in two different object files.
758 $<Name> = comdat SelectionKind
760 The selection kind must be one of the following:
763 The linker may choose any COMDAT key, the choice is arbitrary.
765 The linker may choose any COMDAT key but the sections must contain the
768 The linker will choose the section containing the largest COMDAT key.
770 The linker requires that only section with this COMDAT key exist.
772 The linker may choose any COMDAT key but the sections must contain the
775 Note that the Mach-O platform doesn't support COMDATs and ELF only supports
776 ``any`` as a selection kind.
778 Here is an example of a COMDAT group where a function will only be selected if
779 the COMDAT key's section is the largest:
783 $foo = comdat largest
784 @foo = global i32 2, comdat($foo)
786 define void @bar() comdat($foo) {
790 As a syntactic sugar the ``$name`` can be omitted if the name is the same as
796 @foo = global i32 2, comdat
799 In a COFF object file, this will create a COMDAT section with selection kind
800 ``IMAGE_COMDAT_SELECT_LARGEST`` containing the contents of the ``@foo`` symbol
801 and another COMDAT section with selection kind
802 ``IMAGE_COMDAT_SELECT_ASSOCIATIVE`` which is associated with the first COMDAT
803 section and contains the contents of the ``@bar`` symbol.
805 There are some restrictions on the properties of the global object.
806 It, or an alias to it, must have the same name as the COMDAT group when
808 The contents and size of this object may be used during link-time to determine
809 which COMDAT groups get selected depending on the selection kind.
810 Because the name of the object must match the name of the COMDAT group, the
811 linkage of the global object must not be local; local symbols can get renamed
812 if a collision occurs in the symbol table.
814 The combined use of COMDATS and section attributes may yield surprising results.
821 @g1 = global i32 42, section "sec", comdat($foo)
822 @g2 = global i32 42, section "sec", comdat($bar)
824 From the object file perspective, this requires the creation of two sections
825 with the same name. This is necessary because both globals belong to different
826 COMDAT groups and COMDATs, at the object file level, are represented by
829 Note that certain IR constructs like global variables and functions may
830 create COMDATs in the object file in addition to any which are specified using
831 COMDAT IR. This arises when the code generator is configured to emit globals
832 in individual sections (e.g. when `-data-sections` or `-function-sections`
833 is supplied to `llc`).
835 .. _namedmetadatastructure:
840 Named metadata is a collection of metadata. :ref:`Metadata
841 nodes <metadata>` (but not metadata strings) are the only valid
842 operands for a named metadata.
844 #. Named metadata are represented as a string of characters with the
845 metadata prefix. The rules for metadata names are the same as for
846 identifiers, but quoted names are not allowed. ``"\xx"`` type escapes
847 are still valid, which allows any character to be part of a name.
851 ; Some unnamed metadata nodes, which are referenced by the named metadata.
856 !name = !{!0, !1, !2}
863 The return type and each parameter of a function type may have a set of
864 *parameter attributes* associated with them. Parameter attributes are
865 used to communicate additional information about the result or
866 parameters of a function. Parameter attributes are considered to be part
867 of the function, not of the function type, so functions with different
868 parameter attributes can have the same function type.
870 Parameter attributes are simple keywords that follow the type specified.
871 If multiple parameter attributes are needed, they are space separated.
876 declare i32 @printf(i8* noalias nocapture, ...)
877 declare i32 @atoi(i8 zeroext)
878 declare signext i8 @returns_signed_char()
880 Note that any attributes for the function result (``nounwind``,
881 ``readonly``) come immediately after the argument list.
883 Currently, only the following parameter attributes are defined:
886 This indicates to the code generator that the parameter or return
887 value should be zero-extended to the extent required by the target's
888 ABI (which is usually 32-bits, but is 8-bits for a i1 on x86-64) by
889 the caller (for a parameter) or the callee (for a return value).
891 This indicates to the code generator that the parameter or return
892 value should be sign-extended to the extent required by the target's
893 ABI (which is usually 32-bits) by the caller (for a parameter) or
894 the callee (for a return value).
896 This indicates that this parameter or return value should be treated
897 in a special target-dependent fashion while emitting code for
898 a function call or return (usually, by putting it in a register as
899 opposed to memory, though some targets use it to distinguish between
900 two different kinds of registers). Use of this attribute is
903 This indicates that the pointer parameter should really be passed by
904 value to the function. The attribute implies that a hidden copy of
905 the pointee is made between the caller and the callee, so the callee
906 is unable to modify the value in the caller. This attribute is only
907 valid on LLVM pointer arguments. It is generally used to pass
908 structs and arrays by value, but is also valid on pointers to
909 scalars. The copy is considered to belong to the caller not the
910 callee (for example, ``readonly`` functions should not write to
911 ``byval`` parameters). This is not a valid attribute for return
914 The byval attribute also supports specifying an alignment with the
915 align attribute. It indicates the alignment of the stack slot to
916 form and the known alignment of the pointer specified to the call
917 site. If the alignment is not specified, then the code generator
918 makes a target-specific assumption.
924 The ``inalloca`` argument attribute allows the caller to take the
925 address of outgoing stack arguments. An ``inalloca`` argument must
926 be a pointer to stack memory produced by an ``alloca`` instruction.
927 The alloca, or argument allocation, must also be tagged with the
928 inalloca keyword. Only the last argument may have the ``inalloca``
929 attribute, and that argument is guaranteed to be passed in memory.
931 An argument allocation may be used by a call at most once because
932 the call may deallocate it. The ``inalloca`` attribute cannot be
933 used in conjunction with other attributes that affect argument
934 storage, like ``inreg``, ``nest``, ``sret``, or ``byval``. The
935 ``inalloca`` attribute also disables LLVM's implicit lowering of
936 large aggregate return values, which means that frontend authors
937 must lower them with ``sret`` pointers.
939 When the call site is reached, the argument allocation must have
940 been the most recent stack allocation that is still live, or the
941 results are undefined. It is possible to allocate additional stack
942 space after an argument allocation and before its call site, but it
943 must be cleared off with :ref:`llvm.stackrestore
946 See :doc:`InAlloca` for more information on how to use this
950 This indicates that the pointer parameter specifies the address of a
951 structure that is the return value of the function in the source
952 program. This pointer must be guaranteed by the caller to be valid:
953 loads and stores to the structure may be assumed by the callee
954 not to trap and to be properly aligned. This may only be applied to
955 the first parameter. This is not a valid attribute for return
959 This indicates that the pointer value may be assumed by the optimizer to
960 have the specified alignment.
962 Note that this attribute has additional semantics when combined with the
968 This indicates that objects accessed via pointer values
969 :ref:`based <pointeraliasing>` on the argument or return value are not also
970 accessed, during the execution of the function, via pointer values not
971 *based* on the argument or return value. The attribute on a return value
972 also has additional semantics described below. The caller shares the
973 responsibility with the callee for ensuring that these requirements are met.
974 For further details, please see the discussion of the NoAlias response in
975 :ref:`alias analysis <Must, May, or No>`.
977 Note that this definition of ``noalias`` is intentionally similar
978 to the definition of ``restrict`` in C99 for function arguments.
980 For function return values, C99's ``restrict`` is not meaningful,
981 while LLVM's ``noalias`` is. Furthermore, the semantics of the ``noalias``
982 attribute on return values are stronger than the semantics of the attribute
983 when used on function arguments. On function return values, the ``noalias``
984 attribute indicates that the function acts like a system memory allocation
985 function, returning a pointer to allocated storage disjoint from the
986 storage for any other object accessible to the caller.
989 This indicates that the callee does not make any copies of the
990 pointer that outlive the callee itself. This is not a valid
991 attribute for return values.
996 This indicates that the pointer parameter can be excised using the
997 :ref:`trampoline intrinsics <int_trampoline>`. This is not a valid
998 attribute for return values and can only be applied to one parameter.
1001 This indicates that the function always returns the argument as its return
1002 value. This is an optimization hint to the code generator when generating
1003 the caller, allowing tail call optimization and omission of register saves
1004 and restores in some cases; it is not checked or enforced when generating
1005 the callee. The parameter and the function return type must be valid
1006 operands for the :ref:`bitcast instruction <i_bitcast>`. This is not a
1007 valid attribute for return values and can only be applied to one parameter.
1010 This indicates that the parameter or return pointer is not null. This
1011 attribute may only be applied to pointer typed parameters. This is not
1012 checked or enforced by LLVM, the caller must ensure that the pointer
1013 passed in is non-null, or the callee must ensure that the returned pointer
1016 ``dereferenceable(<n>)``
1017 This indicates that the parameter or return pointer is dereferenceable. This
1018 attribute may only be applied to pointer typed parameters. A pointer that
1019 is dereferenceable can be loaded from speculatively without a risk of
1020 trapping. The number of bytes known to be dereferenceable must be provided
1021 in parentheses. It is legal for the number of bytes to be less than the
1022 size of the pointee type. The ``nonnull`` attribute does not imply
1023 dereferenceability (consider a pointer to one element past the end of an
1024 array), however ``dereferenceable(<n>)`` does imply ``nonnull`` in
1025 ``addrspace(0)`` (which is the default address space).
1027 ``dereferenceable_or_null(<n>)``
1028 This indicates that the parameter or return value isn't both
1029 non-null and non-dereferenceable (up to ``<n>`` bytes) at the same
1030 time. All non-null pointers tagged with
1031 ``dereferenceable_or_null(<n>)`` are ``dereferenceable(<n>)``.
1032 For address space 0 ``dereferenceable_or_null(<n>)`` implies that
1033 a pointer is exactly one of ``dereferenceable(<n>)`` or ``null``,
1034 and in other address spaces ``dereferenceable_or_null(<n>)``
1035 implies that a pointer is at least one of ``dereferenceable(<n>)``
1036 or ``null`` (i.e. it may be both ``null`` and
1037 ``dereferenceable(<n>)``). This attribute may only be applied to
1038 pointer typed parameters.
1042 Garbage Collector Strategy Names
1043 --------------------------------
1045 Each function may specify a garbage collector strategy name, which is simply a
1048 .. code-block:: llvm
1050 define void @f() gc "name" { ... }
1052 The supported values of *name* includes those :ref:`built in to LLVM
1053 <builtin-gc-strategies>` and any provided by loaded plugins. Specifying a GC
1054 strategy will cause the compiler to alter its output in order to support the
1055 named garbage collection algorithm. Note that LLVM itself does not contain a
1056 garbage collector, this functionality is restricted to generating machine code
1057 which can interoperate with a collector provided externally.
1064 Prefix data is data associated with a function which the code
1065 generator will emit immediately before the function's entrypoint.
1066 The purpose of this feature is to allow frontends to associate
1067 language-specific runtime metadata with specific functions and make it
1068 available through the function pointer while still allowing the
1069 function pointer to be called.
1071 To access the data for a given function, a program may bitcast the
1072 function pointer to a pointer to the constant's type and dereference
1073 index -1. This implies that the IR symbol points just past the end of
1074 the prefix data. For instance, take the example of a function annotated
1075 with a single ``i32``,
1077 .. code-block:: llvm
1079 define void @f() prefix i32 123 { ... }
1081 The prefix data can be referenced as,
1083 .. code-block:: llvm
1085 %0 = bitcast void* () @f to i32*
1086 %a = getelementptr inbounds i32, i32* %0, i32 -1
1087 %b = load i32, i32* %a
1089 Prefix data is laid out as if it were an initializer for a global variable
1090 of the prefix data's type. The function will be placed such that the
1091 beginning of the prefix data is aligned. This means that if the size
1092 of the prefix data is not a multiple of the alignment size, the
1093 function's entrypoint will not be aligned. If alignment of the
1094 function's entrypoint is desired, padding must be added to the prefix
1097 A function may have prefix data but no body. This has similar semantics
1098 to the ``available_externally`` linkage in that the data may be used by the
1099 optimizers but will not be emitted in the object file.
1106 The ``prologue`` attribute allows arbitrary code (encoded as bytes) to
1107 be inserted prior to the function body. This can be used for enabling
1108 function hot-patching and instrumentation.
1110 To maintain the semantics of ordinary function calls, the prologue data must
1111 have a particular format. Specifically, it must begin with a sequence of
1112 bytes which decode to a sequence of machine instructions, valid for the
1113 module's target, which transfer control to the point immediately succeeding
1114 the prologue data, without performing any other visible action. This allows
1115 the inliner and other passes to reason about the semantics of the function
1116 definition without needing to reason about the prologue data. Obviously this
1117 makes the format of the prologue data highly target dependent.
1119 A trivial example of valid prologue data for the x86 architecture is ``i8 144``,
1120 which encodes the ``nop`` instruction:
1122 .. code-block:: llvm
1124 define void @f() prologue i8 144 { ... }
1126 Generally prologue data can be formed by encoding a relative branch instruction
1127 which skips the metadata, as in this example of valid prologue data for the
1128 x86_64 architecture, where the first two bytes encode ``jmp .+10``:
1130 .. code-block:: llvm
1132 %0 = type <{ i8, i8, i8* }>
1134 define void @f() prologue %0 <{ i8 235, i8 8, i8* @md}> { ... }
1136 A function may have prologue data but no body. This has similar semantics
1137 to the ``available_externally`` linkage in that the data may be used by the
1138 optimizers but will not be emitted in the object file.
1142 Personality Function
1143 --------------------
1145 The ``personality`` attribute permits functions to specify what function
1146 to use for exception handling.
1153 Attribute groups are groups of attributes that are referenced by objects within
1154 the IR. They are important for keeping ``.ll`` files readable, because a lot of
1155 functions will use the same set of attributes. In the degenerative case of a
1156 ``.ll`` file that corresponds to a single ``.c`` file, the single attribute
1157 group will capture the important command line flags used to build that file.
1159 An attribute group is a module-level object. To use an attribute group, an
1160 object references the attribute group's ID (e.g. ``#37``). An object may refer
1161 to more than one attribute group. In that situation, the attributes from the
1162 different groups are merged.
1164 Here is an example of attribute groups for a function that should always be
1165 inlined, has a stack alignment of 4, and which shouldn't use SSE instructions:
1167 .. code-block:: llvm
1169 ; Target-independent attributes:
1170 attributes #0 = { alwaysinline alignstack=4 }
1172 ; Target-dependent attributes:
1173 attributes #1 = { "no-sse" }
1175 ; Function @f has attributes: alwaysinline, alignstack=4, and "no-sse".
1176 define void @f() #0 #1 { ... }
1183 Function attributes are set to communicate additional information about
1184 a function. Function attributes are considered to be part of the
1185 function, not of the function type, so functions with different function
1186 attributes can have the same function type.
1188 Function attributes are simple keywords that follow the type specified.
1189 If multiple attributes are needed, they are space separated. For
1192 .. code-block:: llvm
1194 define void @f() noinline { ... }
1195 define void @f() alwaysinline { ... }
1196 define void @f() alwaysinline optsize { ... }
1197 define void @f() optsize { ... }
1200 This attribute indicates that, when emitting the prologue and
1201 epilogue, the backend should forcibly align the stack pointer.
1202 Specify the desired alignment, which must be a power of two, in
1205 This attribute indicates that the inliner should attempt to inline
1206 this function into callers whenever possible, ignoring any active
1207 inlining size threshold for this caller.
1209 This indicates that the callee function at a call site should be
1210 recognized as a built-in function, even though the function's declaration
1211 uses the ``nobuiltin`` attribute. This is only valid at call sites for
1212 direct calls to functions that are declared with the ``nobuiltin``
1215 This attribute indicates that this function is rarely called. When
1216 computing edge weights, basic blocks post-dominated by a cold
1217 function call are also considered to be cold; and, thus, given low
1220 This attribute indicates that the callee is dependent on a convergent
1221 thread execution pattern under certain parallel execution models.
1222 Transformations that are execution model agnostic may only move or
1223 tranform this call if the final location is control equivalent to its
1224 original position in the program, where control equivalence is defined as
1225 A dominates B and B post-dominates A, or vice versa.
1227 This attribute indicates that the source code contained a hint that
1228 inlining this function is desirable (such as the "inline" keyword in
1229 C/C++). It is just a hint; it imposes no requirements on the
1232 This attribute indicates that the function should be added to a
1233 jump-instruction table at code-generation time, and that all address-taken
1234 references to this function should be replaced with a reference to the
1235 appropriate jump-instruction-table function pointer. Note that this creates
1236 a new pointer for the original function, which means that code that depends
1237 on function-pointer identity can break. So, any function annotated with
1238 ``jumptable`` must also be ``unnamed_addr``.
1240 This attribute suggests that optimization passes and code generator
1241 passes make choices that keep the code size of this function as small
1242 as possible and perform optimizations that may sacrifice runtime
1243 performance in order to minimize the size of the generated code.
1245 This attribute disables prologue / epilogue emission for the
1246 function. This can have very system-specific consequences.
1248 This indicates that the callee function at a call site is not recognized as
1249 a built-in function. LLVM will retain the original call and not replace it
1250 with equivalent code based on the semantics of the built-in function, unless
1251 the call site uses the ``builtin`` attribute. This is valid at call sites
1252 and on function declarations and definitions.
1254 This attribute indicates that calls to the function cannot be
1255 duplicated. A call to a ``noduplicate`` function may be moved
1256 within its parent function, but may not be duplicated within
1257 its parent function.
1259 A function containing a ``noduplicate`` call may still
1260 be an inlining candidate, provided that the call is not
1261 duplicated by inlining. That implies that the function has
1262 internal linkage and only has one call site, so the original
1263 call is dead after inlining.
1265 This attributes disables implicit floating point instructions.
1267 This attribute indicates that the inliner should never inline this
1268 function in any situation. This attribute may not be used together
1269 with the ``alwaysinline`` attribute.
1271 This attribute suppresses lazy symbol binding for the function. This
1272 may make calls to the function faster, at the cost of extra program
1273 startup time if the function is not called during program startup.
1275 This attribute indicates that the code generator should not use a
1276 red zone, even if the target-specific ABI normally permits it.
1278 This function attribute indicates that the function never returns
1279 normally. This produces undefined behavior at runtime if the
1280 function ever does dynamically return.
1282 This function attribute indicates that the function never raises an
1283 exception. If the function does raise an exception, its runtime
1284 behavior is undefined. However, functions marked nounwind may still
1285 trap or generate asynchronous exceptions. Exception handling schemes
1286 that are recognized by LLVM to handle asynchronous exceptions, such
1287 as SEH, will still provide their implementation defined semantics.
1289 This function attribute indicates that the function is not optimized
1290 by any optimization or code generator passes with the
1291 exception of interprocedural optimization passes.
1292 This attribute cannot be used together with the ``alwaysinline``
1293 attribute; this attribute is also incompatible
1294 with the ``minsize`` attribute and the ``optsize`` attribute.
1296 This attribute requires the ``noinline`` attribute to be specified on
1297 the function as well, so the function is never inlined into any caller.
1298 Only functions with the ``alwaysinline`` attribute are valid
1299 candidates for inlining into the body of this function.
1301 This attribute suggests that optimization passes and code generator
1302 passes make choices that keep the code size of this function low,
1303 and otherwise do optimizations specifically to reduce code size as
1304 long as they do not significantly impact runtime performance.
1306 On a function, this attribute indicates that the function computes its
1307 result (or decides to unwind an exception) based strictly on its arguments,
1308 without dereferencing any pointer arguments or otherwise accessing
1309 any mutable state (e.g. memory, control registers, etc) visible to
1310 caller functions. It does not write through any pointer arguments
1311 (including ``byval`` arguments) and never changes any state visible
1312 to callers. This means that it cannot unwind exceptions by calling
1313 the ``C++`` exception throwing methods.
1315 On an argument, this attribute indicates that the function does not
1316 dereference that pointer argument, even though it may read or write the
1317 memory that the pointer points to if accessed through other pointers.
1319 On a function, this attribute indicates that the function does not write
1320 through any pointer arguments (including ``byval`` arguments) or otherwise
1321 modify any state (e.g. memory, control registers, etc) visible to
1322 caller functions. It may dereference pointer arguments and read
1323 state that may be set in the caller. A readonly function always
1324 returns the same value (or unwinds an exception identically) when
1325 called with the same set of arguments and global state. It cannot
1326 unwind an exception by calling the ``C++`` exception throwing
1329 On an argument, this attribute indicates that the function does not write
1330 through this pointer argument, even though it may write to the memory that
1331 the pointer points to.
1333 This attribute indicates that the only memory accesses inside function are
1334 loads and stores from objects pointed to by its pointer-typed arguments,
1335 with arbitrary offsets. Or in other words, all memory operations in the
1336 function can refer to memory only using pointers based on its function
1338 Note that ``argmemonly`` can be used together with ``readonly`` attribute
1339 in order to specify that function reads only from its arguments.
1341 This attribute indicates that this function can return twice. The C
1342 ``setjmp`` is an example of such a function. The compiler disables
1343 some optimizations (like tail calls) in the caller of these
1346 This attribute indicates that
1347 `SafeStack <http://clang.llvm.org/docs/SafeStack.html>`_
1348 protection is enabled for this function.
1350 If a function that has a ``safestack`` attribute is inlined into a
1351 function that doesn't have a ``safestack`` attribute or which has an
1352 ``ssp``, ``sspstrong`` or ``sspreq`` attribute, then the resulting
1353 function will have a ``safestack`` attribute.
1354 ``sanitize_address``
1355 This attribute indicates that AddressSanitizer checks
1356 (dynamic address safety analysis) are enabled for this function.
1358 This attribute indicates that MemorySanitizer checks (dynamic detection
1359 of accesses to uninitialized memory) are enabled for this function.
1361 This attribute indicates that ThreadSanitizer checks
1362 (dynamic thread safety analysis) are enabled for this function.
1364 This attribute indicates that the function should emit a stack
1365 smashing protector. It is in the form of a "canary" --- a random value
1366 placed on the stack before the local variables that's checked upon
1367 return from the function to see if it has been overwritten. A
1368 heuristic is used to determine if a function needs stack protectors
1369 or not. The heuristic used will enable protectors for functions with:
1371 - Character arrays larger than ``ssp-buffer-size`` (default 8).
1372 - Aggregates containing character arrays larger than ``ssp-buffer-size``.
1373 - Calls to alloca() with variable sizes or constant sizes greater than
1374 ``ssp-buffer-size``.
1376 Variables that are identified as requiring a protector will be arranged
1377 on the stack such that they are adjacent to the stack protector guard.
1379 If a function that has an ``ssp`` attribute is inlined into a
1380 function that doesn't have an ``ssp`` attribute, then the resulting
1381 function will have an ``ssp`` attribute.
1383 This attribute indicates that the function should *always* emit a
1384 stack smashing protector. This overrides the ``ssp`` function
1387 Variables that are identified as requiring a protector will be arranged
1388 on the stack such that they are adjacent to the stack protector guard.
1389 The specific layout rules are:
1391 #. Large arrays and structures containing large arrays
1392 (``>= ssp-buffer-size``) are closest to the stack protector.
1393 #. Small arrays and structures containing small arrays
1394 (``< ssp-buffer-size``) are 2nd closest to the protector.
1395 #. Variables that have had their address taken are 3rd closest to the
1398 If a function that has an ``sspreq`` attribute is inlined into a
1399 function that doesn't have an ``sspreq`` attribute or which has an
1400 ``ssp`` or ``sspstrong`` attribute, then the resulting function will have
1401 an ``sspreq`` attribute.
1403 This attribute indicates that the function should emit a stack smashing
1404 protector. This attribute causes a strong heuristic to be used when
1405 determining if a function needs stack protectors. The strong heuristic
1406 will enable protectors for functions with:
1408 - Arrays of any size and type
1409 - Aggregates containing an array of any size and type.
1410 - Calls to alloca().
1411 - Local variables that have had their address taken.
1413 Variables that are identified as requiring a protector will be arranged
1414 on the stack such that they are adjacent to the stack protector guard.
1415 The specific layout rules are:
1417 #. Large arrays and structures containing large arrays
1418 (``>= ssp-buffer-size``) are closest to the stack protector.
1419 #. Small arrays and structures containing small arrays
1420 (``< ssp-buffer-size``) are 2nd closest to the protector.
1421 #. Variables that have had their address taken are 3rd closest to the
1424 This overrides the ``ssp`` function attribute.
1426 If a function that has an ``sspstrong`` attribute is inlined into a
1427 function that doesn't have an ``sspstrong`` attribute, then the
1428 resulting function will have an ``sspstrong`` attribute.
1430 This attribute indicates that the function will delegate to some other
1431 function with a tail call. The prototype of a thunk should not be used for
1432 optimization purposes. The caller is expected to cast the thunk prototype to
1433 match the thunk target prototype.
1435 This attribute indicates that the ABI being targeted requires that
1436 an unwind table entry be produced for this function even if we can
1437 show that no exceptions passes by it. This is normally the case for
1438 the ELF x86-64 abi, but it can be disabled for some compilation
1443 Module-Level Inline Assembly
1444 ----------------------------
1446 Modules may contain "module-level inline asm" blocks, which corresponds
1447 to the GCC "file scope inline asm" blocks. These blocks are internally
1448 concatenated by LLVM and treated as a single unit, but may be separated
1449 in the ``.ll`` file if desired. The syntax is very simple:
1451 .. code-block:: llvm
1453 module asm "inline asm code goes here"
1454 module asm "more can go here"
1456 The strings can contain any character by escaping non-printable
1457 characters. The escape sequence used is simply "\\xx" where "xx" is the
1458 two digit hex code for the number.
1460 Note that the assembly string *must* be parseable by LLVM's integrated assembler
1461 (unless it is disabled), even when emitting a ``.s`` file.
1463 .. _langref_datalayout:
1468 A module may specify a target specific data layout string that specifies
1469 how data is to be laid out in memory. The syntax for the data layout is
1472 .. code-block:: llvm
1474 target datalayout = "layout specification"
1476 The *layout specification* consists of a list of specifications
1477 separated by the minus sign character ('-'). Each specification starts
1478 with a letter and may include other information after the letter to
1479 define some aspect of the data layout. The specifications accepted are
1483 Specifies that the target lays out data in big-endian form. That is,
1484 the bits with the most significance have the lowest address
1487 Specifies that the target lays out data in little-endian form. That
1488 is, the bits with the least significance have the lowest address
1491 Specifies the natural alignment of the stack in bits. Alignment
1492 promotion of stack variables is limited to the natural stack
1493 alignment to avoid dynamic stack realignment. The stack alignment
1494 must be a multiple of 8-bits. If omitted, the natural stack
1495 alignment defaults to "unspecified", which does not prevent any
1496 alignment promotions.
1497 ``p[n]:<size>:<abi>:<pref>``
1498 This specifies the *size* of a pointer and its ``<abi>`` and
1499 ``<pref>``\erred alignments for address space ``n``. All sizes are in
1500 bits. The address space, ``n``, is optional, and if not specified,
1501 denotes the default address space 0. The value of ``n`` must be
1502 in the range [1,2^23).
1503 ``i<size>:<abi>:<pref>``
1504 This specifies the alignment for an integer type of a given bit
1505 ``<size>``. The value of ``<size>`` must be in the range [1,2^23).
1506 ``v<size>:<abi>:<pref>``
1507 This specifies the alignment for a vector type of a given bit
1509 ``f<size>:<abi>:<pref>``
1510 This specifies the alignment for a floating point type of a given bit
1511 ``<size>``. Only values of ``<size>`` that are supported by the target
1512 will work. 32 (float) and 64 (double) are supported on all targets; 80
1513 or 128 (different flavors of long double) are also supported on some
1516 This specifies the alignment for an object of aggregate type.
1518 If present, specifies that llvm names are mangled in the output. The
1521 * ``e``: ELF mangling: Private symbols get a ``.L`` prefix.
1522 * ``m``: Mips mangling: Private symbols get a ``$`` prefix.
1523 * ``o``: Mach-O mangling: Private symbols get ``L`` prefix. Other
1524 symbols get a ``_`` prefix.
1525 * ``w``: Windows COFF prefix: Similar to Mach-O, but stdcall and fastcall
1526 functions also get a suffix based on the frame size.
1527 ``n<size1>:<size2>:<size3>...``
1528 This specifies a set of native integer widths for the target CPU in
1529 bits. For example, it might contain ``n32`` for 32-bit PowerPC,
1530 ``n32:64`` for PowerPC 64, or ``n8:16:32:64`` for X86-64. Elements of
1531 this set are considered to support most general arithmetic operations
1534 On every specification that takes a ``<abi>:<pref>``, specifying the
1535 ``<pref>`` alignment is optional. If omitted, the preceding ``:``
1536 should be omitted too and ``<pref>`` will be equal to ``<abi>``.
1538 When constructing the data layout for a given target, LLVM starts with a
1539 default set of specifications which are then (possibly) overridden by
1540 the specifications in the ``datalayout`` keyword. The default
1541 specifications are given in this list:
1543 - ``E`` - big endian
1544 - ``p:64:64:64`` - 64-bit pointers with 64-bit alignment.
1545 - ``p[n]:64:64:64`` - Other address spaces are assumed to be the
1546 same as the default address space.
1547 - ``S0`` - natural stack alignment is unspecified
1548 - ``i1:8:8`` - i1 is 8-bit (byte) aligned
1549 - ``i8:8:8`` - i8 is 8-bit (byte) aligned
1550 - ``i16:16:16`` - i16 is 16-bit aligned
1551 - ``i32:32:32`` - i32 is 32-bit aligned
1552 - ``i64:32:64`` - i64 has ABI alignment of 32-bits but preferred
1553 alignment of 64-bits
1554 - ``f16:16:16`` - half is 16-bit aligned
1555 - ``f32:32:32`` - float is 32-bit aligned
1556 - ``f64:64:64`` - double is 64-bit aligned
1557 - ``f128:128:128`` - quad is 128-bit aligned
1558 - ``v64:64:64`` - 64-bit vector is 64-bit aligned
1559 - ``v128:128:128`` - 128-bit vector is 128-bit aligned
1560 - ``a:0:64`` - aggregates are 64-bit aligned
1562 When LLVM is determining the alignment for a given type, it uses the
1565 #. If the type sought is an exact match for one of the specifications,
1566 that specification is used.
1567 #. If no match is found, and the type sought is an integer type, then
1568 the smallest integer type that is larger than the bitwidth of the
1569 sought type is used. If none of the specifications are larger than
1570 the bitwidth then the largest integer type is used. For example,
1571 given the default specifications above, the i7 type will use the
1572 alignment of i8 (next largest) while both i65 and i256 will use the
1573 alignment of i64 (largest specified).
1574 #. If no match is found, and the type sought is a vector type, then the
1575 largest vector type that is smaller than the sought vector type will
1576 be used as a fall back. This happens because <128 x double> can be
1577 implemented in terms of 64 <2 x double>, for example.
1579 The function of the data layout string may not be what you expect.
1580 Notably, this is not a specification from the frontend of what alignment
1581 the code generator should use.
1583 Instead, if specified, the target data layout is required to match what
1584 the ultimate *code generator* expects. This string is used by the
1585 mid-level optimizers to improve code, and this only works if it matches
1586 what the ultimate code generator uses. There is no way to generate IR
1587 that does not embed this target-specific detail into the IR. If you
1588 don't specify the string, the default specifications will be used to
1589 generate a Data Layout and the optimization phases will operate
1590 accordingly and introduce target specificity into the IR with respect to
1591 these default specifications.
1598 A module may specify a target triple string that describes the target
1599 host. The syntax for the target triple is simply:
1601 .. code-block:: llvm
1603 target triple = "x86_64-apple-macosx10.7.0"
1605 The *target triple* string consists of a series of identifiers delimited
1606 by the minus sign character ('-'). The canonical forms are:
1610 ARCHITECTURE-VENDOR-OPERATING_SYSTEM
1611 ARCHITECTURE-VENDOR-OPERATING_SYSTEM-ENVIRONMENT
1613 This information is passed along to the backend so that it generates
1614 code for the proper architecture. It's possible to override this on the
1615 command line with the ``-mtriple`` command line option.
1617 .. _pointeraliasing:
1619 Pointer Aliasing Rules
1620 ----------------------
1622 Any memory access must be done through a pointer value associated with
1623 an address range of the memory access, otherwise the behavior is
1624 undefined. Pointer values are associated with address ranges according
1625 to the following rules:
1627 - A pointer value is associated with the addresses associated with any
1628 value it is *based* on.
1629 - An address of a global variable is associated with the address range
1630 of the variable's storage.
1631 - The result value of an allocation instruction is associated with the
1632 address range of the allocated storage.
1633 - A null pointer in the default address-space is associated with no
1635 - An integer constant other than zero or a pointer value returned from
1636 a function not defined within LLVM may be associated with address
1637 ranges allocated through mechanisms other than those provided by
1638 LLVM. Such ranges shall not overlap with any ranges of addresses
1639 allocated by mechanisms provided by LLVM.
1641 A pointer value is *based* on another pointer value according to the
1644 - A pointer value formed from a ``getelementptr`` operation is *based*
1645 on the first value operand of the ``getelementptr``.
1646 - The result value of a ``bitcast`` is *based* on the operand of the
1648 - A pointer value formed by an ``inttoptr`` is *based* on all pointer
1649 values that contribute (directly or indirectly) to the computation of
1650 the pointer's value.
1651 - The "*based* on" relationship is transitive.
1653 Note that this definition of *"based"* is intentionally similar to the
1654 definition of *"based"* in C99, though it is slightly weaker.
1656 LLVM IR does not associate types with memory. The result type of a
1657 ``load`` merely indicates the size and alignment of the memory from
1658 which to load, as well as the interpretation of the value. The first
1659 operand type of a ``store`` similarly only indicates the size and
1660 alignment of the store.
1662 Consequently, type-based alias analysis, aka TBAA, aka
1663 ``-fstrict-aliasing``, is not applicable to general unadorned LLVM IR.
1664 :ref:`Metadata <metadata>` may be used to encode additional information
1665 which specialized optimization passes may use to implement type-based
1670 Volatile Memory Accesses
1671 ------------------------
1673 Certain memory accesses, such as :ref:`load <i_load>`'s,
1674 :ref:`store <i_store>`'s, and :ref:`llvm.memcpy <int_memcpy>`'s may be
1675 marked ``volatile``. The optimizers must not change the number of
1676 volatile operations or change their order of execution relative to other
1677 volatile operations. The optimizers *may* change the order of volatile
1678 operations relative to non-volatile operations. This is not Java's
1679 "volatile" and has no cross-thread synchronization behavior.
1681 IR-level volatile loads and stores cannot safely be optimized into
1682 llvm.memcpy or llvm.memmove intrinsics even when those intrinsics are
1683 flagged volatile. Likewise, the backend should never split or merge
1684 target-legal volatile load/store instructions.
1686 .. admonition:: Rationale
1688 Platforms may rely on volatile loads and stores of natively supported
1689 data width to be executed as single instruction. For example, in C
1690 this holds for an l-value of volatile primitive type with native
1691 hardware support, but not necessarily for aggregate types. The
1692 frontend upholds these expectations, which are intentionally
1693 unspecified in the IR. The rules above ensure that IR transformations
1694 do not violate the frontend's contract with the language.
1698 Memory Model for Concurrent Operations
1699 --------------------------------------
1701 The LLVM IR does not define any way to start parallel threads of
1702 execution or to register signal handlers. Nonetheless, there are
1703 platform-specific ways to create them, and we define LLVM IR's behavior
1704 in their presence. This model is inspired by the C++0x memory model.
1706 For a more informal introduction to this model, see the :doc:`Atomics`.
1708 We define a *happens-before* partial order as the least partial order
1711 - Is a superset of single-thread program order, and
1712 - When a *synchronizes-with* ``b``, includes an edge from ``a`` to
1713 ``b``. *Synchronizes-with* pairs are introduced by platform-specific
1714 techniques, like pthread locks, thread creation, thread joining,
1715 etc., and by atomic instructions. (See also :ref:`Atomic Memory Ordering
1716 Constraints <ordering>`).
1718 Note that program order does not introduce *happens-before* edges
1719 between a thread and signals executing inside that thread.
1721 Every (defined) read operation (load instructions, memcpy, atomic
1722 loads/read-modify-writes, etc.) R reads a series of bytes written by
1723 (defined) write operations (store instructions, atomic
1724 stores/read-modify-writes, memcpy, etc.). For the purposes of this
1725 section, initialized globals are considered to have a write of the
1726 initializer which is atomic and happens before any other read or write
1727 of the memory in question. For each byte of a read R, R\ :sub:`byte`
1728 may see any write to the same byte, except:
1730 - If write\ :sub:`1` happens before write\ :sub:`2`, and
1731 write\ :sub:`2` happens before R\ :sub:`byte`, then
1732 R\ :sub:`byte` does not see write\ :sub:`1`.
1733 - If R\ :sub:`byte` happens before write\ :sub:`3`, then
1734 R\ :sub:`byte` does not see write\ :sub:`3`.
1736 Given that definition, R\ :sub:`byte` is defined as follows:
1738 - If R is volatile, the result is target-dependent. (Volatile is
1739 supposed to give guarantees which can support ``sig_atomic_t`` in
1740 C/C++, and may be used for accesses to addresses that do not behave
1741 like normal memory. It does not generally provide cross-thread
1743 - Otherwise, if there is no write to the same byte that happens before
1744 R\ :sub:`byte`, R\ :sub:`byte` returns ``undef`` for that byte.
1745 - Otherwise, if R\ :sub:`byte` may see exactly one write,
1746 R\ :sub:`byte` returns the value written by that write.
1747 - Otherwise, if R is atomic, and all the writes R\ :sub:`byte` may
1748 see are atomic, it chooses one of the values written. See the :ref:`Atomic
1749 Memory Ordering Constraints <ordering>` section for additional
1750 constraints on how the choice is made.
1751 - Otherwise R\ :sub:`byte` returns ``undef``.
1753 R returns the value composed of the series of bytes it read. This
1754 implies that some bytes within the value may be ``undef`` **without**
1755 the entire value being ``undef``. Note that this only defines the
1756 semantics of the operation; it doesn't mean that targets will emit more
1757 than one instruction to read the series of bytes.
1759 Note that in cases where none of the atomic intrinsics are used, this
1760 model places only one restriction on IR transformations on top of what
1761 is required for single-threaded execution: introducing a store to a byte
1762 which might not otherwise be stored is not allowed in general.
1763 (Specifically, in the case where another thread might write to and read
1764 from an address, introducing a store can change a load that may see
1765 exactly one write into a load that may see multiple writes.)
1769 Atomic Memory Ordering Constraints
1770 ----------------------------------
1772 Atomic instructions (:ref:`cmpxchg <i_cmpxchg>`,
1773 :ref:`atomicrmw <i_atomicrmw>`, :ref:`fence <i_fence>`,
1774 :ref:`atomic load <i_load>`, and :ref:`atomic store <i_store>`) take
1775 ordering parameters that determine which other atomic instructions on
1776 the same address they *synchronize with*. These semantics are borrowed
1777 from Java and C++0x, but are somewhat more colloquial. If these
1778 descriptions aren't precise enough, check those specs (see spec
1779 references in the :doc:`atomics guide <Atomics>`).
1780 :ref:`fence <i_fence>` instructions treat these orderings somewhat
1781 differently since they don't take an address. See that instruction's
1782 documentation for details.
1784 For a simpler introduction to the ordering constraints, see the
1788 The set of values that can be read is governed by the happens-before
1789 partial order. A value cannot be read unless some operation wrote
1790 it. This is intended to provide a guarantee strong enough to model
1791 Java's non-volatile shared variables. This ordering cannot be
1792 specified for read-modify-write operations; it is not strong enough
1793 to make them atomic in any interesting way.
1795 In addition to the guarantees of ``unordered``, there is a single
1796 total order for modifications by ``monotonic`` operations on each
1797 address. All modification orders must be compatible with the
1798 happens-before order. There is no guarantee that the modification
1799 orders can be combined to a global total order for the whole program
1800 (and this often will not be possible). The read in an atomic
1801 read-modify-write operation (:ref:`cmpxchg <i_cmpxchg>` and
1802 :ref:`atomicrmw <i_atomicrmw>`) reads the value in the modification
1803 order immediately before the value it writes. If one atomic read
1804 happens before another atomic read of the same address, the later
1805 read must see the same value or a later value in the address's
1806 modification order. This disallows reordering of ``monotonic`` (or
1807 stronger) operations on the same address. If an address is written
1808 ``monotonic``-ally by one thread, and other threads ``monotonic``-ally
1809 read that address repeatedly, the other threads must eventually see
1810 the write. This corresponds to the C++0x/C1x
1811 ``memory_order_relaxed``.
1813 In addition to the guarantees of ``monotonic``, a
1814 *synchronizes-with* edge may be formed with a ``release`` operation.
1815 This is intended to model C++'s ``memory_order_acquire``.
1817 In addition to the guarantees of ``monotonic``, if this operation
1818 writes a value which is subsequently read by an ``acquire``
1819 operation, it *synchronizes-with* that operation. (This isn't a
1820 complete description; see the C++0x definition of a release
1821 sequence.) This corresponds to the C++0x/C1x
1822 ``memory_order_release``.
1823 ``acq_rel`` (acquire+release)
1824 Acts as both an ``acquire`` and ``release`` operation on its
1825 address. This corresponds to the C++0x/C1x ``memory_order_acq_rel``.
1826 ``seq_cst`` (sequentially consistent)
1827 In addition to the guarantees of ``acq_rel`` (``acquire`` for an
1828 operation that only reads, ``release`` for an operation that only
1829 writes), there is a global total order on all
1830 sequentially-consistent operations on all addresses, which is
1831 consistent with the *happens-before* partial order and with the
1832 modification orders of all the affected addresses. Each
1833 sequentially-consistent read sees the last preceding write to the
1834 same address in this global order. This corresponds to the C++0x/C1x
1835 ``memory_order_seq_cst`` and Java volatile.
1839 If an atomic operation is marked ``singlethread``, it only *synchronizes
1840 with* or participates in modification and seq\_cst total orderings with
1841 other operations running in the same thread (for example, in signal
1849 LLVM IR floating-point binary ops (:ref:`fadd <i_fadd>`,
1850 :ref:`fsub <i_fsub>`, :ref:`fmul <i_fmul>`, :ref:`fdiv <i_fdiv>`,
1851 :ref:`frem <i_frem>`, :ref:`fcmp <i_fcmp>`) have the following flags that can
1852 be set to enable otherwise unsafe floating point operations
1855 No NaNs - Allow optimizations to assume the arguments and result are not
1856 NaN. Such optimizations are required to retain defined behavior over
1857 NaNs, but the value of the result is undefined.
1860 No Infs - Allow optimizations to assume the arguments and result are not
1861 +/-Inf. Such optimizations are required to retain defined behavior over
1862 +/-Inf, but the value of the result is undefined.
1865 No Signed Zeros - Allow optimizations to treat the sign of a zero
1866 argument or result as insignificant.
1869 Allow Reciprocal - Allow optimizations to use the reciprocal of an
1870 argument rather than perform division.
1873 Fast - Allow algebraically equivalent transformations that may
1874 dramatically change results in floating point (e.g. reassociate). This
1875 flag implies all the others.
1879 Use-list Order Directives
1880 -------------------------
1882 Use-list directives encode the in-memory order of each use-list, allowing the
1883 order to be recreated. ``<order-indexes>`` is a comma-separated list of
1884 indexes that are assigned to the referenced value's uses. The referenced
1885 value's use-list is immediately sorted by these indexes.
1887 Use-list directives may appear at function scope or global scope. They are not
1888 instructions, and have no effect on the semantics of the IR. When they're at
1889 function scope, they must appear after the terminator of the final basic block.
1891 If basic blocks have their address taken via ``blockaddress()`` expressions,
1892 ``uselistorder_bb`` can be used to reorder their use-lists from outside their
1899 uselistorder <ty> <value>, { <order-indexes> }
1900 uselistorder_bb @function, %block { <order-indexes> }
1906 define void @foo(i32 %arg1, i32 %arg2) {
1908 ; ... instructions ...
1910 ; ... instructions ...
1912 ; At function scope.
1913 uselistorder i32 %arg1, { 1, 0, 2 }
1914 uselistorder label %bb, { 1, 0 }
1918 uselistorder i32* @global, { 1, 2, 0 }
1919 uselistorder i32 7, { 1, 0 }
1920 uselistorder i32 (i32) @bar, { 1, 0 }
1921 uselistorder_bb @foo, %bb, { 5, 1, 3, 2, 0, 4 }
1928 The LLVM type system is one of the most important features of the
1929 intermediate representation. Being typed enables a number of
1930 optimizations to be performed on the intermediate representation
1931 directly, without having to do extra analyses on the side before the
1932 transformation. A strong type system makes it easier to read the
1933 generated code and enables novel analyses and transformations that are
1934 not feasible to perform on normal three address code representations.
1944 The void type does not represent any value and has no size.
1962 The function type can be thought of as a function signature. It consists of a
1963 return type and a list of formal parameter types. The return type of a function
1964 type is a void type or first class type --- except for :ref:`label <t_label>`
1965 and :ref:`metadata <t_metadata>` types.
1971 <returntype> (<parameter list>)
1973 ...where '``<parameter list>``' is a comma-separated list of type
1974 specifiers. Optionally, the parameter list may include a type ``...``, which
1975 indicates that the function takes a variable number of arguments. Variable
1976 argument functions can access their arguments with the :ref:`variable argument
1977 handling intrinsic <int_varargs>` functions. '``<returntype>``' is any type
1978 except :ref:`label <t_label>` and :ref:`metadata <t_metadata>`.
1982 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1983 | ``i32 (i32)`` | function taking an ``i32``, returning an ``i32`` |
1984 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1985 | ``float (i16, i32 *) *`` | :ref:`Pointer <t_pointer>` to a function that takes an ``i16`` and a :ref:`pointer <t_pointer>` to ``i32``, returning ``float``. |
1986 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1987 | ``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. |
1988 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1989 | ``{i32, i32} (i32)`` | A function taking an ``i32``, returning a :ref:`structure <t_struct>` containing two ``i32`` values |
1990 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1997 The :ref:`first class <t_firstclass>` types are perhaps the most important.
1998 Values of these types are the only ones which can be produced by
2006 These are the types that are valid in registers from CodeGen's perspective.
2015 The integer type is a very simple type that simply specifies an
2016 arbitrary bit width for the integer type desired. Any bit width from 1
2017 bit to 2\ :sup:`23`\ -1 (about 8 million) can be specified.
2025 The number of bits the integer will occupy is specified by the ``N``
2031 +----------------+------------------------------------------------+
2032 | ``i1`` | a single-bit integer. |
2033 +----------------+------------------------------------------------+
2034 | ``i32`` | a 32-bit integer. |
2035 +----------------+------------------------------------------------+
2036 | ``i1942652`` | a really big integer of over 1 million bits. |
2037 +----------------+------------------------------------------------+
2041 Floating Point Types
2042 """"""""""""""""""""
2051 - 16-bit floating point value
2054 - 32-bit floating point value
2057 - 64-bit floating point value
2060 - 128-bit floating point value (112-bit mantissa)
2063 - 80-bit floating point value (X87)
2066 - 128-bit floating point value (two 64-bits)
2073 The x86_mmx type represents a value held in an MMX register on an x86
2074 machine. The operations allowed on it are quite limited: parameters and
2075 return values, load and store, and bitcast. User-specified MMX
2076 instructions are represented as intrinsic or asm calls with arguments
2077 and/or results of this type. There are no arrays, vectors or constants
2094 The pointer type is used to specify memory locations. Pointers are
2095 commonly used to reference objects in memory.
2097 Pointer types may have an optional address space attribute defining the
2098 numbered address space where the pointed-to object resides. The default
2099 address space is number zero. The semantics of non-zero address spaces
2100 are target-specific.
2102 Note that LLVM does not permit pointers to void (``void*``) nor does it
2103 permit pointers to labels (``label*``). Use ``i8*`` instead.
2113 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2114 | ``[4 x i32]*`` | A :ref:`pointer <t_pointer>` to :ref:`array <t_array>` of four ``i32`` values. |
2115 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2116 | ``i32 (i32*) *`` | A :ref:`pointer <t_pointer>` to a :ref:`function <t_function>` that takes an ``i32*``, returning an ``i32``. |
2117 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2118 | ``i32 addrspace(5)*`` | A :ref:`pointer <t_pointer>` to an ``i32`` value that resides in address space #5. |
2119 +-------------------------+--------------------------------------------------------------------------------------------------------------+
2128 A vector type is a simple derived type that represents a vector of
2129 elements. Vector types are used when multiple primitive data are
2130 operated in parallel using a single instruction (SIMD). A vector type
2131 requires a size (number of elements) and an underlying primitive data
2132 type. Vector types are considered :ref:`first class <t_firstclass>`.
2138 < <# elements> x <elementtype> >
2140 The number of elements is a constant integer value larger than 0;
2141 elementtype may be any integer, floating point or pointer type. Vectors
2142 of size zero are not allowed.
2146 +-------------------+--------------------------------------------------+
2147 | ``<4 x i32>`` | Vector of 4 32-bit integer values. |
2148 +-------------------+--------------------------------------------------+
2149 | ``<8 x float>`` | Vector of 8 32-bit floating-point values. |
2150 +-------------------+--------------------------------------------------+
2151 | ``<2 x i64>`` | Vector of 2 64-bit integer values. |
2152 +-------------------+--------------------------------------------------+
2153 | ``<4 x i64*>`` | Vector of 4 pointers to 64-bit integer values. |
2154 +-------------------+--------------------------------------------------+
2163 The label type represents code labels.
2178 The token type is used when a value is associated with an instruction
2179 but all uses of the value must not attempt to introspect or obscure it.
2180 As such, it is not appropriate to have a :ref:`phi <i_phi>` or
2181 :ref:`select <i_select>` of type token.
2198 The metadata type represents embedded metadata. No derived types may be
2199 created from metadata except for :ref:`function <t_function>` arguments.
2212 Aggregate Types are a subset of derived types that can contain multiple
2213 member types. :ref:`Arrays <t_array>` and :ref:`structs <t_struct>` are
2214 aggregate types. :ref:`Vectors <t_vector>` are not considered to be
2224 The array type is a very simple derived type that arranges elements
2225 sequentially in memory. The array type requires a size (number of
2226 elements) and an underlying data type.
2232 [<# elements> x <elementtype>]
2234 The number of elements is a constant integer value; ``elementtype`` may
2235 be any type with a size.
2239 +------------------+--------------------------------------+
2240 | ``[40 x i32]`` | Array of 40 32-bit integer values. |
2241 +------------------+--------------------------------------+
2242 | ``[41 x i32]`` | Array of 41 32-bit integer values. |
2243 +------------------+--------------------------------------+
2244 | ``[4 x i8]`` | Array of 4 8-bit integer values. |
2245 +------------------+--------------------------------------+
2247 Here are some examples of multidimensional arrays:
2249 +-----------------------------+----------------------------------------------------------+
2250 | ``[3 x [4 x i32]]`` | 3x4 array of 32-bit integer values. |
2251 +-----------------------------+----------------------------------------------------------+
2252 | ``[12 x [10 x float]]`` | 12x10 array of single precision floating point values. |
2253 +-----------------------------+----------------------------------------------------------+
2254 | ``[2 x [3 x [4 x i16]]]`` | 2x3x4 array of 16-bit integer values. |
2255 +-----------------------------+----------------------------------------------------------+
2257 There is no restriction on indexing beyond the end of the array implied
2258 by a static type (though there are restrictions on indexing beyond the
2259 bounds of an allocated object in some cases). This means that
2260 single-dimension 'variable sized array' addressing can be implemented in
2261 LLVM with a zero length array type. An implementation of 'pascal style
2262 arrays' in LLVM could use the type "``{ i32, [0 x float]}``", for
2272 The structure type is used to represent a collection of data members
2273 together in memory. The elements of a structure may be any type that has
2276 Structures in memory are accessed using '``load``' and '``store``' by
2277 getting a pointer to a field with the '``getelementptr``' instruction.
2278 Structures in registers are accessed using the '``extractvalue``' and
2279 '``insertvalue``' instructions.
2281 Structures may optionally be "packed" structures, which indicate that
2282 the alignment of the struct is one byte, and that there is no padding
2283 between the elements. In non-packed structs, padding between field types
2284 is inserted as defined by the DataLayout string in the module, which is
2285 required to match what the underlying code generator expects.
2287 Structures can either be "literal" or "identified". A literal structure
2288 is defined inline with other types (e.g. ``{i32, i32}*``) whereas
2289 identified types are always defined at the top level with a name.
2290 Literal types are uniqued by their contents and can never be recursive
2291 or opaque since there is no way to write one. Identified types can be
2292 recursive, can be opaqued, and are never uniqued.
2298 %T1 = type { <type list> } ; Identified normal struct type
2299 %T2 = type <{ <type list> }> ; Identified packed struct type
2303 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2304 | ``{ i32, i32, i32 }`` | A triple of three ``i32`` values |
2305 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2306 | ``{ 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``. |
2307 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2308 | ``<{ i8, i32 }>`` | A packed struct known to be 5 bytes in size. |
2309 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
2313 Opaque Structure Types
2314 """"""""""""""""""""""
2318 Opaque structure types are used to represent named structure types that
2319 do not have a body specified. This corresponds (for example) to the C
2320 notion of a forward declared structure.
2331 +--------------+-------------------+
2332 | ``opaque`` | An opaque type. |
2333 +--------------+-------------------+
2340 LLVM has several different basic types of constants. This section
2341 describes them all and their syntax.
2346 **Boolean constants**
2347 The two strings '``true``' and '``false``' are both valid constants
2349 **Integer constants**
2350 Standard integers (such as '4') are constants of the
2351 :ref:`integer <t_integer>` type. Negative numbers may be used with
2353 **Floating point constants**
2354 Floating point constants use standard decimal notation (e.g.
2355 123.421), exponential notation (e.g. 1.23421e+2), or a more precise
2356 hexadecimal notation (see below). The assembler requires the exact
2357 decimal value of a floating-point constant. For example, the
2358 assembler accepts 1.25 but rejects 1.3 because 1.3 is a repeating
2359 decimal in binary. Floating point constants must have a :ref:`floating
2360 point <t_floating>` type.
2361 **Null pointer constants**
2362 The identifier '``null``' is recognized as a null pointer constant
2363 and must be of :ref:`pointer type <t_pointer>`.
2365 The one non-intuitive notation for constants is the hexadecimal form of
2366 floating point constants. For example, the form
2367 '``double 0x432ff973cafa8000``' is equivalent to (but harder to read
2368 than) '``double 4.5e+15``'. The only time hexadecimal floating point
2369 constants are required (and the only time that they are generated by the
2370 disassembler) is when a floating point constant must be emitted but it
2371 cannot be represented as a decimal floating point number in a reasonable
2372 number of digits. For example, NaN's, infinities, and other special
2373 values are represented in their IEEE hexadecimal format so that assembly
2374 and disassembly do not cause any bits to change in the constants.
2376 When using the hexadecimal form, constants of types half, float, and
2377 double are represented using the 16-digit form shown above (which
2378 matches the IEEE754 representation for double); half and float values
2379 must, however, be exactly representable as IEEE 754 half and single
2380 precision, respectively. Hexadecimal format is always used for long
2381 double, and there are three forms of long double. The 80-bit format used
2382 by x86 is represented as ``0xK`` followed by 20 hexadecimal digits. The
2383 128-bit format used by PowerPC (two adjacent doubles) is represented by
2384 ``0xM`` followed by 32 hexadecimal digits. The IEEE 128-bit format is
2385 represented by ``0xL`` followed by 32 hexadecimal digits. Long doubles
2386 will only work if they match the long double format on your target.
2387 The IEEE 16-bit format (half precision) is represented by ``0xH``
2388 followed by 4 hexadecimal digits. All hexadecimal formats are big-endian
2389 (sign bit at the left).
2391 There are no constants of type x86_mmx.
2393 .. _complexconstants:
2398 Complex constants are a (potentially recursive) combination of simple
2399 constants and smaller complex constants.
2401 **Structure constants**
2402 Structure constants are represented with notation similar to
2403 structure type definitions (a comma separated list of elements,
2404 surrounded by braces (``{}``)). For example:
2405 "``{ i32 4, float 17.0, i32* @G }``", where "``@G``" is declared as
2406 "``@G = external global i32``". Structure constants must have
2407 :ref:`structure type <t_struct>`, and the number and types of elements
2408 must match those specified by the type.
2410 Array constants are represented with notation similar to array type
2411 definitions (a comma separated list of elements, surrounded by
2412 square brackets (``[]``)). For example:
2413 "``[ i32 42, i32 11, i32 74 ]``". Array constants must have
2414 :ref:`array type <t_array>`, and the number and types of elements must
2415 match those specified by the type. As a special case, character array
2416 constants may also be represented as a double-quoted string using the ``c``
2417 prefix. For example: "``c"Hello World\0A\00"``".
2418 **Vector constants**
2419 Vector constants are represented with notation similar to vector
2420 type definitions (a comma separated list of elements, surrounded by
2421 less-than/greater-than's (``<>``)). For example:
2422 "``< i32 42, i32 11, i32 74, i32 100 >``". Vector constants
2423 must have :ref:`vector type <t_vector>`, and the number and types of
2424 elements must match those specified by the type.
2425 **Zero initialization**
2426 The string '``zeroinitializer``' can be used to zero initialize a
2427 value to zero of *any* type, including scalar and
2428 :ref:`aggregate <t_aggregate>` types. This is often used to avoid
2429 having to print large zero initializers (e.g. for large arrays) and
2430 is always exactly equivalent to using explicit zero initializers.
2432 A metadata node is a constant tuple without types. For example:
2433 "``!{!0, !{!2, !0}, !"test"}``". Metadata can reference constant values,
2434 for example: "``!{!0, i32 0, i8* @global, i64 (i64)* @function, !"str"}``".
2435 Unlike other typed constants that are meant to be interpreted as part of
2436 the instruction stream, metadata is a place to attach additional
2437 information such as debug info.
2439 Global Variable and Function Addresses
2440 --------------------------------------
2442 The addresses of :ref:`global variables <globalvars>` and
2443 :ref:`functions <functionstructure>` are always implicitly valid
2444 (link-time) constants. These constants are explicitly referenced when
2445 the :ref:`identifier for the global <identifiers>` is used and always have
2446 :ref:`pointer <t_pointer>` type. For example, the following is a legal LLVM
2449 .. code-block:: llvm
2453 @Z = global [2 x i32*] [ i32* @X, i32* @Y ]
2460 The string '``undef``' can be used anywhere a constant is expected, and
2461 indicates that the user of the value may receive an unspecified
2462 bit-pattern. Undefined values may be of any type (other than '``label``'
2463 or '``void``') and be used anywhere a constant is permitted.
2465 Undefined values are useful because they indicate to the compiler that
2466 the program is well defined no matter what value is used. This gives the
2467 compiler more freedom to optimize. Here are some examples of
2468 (potentially surprising) transformations that are valid (in pseudo IR):
2470 .. code-block:: llvm
2480 This is safe because all of the output bits are affected by the undef
2481 bits. Any output bit can have a zero or one depending on the input bits.
2483 .. code-block:: llvm
2494 These logical operations have bits that are not always affected by the
2495 input. For example, if ``%X`` has a zero bit, then the output of the
2496 '``and``' operation will always be a zero for that bit, no matter what
2497 the corresponding bit from the '``undef``' is. As such, it is unsafe to
2498 optimize or assume that the result of the '``and``' is '``undef``'.
2499 However, it is safe to assume that all bits of the '``undef``' could be
2500 0, and optimize the '``and``' to 0. Likewise, it is safe to assume that
2501 all the bits of the '``undef``' operand to the '``or``' could be set,
2502 allowing the '``or``' to be folded to -1.
2504 .. code-block:: llvm
2506 %A = select undef, %X, %Y
2507 %B = select undef, 42, %Y
2508 %C = select %X, %Y, undef
2518 This set of examples shows that undefined '``select``' (and conditional
2519 branch) conditions can go *either way*, but they have to come from one
2520 of the two operands. In the ``%A`` example, if ``%X`` and ``%Y`` were
2521 both known to have a clear low bit, then ``%A`` would have to have a
2522 cleared low bit. However, in the ``%C`` example, the optimizer is
2523 allowed to assume that the '``undef``' operand could be the same as
2524 ``%Y``, allowing the whole '``select``' to be eliminated.
2526 .. code-block:: llvm
2528 %A = xor undef, undef
2545 This example points out that two '``undef``' operands are not
2546 necessarily the same. This can be surprising to people (and also matches
2547 C semantics) where they assume that "``X^X``" is always zero, even if
2548 ``X`` is undefined. This isn't true for a number of reasons, but the
2549 short answer is that an '``undef``' "variable" can arbitrarily change
2550 its value over its "live range". This is true because the variable
2551 doesn't actually *have a live range*. Instead, the value is logically
2552 read from arbitrary registers that happen to be around when needed, so
2553 the value is not necessarily consistent over time. In fact, ``%A`` and
2554 ``%C`` need to have the same semantics or the core LLVM "replace all
2555 uses with" concept would not hold.
2557 .. code-block:: llvm
2565 These examples show the crucial difference between an *undefined value*
2566 and *undefined behavior*. An undefined value (like '``undef``') is
2567 allowed to have an arbitrary bit-pattern. This means that the ``%A``
2568 operation can be constant folded to '``undef``', because the '``undef``'
2569 could be an SNaN, and ``fdiv`` is not (currently) defined on SNaN's.
2570 However, in the second example, we can make a more aggressive
2571 assumption: because the ``undef`` is allowed to be an arbitrary value,
2572 we are allowed to assume that it could be zero. Since a divide by zero
2573 has *undefined behavior*, we are allowed to assume that the operation
2574 does not execute at all. This allows us to delete the divide and all
2575 code after it. Because the undefined operation "can't happen", the
2576 optimizer can assume that it occurs in dead code.
2578 .. code-block:: llvm
2580 a: store undef -> %X
2581 b: store %X -> undef
2586 These examples reiterate the ``fdiv`` example: a store *of* an undefined
2587 value can be assumed to not have any effect; we can assume that the
2588 value is overwritten with bits that happen to match what was already
2589 there. However, a store *to* an undefined location could clobber
2590 arbitrary memory, therefore, it has undefined behavior.
2597 Poison values are similar to :ref:`undef values <undefvalues>`, however
2598 they also represent the fact that an instruction or constant expression
2599 that cannot evoke side effects has nevertheless detected a condition
2600 that results in undefined behavior.
2602 There is currently no way of representing a poison value in the IR; they
2603 only exist when produced by operations such as :ref:`add <i_add>` with
2606 Poison value behavior is defined in terms of value *dependence*:
2608 - Values other than :ref:`phi <i_phi>` nodes depend on their operands.
2609 - :ref:`Phi <i_phi>` nodes depend on the operand corresponding to
2610 their dynamic predecessor basic block.
2611 - Function arguments depend on the corresponding actual argument values
2612 in the dynamic callers of their functions.
2613 - :ref:`Call <i_call>` instructions depend on the :ref:`ret <i_ret>`
2614 instructions that dynamically transfer control back to them.
2615 - :ref:`Invoke <i_invoke>` instructions depend on the
2616 :ref:`ret <i_ret>`, :ref:`resume <i_resume>`, or exception-throwing
2617 call instructions that dynamically transfer control back to them.
2618 - Non-volatile loads and stores depend on the most recent stores to all
2619 of the referenced memory addresses, following the order in the IR
2620 (including loads and stores implied by intrinsics such as
2621 :ref:`@llvm.memcpy <int_memcpy>`.)
2622 - An instruction with externally visible side effects depends on the
2623 most recent preceding instruction with externally visible side
2624 effects, following the order in the IR. (This includes :ref:`volatile
2625 operations <volatile>`.)
2626 - An instruction *control-depends* on a :ref:`terminator
2627 instruction <terminators>` if the terminator instruction has
2628 multiple successors and the instruction is always executed when
2629 control transfers to one of the successors, and may not be executed
2630 when control is transferred to another.
2631 - Additionally, an instruction also *control-depends* on a terminator
2632 instruction if the set of instructions it otherwise depends on would
2633 be different if the terminator had transferred control to a different
2635 - Dependence is transitive.
2637 Poison values have the same behavior as :ref:`undef values <undefvalues>`,
2638 with the additional effect that any instruction that has a *dependence*
2639 on a poison value has undefined behavior.
2641 Here are some examples:
2643 .. code-block:: llvm
2646 %poison = sub nuw i32 0, 1 ; Results in a poison value.
2647 %still_poison = and i32 %poison, 0 ; 0, but also poison.
2648 %poison_yet_again = getelementptr i32, i32* @h, i32 %still_poison
2649 store i32 0, i32* %poison_yet_again ; memory at @h[0] is poisoned
2651 store i32 %poison, i32* @g ; Poison value stored to memory.
2652 %poison2 = load i32, i32* @g ; Poison value loaded back from memory.
2654 store volatile i32 %poison, i32* @g ; External observation; undefined behavior.
2656 %narrowaddr = bitcast i32* @g to i16*
2657 %wideaddr = bitcast i32* @g to i64*
2658 %poison3 = load i16, i16* %narrowaddr ; Returns a poison value.
2659 %poison4 = load i64, i64* %wideaddr ; Returns a poison value.
2661 %cmp = icmp slt i32 %poison, 0 ; Returns a poison value.
2662 br i1 %cmp, label %true, label %end ; Branch to either destination.
2665 store volatile i32 0, i32* @g ; This is control-dependent on %cmp, so
2666 ; it has undefined behavior.
2670 %p = phi i32 [ 0, %entry ], [ 1, %true ]
2671 ; Both edges into this PHI are
2672 ; control-dependent on %cmp, so this
2673 ; always results in a poison value.
2675 store volatile i32 0, i32* @g ; This would depend on the store in %true
2676 ; if %cmp is true, or the store in %entry
2677 ; otherwise, so this is undefined behavior.
2679 br i1 %cmp, label %second_true, label %second_end
2680 ; The same branch again, but this time the
2681 ; true block doesn't have side effects.
2688 store volatile i32 0, i32* @g ; This time, the instruction always depends
2689 ; on the store in %end. Also, it is
2690 ; control-equivalent to %end, so this is
2691 ; well-defined (ignoring earlier undefined
2692 ; behavior in this example).
2696 Addresses of Basic Blocks
2697 -------------------------
2699 ``blockaddress(@function, %block)``
2701 The '``blockaddress``' constant computes the address of the specified
2702 basic block in the specified function, and always has an ``i8*`` type.
2703 Taking the address of the entry block is illegal.
2705 This value only has defined behavior when used as an operand to the
2706 ':ref:`indirectbr <i_indirectbr>`' instruction, or for comparisons
2707 against null. Pointer equality tests between labels addresses results in
2708 undefined behavior --- though, again, comparison against null is ok, and
2709 no label is equal to the null pointer. This may be passed around as an
2710 opaque pointer sized value as long as the bits are not inspected. This
2711 allows ``ptrtoint`` and arithmetic to be performed on these values so
2712 long as the original value is reconstituted before the ``indirectbr``
2715 Finally, some targets may provide defined semantics when using the value
2716 as the operand to an inline assembly, but that is target specific.
2720 Constant Expressions
2721 --------------------
2723 Constant expressions are used to allow expressions involving other
2724 constants to be used as constants. Constant expressions may be of any
2725 :ref:`first class <t_firstclass>` type and may involve any LLVM operation
2726 that does not have side effects (e.g. load and call are not supported).
2727 The following is the syntax for constant expressions:
2729 ``trunc (CST to TYPE)``
2730 Truncate a constant to another type. The bit size of CST must be
2731 larger than the bit size of TYPE. Both types must be integers.
2732 ``zext (CST to TYPE)``
2733 Zero extend a constant to another type. The bit size of CST must be
2734 smaller than the bit size of TYPE. Both types must be integers.
2735 ``sext (CST to TYPE)``
2736 Sign extend a constant to another type. The bit size of CST must be
2737 smaller than the bit size of TYPE. Both types must be integers.
2738 ``fptrunc (CST to TYPE)``
2739 Truncate a floating point constant to another floating point type.
2740 The size of CST must be larger than the size of TYPE. Both types
2741 must be floating point.
2742 ``fpext (CST to TYPE)``
2743 Floating point extend a constant to another type. The size of CST
2744 must be smaller or equal to the size of TYPE. Both types must be
2746 ``fptoui (CST to TYPE)``
2747 Convert a floating point constant to the corresponding unsigned
2748 integer constant. TYPE must be a scalar or vector integer type. CST
2749 must be of scalar or vector floating point type. Both CST and TYPE
2750 must be scalars, or vectors of the same number of elements. If the
2751 value won't fit in the integer type, the results are undefined.
2752 ``fptosi (CST to TYPE)``
2753 Convert a floating point constant to the corresponding signed
2754 integer constant. TYPE must be a scalar or vector integer type. CST
2755 must be of scalar or vector floating point type. Both CST and TYPE
2756 must be scalars, or vectors of the same number of elements. If the
2757 value won't fit in the integer type, the results are undefined.
2758 ``uitofp (CST to TYPE)``
2759 Convert an unsigned integer constant to the corresponding floating
2760 point constant. TYPE must be a scalar or vector floating point type.
2761 CST must be of scalar or vector integer type. Both CST and TYPE must
2762 be scalars, or vectors of the same number of elements. If the value
2763 won't fit in the floating point type, the results are undefined.
2764 ``sitofp (CST to TYPE)``
2765 Convert a signed integer constant to the corresponding floating
2766 point constant. TYPE must be a scalar or vector floating point type.
2767 CST must be of scalar or vector integer type. Both CST and TYPE must
2768 be scalars, or vectors of the same number of elements. If the value
2769 won't fit in the floating point type, the results are undefined.
2770 ``ptrtoint (CST to TYPE)``
2771 Convert a pointer typed constant to the corresponding integer
2772 constant. ``TYPE`` must be an integer type. ``CST`` must be of
2773 pointer type. The ``CST`` value is zero extended, truncated, or
2774 unchanged to make it fit in ``TYPE``.
2775 ``inttoptr (CST to TYPE)``
2776 Convert an integer constant to a pointer constant. TYPE must be a
2777 pointer type. CST must be of integer type. The CST value is zero
2778 extended, truncated, or unchanged to make it fit in a pointer size.
2779 This one is *really* dangerous!
2780 ``bitcast (CST to TYPE)``
2781 Convert a constant, CST, to another TYPE. The constraints of the
2782 operands are the same as those for the :ref:`bitcast
2783 instruction <i_bitcast>`.
2784 ``addrspacecast (CST to TYPE)``
2785 Convert a constant pointer or constant vector of pointer, CST, to another
2786 TYPE in a different address space. The constraints of the operands are the
2787 same as those for the :ref:`addrspacecast instruction <i_addrspacecast>`.
2788 ``getelementptr (TY, CSTPTR, IDX0, IDX1, ...)``, ``getelementptr inbounds (TY, CSTPTR, IDX0, IDX1, ...)``
2789 Perform the :ref:`getelementptr operation <i_getelementptr>` on
2790 constants. As with the :ref:`getelementptr <i_getelementptr>`
2791 instruction, the index list may have zero or more indexes, which are
2792 required to make sense for the type of "pointer to TY".
2793 ``select (COND, VAL1, VAL2)``
2794 Perform the :ref:`select operation <i_select>` on constants.
2795 ``icmp COND (VAL1, VAL2)``
2796 Performs the :ref:`icmp operation <i_icmp>` on constants.
2797 ``fcmp COND (VAL1, VAL2)``
2798 Performs the :ref:`fcmp operation <i_fcmp>` on constants.
2799 ``extractelement (VAL, IDX)``
2800 Perform the :ref:`extractelement operation <i_extractelement>` on
2802 ``insertelement (VAL, ELT, IDX)``
2803 Perform the :ref:`insertelement operation <i_insertelement>` on
2805 ``shufflevector (VEC1, VEC2, IDXMASK)``
2806 Perform the :ref:`shufflevector operation <i_shufflevector>` on
2808 ``extractvalue (VAL, IDX0, IDX1, ...)``
2809 Perform the :ref:`extractvalue operation <i_extractvalue>` on
2810 constants. The index list is interpreted in a similar manner as
2811 indices in a ':ref:`getelementptr <i_getelementptr>`' operation. At
2812 least one index value must be specified.
2813 ``insertvalue (VAL, ELT, IDX0, IDX1, ...)``
2814 Perform the :ref:`insertvalue operation <i_insertvalue>` on constants.
2815 The index list is interpreted in a similar manner as indices in a
2816 ':ref:`getelementptr <i_getelementptr>`' operation. At least one index
2817 value must be specified.
2818 ``OPCODE (LHS, RHS)``
2819 Perform the specified operation of the LHS and RHS constants. OPCODE
2820 may be any of the :ref:`binary <binaryops>` or :ref:`bitwise
2821 binary <bitwiseops>` operations. The constraints on operands are
2822 the same as those for the corresponding instruction (e.g. no bitwise
2823 operations on floating point values are allowed).
2830 Inline Assembler Expressions
2831 ----------------------------
2833 LLVM supports inline assembler expressions (as opposed to :ref:`Module-Level
2834 Inline Assembly <moduleasm>`) through the use of a special value. This value
2835 represents the inline assembler as a template string (containing the
2836 instructions to emit), a list of operand constraints (stored as a string), a
2837 flag that indicates whether or not the inline asm expression has side effects,
2838 and a flag indicating whether the function containing the asm needs to align its
2839 stack conservatively.
2841 The template string supports argument substitution of the operands using "``$``"
2842 followed by a number, to indicate substitution of the given register/memory
2843 location, as specified by the constraint string. "``${NUM:MODIFIER}``" may also
2844 be used, where ``MODIFIER`` is a target-specific annotation for how to print the
2845 operand (See :ref:`inline-asm-modifiers`).
2847 A literal "``$``" may be included by using "``$$``" in the template. To include
2848 other special characters into the output, the usual "``\XX``" escapes may be
2849 used, just as in other strings. Note that after template substitution, the
2850 resulting assembly string is parsed by LLVM's integrated assembler unless it is
2851 disabled -- even when emitting a ``.s`` file -- and thus must contain assembly
2852 syntax known to LLVM.
2854 LLVM's support for inline asm is modeled closely on the requirements of Clang's
2855 GCC-compatible inline-asm support. Thus, the feature-set and the constraint and
2856 modifier codes listed here are similar or identical to those in GCC's inline asm
2857 support. However, to be clear, the syntax of the template and constraint strings
2858 described here is *not* the same as the syntax accepted by GCC and Clang, and,
2859 while most constraint letters are passed through as-is by Clang, some get
2860 translated to other codes when converting from the C source to the LLVM
2863 An example inline assembler expression is:
2865 .. code-block:: llvm
2867 i32 (i32) asm "bswap $0", "=r,r"
2869 Inline assembler expressions may **only** be used as the callee operand
2870 of a :ref:`call <i_call>` or an :ref:`invoke <i_invoke>` instruction.
2871 Thus, typically we have:
2873 .. code-block:: llvm
2875 %X = call i32 asm "bswap $0", "=r,r"(i32 %Y)
2877 Inline asms with side effects not visible in the constraint list must be
2878 marked as having side effects. This is done through the use of the
2879 '``sideeffect``' keyword, like so:
2881 .. code-block:: llvm
2883 call void asm sideeffect "eieio", ""()
2885 In some cases inline asms will contain code that will not work unless
2886 the stack is aligned in some way, such as calls or SSE instructions on
2887 x86, yet will not contain code that does that alignment within the asm.
2888 The compiler should make conservative assumptions about what the asm
2889 might contain and should generate its usual stack alignment code in the
2890 prologue if the '``alignstack``' keyword is present:
2892 .. code-block:: llvm
2894 call void asm alignstack "eieio", ""()
2896 Inline asms also support using non-standard assembly dialects. The
2897 assumed dialect is ATT. When the '``inteldialect``' keyword is present,
2898 the inline asm is using the Intel dialect. Currently, ATT and Intel are
2899 the only supported dialects. An example is:
2901 .. code-block:: llvm
2903 call void asm inteldialect "eieio", ""()
2905 If multiple keywords appear the '``sideeffect``' keyword must come
2906 first, the '``alignstack``' keyword second and the '``inteldialect``'
2909 Inline Asm Constraint String
2910 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2912 The constraint list is a comma-separated string, each element containing one or
2913 more constraint codes.
2915 For each element in the constraint list an appropriate register or memory
2916 operand will be chosen, and it will be made available to assembly template
2917 string expansion as ``$0`` for the first constraint in the list, ``$1`` for the
2920 There are three different types of constraints, which are distinguished by a
2921 prefix symbol in front of the constraint code: Output, Input, and Clobber. The
2922 constraints must always be given in that order: outputs first, then inputs, then
2923 clobbers. They cannot be intermingled.
2925 There are also three different categories of constraint codes:
2927 - Register constraint. This is either a register class, or a fixed physical
2928 register. This kind of constraint will allocate a register, and if necessary,
2929 bitcast the argument or result to the appropriate type.
2930 - Memory constraint. This kind of constraint is for use with an instruction
2931 taking a memory operand. Different constraints allow for different addressing
2932 modes used by the target.
2933 - Immediate value constraint. This kind of constraint is for an integer or other
2934 immediate value which can be rendered directly into an instruction. The
2935 various target-specific constraints allow the selection of a value in the
2936 proper range for the instruction you wish to use it with.
2941 Output constraints are specified by an "``=``" prefix (e.g. "``=r``"). This
2942 indicates that the assembly will write to this operand, and the operand will
2943 then be made available as a return value of the ``asm`` expression. Output
2944 constraints do not consume an argument from the call instruction. (Except, see
2945 below about indirect outputs).
2947 Normally, it is expected that no output locations are written to by the assembly
2948 expression until *all* of the inputs have been read. As such, LLVM may assign
2949 the same register to an output and an input. If this is not safe (e.g. if the
2950 assembly contains two instructions, where the first writes to one output, and
2951 the second reads an input and writes to a second output), then the "``&``"
2952 modifier must be used (e.g. "``=&r``") to specify that the output is an
2953 "early-clobber" output. Marking an ouput as "early-clobber" ensures that LLVM
2954 will not use the same register for any inputs (other than an input tied to this
2960 Input constraints do not have a prefix -- just the constraint codes. Each input
2961 constraint will consume one argument from the call instruction. It is not
2962 permitted for the asm to write to any input register or memory location (unless
2963 that input is tied to an output). Note also that multiple inputs may all be
2964 assigned to the same register, if LLVM can determine that they necessarily all
2965 contain the same value.
2967 Instead of providing a Constraint Code, input constraints may also "tie"
2968 themselves to an output constraint, by providing an integer as the constraint
2969 string. Tied inputs still consume an argument from the call instruction, and
2970 take up a position in the asm template numbering as is usual -- they will simply
2971 be constrained to always use the same register as the output they've been tied
2972 to. For example, a constraint string of "``=r,0``" says to assign a register for
2973 output, and use that register as an input as well (it being the 0'th
2976 It is permitted to tie an input to an "early-clobber" output. In that case, no
2977 *other* input may share the same register as the input tied to the early-clobber
2978 (even when the other input has the same value).
2980 You may only tie an input to an output which has a register constraint, not a
2981 memory constraint. Only a single input may be tied to an output.
2983 There is also an "interesting" feature which deserves a bit of explanation: if a
2984 register class constraint allocates a register which is too small for the value
2985 type operand provided as input, the input value will be split into multiple
2986 registers, and all of them passed to the inline asm.
2988 However, this feature is often not as useful as you might think.
2990 Firstly, the registers are *not* guaranteed to be consecutive. So, on those
2991 architectures that have instructions which operate on multiple consecutive
2992 instructions, this is not an appropriate way to support them. (e.g. the 32-bit
2993 SparcV8 has a 64-bit load, which instruction takes a single 32-bit register. The
2994 hardware then loads into both the named register, and the next register. This
2995 feature of inline asm would not be useful to support that.)
2997 A few of the targets provide a template string modifier allowing explicit access
2998 to the second register of a two-register operand (e.g. MIPS ``L``, ``M``, and
2999 ``D``). On such an architecture, you can actually access the second allocated
3000 register (yet, still, not any subsequent ones). But, in that case, you're still
3001 probably better off simply splitting the value into two separate operands, for
3002 clarity. (e.g. see the description of the ``A`` constraint on X86, which,
3003 despite existing only for use with this feature, is not really a good idea to
3006 Indirect inputs and outputs
3007 """""""""""""""""""""""""""
3009 Indirect output or input constraints can be specified by the "``*``" modifier
3010 (which goes after the "``=``" in case of an output). This indicates that the asm
3011 will write to or read from the contents of an *address* provided as an input
3012 argument. (Note that in this way, indirect outputs act more like an *input* than
3013 an output: just like an input, they consume an argument of the call expression,
3014 rather than producing a return value. An indirect output constraint is an
3015 "output" only in that the asm is expected to write to the contents of the input
3016 memory location, instead of just read from it).
3018 This is most typically used for memory constraint, e.g. "``=*m``", to pass the
3019 address of a variable as a value.
3021 It is also possible to use an indirect *register* constraint, but only on output
3022 (e.g. "``=*r``"). This will cause LLVM to allocate a register for an output
3023 value normally, and then, separately emit a store to the address provided as
3024 input, after the provided inline asm. (It's not clear what value this
3025 functionality provides, compared to writing the store explicitly after the asm
3026 statement, and it can only produce worse code, since it bypasses many
3027 optimization passes. I would recommend not using it.)
3033 A clobber constraint is indicated by a "``~``" prefix. A clobber does not
3034 consume an input operand, nor generate an output. Clobbers cannot use any of the
3035 general constraint code letters -- they may use only explicit register
3036 constraints, e.g. "``~{eax}``". The one exception is that a clobber string of
3037 "``~{memory}``" indicates that the assembly writes to arbitrary undeclared
3038 memory locations -- not only the memory pointed to by a declared indirect
3044 After a potential prefix comes constraint code, or codes.
3046 A Constraint Code is either a single letter (e.g. "``r``"), a "``^``" character
3047 followed by two letters (e.g. "``^wc``"), or "``{``" register-name "``}``"
3050 The one and two letter constraint codes are typically chosen to be the same as
3051 GCC's constraint codes.
3053 A single constraint may include one or more than constraint code in it, leaving
3054 it up to LLVM to choose which one to use. This is included mainly for
3055 compatibility with the translation of GCC inline asm coming from clang.
3057 There are two ways to specify alternatives, and either or both may be used in an
3058 inline asm constraint list:
3060 1) Append the codes to each other, making a constraint code set. E.g. "``im``"
3061 or "``{eax}m``". This means "choose any of the options in the set". The
3062 choice of constraint is made independently for each constraint in the
3065 2) Use "``|``" between constraint code sets, creating alternatives. Every
3066 constraint in the constraint list must have the same number of alternative
3067 sets. With this syntax, the same alternative in *all* of the items in the
3068 constraint list will be chosen together.
3070 Putting those together, you might have a two operand constraint string like
3071 ``"rm|r,ri|rm"``. This indicates that if operand 0 is ``r`` or ``m``, then
3072 operand 1 may be one of ``r`` or ``i``. If operand 0 is ``r``, then operand 1
3073 may be one of ``r`` or ``m``. But, operand 0 and 1 cannot both be of type m.
3075 However, the use of either of the alternatives features is *NOT* recommended, as
3076 LLVM is not able to make an intelligent choice about which one to use. (At the
3077 point it currently needs to choose, not enough information is available to do so
3078 in a smart way.) Thus, it simply tries to make a choice that's most likely to
3079 compile, not one that will be optimal performance. (e.g., given "``rm``", it'll
3080 always choose to use memory, not registers). And, if given multiple registers,
3081 or multiple register classes, it will simply choose the first one. (In fact, it
3082 doesn't currently even ensure explicitly specified physical registers are
3083 unique, so specifying multiple physical registers as alternatives, like
3084 ``{r11}{r12},{r11}{r12}``, will assign r11 to both operands, not at all what was
3087 Supported Constraint Code List
3088 """"""""""""""""""""""""""""""
3090 The constraint codes are, in general, expected to behave the same way they do in
3091 GCC. LLVM's support is often implemented on an 'as-needed' basis, to support C
3092 inline asm code which was supported by GCC. A mismatch in behavior between LLVM
3093 and GCC likely indicates a bug in LLVM.
3095 Some constraint codes are typically supported by all targets:
3097 - ``r``: A register in the target's general purpose register class.
3098 - ``m``: A memory address operand. It is target-specific what addressing modes
3099 are supported, typical examples are register, or register + register offset,
3100 or register + immediate offset (of some target-specific size).
3101 - ``i``: An integer constant (of target-specific width). Allows either a simple
3102 immediate, or a relocatable value.
3103 - ``n``: An integer constant -- *not* including relocatable values.
3104 - ``s``: An integer constant, but allowing *only* relocatable values.
3105 - ``X``: Allows an operand of any kind, no constraint whatsoever. Typically
3106 useful to pass a label for an asm branch or call.
3108 .. FIXME: but that surely isn't actually okay to jump out of an asm
3109 block without telling llvm about the control transfer???)
3111 - ``{register-name}``: Requires exactly the named physical register.
3113 Other constraints are target-specific:
3117 - ``z``: An immediate integer 0. Outputs ``WZR`` or ``XZR``, as appropriate.
3118 - ``I``: An immediate integer valid for an ``ADD`` or ``SUB`` instruction,
3119 i.e. 0 to 4095 with optional shift by 12.
3120 - ``J``: An immediate integer that, when negated, is valid for an ``ADD`` or
3121 ``SUB`` instruction, i.e. -1 to -4095 with optional left shift by 12.
3122 - ``K``: An immediate integer that is valid for the 'bitmask immediate 32' of a
3123 logical instruction like ``AND``, ``EOR``, or ``ORR`` with a 32-bit register.
3124 - ``L``: An immediate integer that is valid for the 'bitmask immediate 64' of a
3125 logical instruction like ``AND``, ``EOR``, or ``ORR`` with a 64-bit register.
3126 - ``M``: An immediate integer for use with the ``MOV`` assembly alias on a
3127 32-bit register. This is a superset of ``K``: in addition to the bitmask
3128 immediate, also allows immediate integers which can be loaded with a single
3129 ``MOVZ`` or ``MOVL`` instruction.
3130 - ``N``: An immediate integer for use with the ``MOV`` assembly alias on a
3131 64-bit register. This is a superset of ``L``.
3132 - ``Q``: Memory address operand must be in a single register (no
3133 offsets). (However, LLVM currently does this for the ``m`` constraint as
3135 - ``r``: A 32 or 64-bit integer register (W* or X*).
3136 - ``w``: A 32, 64, or 128-bit floating-point/SIMD register.
3137 - ``x``: A lower 128-bit floating-point/SIMD register (``V0`` to ``V15``).
3141 - ``r``: A 32 or 64-bit integer register.
3142 - ``[0-9]v``: The 32-bit VGPR register, number 0-9.
3143 - ``[0-9]s``: The 32-bit SGPR register, number 0-9.
3148 - ``Q``, ``Um``, ``Un``, ``Uq``, ``Us``, ``Ut``, ``Uv``, ``Uy``: Memory address
3149 operand. Treated the same as operand ``m``, at the moment.
3151 ARM and ARM's Thumb2 mode:
3153 - ``j``: An immediate integer between 0 and 65535 (valid for ``MOVW``)
3154 - ``I``: An immediate integer valid for a data-processing instruction.
3155 - ``J``: An immediate integer between -4095 and 4095.
3156 - ``K``: An immediate integer whose bitwise inverse is valid for a
3157 data-processing instruction. (Can be used with template modifier "``B``" to
3158 print the inverted value).
3159 - ``L``: An immediate integer whose negation is valid for a data-processing
3160 instruction. (Can be used with template modifier "``n``" to print the negated
3162 - ``M``: A power of two or a integer between 0 and 32.
3163 - ``N``: Invalid immediate constraint.
3164 - ``O``: Invalid immediate constraint.
3165 - ``r``: A general-purpose 32-bit integer register (``r0-r15``).
3166 - ``l``: In Thumb2 mode, low 32-bit GPR registers (``r0-r7``). In ARM mode, same
3168 - ``h``: In Thumb2 mode, a high 32-bit GPR register (``r8-r15``). In ARM mode,
3170 - ``w``: A 32, 64, or 128-bit floating-point/SIMD register: ``s0-s31``,
3171 ``d0-d31``, or ``q0-q15``.
3172 - ``x``: A 32, 64, or 128-bit floating-point/SIMD register: ``s0-s15``,
3173 ``d0-d7``, or ``q0-q3``.
3174 - ``t``: A floating-point/SIMD register, only supports 32-bit values:
3179 - ``I``: An immediate integer between 0 and 255.
3180 - ``J``: An immediate integer between -255 and -1.
3181 - ``K``: An immediate integer between 0 and 255, with optional left-shift by
3183 - ``L``: An immediate integer between -7 and 7.
3184 - ``M``: An immediate integer which is a multiple of 4 between 0 and 1020.
3185 - ``N``: An immediate integer between 0 and 31.
3186 - ``O``: An immediate integer which is a multiple of 4 between -508 and 508.
3187 - ``r``: A low 32-bit GPR register (``r0-r7``).
3188 - ``l``: A low 32-bit GPR register (``r0-r7``).
3189 - ``h``: A high GPR register (``r0-r7``).
3190 - ``w``: A 32, 64, or 128-bit floating-point/SIMD register: ``s0-s31``,
3191 ``d0-d31``, or ``q0-q15``.
3192 - ``x``: A 32, 64, or 128-bit floating-point/SIMD register: ``s0-s15``,
3193 ``d0-d7``, or ``q0-q3``.
3194 - ``t``: A floating-point/SIMD register, only supports 32-bit values:
3200 - ``o``, ``v``: A memory address operand, treated the same as constraint ``m``,
3202 - ``r``: A 32 or 64-bit register.
3206 - ``r``: An 8 or 16-bit register.
3210 - ``I``: An immediate signed 16-bit integer.
3211 - ``J``: An immediate integer zero.
3212 - ``K``: An immediate unsigned 16-bit integer.
3213 - ``L``: An immediate 32-bit integer, where the lower 16 bits are 0.
3214 - ``N``: An immediate integer between -65535 and -1.
3215 - ``O``: An immediate signed 15-bit integer.
3216 - ``P``: An immediate integer between 1 and 65535.
3217 - ``m``: A memory address operand. In MIPS-SE mode, allows a base address
3218 register plus 16-bit immediate offset. In MIPS mode, just a base register.
3219 - ``R``: A memory address operand. In MIPS-SE mode, allows a base address
3220 register plus a 9-bit signed offset. In MIPS mode, the same as constraint
3222 - ``ZC``: A memory address operand, suitable for use in a ``pref``, ``ll``, or
3223 ``sc`` instruction on the given subtarget (details vary).
3224 - ``r``, ``d``, ``y``: A 32 or 64-bit GPR register.
3225 - ``f``: A 32 or 64-bit FPU register (``F0-F31``), or a 128-bit MSA register
3226 (``W0-W31``). In the case of MSA registers, it is recommended to use the ``w``
3227 argument modifier for compatibility with GCC.
3228 - ``c``: A 32-bit or 64-bit GPR register suitable for indirect jump (always
3230 - ``l``: The ``lo`` register, 32 or 64-bit.
3235 - ``b``: A 1-bit integer register.
3236 - ``c`` or ``h``: A 16-bit integer register.
3237 - ``r``: A 32-bit integer register.
3238 - ``l`` or ``N``: A 64-bit integer register.
3239 - ``f``: A 32-bit float register.
3240 - ``d``: A 64-bit float register.
3245 - ``I``: An immediate signed 16-bit integer.
3246 - ``J``: An immediate unsigned 16-bit integer, shifted left 16 bits.
3247 - ``K``: An immediate unsigned 16-bit integer.
3248 - ``L``: An immediate signed 16-bit integer, shifted left 16 bits.
3249 - ``M``: An immediate integer greater than 31.
3250 - ``N``: An immediate integer that is an exact power of 2.
3251 - ``O``: The immediate integer constant 0.
3252 - ``P``: An immediate integer constant whose negation is a signed 16-bit
3254 - ``es``, ``o``, ``Q``, ``Z``, ``Zy``: A memory address operand, currently
3255 treated the same as ``m``.
3256 - ``r``: A 32 or 64-bit integer register.
3257 - ``b``: A 32 or 64-bit integer register, excluding ``R0`` (that is:
3259 - ``f``: A 32 or 64-bit float register (``F0-F31``), or when QPX is enabled, a
3260 128 or 256-bit QPX register (``Q0-Q31``; aliases the ``F`` registers).
3261 - ``v``: For ``4 x f32`` or ``4 x f64`` types, when QPX is enabled, a
3262 128 or 256-bit QPX register (``Q0-Q31``), otherwise a 128-bit
3263 altivec vector register (``V0-V31``).
3265 .. FIXME: is this a bug that v accepts QPX registers? I think this
3266 is supposed to only use the altivec vector registers?
3268 - ``y``: Condition register (``CR0-CR7``).
3269 - ``wc``: An individual CR bit in a CR register.
3270 - ``wa``, ``wd``, ``wf``: Any 128-bit VSX vector register, from the full VSX
3271 register set (overlapping both the floating-point and vector register files).
3272 - ``ws``: A 32 or 64-bit floating point register, from the full VSX register
3277 - ``I``: An immediate 13-bit signed integer.
3278 - ``r``: A 32-bit integer register.
3282 - ``I``: An immediate unsigned 8-bit integer.
3283 - ``J``: An immediate unsigned 12-bit integer.
3284 - ``K``: An immediate signed 16-bit integer.
3285 - ``L``: An immediate signed 20-bit integer.
3286 - ``M``: An immediate integer 0x7fffffff.
3287 - ``Q``, ``R``, ``S``, ``T``: A memory address operand, treated the same as
3288 ``m``, at the moment.
3289 - ``r`` or ``d``: A 32, 64, or 128-bit integer register.
3290 - ``a``: A 32, 64, or 128-bit integer address register (excludes R0, which in an
3291 address context evaluates as zero).
3292 - ``h``: A 32-bit value in the high part of a 64bit data register
3294 - ``f``: A 32, 64, or 128-bit floating point register.
3298 - ``I``: An immediate integer between 0 and 31.
3299 - ``J``: An immediate integer between 0 and 64.
3300 - ``K``: An immediate signed 8-bit integer.
3301 - ``L``: An immediate integer, 0xff or 0xffff or (in 64-bit mode only)
3303 - ``M``: An immediate integer between 0 and 3.
3304 - ``N``: An immediate unsigned 8-bit integer.
3305 - ``O``: An immediate integer between 0 and 127.
3306 - ``e``: An immediate 32-bit signed integer.
3307 - ``Z``: An immediate 32-bit unsigned integer.
3308 - ``o``, ``v``: Treated the same as ``m``, at the moment.
3309 - ``q``: An 8, 16, 32, or 64-bit register which can be accessed as an 8-bit
3310 ``l`` integer register. On X86-32, this is the ``a``, ``b``, ``c``, and ``d``
3311 registers, and on X86-64, it is all of the integer registers.
3312 - ``Q``: An 8, 16, 32, or 64-bit register which can be accessed as an 8-bit
3313 ``h`` integer register. This is the ``a``, ``b``, ``c``, and ``d`` registers.
3314 - ``r`` or ``l``: An 8, 16, 32, or 64-bit integer register.
3315 - ``R``: An 8, 16, 32, or 64-bit "legacy" integer register -- one which has
3316 existed since i386, and can be accessed without the REX prefix.
3317 - ``f``: A 32, 64, or 80-bit '387 FPU stack pseudo-register.
3318 - ``y``: A 64-bit MMX register, if MMX is enabled.
3319 - ``x``: If SSE is enabled: a 32 or 64-bit scalar operand, or 128-bit vector
3320 operand in a SSE register. If AVX is also enabled, can also be a 256-bit
3321 vector operand in an AVX register. If AVX-512 is also enabled, can also be a
3322 512-bit vector operand in an AVX512 register, Otherwise, an error.
3323 - ``Y``: The same as ``x``, if *SSE2* is enabled, otherwise an error.
3324 - ``A``: Special case: allocates EAX first, then EDX, for a single operand (in
3325 32-bit mode, a 64-bit integer operand will get split into two registers). It
3326 is not recommended to use this constraint, as in 64-bit mode, the 64-bit
3327 operand will get allocated only to RAX -- if two 32-bit operands are needed,
3328 you're better off splitting it yourself, before passing it to the asm
3333 - ``r``: A 32-bit integer register.
3336 .. _inline-asm-modifiers:
3338 Asm template argument modifiers
3339 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3341 In the asm template string, modifiers can be used on the operand reference, like
3344 The modifiers are, in general, expected to behave the same way they do in
3345 GCC. LLVM's support is often implemented on an 'as-needed' basis, to support C
3346 inline asm code which was supported by GCC. A mismatch in behavior between LLVM
3347 and GCC likely indicates a bug in LLVM.
3351 - ``c``: Print an immediate integer constant unadorned, without
3352 the target-specific immediate punctuation (e.g. no ``$`` prefix).
3353 - ``n``: Negate and print immediate integer constant unadorned, without the
3354 target-specific immediate punctuation (e.g. no ``$`` prefix).
3355 - ``l``: Print as an unadorned label, without the target-specific label
3356 punctuation (e.g. no ``$`` prefix).
3360 - ``w``: Print a GPR register with a ``w*`` name instead of ``x*`` name. E.g.,
3361 instead of ``x30``, print ``w30``.
3362 - ``x``: Print a GPR register with a ``x*`` name. (this is the default, anyhow).
3363 - ``b``, ``h``, ``s``, ``d``, ``q``: Print a floating-point/SIMD register with a
3364 ``b*``, ``h*``, ``s*``, ``d*``, or ``q*`` name, rather than the default of
3373 - ``a``: Print an operand as an address (with ``[`` and ``]`` surrounding a
3377 - ``y``: Print a VFP single-precision register as an indexed double (e.g. print
3378 as ``d4[1]`` instead of ``s9``)
3379 - ``B``: Bitwise invert and print an immediate integer constant without ``#``
3381 - ``L``: Print the low 16-bits of an immediate integer constant.
3382 - ``M``: Print as a register set suitable for ldm/stm. Also prints *all*
3383 register operands subsequent to the specified one (!), so use carefully.
3384 - ``Q``: Print the low-order register of a register-pair, or the low-order
3385 register of a two-register operand.
3386 - ``R``: Print the high-order register of a register-pair, or the high-order
3387 register of a two-register operand.
3388 - ``H``: Print the second register of a register-pair. (On a big-endian system,
3389 ``H`` is equivalent to ``Q``, and on little-endian system, ``H`` is equivalent
3392 .. FIXME: H doesn't currently support printing the second register
3393 of a two-register operand.
3395 - ``e``: Print the low doubleword register of a NEON quad register.
3396 - ``f``: Print the high doubleword register of a NEON quad register.
3397 - ``m``: Print the base register of a memory operand without the ``[`` and ``]``
3402 - ``L``: Print the second register of a two-register operand. Requires that it
3403 has been allocated consecutively to the first.
3405 .. FIXME: why is it restricted to consecutive ones? And there's
3406 nothing that ensures that happens, is there?
3408 - ``I``: Print the letter 'i' if the operand is an integer constant, otherwise
3409 nothing. Used to print 'addi' vs 'add' instructions.
3413 No additional modifiers.
3417 - ``X``: Print an immediate integer as hexadecimal
3418 - ``x``: Print the low 16 bits of an immediate integer as hexadecimal.
3419 - ``d``: Print an immediate integer as decimal.
3420 - ``m``: Subtract one and print an immediate integer as decimal.
3421 - ``z``: Print $0 if an immediate zero, otherwise print normally.
3422 - ``L``: Print the low-order register of a two-register operand, or prints the
3423 address of the low-order word of a double-word memory operand.
3425 .. FIXME: L seems to be missing memory operand support.
3427 - ``M``: Print the high-order register of a two-register operand, or prints the
3428 address of the high-order word of a double-word memory operand.
3430 .. FIXME: M seems to be missing memory operand support.
3432 - ``D``: Print the second register of a two-register operand, or prints the
3433 second word of a double-word memory operand. (On a big-endian system, ``D`` is
3434 equivalent to ``L``, and on little-endian system, ``D`` is equivalent to
3436 - ``w``: No effect. Provided for compatibility with GCC which requires this
3437 modifier in order to print MSA registers (``W0-W31``) with the ``f``
3446 - ``L``: Print the second register of a two-register operand. Requires that it
3447 has been allocated consecutively to the first.
3449 .. FIXME: why is it restricted to consecutive ones? And there's
3450 nothing that ensures that happens, is there?
3452 - ``I``: Print the letter 'i' if the operand is an integer constant, otherwise
3453 nothing. Used to print 'addi' vs 'add' instructions.
3454 - ``y``: For a memory operand, prints formatter for a two-register X-form
3455 instruction. (Currently always prints ``r0,OPERAND``).
3456 - ``U``: Prints 'u' if the memory operand is an update form, and nothing
3457 otherwise. (NOTE: LLVM does not support update form, so this will currently
3458 always print nothing)
3459 - ``X``: Prints 'x' if the memory operand is an indexed form. (NOTE: LLVM does
3460 not support indexed form, so this will currently always print nothing)
3468 SystemZ implements only ``n``, and does *not* support any of the other
3469 target-independent modifiers.
3473 - ``c``: Print an unadorned integer or symbol name. (The latter is
3474 target-specific behavior for this typically target-independent modifier).
3475 - ``A``: Print a register name with a '``*``' before it.
3476 - ``b``: Print an 8-bit register name (e.g. ``al``); do nothing on a memory
3478 - ``h``: Print the upper 8-bit register name (e.g. ``ah``); do nothing on a
3480 - ``w``: Print the 16-bit register name (e.g. ``ax``); do nothing on a memory
3482 - ``k``: Print the 32-bit register name (e.g. ``eax``); do nothing on a memory
3484 - ``q``: Print the 64-bit register name (e.g. ``rax``), if 64-bit registers are
3485 available, otherwise the 32-bit register name; do nothing on a memory operand.
3486 - ``n``: Negate and print an unadorned integer, or, for operands other than an
3487 immediate integer (e.g. a relocatable symbol expression), print a '-' before
3488 the operand. (The behavior for relocatable symbol expressions is a
3489 target-specific behavior for this typically target-independent modifier)
3490 - ``H``: Print a memory reference with additional offset +8.
3491 - ``P``: Print a memory reference or operand for use as the argument of a call
3492 instruction. (E.g. omit ``(rip)``, even though it's PC-relative.)
3496 No additional modifiers.
3502 The call instructions that wrap inline asm nodes may have a
3503 "``!srcloc``" MDNode attached to it that contains a list of constant
3504 integers. If present, the code generator will use the integer as the
3505 location cookie value when report errors through the ``LLVMContext``
3506 error reporting mechanisms. This allows a front-end to correlate backend
3507 errors that occur with inline asm back to the source code that produced
3510 .. code-block:: llvm
3512 call void asm sideeffect "something bad", ""(), !srcloc !42
3514 !42 = !{ i32 1234567 }
3516 It is up to the front-end to make sense of the magic numbers it places
3517 in the IR. If the MDNode contains multiple constants, the code generator
3518 will use the one that corresponds to the line of the asm that the error
3526 LLVM IR allows metadata to be attached to instructions in the program
3527 that can convey extra information about the code to the optimizers and
3528 code generator. One example application of metadata is source-level
3529 debug information. There are two metadata primitives: strings and nodes.
3531 Metadata does not have a type, and is not a value. If referenced from a
3532 ``call`` instruction, it uses the ``metadata`` type.
3534 All metadata are identified in syntax by a exclamation point ('``!``').
3536 .. _metadata-string:
3538 Metadata Nodes and Metadata Strings
3539 -----------------------------------
3541 A metadata string is a string surrounded by double quotes. It can
3542 contain any character by escaping non-printable characters with
3543 "``\xx``" where "``xx``" is the two digit hex code. For example:
3546 Metadata nodes are represented with notation similar to structure
3547 constants (a comma separated list of elements, surrounded by braces and
3548 preceded by an exclamation point). Metadata nodes can have any values as
3549 their operand. For example:
3551 .. code-block:: llvm
3553 !{ !"test\00", i32 10}
3555 Metadata nodes that aren't uniqued use the ``distinct`` keyword. For example:
3557 .. code-block:: llvm
3559 !0 = distinct !{!"test\00", i32 10}
3561 ``distinct`` nodes are useful when nodes shouldn't be merged based on their
3562 content. They can also occur when transformations cause uniquing collisions
3563 when metadata operands change.
3565 A :ref:`named metadata <namedmetadatastructure>` is a collection of
3566 metadata nodes, which can be looked up in the module symbol table. For
3569 .. code-block:: llvm
3573 Metadata can be used as function arguments. Here ``llvm.dbg.value``
3574 function is using two metadata arguments:
3576 .. code-block:: llvm
3578 call void @llvm.dbg.value(metadata !24, i64 0, metadata !25)
3580 Metadata can be attached with an instruction. Here metadata ``!21`` is
3581 attached to the ``add`` instruction using the ``!dbg`` identifier:
3583 .. code-block:: llvm
3585 %indvar.next = add i64 %indvar, 1, !dbg !21
3587 More information about specific metadata nodes recognized by the
3588 optimizers and code generator is found below.
3590 .. _specialized-metadata:
3592 Specialized Metadata Nodes
3593 ^^^^^^^^^^^^^^^^^^^^^^^^^^
3595 Specialized metadata nodes are custom data structures in metadata (as opposed
3596 to generic tuples). Their fields are labelled, and can be specified in any
3599 These aren't inherently debug info centric, but currently all the specialized
3600 metadata nodes are related to debug info.
3607 ``DICompileUnit`` nodes represent a compile unit. The ``enums:``,
3608 ``retainedTypes:``, ``subprograms:``, ``globals:`` and ``imports:`` fields are
3609 tuples containing the debug info to be emitted along with the compile unit,
3610 regardless of code optimizations (some nodes are only emitted if there are
3611 references to them from instructions).
3613 .. code-block:: llvm
3615 !0 = !DICompileUnit(language: DW_LANG_C99, file: !1, producer: "clang",
3616 isOptimized: true, flags: "-O2", runtimeVersion: 2,
3617 splitDebugFilename: "abc.debug", emissionKind: 1,
3618 enums: !2, retainedTypes: !3, subprograms: !4,
3619 globals: !5, imports: !6)
3621 Compile unit descriptors provide the root scope for objects declared in a
3622 specific compilation unit. File descriptors are defined using this scope.
3623 These descriptors are collected by a named metadata ``!llvm.dbg.cu``. They
3624 keep track of subprograms, global variables, type information, and imported
3625 entities (declarations and namespaces).
3632 ``DIFile`` nodes represent files. The ``filename:`` can include slashes.
3634 .. code-block:: llvm
3636 !0 = !DIFile(filename: "path/to/file", directory: "/path/to/dir")
3638 Files are sometimes used in ``scope:`` fields, and are the only valid target
3639 for ``file:`` fields.
3646 ``DIBasicType`` nodes represent primitive types, such as ``int``, ``bool`` and
3647 ``float``. ``tag:`` defaults to ``DW_TAG_base_type``.
3649 .. code-block:: llvm
3651 !0 = !DIBasicType(name: "unsigned char", size: 8, align: 8,
3652 encoding: DW_ATE_unsigned_char)
3653 !1 = !DIBasicType(tag: DW_TAG_unspecified_type, name: "decltype(nullptr)")
3655 The ``encoding:`` describes the details of the type. Usually it's one of the
3658 .. code-block:: llvm
3664 DW_ATE_signed_char = 6
3666 DW_ATE_unsigned_char = 8
3668 .. _DISubroutineType:
3673 ``DISubroutineType`` nodes represent subroutine types. Their ``types:`` field
3674 refers to a tuple; the first operand is the return type, while the rest are the
3675 types of the formal arguments in order. If the first operand is ``null``, that
3676 represents a function with no return value (such as ``void foo() {}`` in C++).
3678 .. code-block:: llvm
3680 !0 = !BasicType(name: "int", size: 32, align: 32, DW_ATE_signed)
3681 !1 = !BasicType(name: "char", size: 8, align: 8, DW_ATE_signed_char)
3682 !2 = !DISubroutineType(types: !{null, !0, !1}) ; void (int, char)
3689 ``DIDerivedType`` nodes represent types derived from other types, such as
3692 .. code-block:: llvm
3694 !0 = !DIBasicType(name: "unsigned char", size: 8, align: 8,
3695 encoding: DW_ATE_unsigned_char)
3696 !1 = !DIDerivedType(tag: DW_TAG_pointer_type, baseType: !0, size: 32,
3699 The following ``tag:`` values are valid:
3701 .. code-block:: llvm
3703 DW_TAG_formal_parameter = 5
3705 DW_TAG_pointer_type = 15
3706 DW_TAG_reference_type = 16
3708 DW_TAG_ptr_to_member_type = 31
3709 DW_TAG_const_type = 38
3710 DW_TAG_volatile_type = 53
3711 DW_TAG_restrict_type = 55
3713 ``DW_TAG_member`` is used to define a member of a :ref:`composite type
3714 <DICompositeType>` or :ref:`subprogram <DISubprogram>`. The type of the member
3715 is the ``baseType:``. The ``offset:`` is the member's bit offset.
3716 ``DW_TAG_formal_parameter`` is used to define a member which is a formal
3717 argument of a subprogram.
3719 ``DW_TAG_typedef`` is used to provide a name for the ``baseType:``.
3721 ``DW_TAG_pointer_type``, ``DW_TAG_reference_type``, ``DW_TAG_const_type``,
3722 ``DW_TAG_volatile_type`` and ``DW_TAG_restrict_type`` are used to qualify the
3725 Note that the ``void *`` type is expressed as a type derived from NULL.
3727 .. _DICompositeType:
3732 ``DICompositeType`` nodes represent types composed of other types, like
3733 structures and unions. ``elements:`` points to a tuple of the composed types.
3735 If the source language supports ODR, the ``identifier:`` field gives the unique
3736 identifier used for type merging between modules. When specified, other types
3737 can refer to composite types indirectly via a :ref:`metadata string
3738 <metadata-string>` that matches their identifier.
3740 .. code-block:: llvm
3742 !0 = !DIEnumerator(name: "SixKind", value: 7)
3743 !1 = !DIEnumerator(name: "SevenKind", value: 7)
3744 !2 = !DIEnumerator(name: "NegEightKind", value: -8)
3745 !3 = !DICompositeType(tag: DW_TAG_enumeration_type, name: "Enum", file: !12,
3746 line: 2, size: 32, align: 32, identifier: "_M4Enum",
3747 elements: !{!0, !1, !2})
3749 The following ``tag:`` values are valid:
3751 .. code-block:: llvm
3753 DW_TAG_array_type = 1
3754 DW_TAG_class_type = 2
3755 DW_TAG_enumeration_type = 4
3756 DW_TAG_structure_type = 19
3757 DW_TAG_union_type = 23
3758 DW_TAG_subroutine_type = 21
3759 DW_TAG_inheritance = 28
3762 For ``DW_TAG_array_type``, the ``elements:`` should be :ref:`subrange
3763 descriptors <DISubrange>`, each representing the range of subscripts at that
3764 level of indexing. The ``DIFlagVector`` flag to ``flags:`` indicates that an
3765 array type is a native packed vector.
3767 For ``DW_TAG_enumeration_type``, the ``elements:`` should be :ref:`enumerator
3768 descriptors <DIEnumerator>`, each representing the definition of an enumeration
3769 value for the set. All enumeration type descriptors are collected in the
3770 ``enums:`` field of the :ref:`compile unit <DICompileUnit>`.
3772 For ``DW_TAG_structure_type``, ``DW_TAG_class_type``, and
3773 ``DW_TAG_union_type``, the ``elements:`` should be :ref:`derived types
3774 <DIDerivedType>` with ``tag: DW_TAG_member`` or ``tag: DW_TAG_inheritance``.
3781 ``DISubrange`` nodes are the elements for ``DW_TAG_array_type`` variants of
3782 :ref:`DICompositeType`. ``count: -1`` indicates an empty array.
3784 .. code-block:: llvm
3786 !0 = !DISubrange(count: 5, lowerBound: 0) ; array counting from 0
3787 !1 = !DISubrange(count: 5, lowerBound: 1) ; array counting from 1
3788 !2 = !DISubrange(count: -1) ; empty array.
3795 ``DIEnumerator`` nodes are the elements for ``DW_TAG_enumeration_type``
3796 variants of :ref:`DICompositeType`.
3798 .. code-block:: llvm
3800 !0 = !DIEnumerator(name: "SixKind", value: 7)
3801 !1 = !DIEnumerator(name: "SevenKind", value: 7)
3802 !2 = !DIEnumerator(name: "NegEightKind", value: -8)
3804 DITemplateTypeParameter
3805 """""""""""""""""""""""
3807 ``DITemplateTypeParameter`` nodes represent type parameters to generic source
3808 language constructs. They are used (optionally) in :ref:`DICompositeType` and
3809 :ref:`DISubprogram` ``templateParams:`` fields.
3811 .. code-block:: llvm
3813 !0 = !DITemplateTypeParameter(name: "Ty", type: !1)
3815 DITemplateValueParameter
3816 """"""""""""""""""""""""
3818 ``DITemplateValueParameter`` nodes represent value parameters to generic source
3819 language constructs. ``tag:`` defaults to ``DW_TAG_template_value_parameter``,
3820 but if specified can also be set to ``DW_TAG_GNU_template_template_param`` or
3821 ``DW_TAG_GNU_template_param_pack``. They are used (optionally) in
3822 :ref:`DICompositeType` and :ref:`DISubprogram` ``templateParams:`` fields.
3824 .. code-block:: llvm
3826 !0 = !DITemplateValueParameter(name: "Ty", type: !1, value: i32 7)
3831 ``DINamespace`` nodes represent namespaces in the source language.
3833 .. code-block:: llvm
3835 !0 = !DINamespace(name: "myawesomeproject", scope: !1, file: !2, line: 7)
3840 ``DIGlobalVariable`` nodes represent global variables in the source language.
3842 .. code-block:: llvm
3844 !0 = !DIGlobalVariable(name: "foo", linkageName: "foo", scope: !1,
3845 file: !2, line: 7, type: !3, isLocal: true,
3846 isDefinition: false, variable: i32* @foo,
3849 All global variables should be referenced by the `globals:` field of a
3850 :ref:`compile unit <DICompileUnit>`.
3857 ``DISubprogram`` nodes represent functions from the source language. The
3858 ``variables:`` field points at :ref:`variables <DILocalVariable>` that must be
3859 retained, even if their IR counterparts are optimized out of the IR. The
3860 ``type:`` field must point at an :ref:`DISubroutineType`.
3862 .. code-block:: llvm
3864 !0 = !DISubprogram(name: "foo", linkageName: "_Zfoov", scope: !1,
3865 file: !2, line: 7, type: !3, isLocal: true,
3866 isDefinition: false, scopeLine: 8, containingType: !4,
3867 virtuality: DW_VIRTUALITY_pure_virtual, virtualIndex: 10,
3868 flags: DIFlagPrototyped, isOptimized: true,
3869 function: void ()* @_Z3foov,
3870 templateParams: !5, declaration: !6, variables: !7)
3877 ``DILexicalBlock`` nodes describe nested blocks within a :ref:`subprogram
3878 <DISubprogram>`. The line number and column numbers are used to dinstinguish
3879 two lexical blocks at same depth. They are valid targets for ``scope:``
3882 .. code-block:: llvm
3884 !0 = distinct !DILexicalBlock(scope: !1, file: !2, line: 7, column: 35)
3886 Usually lexical blocks are ``distinct`` to prevent node merging based on
3889 .. _DILexicalBlockFile:
3894 ``DILexicalBlockFile`` nodes are used to discriminate between sections of a
3895 :ref:`lexical block <DILexicalBlock>`. The ``file:`` field can be changed to
3896 indicate textual inclusion, or the ``discriminator:`` field can be used to
3897 discriminate between control flow within a single block in the source language.
3899 .. code-block:: llvm
3901 !0 = !DILexicalBlock(scope: !3, file: !4, line: 7, column: 35)
3902 !1 = !DILexicalBlockFile(scope: !0, file: !4, discriminator: 0)
3903 !2 = !DILexicalBlockFile(scope: !0, file: !4, discriminator: 1)
3910 ``DILocation`` nodes represent source debug locations. The ``scope:`` field is
3911 mandatory, and points at an :ref:`DILexicalBlockFile`, an
3912 :ref:`DILexicalBlock`, or an :ref:`DISubprogram`.
3914 .. code-block:: llvm
3916 !0 = !DILocation(line: 2900, column: 42, scope: !1, inlinedAt: !2)
3918 .. _DILocalVariable:
3923 ``DILocalVariable`` nodes represent local variables in the source language. If
3924 the ``arg:`` field is set to non-zero, then this variable is a subprogram
3925 parameter, and it will be included in the ``variables:`` field of its
3926 :ref:`DISubprogram`.
3928 .. code-block:: llvm
3930 !0 = !DILocalVariable(name: "this", arg: 1, scope: !3, file: !2, line: 7,
3931 type: !3, flags: DIFlagArtificial)
3932 !1 = !DILocalVariable(name: "x", arg: 2, scope: !4, file: !2, line: 7,
3934 !2 = !DILocalVariable(name: "y", scope: !5, file: !2, line: 7, type: !3)
3939 ``DIExpression`` nodes represent DWARF expression sequences. They are used in
3940 :ref:`debug intrinsics<dbg_intrinsics>` (such as ``llvm.dbg.declare``) to
3941 describe how the referenced LLVM variable relates to the source language
3944 The current supported vocabulary is limited:
3946 - ``DW_OP_deref`` dereferences the working expression.
3947 - ``DW_OP_plus, 93`` adds ``93`` to the working expression.
3948 - ``DW_OP_bit_piece, 16, 8`` specifies the offset and size (``16`` and ``8``
3949 here, respectively) of the variable piece from the working expression.
3951 .. code-block:: llvm
3953 !0 = !DIExpression(DW_OP_deref)
3954 !1 = !DIExpression(DW_OP_plus, 3)
3955 !2 = !DIExpression(DW_OP_bit_piece, 3, 7)
3956 !3 = !DIExpression(DW_OP_deref, DW_OP_plus, 3, DW_OP_bit_piece, 3, 7)
3961 ``DIObjCProperty`` nodes represent Objective-C property nodes.
3963 .. code-block:: llvm
3965 !3 = !DIObjCProperty(name: "foo", file: !1, line: 7, setter: "setFoo",
3966 getter: "getFoo", attributes: 7, type: !2)
3971 ``DIImportedEntity`` nodes represent entities (such as modules) imported into a
3974 .. code-block:: llvm
3976 !2 = !DIImportedEntity(tag: DW_TAG_imported_module, name: "foo", scope: !0,
3977 entity: !1, line: 7)
3982 In LLVM IR, memory does not have types, so LLVM's own type system is not
3983 suitable for doing TBAA. Instead, metadata is added to the IR to
3984 describe a type system of a higher level language. This can be used to
3985 implement typical C/C++ TBAA, but it can also be used to implement
3986 custom alias analysis behavior for other languages.
3988 The current metadata format is very simple. TBAA metadata nodes have up
3989 to three fields, e.g.:
3991 .. code-block:: llvm
3993 !0 = !{ !"an example type tree" }
3994 !1 = !{ !"int", !0 }
3995 !2 = !{ !"float", !0 }
3996 !3 = !{ !"const float", !2, i64 1 }
3998 The first field is an identity field. It can be any value, usually a
3999 metadata string, which uniquely identifies the type. The most important
4000 name in the tree is the name of the root node. Two trees with different
4001 root node names are entirely disjoint, even if they have leaves with
4004 The second field identifies the type's parent node in the tree, or is
4005 null or omitted for a root node. A type is considered to alias all of
4006 its descendants and all of its ancestors in the tree. Also, a type is
4007 considered to alias all types in other trees, so that bitcode produced
4008 from multiple front-ends is handled conservatively.
4010 If the third field is present, it's an integer which if equal to 1
4011 indicates that the type is "constant" (meaning
4012 ``pointsToConstantMemory`` should return true; see `other useful
4013 AliasAnalysis methods <AliasAnalysis.html#OtherItfs>`_).
4015 '``tbaa.struct``' Metadata
4016 ^^^^^^^^^^^^^^^^^^^^^^^^^^
4018 The :ref:`llvm.memcpy <int_memcpy>` is often used to implement
4019 aggregate assignment operations in C and similar languages, however it
4020 is defined to copy a contiguous region of memory, which is more than
4021 strictly necessary for aggregate types which contain holes due to
4022 padding. Also, it doesn't contain any TBAA information about the fields
4025 ``!tbaa.struct`` metadata can describe which memory subregions in a
4026 memcpy are padding and what the TBAA tags of the struct are.
4028 The current metadata format is very simple. ``!tbaa.struct`` metadata
4029 nodes are a list of operands which are in conceptual groups of three.
4030 For each group of three, the first operand gives the byte offset of a
4031 field in bytes, the second gives its size in bytes, and the third gives
4034 .. code-block:: llvm
4036 !4 = !{ i64 0, i64 4, !1, i64 8, i64 4, !2 }
4038 This describes a struct with two fields. The first is at offset 0 bytes
4039 with size 4 bytes, and has tbaa tag !1. The second is at offset 8 bytes
4040 and has size 4 bytes and has tbaa tag !2.
4042 Note that the fields need not be contiguous. In this example, there is a
4043 4 byte gap between the two fields. This gap represents padding which
4044 does not carry useful data and need not be preserved.
4046 '``noalias``' and '``alias.scope``' Metadata
4047 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4049 ``noalias`` and ``alias.scope`` metadata provide the ability to specify generic
4050 noalias memory-access sets. This means that some collection of memory access
4051 instructions (loads, stores, memory-accessing calls, etc.) that carry
4052 ``noalias`` metadata can specifically be specified not to alias with some other
4053 collection of memory access instructions that carry ``alias.scope`` metadata.
4054 Each type of metadata specifies a list of scopes where each scope has an id and
4055 a domain. When evaluating an aliasing query, if for some domain, the set
4056 of scopes with that domain in one instruction's ``alias.scope`` list is a
4057 subset of (or equal to) the set of scopes for that domain in another
4058 instruction's ``noalias`` list, then the two memory accesses are assumed not to
4061 The metadata identifying each domain is itself a list containing one or two
4062 entries. The first entry is the name of the domain. Note that if the name is a
4063 string then it can be combined accross functions and translation units. A
4064 self-reference can be used to create globally unique domain names. A
4065 descriptive string may optionally be provided as a second list entry.
4067 The metadata identifying each scope is also itself a list containing two or
4068 three entries. The first entry is the name of the scope. Note that if the name
4069 is a string then it can be combined accross functions and translation units. A
4070 self-reference can be used to create globally unique scope names. A metadata
4071 reference to the scope's domain is the second entry. A descriptive string may
4072 optionally be provided as a third list entry.
4076 .. code-block:: llvm
4078 ; Two scope domains:
4082 ; Some scopes in these domains:
4088 !5 = !{!4} ; A list containing only scope !4
4092 ; These two instructions don't alias:
4093 %0 = load float, float* %c, align 4, !alias.scope !5
4094 store float %0, float* %arrayidx.i, align 4, !noalias !5
4096 ; These two instructions also don't alias (for domain !1, the set of scopes
4097 ; in the !alias.scope equals that in the !noalias list):
4098 %2 = load float, float* %c, align 4, !alias.scope !5
4099 store float %2, float* %arrayidx.i2, align 4, !noalias !6
4101 ; These two instructions may alias (for domain !0, the set of scopes in
4102 ; the !noalias list is not a superset of, or equal to, the scopes in the
4103 ; !alias.scope list):
4104 %2 = load float, float* %c, align 4, !alias.scope !6
4105 store float %0, float* %arrayidx.i, align 4, !noalias !7
4107 '``fpmath``' Metadata
4108 ^^^^^^^^^^^^^^^^^^^^^
4110 ``fpmath`` metadata may be attached to any instruction of floating point
4111 type. It can be used to express the maximum acceptable error in the
4112 result of that instruction, in ULPs, thus potentially allowing the
4113 compiler to use a more efficient but less accurate method of computing
4114 it. ULP is defined as follows:
4116 If ``x`` is a real number that lies between two finite consecutive
4117 floating-point numbers ``a`` and ``b``, without being equal to one
4118 of them, then ``ulp(x) = |b - a|``, otherwise ``ulp(x)`` is the
4119 distance between the two non-equal finite floating-point numbers
4120 nearest ``x``. Moreover, ``ulp(NaN)`` is ``NaN``.
4122 The metadata node shall consist of a single positive floating point
4123 number representing the maximum relative error, for example:
4125 .. code-block:: llvm
4127 !0 = !{ float 2.5 } ; maximum acceptable inaccuracy is 2.5 ULPs
4131 '``range``' Metadata
4132 ^^^^^^^^^^^^^^^^^^^^
4134 ``range`` metadata may be attached only to ``load``, ``call`` and ``invoke`` of
4135 integer types. It expresses the possible ranges the loaded value or the value
4136 returned by the called function at this call site is in. The ranges are
4137 represented with a flattened list of integers. The loaded value or the value
4138 returned is known to be in the union of the ranges defined by each consecutive
4139 pair. Each pair has the following properties:
4141 - The type must match the type loaded by the instruction.
4142 - The pair ``a,b`` represents the range ``[a,b)``.
4143 - Both ``a`` and ``b`` are constants.
4144 - The range is allowed to wrap.
4145 - The range should not represent the full or empty set. That is,
4148 In addition, the pairs must be in signed order of the lower bound and
4149 they must be non-contiguous.
4153 .. code-block:: llvm
4155 %a = load i8, i8* %x, align 1, !range !0 ; Can only be 0 or 1
4156 %b = load i8, i8* %y, align 1, !range !1 ; Can only be 255 (-1), 0 or 1
4157 %c = call i8 @foo(), !range !2 ; Can only be 0, 1, 3, 4 or 5
4158 %d = invoke i8 @bar() to label %cont
4159 unwind label %lpad, !range !3 ; Can only be -2, -1, 3, 4 or 5
4161 !0 = !{ i8 0, i8 2 }
4162 !1 = !{ i8 255, i8 2 }
4163 !2 = !{ i8 0, i8 2, i8 3, i8 6 }
4164 !3 = !{ i8 -2, i8 0, i8 3, i8 6 }
4169 It is sometimes useful to attach information to loop constructs. Currently,
4170 loop metadata is implemented as metadata attached to the branch instruction
4171 in the loop latch block. This type of metadata refer to a metadata node that is
4172 guaranteed to be separate for each loop. The loop identifier metadata is
4173 specified with the name ``llvm.loop``.
4175 The loop identifier metadata is implemented using a metadata that refers to
4176 itself to avoid merging it with any other identifier metadata, e.g.,
4177 during module linkage or function inlining. That is, each loop should refer
4178 to their own identification metadata even if they reside in separate functions.
4179 The following example contains loop identifier metadata for two separate loop
4182 .. code-block:: llvm
4187 The loop identifier metadata can be used to specify additional
4188 per-loop metadata. Any operands after the first operand can be treated
4189 as user-defined metadata. For example the ``llvm.loop.unroll.count``
4190 suggests an unroll factor to the loop unroller:
4192 .. code-block:: llvm
4194 br i1 %exitcond, label %._crit_edge, label %.lr.ph, !llvm.loop !0
4197 !1 = !{!"llvm.loop.unroll.count", i32 4}
4199 '``llvm.loop.vectorize``' and '``llvm.loop.interleave``'
4200 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4202 Metadata prefixed with ``llvm.loop.vectorize`` or ``llvm.loop.interleave`` are
4203 used to control per-loop vectorization and interleaving parameters such as
4204 vectorization width and interleave count. These metadata should be used in
4205 conjunction with ``llvm.loop`` loop identification metadata. The
4206 ``llvm.loop.vectorize`` and ``llvm.loop.interleave`` metadata are only
4207 optimization hints and the optimizer will only interleave and vectorize loops if
4208 it believes it is safe to do so. The ``llvm.mem.parallel_loop_access`` metadata
4209 which contains information about loop-carried memory dependencies can be helpful
4210 in determining the safety of these transformations.
4212 '``llvm.loop.interleave.count``' Metadata
4213 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4215 This metadata suggests an interleave count to the loop interleaver.
4216 The first operand is the string ``llvm.loop.interleave.count`` and the
4217 second operand is an integer specifying the interleave count. For
4220 .. code-block:: llvm
4222 !0 = !{!"llvm.loop.interleave.count", i32 4}
4224 Note that setting ``llvm.loop.interleave.count`` to 1 disables interleaving
4225 multiple iterations of the loop. If ``llvm.loop.interleave.count`` is set to 0
4226 then the interleave count will be determined automatically.
4228 '``llvm.loop.vectorize.enable``' Metadata
4229 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4231 This metadata selectively enables or disables vectorization for the loop. The
4232 first operand is the string ``llvm.loop.vectorize.enable`` and the second operand
4233 is a bit. If the bit operand value is 1 vectorization is enabled. A value of
4234 0 disables vectorization:
4236 .. code-block:: llvm
4238 !0 = !{!"llvm.loop.vectorize.enable", i1 0}
4239 !1 = !{!"llvm.loop.vectorize.enable", i1 1}
4241 '``llvm.loop.vectorize.width``' Metadata
4242 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4244 This metadata sets the target width of the vectorizer. The first
4245 operand is the string ``llvm.loop.vectorize.width`` and the second
4246 operand is an integer specifying the width. For example:
4248 .. code-block:: llvm
4250 !0 = !{!"llvm.loop.vectorize.width", i32 4}
4252 Note that setting ``llvm.loop.vectorize.width`` to 1 disables
4253 vectorization of the loop. If ``llvm.loop.vectorize.width`` is set to
4254 0 or if the loop does not have this metadata the width will be
4255 determined automatically.
4257 '``llvm.loop.unroll``'
4258 ^^^^^^^^^^^^^^^^^^^^^^
4260 Metadata prefixed with ``llvm.loop.unroll`` are loop unrolling
4261 optimization hints such as the unroll factor. ``llvm.loop.unroll``
4262 metadata should be used in conjunction with ``llvm.loop`` loop
4263 identification metadata. The ``llvm.loop.unroll`` metadata are only
4264 optimization hints and the unrolling will only be performed if the
4265 optimizer believes it is safe to do so.
4267 '``llvm.loop.unroll.count``' Metadata
4268 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4270 This metadata suggests an unroll factor to the loop unroller. The
4271 first operand is the string ``llvm.loop.unroll.count`` and the second
4272 operand is a positive integer specifying the unroll factor. For
4275 .. code-block:: llvm
4277 !0 = !{!"llvm.loop.unroll.count", i32 4}
4279 If the trip count of the loop is less than the unroll count the loop
4280 will be partially unrolled.
4282 '``llvm.loop.unroll.disable``' Metadata
4283 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4285 This metadata disables loop unrolling. The metadata has a single operand
4286 which is the string ``llvm.loop.unroll.disable``. For example:
4288 .. code-block:: llvm
4290 !0 = !{!"llvm.loop.unroll.disable"}
4292 '``llvm.loop.unroll.runtime.disable``' Metadata
4293 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4295 This metadata disables runtime loop unrolling. The metadata has a single
4296 operand which is the string ``llvm.loop.unroll.runtime.disable``. For example:
4298 .. code-block:: llvm
4300 !0 = !{!"llvm.loop.unroll.runtime.disable"}
4302 '``llvm.loop.unroll.enable``' Metadata
4303 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4305 This metadata suggests that the loop should be fully unrolled if the trip count
4306 is known at compile time and partially unrolled if the trip count is not known
4307 at compile time. The metadata has a single operand which is the string
4308 ``llvm.loop.unroll.enable``. For example:
4310 .. code-block:: llvm
4312 !0 = !{!"llvm.loop.unroll.enable"}
4314 '``llvm.loop.unroll.full``' Metadata
4315 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4317 This metadata suggests that the loop should be unrolled fully. The
4318 metadata has a single operand which is the string ``llvm.loop.unroll.full``.
4321 .. code-block:: llvm
4323 !0 = !{!"llvm.loop.unroll.full"}
4328 Metadata types used to annotate memory accesses with information helpful
4329 for optimizations are prefixed with ``llvm.mem``.
4331 '``llvm.mem.parallel_loop_access``' Metadata
4332 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4334 The ``llvm.mem.parallel_loop_access`` metadata refers to a loop identifier,
4335 or metadata containing a list of loop identifiers for nested loops.
4336 The metadata is attached to memory accessing instructions and denotes that
4337 no loop carried memory dependence exist between it and other instructions denoted
4338 with the same loop identifier.
4340 Precisely, given two instructions ``m1`` and ``m2`` that both have the
4341 ``llvm.mem.parallel_loop_access`` metadata, with ``L1`` and ``L2`` being the
4342 set of loops associated with that metadata, respectively, then there is no loop
4343 carried dependence between ``m1`` and ``m2`` for loops in both ``L1`` and
4346 As a special case, if all memory accessing instructions in a loop have
4347 ``llvm.mem.parallel_loop_access`` metadata that refers to that loop, then the
4348 loop has no loop carried memory dependences and is considered to be a parallel
4351 Note that if not all memory access instructions have such metadata referring to
4352 the loop, then the loop is considered not being trivially parallel. Additional
4353 memory dependence analysis is required to make that determination. As a fail
4354 safe mechanism, this causes loops that were originally parallel to be considered
4355 sequential (if optimization passes that are unaware of the parallel semantics
4356 insert new memory instructions into the loop body).
4358 Example of a loop that is considered parallel due to its correct use of
4359 both ``llvm.loop`` and ``llvm.mem.parallel_loop_access``
4360 metadata types that refer to the same loop identifier metadata.
4362 .. code-block:: llvm
4366 %val0 = load i32, i32* %arrayidx, !llvm.mem.parallel_loop_access !0
4368 store i32 %val0, i32* %arrayidx1, !llvm.mem.parallel_loop_access !0
4370 br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !0
4376 It is also possible to have nested parallel loops. In that case the
4377 memory accesses refer to a list of loop identifier metadata nodes instead of
4378 the loop identifier metadata node directly:
4380 .. code-block:: llvm
4384 %val1 = load i32, i32* %arrayidx3, !llvm.mem.parallel_loop_access !2
4386 br label %inner.for.body
4390 %val0 = load i32, i32* %arrayidx1, !llvm.mem.parallel_loop_access !0
4392 store i32 %val0, i32* %arrayidx2, !llvm.mem.parallel_loop_access !0
4394 br i1 %exitcond, label %inner.for.end, label %inner.for.body, !llvm.loop !1
4398 store i32 %val1, i32* %arrayidx4, !llvm.mem.parallel_loop_access !2
4400 br i1 %exitcond, label %outer.for.end, label %outer.for.body, !llvm.loop !2
4402 outer.for.end: ; preds = %for.body
4404 !0 = !{!1, !2} ; a list of loop identifiers
4405 !1 = !{!1} ; an identifier for the inner loop
4406 !2 = !{!2} ; an identifier for the outer loop
4411 The ``llvm.bitsets`` global metadata is used to implement
4412 :doc:`bitsets <BitSets>`.
4414 Module Flags Metadata
4415 =====================
4417 Information about the module as a whole is difficult to convey to LLVM's
4418 subsystems. The LLVM IR isn't sufficient to transmit this information.
4419 The ``llvm.module.flags`` named metadata exists in order to facilitate
4420 this. These flags are in the form of key / value pairs --- much like a
4421 dictionary --- making it easy for any subsystem who cares about a flag to
4424 The ``llvm.module.flags`` metadata contains a list of metadata triplets.
4425 Each triplet has the following form:
4427 - The first element is a *behavior* flag, which specifies the behavior
4428 when two (or more) modules are merged together, and it encounters two
4429 (or more) metadata with the same ID. The supported behaviors are
4431 - The second element is a metadata string that is a unique ID for the
4432 metadata. Each module may only have one flag entry for each unique ID (not
4433 including entries with the **Require** behavior).
4434 - The third element is the value of the flag.
4436 When two (or more) modules are merged together, the resulting
4437 ``llvm.module.flags`` metadata is the union of the modules' flags. That is, for
4438 each unique metadata ID string, there will be exactly one entry in the merged
4439 modules ``llvm.module.flags`` metadata table, and the value for that entry will
4440 be determined by the merge behavior flag, as described below. The only exception
4441 is that entries with the *Require* behavior are always preserved.
4443 The following behaviors are supported:
4454 Emits an error if two values disagree, otherwise the resulting value
4455 is that of the operands.
4459 Emits a warning if two values disagree. The result value will be the
4460 operand for the flag from the first module being linked.
4464 Adds a requirement that another module flag be present and have a
4465 specified value after linking is performed. The value must be a
4466 metadata pair, where the first element of the pair is the ID of the
4467 module flag to be restricted, and the second element of the pair is
4468 the value the module flag should be restricted to. This behavior can
4469 be used to restrict the allowable results (via triggering of an
4470 error) of linking IDs with the **Override** behavior.
4474 Uses the specified value, regardless of the behavior or value of the
4475 other module. If both modules specify **Override**, but the values
4476 differ, an error will be emitted.
4480 Appends the two values, which are required to be metadata nodes.
4484 Appends the two values, which are required to be metadata
4485 nodes. However, duplicate entries in the second list are dropped
4486 during the append operation.
4488 It is an error for a particular unique flag ID to have multiple behaviors,
4489 except in the case of **Require** (which adds restrictions on another metadata
4490 value) or **Override**.
4492 An example of module flags:
4494 .. code-block:: llvm
4496 !0 = !{ i32 1, !"foo", i32 1 }
4497 !1 = !{ i32 4, !"bar", i32 37 }
4498 !2 = !{ i32 2, !"qux", i32 42 }
4499 !3 = !{ i32 3, !"qux",
4504 !llvm.module.flags = !{ !0, !1, !2, !3 }
4506 - Metadata ``!0`` has the ID ``!"foo"`` and the value '1'. The behavior
4507 if two or more ``!"foo"`` flags are seen is to emit an error if their
4508 values are not equal.
4510 - Metadata ``!1`` has the ID ``!"bar"`` and the value '37'. The
4511 behavior if two or more ``!"bar"`` flags are seen is to use the value
4514 - Metadata ``!2`` has the ID ``!"qux"`` and the value '42'. The
4515 behavior if two or more ``!"qux"`` flags are seen is to emit a
4516 warning if their values are not equal.
4518 - Metadata ``!3`` has the ID ``!"qux"`` and the value:
4524 The behavior is to emit an error if the ``llvm.module.flags`` does not
4525 contain a flag with the ID ``!"foo"`` that has the value '1' after linking is
4528 Objective-C Garbage Collection Module Flags Metadata
4529 ----------------------------------------------------
4531 On the Mach-O platform, Objective-C stores metadata about garbage
4532 collection in a special section called "image info". The metadata
4533 consists of a version number and a bitmask specifying what types of
4534 garbage collection are supported (if any) by the file. If two or more
4535 modules are linked together their garbage collection metadata needs to
4536 be merged rather than appended together.
4538 The Objective-C garbage collection module flags metadata consists of the
4539 following key-value pairs:
4548 * - ``Objective-C Version``
4549 - **[Required]** --- The Objective-C ABI version. Valid values are 1 and 2.
4551 * - ``Objective-C Image Info Version``
4552 - **[Required]** --- The version of the image info section. Currently
4555 * - ``Objective-C Image Info Section``
4556 - **[Required]** --- The section to place the metadata. Valid values are
4557 ``"__OBJC, __image_info, regular"`` for Objective-C ABI version 1, and
4558 ``"__DATA,__objc_imageinfo, regular, no_dead_strip"`` for
4559 Objective-C ABI version 2.
4561 * - ``Objective-C Garbage Collection``
4562 - **[Required]** --- Specifies whether garbage collection is supported or
4563 not. Valid values are 0, for no garbage collection, and 2, for garbage
4564 collection supported.
4566 * - ``Objective-C GC Only``
4567 - **[Optional]** --- Specifies that only garbage collection is supported.
4568 If present, its value must be 6. This flag requires that the
4569 ``Objective-C Garbage Collection`` flag have the value 2.
4571 Some important flag interactions:
4573 - If a module with ``Objective-C Garbage Collection`` set to 0 is
4574 merged with a module with ``Objective-C Garbage Collection`` set to
4575 2, then the resulting module has the
4576 ``Objective-C Garbage Collection`` flag set to 0.
4577 - A module with ``Objective-C Garbage Collection`` set to 0 cannot be
4578 merged with a module with ``Objective-C GC Only`` set to 6.
4580 Automatic Linker Flags Module Flags Metadata
4581 --------------------------------------------
4583 Some targets support embedding flags to the linker inside individual object
4584 files. Typically this is used in conjunction with language extensions which
4585 allow source files to explicitly declare the libraries they depend on, and have
4586 these automatically be transmitted to the linker via object files.
4588 These flags are encoded in the IR using metadata in the module flags section,
4589 using the ``Linker Options`` key. The merge behavior for this flag is required
4590 to be ``AppendUnique``, and the value for the key is expected to be a metadata
4591 node which should be a list of other metadata nodes, each of which should be a
4592 list of metadata strings defining linker options.
4594 For example, the following metadata section specifies two separate sets of
4595 linker options, presumably to link against ``libz`` and the ``Cocoa``
4598 !0 = !{ i32 6, !"Linker Options",
4601 !{ !"-framework", !"Cocoa" } } }
4602 !llvm.module.flags = !{ !0 }
4604 The metadata encoding as lists of lists of options, as opposed to a collapsed
4605 list of options, is chosen so that the IR encoding can use multiple option
4606 strings to specify e.g., a single library, while still having that specifier be
4607 preserved as an atomic element that can be recognized by a target specific
4608 assembly writer or object file emitter.
4610 Each individual option is required to be either a valid option for the target's
4611 linker, or an option that is reserved by the target specific assembly writer or
4612 object file emitter. No other aspect of these options is defined by the IR.
4614 C type width Module Flags Metadata
4615 ----------------------------------
4617 The ARM backend emits a section into each generated object file describing the
4618 options that it was compiled with (in a compiler-independent way) to prevent
4619 linking incompatible objects, and to allow automatic library selection. Some
4620 of these options are not visible at the IR level, namely wchar_t width and enum
4623 To pass this information to the backend, these options are encoded in module
4624 flags metadata, using the following key-value pairs:
4634 - * 0 --- sizeof(wchar_t) == 4
4635 * 1 --- sizeof(wchar_t) == 2
4638 - * 0 --- Enums are at least as large as an ``int``.
4639 * 1 --- Enums are stored in the smallest integer type which can
4640 represent all of its values.
4642 For example, the following metadata section specifies that the module was
4643 compiled with a ``wchar_t`` width of 4 bytes, and the underlying type of an
4644 enum is the smallest type which can represent all of its values::
4646 !llvm.module.flags = !{!0, !1}
4647 !0 = !{i32 1, !"short_wchar", i32 1}
4648 !1 = !{i32 1, !"short_enum", i32 0}
4650 .. _intrinsicglobalvariables:
4652 Intrinsic Global Variables
4653 ==========================
4655 LLVM has a number of "magic" global variables that contain data that
4656 affect code generation or other IR semantics. These are documented here.
4657 All globals of this sort should have a section specified as
4658 "``llvm.metadata``". This section and all globals that start with
4659 "``llvm.``" are reserved for use by LLVM.
4663 The '``llvm.used``' Global Variable
4664 -----------------------------------
4666 The ``@llvm.used`` global is an array which has
4667 :ref:`appending linkage <linkage_appending>`. This array contains a list of
4668 pointers to named global variables, functions and aliases which may optionally
4669 have a pointer cast formed of bitcast or getelementptr. For example, a legal
4672 .. code-block:: llvm
4677 @llvm.used = appending global [2 x i8*] [
4679 i8* bitcast (i32* @Y to i8*)
4680 ], section "llvm.metadata"
4682 If a symbol appears in the ``@llvm.used`` list, then the compiler, assembler,
4683 and linker are required to treat the symbol as if there is a reference to the
4684 symbol that it cannot see (which is why they have to be named). For example, if
4685 a variable has internal linkage and no references other than that from the
4686 ``@llvm.used`` list, it cannot be deleted. This is commonly used to represent
4687 references from inline asms and other things the compiler cannot "see", and
4688 corresponds to "``attribute((used))``" in GNU C.
4690 On some targets, the code generator must emit a directive to the
4691 assembler or object file to prevent the assembler and linker from
4692 molesting the symbol.
4694 .. _gv_llvmcompilerused:
4696 The '``llvm.compiler.used``' Global Variable
4697 --------------------------------------------
4699 The ``@llvm.compiler.used`` directive is the same as the ``@llvm.used``
4700 directive, except that it only prevents the compiler from touching the
4701 symbol. On targets that support it, this allows an intelligent linker to
4702 optimize references to the symbol without being impeded as it would be
4705 This is a rare construct that should only be used in rare circumstances,
4706 and should not be exposed to source languages.
4708 .. _gv_llvmglobalctors:
4710 The '``llvm.global_ctors``' Global Variable
4711 -------------------------------------------
4713 .. code-block:: llvm
4715 %0 = type { i32, void ()*, i8* }
4716 @llvm.global_ctors = appending global [1 x %0] [%0 { i32 65535, void ()* @ctor, i8* @data }]
4718 The ``@llvm.global_ctors`` array contains a list of constructor
4719 functions, priorities, and an optional associated global or function.
4720 The functions referenced by this array will be called in ascending order
4721 of priority (i.e. lowest first) when the module is loaded. The order of
4722 functions with the same priority is not defined.
4724 If the third field is present, non-null, and points to a global variable
4725 or function, the initializer function will only run if the associated
4726 data from the current module is not discarded.
4728 .. _llvmglobaldtors:
4730 The '``llvm.global_dtors``' Global Variable
4731 -------------------------------------------
4733 .. code-block:: llvm
4735 %0 = type { i32, void ()*, i8* }
4736 @llvm.global_dtors = appending global [1 x %0] [%0 { i32 65535, void ()* @dtor, i8* @data }]
4738 The ``@llvm.global_dtors`` array contains a list of destructor
4739 functions, priorities, and an optional associated global or function.
4740 The functions referenced by this array will be called in descending
4741 order of priority (i.e. highest first) when the module is unloaded. The
4742 order of functions with the same priority is not defined.
4744 If the third field is present, non-null, and points to a global variable
4745 or function, the destructor function will only run if the associated
4746 data from the current module is not discarded.
4748 Instruction Reference
4749 =====================
4751 The LLVM instruction set consists of several different classifications
4752 of instructions: :ref:`terminator instructions <terminators>`, :ref:`binary
4753 instructions <binaryops>`, :ref:`bitwise binary
4754 instructions <bitwiseops>`, :ref:`memory instructions <memoryops>`, and
4755 :ref:`other instructions <otherops>`.
4759 Terminator Instructions
4760 -----------------------
4762 As mentioned :ref:`previously <functionstructure>`, every basic block in a
4763 program ends with a "Terminator" instruction, which indicates which
4764 block should be executed after the current block is finished. These
4765 terminator instructions typically yield a '``void``' value: they produce
4766 control flow, not values (the one exception being the
4767 ':ref:`invoke <i_invoke>`' instruction).
4769 The terminator instructions are: ':ref:`ret <i_ret>`',
4770 ':ref:`br <i_br>`', ':ref:`switch <i_switch>`',
4771 ':ref:`indirectbr <i_indirectbr>`', ':ref:`invoke <i_invoke>`',
4772 ':ref:`resume <i_resume>`', ':ref:`catchpad <i_catchpad>`',
4773 ':ref:`catchendpad <i_catchendpad>`',
4774 ':ref:`catchret <i_catchret>`',
4775 ':ref:`cleanupret <i_cleanupret>`',
4776 ':ref:`terminatepad <i_terminatepad>`',
4777 and ':ref:`unreachable <i_unreachable>`'.
4781 '``ret``' Instruction
4782 ^^^^^^^^^^^^^^^^^^^^^
4789 ret <type> <value> ; Return a value from a non-void function
4790 ret void ; Return from void function
4795 The '``ret``' instruction is used to return control flow (and optionally
4796 a value) from a function back to the caller.
4798 There are two forms of the '``ret``' instruction: one that returns a
4799 value and then causes control flow, and one that just causes control
4805 The '``ret``' instruction optionally accepts a single argument, the
4806 return value. The type of the return value must be a ':ref:`first
4807 class <t_firstclass>`' type.
4809 A function is not :ref:`well formed <wellformed>` if it it has a non-void
4810 return type and contains a '``ret``' instruction with no return value or
4811 a return value with a type that does not match its type, or if it has a
4812 void return type and contains a '``ret``' instruction with a return
4818 When the '``ret``' instruction is executed, control flow returns back to
4819 the calling function's context. If the caller is a
4820 ":ref:`call <i_call>`" instruction, execution continues at the
4821 instruction after the call. If the caller was an
4822 ":ref:`invoke <i_invoke>`" instruction, execution continues at the
4823 beginning of the "normal" destination block. If the instruction returns
4824 a value, that value shall set the call or invoke instruction's return
4830 .. code-block:: llvm
4832 ret i32 5 ; Return an integer value of 5
4833 ret void ; Return from a void function
4834 ret { i32, i8 } { i32 4, i8 2 } ; Return a struct of values 4 and 2
4838 '``br``' Instruction
4839 ^^^^^^^^^^^^^^^^^^^^
4846 br i1 <cond>, label <iftrue>, label <iffalse>
4847 br label <dest> ; Unconditional branch
4852 The '``br``' instruction is used to cause control flow to transfer to a
4853 different basic block in the current function. There are two forms of
4854 this instruction, corresponding to a conditional branch and an
4855 unconditional branch.
4860 The conditional branch form of the '``br``' instruction takes a single
4861 '``i1``' value and two '``label``' values. The unconditional form of the
4862 '``br``' instruction takes a single '``label``' value as a target.
4867 Upon execution of a conditional '``br``' instruction, the '``i1``'
4868 argument is evaluated. If the value is ``true``, control flows to the
4869 '``iftrue``' ``label`` argument. If "cond" is ``false``, control flows
4870 to the '``iffalse``' ``label`` argument.
4875 .. code-block:: llvm
4878 %cond = icmp eq i32 %a, %b
4879 br i1 %cond, label %IfEqual, label %IfUnequal
4887 '``switch``' Instruction
4888 ^^^^^^^^^^^^^^^^^^^^^^^^
4895 switch <intty> <value>, label <defaultdest> [ <intty> <val>, label <dest> ... ]
4900 The '``switch``' instruction is used to transfer control flow to one of
4901 several different places. It is a generalization of the '``br``'
4902 instruction, allowing a branch to occur to one of many possible
4908 The '``switch``' instruction uses three parameters: an integer
4909 comparison value '``value``', a default '``label``' destination, and an
4910 array of pairs of comparison value constants and '``label``'s. The table
4911 is not allowed to contain duplicate constant entries.
4916 The ``switch`` instruction specifies a table of values and destinations.
4917 When the '``switch``' instruction is executed, this table is searched
4918 for the given value. If the value is found, control flow is transferred
4919 to the corresponding destination; otherwise, control flow is transferred
4920 to the default destination.
4925 Depending on properties of the target machine and the particular
4926 ``switch`` instruction, this instruction may be code generated in
4927 different ways. For example, it could be generated as a series of
4928 chained conditional branches or with a lookup table.
4933 .. code-block:: llvm
4935 ; Emulate a conditional br instruction
4936 %Val = zext i1 %value to i32
4937 switch i32 %Val, label %truedest [ i32 0, label %falsedest ]
4939 ; Emulate an unconditional br instruction
4940 switch i32 0, label %dest [ ]
4942 ; Implement a jump table:
4943 switch i32 %val, label %otherwise [ i32 0, label %onzero
4945 i32 2, label %ontwo ]
4949 '``indirectbr``' Instruction
4950 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4957 indirectbr <somety>* <address>, [ label <dest1>, label <dest2>, ... ]
4962 The '``indirectbr``' instruction implements an indirect branch to a
4963 label within the current function, whose address is specified by
4964 "``address``". Address must be derived from a
4965 :ref:`blockaddress <blockaddress>` constant.
4970 The '``address``' argument is the address of the label to jump to. The
4971 rest of the arguments indicate the full set of possible destinations
4972 that the address may point to. Blocks are allowed to occur multiple
4973 times in the destination list, though this isn't particularly useful.
4975 This destination list is required so that dataflow analysis has an
4976 accurate understanding of the CFG.
4981 Control transfers to the block specified in the address argument. All
4982 possible destination blocks must be listed in the label list, otherwise
4983 this instruction has undefined behavior. This implies that jumps to
4984 labels defined in other functions have undefined behavior as well.
4989 This is typically implemented with a jump through a register.
4994 .. code-block:: llvm
4996 indirectbr i8* %Addr, [ label %bb1, label %bb2, label %bb3 ]
5000 '``invoke``' Instruction
5001 ^^^^^^^^^^^^^^^^^^^^^^^^
5008 <result> = invoke [cconv] [ret attrs] <ptr to function ty> <function ptr val>(<function args>) [fn attrs]
5009 to label <normal label> unwind label <exception label>
5014 The '``invoke``' instruction causes control to transfer to a specified
5015 function, with the possibility of control flow transfer to either the
5016 '``normal``' label or the '``exception``' label. If the callee function
5017 returns with the "``ret``" instruction, control flow will return to the
5018 "normal" label. If the callee (or any indirect callees) returns via the
5019 ":ref:`resume <i_resume>`" instruction or other exception handling
5020 mechanism, control is interrupted and continued at the dynamically
5021 nearest "exception" label.
5023 The '``exception``' label is a `landing
5024 pad <ExceptionHandling.html#overview>`_ for the exception. As such,
5025 '``exception``' label is required to have the
5026 ":ref:`landingpad <i_landingpad>`" instruction, which contains the
5027 information about the behavior of the program after unwinding happens,
5028 as its first non-PHI instruction. The restrictions on the
5029 "``landingpad``" instruction's tightly couples it to the "``invoke``"
5030 instruction, so that the important information contained within the
5031 "``landingpad``" instruction can't be lost through normal code motion.
5036 This instruction requires several arguments:
5038 #. The optional "cconv" marker indicates which :ref:`calling
5039 convention <callingconv>` the call should use. If none is
5040 specified, the call defaults to using C calling conventions.
5041 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
5042 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
5044 #. '``ptr to function ty``': shall be the signature of the pointer to
5045 function value being invoked. In most cases, this is a direct
5046 function invocation, but indirect ``invoke``'s are just as possible,
5047 branching off an arbitrary pointer to function value.
5048 #. '``function ptr val``': An LLVM value containing a pointer to a
5049 function to be invoked.
5050 #. '``function args``': argument list whose types match the function
5051 signature argument types and parameter attributes. All arguments must
5052 be of :ref:`first class <t_firstclass>` type. If the function signature
5053 indicates the function accepts a variable number of arguments, the
5054 extra arguments can be specified.
5055 #. '``normal label``': the label reached when the called function
5056 executes a '``ret``' instruction.
5057 #. '``exception label``': the label reached when a callee returns via
5058 the :ref:`resume <i_resume>` instruction or other exception handling
5060 #. The optional :ref:`function attributes <fnattrs>` list. Only
5061 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
5062 attributes are valid here.
5067 This instruction is designed to operate as a standard '``call``'
5068 instruction in most regards. The primary difference is that it
5069 establishes an association with a label, which is used by the runtime
5070 library to unwind the stack.
5072 This instruction is used in languages with destructors to ensure that
5073 proper cleanup is performed in the case of either a ``longjmp`` or a
5074 thrown exception. Additionally, this is important for implementation of
5075 '``catch``' clauses in high-level languages that support them.
5077 For the purposes of the SSA form, the definition of the value returned
5078 by the '``invoke``' instruction is deemed to occur on the edge from the
5079 current block to the "normal" label. If the callee unwinds then no
5080 return value is available.
5085 .. code-block:: llvm
5087 %retval = invoke i32 @Test(i32 15) to label %Continue
5088 unwind label %TestCleanup ; i32:retval set
5089 %retval = invoke coldcc i32 %Testfnptr(i32 15) to label %Continue
5090 unwind label %TestCleanup ; i32:retval set
5094 '``resume``' Instruction
5095 ^^^^^^^^^^^^^^^^^^^^^^^^
5102 resume <type> <value>
5107 The '``resume``' instruction is a terminator instruction that has no
5113 The '``resume``' instruction requires one argument, which must have the
5114 same type as the result of any '``landingpad``' instruction in the same
5120 The '``resume``' instruction resumes propagation of an existing
5121 (in-flight) exception whose unwinding was interrupted with a
5122 :ref:`landingpad <i_landingpad>` instruction.
5127 .. code-block:: llvm
5129 resume { i8*, i32 } %exn
5133 '``catchpad``' Instruction
5134 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5141 <resultval> = catchpad <resultty> [<args>*]
5142 to label <normal label> unwind label <exception label>
5147 The '``catchpad``' instruction is used by `LLVM's exception handling
5148 system <ExceptionHandling.html#overview>`_ to specify that a basic block
5149 is a catch block --- one where a personality routine attempts to transfer
5150 control to catch an exception.
5151 The ``args`` correspond to whatever information the personality
5152 routine requires to know if this is an appropriate place to catch the
5153 exception. Control is tranfered to the ``exception`` label if the
5154 ``catchpad`` is not an appropriate handler for the in-flight exception.
5155 The ``normal`` label should contain the code found in the ``catch``
5156 portion of a ``try``/``catch`` sequence. It defines values supplied by
5157 the :ref:`personality function <personalityfn>` upon re-entry to the
5158 function. The ``resultval`` has the type ``resultty``.
5163 The instruction takes a list of arbitrary values which are interpreted
5164 by the :ref:`personality function <personalityfn>`.
5166 The ``catchpad`` must be provided a ``normal`` label to transfer control
5167 to if the ``catchpad`` matches the exception and an ``exception``
5168 label to transfer control to if it doesn't.
5173 The '``catchpad``' instruction defines the values which are set by the
5174 :ref:`personality function <personalityfn>` upon re-entry to the function, and
5175 therefore the "result type" of the ``catchpad`` instruction. As with
5176 calling conventions, how the personality function results are
5177 represented in LLVM IR is target specific.
5179 When the call stack is being unwound due to an exception being thrown,
5180 the exception is compared against the ``args``. If it doesn't match,
5181 then control is transfered to the ``exception`` basic block.
5183 The ``catchpad`` instruction has several restrictions:
5185 - A catch block is a basic block which is the unwind destination of
5186 an exceptional instruction.
5187 - A catch block must have a '``catchpad``' instruction as its
5188 first non-PHI instruction.
5189 - A catch block's ``exception`` edge must refer to a catch block or a
5191 - There can be only one '``catchpad``' instruction within the
5193 - A basic block that is not a catch block may not include a
5194 '``catchpad``' instruction.
5195 - It is undefined behavior for control to transfer from a ``catchpad`` to a
5196 ``cleanupret`` without first executing a ``catchret`` and a subsequent
5198 - It is undefined behavior for control to transfer from a ``catchpad`` to a
5199 ``ret`` without first executing a ``catchret``.
5204 .. code-block:: llvm
5206 ;; A catch block which can catch an integer.
5207 %res = catchpad { i8*, i32 } [i8** @_ZTIi]
5208 to label %int.handler unwind label %terminate
5212 '``catchendpad``' Instruction
5213 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5220 catchendpad unwind label <nextaction>
5221 catchendpad unwind to caller
5226 The '``catchendpad``' instruction is used by `LLVM's exception handling
5227 system <ExceptionHandling.html#overview>`_ to communicate to the
5228 :ref:`personality function <personalityfn>` which invokes are associated
5229 with a chain of :ref:`catchpad <i_catchpad>` instructions.
5231 The ``nextaction`` label indicates where control should transfer to if
5232 none of the ``catchpad`` instructions are suitable for catching the
5233 in-flight exception.
5235 If a ``nextaction`` label is not present, the instruction unwinds out of
5236 its parent function. The
5237 :ref:`personality function <personalityfn>` will continue processing
5238 exception handling actions in the caller.
5243 The instruction optionally takes a label, ``nextaction``, indicating
5244 where control should transfer to if none of the preceding
5245 ``catchpad`` instructions are suitable for the in-flight exception.
5250 When the call stack is being unwound due to an exception being thrown
5251 and none of the constituent ``catchpad`` instructions match, then
5252 control is transfered to ``nextaction`` if it is present. If it is not
5253 present, control is transfered to the caller.
5255 The ``catchendpad`` instruction has several restrictions:
5257 - A catch-end block is a basic block which is the unwind destination of
5258 an exceptional instruction.
5259 - A catch-end block must have a '``catchendpad``' instruction as its
5260 first non-PHI instruction.
5261 - There can be only one '``catchendpad``' instruction within the
5263 - A basic block that is not a catch-end block may not include a
5264 '``catchendpad``' instruction.
5265 - Exactly one catch block may unwind to a ``catchendpad``.
5266 - The unwind target of invokes between a ``catchpad`` and a
5267 corresponding ``catchret`` must be its ``catchendpad``.
5272 .. code-block:: llvm
5274 catchendpad unwind label %terminate
5275 catchendpad unwind to caller
5279 '``catchret``' Instruction
5280 ^^^^^^^^^^^^^^^^^^^^^^^^^^
5287 catchret label <normal>
5292 The '``catchret``' instruction is a terminator instruction that has a
5299 The '``catchret``' instruction requires one argument which specifies
5300 where control will transfer to next.
5305 The '``catchret``' instruction ends the existing (in-flight) exception
5306 whose unwinding was interrupted with a
5307 :ref:`catchpad <i_catchpad>` instruction.
5308 The :ref:`personality function <personalityfn>` gets a chance to execute
5309 arbitrary code to, for example, run a C++ destructor.
5310 Control then transfers to ``normal``.
5315 .. code-block:: llvm
5317 catchret label %continue
5321 '``cleanupret``' Instruction
5322 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5329 cleanupret <type> <value> unwind label <continue>
5330 cleanupret <type> <value> unwind to caller
5335 The '``cleanupret``' instruction is a terminator instruction that has
5336 an optional successor.
5342 The '``cleanupret``' instruction requires one argument, which must have the
5343 same type as the result of any '``cleanuppad``' instruction in the same
5344 function. It also has an optional successor, ``continue``.
5349 The '``cleanupret``' instruction indicates to the
5350 :ref:`personality function <personalityfn>` that one
5351 :ref:`cleanuppad <i_cleanuppad>` it transferred control to has ended.
5352 It transfers control to ``continue`` or unwinds out of the function.
5357 .. code-block:: llvm
5359 cleanupret void unwind to caller
5360 cleanupret { i8*, i32 } %exn unwind label %continue
5364 '``terminatepad``' Instruction
5365 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5372 terminatepad [<args>*] unwind label <exception label>
5373 terminatepad [<args>*] unwind to caller
5378 The '``terminatepad``' instruction is used by `LLVM's exception handling
5379 system <ExceptionHandling.html#overview>`_ to specify that a basic block
5380 is a terminate block --- one where a personality routine may decide to
5381 terminate the program.
5382 The ``args`` correspond to whatever information the personality
5383 routine requires to know if this is an appropriate place to terminate the
5384 program. Control is transferred to the ``exception`` label if the
5385 personality routine decides not to terminate the program for the
5386 in-flight exception.
5391 The instruction takes a list of arbitrary values which are interpreted
5392 by the :ref:`personality function <personalityfn>`.
5394 The ``terminatepad`` may be given an ``exception`` label to
5395 transfer control to if the in-flight exception matches the ``args``.
5400 When the call stack is being unwound due to an exception being thrown,
5401 the exception is compared against the ``args``. If it matches,
5402 then control is transfered to the ``exception`` basic block. Otherwise,
5403 the program is terminated via personality-specific means. Typically,
5404 the first argument to ``terminatepad`` specifies what function the
5405 personality should defer to in order to terminate the program.
5407 The ``terminatepad`` instruction has several restrictions:
5409 - A terminate block is a basic block which is the unwind destination of
5410 an exceptional instruction.
5411 - A terminate block must have a '``terminatepad``' instruction as its
5412 first non-PHI instruction.
5413 - There can be only one '``terminatepad``' instruction within the
5415 - A basic block that is not a terminate block may not include a
5416 '``terminatepad``' instruction.
5421 .. code-block:: llvm
5423 ;; A terminate block which only permits integers.
5424 terminatepad [i8** @_ZTIi] unwind label %continue
5428 '``unreachable``' Instruction
5429 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5441 The '``unreachable``' instruction has no defined semantics. This
5442 instruction is used to inform the optimizer that a particular portion of
5443 the code is not reachable. This can be used to indicate that the code
5444 after a no-return function cannot be reached, and other facts.
5449 The '``unreachable``' instruction has no defined semantics.
5456 Binary operators are used to do most of the computation in a program.
5457 They require two operands of the same type, execute an operation on
5458 them, and produce a single value. The operands might represent multiple
5459 data, as is the case with the :ref:`vector <t_vector>` data type. The
5460 result value has the same type as its operands.
5462 There are several different binary operators:
5466 '``add``' Instruction
5467 ^^^^^^^^^^^^^^^^^^^^^
5474 <result> = add <ty> <op1>, <op2> ; yields ty:result
5475 <result> = add nuw <ty> <op1>, <op2> ; yields ty:result
5476 <result> = add nsw <ty> <op1>, <op2> ; yields ty:result
5477 <result> = add nuw nsw <ty> <op1>, <op2> ; yields ty:result
5482 The '``add``' instruction returns the sum of its two operands.
5487 The two arguments to the '``add``' instruction must be
5488 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
5489 arguments must have identical types.
5494 The value produced is the integer sum of the two operands.
5496 If the sum has unsigned overflow, the result returned is the
5497 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
5500 Because LLVM integers use a two's complement representation, this
5501 instruction is appropriate for both signed and unsigned integers.
5503 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
5504 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
5505 result value of the ``add`` is a :ref:`poison value <poisonvalues>` if
5506 unsigned and/or signed overflow, respectively, occurs.
5511 .. code-block:: llvm
5513 <result> = add i32 4, %var ; yields i32:result = 4 + %var
5517 '``fadd``' Instruction
5518 ^^^^^^^^^^^^^^^^^^^^^^
5525 <result> = fadd [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
5530 The '``fadd``' instruction returns the sum of its two operands.
5535 The two arguments to the '``fadd``' instruction must be :ref:`floating
5536 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
5537 Both arguments must have identical types.
5542 The value produced is the floating point sum of the two operands. This
5543 instruction can also take any number of :ref:`fast-math flags <fastmath>`,
5544 which are optimization hints to enable otherwise unsafe floating point
5550 .. code-block:: llvm
5552 <result> = fadd float 4.0, %var ; yields float:result = 4.0 + %var
5554 '``sub``' Instruction
5555 ^^^^^^^^^^^^^^^^^^^^^
5562 <result> = sub <ty> <op1>, <op2> ; yields ty:result
5563 <result> = sub nuw <ty> <op1>, <op2> ; yields ty:result
5564 <result> = sub nsw <ty> <op1>, <op2> ; yields ty:result
5565 <result> = sub nuw nsw <ty> <op1>, <op2> ; yields ty:result
5570 The '``sub``' instruction returns the difference of its two operands.
5572 Note that the '``sub``' instruction is used to represent the '``neg``'
5573 instruction present in most other intermediate representations.
5578 The two arguments to the '``sub``' instruction must be
5579 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
5580 arguments must have identical types.
5585 The value produced is the integer difference of the two operands.
5587 If the difference has unsigned overflow, the result returned is the
5588 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
5591 Because LLVM integers use a two's complement representation, this
5592 instruction is appropriate for both signed and unsigned integers.
5594 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
5595 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
5596 result value of the ``sub`` is a :ref:`poison value <poisonvalues>` if
5597 unsigned and/or signed overflow, respectively, occurs.
5602 .. code-block:: llvm
5604 <result> = sub i32 4, %var ; yields i32:result = 4 - %var
5605 <result> = sub i32 0, %val ; yields i32:result = -%var
5609 '``fsub``' Instruction
5610 ^^^^^^^^^^^^^^^^^^^^^^
5617 <result> = fsub [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
5622 The '``fsub``' instruction returns the difference of its two operands.
5624 Note that the '``fsub``' instruction is used to represent the '``fneg``'
5625 instruction present in most other intermediate representations.
5630 The two arguments to the '``fsub``' instruction must be :ref:`floating
5631 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
5632 Both arguments must have identical types.
5637 The value produced is the floating point difference of the two operands.
5638 This instruction can also take any number of :ref:`fast-math
5639 flags <fastmath>`, which are optimization hints to enable otherwise
5640 unsafe floating point optimizations:
5645 .. code-block:: llvm
5647 <result> = fsub float 4.0, %var ; yields float:result = 4.0 - %var
5648 <result> = fsub float -0.0, %val ; yields float:result = -%var
5650 '``mul``' Instruction
5651 ^^^^^^^^^^^^^^^^^^^^^
5658 <result> = mul <ty> <op1>, <op2> ; yields ty:result
5659 <result> = mul nuw <ty> <op1>, <op2> ; yields ty:result
5660 <result> = mul nsw <ty> <op1>, <op2> ; yields ty:result
5661 <result> = mul nuw nsw <ty> <op1>, <op2> ; yields ty:result
5666 The '``mul``' instruction returns the product of its two operands.
5671 The two arguments to the '``mul``' instruction must be
5672 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
5673 arguments must have identical types.
5678 The value produced is the integer product of the two operands.
5680 If the result of the multiplication has unsigned overflow, the result
5681 returned is the mathematical result modulo 2\ :sup:`n`\ , where n is the
5682 bit width of the result.
5684 Because LLVM integers use a two's complement representation, and the
5685 result is the same width as the operands, this instruction returns the
5686 correct result for both signed and unsigned integers. If a full product
5687 (e.g. ``i32`` * ``i32`` -> ``i64``) is needed, the operands should be
5688 sign-extended or zero-extended as appropriate to the width of the full
5691 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
5692 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
5693 result value of the ``mul`` is a :ref:`poison value <poisonvalues>` if
5694 unsigned and/or signed overflow, respectively, occurs.
5699 .. code-block:: llvm
5701 <result> = mul i32 4, %var ; yields i32:result = 4 * %var
5705 '``fmul``' Instruction
5706 ^^^^^^^^^^^^^^^^^^^^^^
5713 <result> = fmul [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
5718 The '``fmul``' instruction returns the product of its two operands.
5723 The two arguments to the '``fmul``' instruction must be :ref:`floating
5724 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
5725 Both arguments must have identical types.
5730 The value produced is the floating point product of the two operands.
5731 This instruction can also take any number of :ref:`fast-math
5732 flags <fastmath>`, which are optimization hints to enable otherwise
5733 unsafe floating point optimizations:
5738 .. code-block:: llvm
5740 <result> = fmul float 4.0, %var ; yields float:result = 4.0 * %var
5742 '``udiv``' Instruction
5743 ^^^^^^^^^^^^^^^^^^^^^^
5750 <result> = udiv <ty> <op1>, <op2> ; yields ty:result
5751 <result> = udiv exact <ty> <op1>, <op2> ; yields ty:result
5756 The '``udiv``' instruction returns the quotient of its two operands.
5761 The two arguments to the '``udiv``' instruction must be
5762 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
5763 arguments must have identical types.
5768 The value produced is the unsigned integer quotient of the two operands.
5770 Note that unsigned integer division and signed integer division are
5771 distinct operations; for signed integer division, use '``sdiv``'.
5773 Division by zero leads to undefined behavior.
5775 If the ``exact`` keyword is present, the result value of the ``udiv`` is
5776 a :ref:`poison value <poisonvalues>` if %op1 is not a multiple of %op2 (as
5777 such, "((a udiv exact b) mul b) == a").
5782 .. code-block:: llvm
5784 <result> = udiv i32 4, %var ; yields i32:result = 4 / %var
5786 '``sdiv``' Instruction
5787 ^^^^^^^^^^^^^^^^^^^^^^
5794 <result> = sdiv <ty> <op1>, <op2> ; yields ty:result
5795 <result> = sdiv exact <ty> <op1>, <op2> ; yields ty:result
5800 The '``sdiv``' instruction returns the quotient of its two operands.
5805 The two arguments to the '``sdiv``' instruction must be
5806 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
5807 arguments must have identical types.
5812 The value produced is the signed integer quotient of the two operands
5813 rounded towards zero.
5815 Note that signed integer division and unsigned integer division are
5816 distinct operations; for unsigned integer division, use '``udiv``'.
5818 Division by zero leads to undefined behavior. Overflow also leads to
5819 undefined behavior; this is a rare case, but can occur, for example, by
5820 doing a 32-bit division of -2147483648 by -1.
5822 If the ``exact`` keyword is present, the result value of the ``sdiv`` is
5823 a :ref:`poison value <poisonvalues>` if the result would be rounded.
5828 .. code-block:: llvm
5830 <result> = sdiv i32 4, %var ; yields i32:result = 4 / %var
5834 '``fdiv``' Instruction
5835 ^^^^^^^^^^^^^^^^^^^^^^
5842 <result> = fdiv [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
5847 The '``fdiv``' instruction returns the quotient of its two operands.
5852 The two arguments to the '``fdiv``' instruction must be :ref:`floating
5853 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
5854 Both arguments must have identical types.
5859 The value produced is the floating point quotient of the two operands.
5860 This instruction can also take any number of :ref:`fast-math
5861 flags <fastmath>`, which are optimization hints to enable otherwise
5862 unsafe floating point optimizations:
5867 .. code-block:: llvm
5869 <result> = fdiv float 4.0, %var ; yields float:result = 4.0 / %var
5871 '``urem``' Instruction
5872 ^^^^^^^^^^^^^^^^^^^^^^
5879 <result> = urem <ty> <op1>, <op2> ; yields ty:result
5884 The '``urem``' instruction returns the remainder from the unsigned
5885 division of its two arguments.
5890 The two arguments to the '``urem``' instruction must be
5891 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
5892 arguments must have identical types.
5897 This instruction returns the unsigned integer *remainder* of a division.
5898 This instruction always performs an unsigned division to get the
5901 Note that unsigned integer remainder and signed integer remainder are
5902 distinct operations; for signed integer remainder, use '``srem``'.
5904 Taking the remainder of a division by zero leads to undefined behavior.
5909 .. code-block:: llvm
5911 <result> = urem i32 4, %var ; yields i32:result = 4 % %var
5913 '``srem``' Instruction
5914 ^^^^^^^^^^^^^^^^^^^^^^
5921 <result> = srem <ty> <op1>, <op2> ; yields ty:result
5926 The '``srem``' instruction returns the remainder from the signed
5927 division of its two operands. This instruction can also take
5928 :ref:`vector <t_vector>` versions of the values in which case the elements
5934 The two arguments to the '``srem``' instruction must be
5935 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
5936 arguments must have identical types.
5941 This instruction returns the *remainder* of a division (where the result
5942 is either zero or has the same sign as the dividend, ``op1``), not the
5943 *modulo* operator (where the result is either zero or has the same sign
5944 as the divisor, ``op2``) of a value. For more information about the
5945 difference, see `The Math
5946 Forum <http://mathforum.org/dr.math/problems/anne.4.28.99.html>`_. For a
5947 table of how this is implemented in various languages, please see
5949 operation <http://en.wikipedia.org/wiki/Modulo_operation>`_.
5951 Note that signed integer remainder and unsigned integer remainder are
5952 distinct operations; for unsigned integer remainder, use '``urem``'.
5954 Taking the remainder of a division by zero leads to undefined behavior.
5955 Overflow also leads to undefined behavior; this is a rare case, but can
5956 occur, for example, by taking the remainder of a 32-bit division of
5957 -2147483648 by -1. (The remainder doesn't actually overflow, but this
5958 rule lets srem be implemented using instructions that return both the
5959 result of the division and the remainder.)
5964 .. code-block:: llvm
5966 <result> = srem i32 4, %var ; yields i32:result = 4 % %var
5970 '``frem``' Instruction
5971 ^^^^^^^^^^^^^^^^^^^^^^
5978 <result> = frem [fast-math flags]* <ty> <op1>, <op2> ; yields ty:result
5983 The '``frem``' instruction returns the remainder from the division of
5989 The two arguments to the '``frem``' instruction must be :ref:`floating
5990 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
5991 Both arguments must have identical types.
5996 This instruction returns the *remainder* of a division. The remainder
5997 has the same sign as the dividend. This instruction can also take any
5998 number of :ref:`fast-math flags <fastmath>`, which are optimization hints
5999 to enable otherwise unsafe floating point optimizations:
6004 .. code-block:: llvm
6006 <result> = frem float 4.0, %var ; yields float:result = 4.0 % %var
6010 Bitwise Binary Operations
6011 -------------------------
6013 Bitwise binary operators are used to do various forms of bit-twiddling
6014 in a program. They are generally very efficient instructions and can
6015 commonly be strength reduced from other instructions. They require two
6016 operands of the same type, execute an operation on them, and produce a
6017 single value. The resulting value is the same type as its operands.
6019 '``shl``' Instruction
6020 ^^^^^^^^^^^^^^^^^^^^^
6027 <result> = shl <ty> <op1>, <op2> ; yields ty:result
6028 <result> = shl nuw <ty> <op1>, <op2> ; yields ty:result
6029 <result> = shl nsw <ty> <op1>, <op2> ; yields ty:result
6030 <result> = shl nuw nsw <ty> <op1>, <op2> ; yields ty:result
6035 The '``shl``' instruction returns the first operand shifted to the left
6036 a specified number of bits.
6041 Both arguments to the '``shl``' instruction must be the same
6042 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
6043 '``op2``' is treated as an unsigned value.
6048 The value produced is ``op1`` \* 2\ :sup:`op2` mod 2\ :sup:`n`,
6049 where ``n`` is the width of the result. If ``op2`` is (statically or
6050 dynamically) equal to or larger than the number of bits in
6051 ``op1``, the result is undefined. If the arguments are vectors, each
6052 vector element of ``op1`` is shifted by the corresponding shift amount
6055 If the ``nuw`` keyword is present, then the shift produces a :ref:`poison
6056 value <poisonvalues>` if it shifts out any non-zero bits. If the
6057 ``nsw`` keyword is present, then the shift produces a :ref:`poison
6058 value <poisonvalues>` if it shifts out any bits that disagree with the
6059 resultant sign bit. As such, NUW/NSW have the same semantics as they
6060 would if the shift were expressed as a mul instruction with the same
6061 nsw/nuw bits in (mul %op1, (shl 1, %op2)).
6066 .. code-block:: llvm
6068 <result> = shl i32 4, %var ; yields i32: 4 << %var
6069 <result> = shl i32 4, 2 ; yields i32: 16
6070 <result> = shl i32 1, 10 ; yields i32: 1024
6071 <result> = shl i32 1, 32 ; undefined
6072 <result> = shl <2 x i32> < i32 1, i32 1>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 2, i32 4>
6074 '``lshr``' Instruction
6075 ^^^^^^^^^^^^^^^^^^^^^^
6082 <result> = lshr <ty> <op1>, <op2> ; yields ty:result
6083 <result> = lshr exact <ty> <op1>, <op2> ; yields ty:result
6088 The '``lshr``' instruction (logical shift right) returns the first
6089 operand shifted to the right a specified number of bits with zero fill.
6094 Both arguments to the '``lshr``' instruction must be the same
6095 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
6096 '``op2``' is treated as an unsigned value.
6101 This instruction always performs a logical shift right operation. The
6102 most significant bits of the result will be filled with zero bits after
6103 the shift. If ``op2`` is (statically or dynamically) equal to or larger
6104 than the number of bits in ``op1``, the result is undefined. If the
6105 arguments are vectors, each vector element of ``op1`` is shifted by the
6106 corresponding shift amount in ``op2``.
6108 If the ``exact`` keyword is present, the result value of the ``lshr`` is
6109 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
6115 .. code-block:: llvm
6117 <result> = lshr i32 4, 1 ; yields i32:result = 2
6118 <result> = lshr i32 4, 2 ; yields i32:result = 1
6119 <result> = lshr i8 4, 3 ; yields i8:result = 0
6120 <result> = lshr i8 -2, 1 ; yields i8:result = 0x7F
6121 <result> = lshr i32 1, 32 ; undefined
6122 <result> = lshr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 0x7FFFFFFF, i32 1>
6124 '``ashr``' Instruction
6125 ^^^^^^^^^^^^^^^^^^^^^^
6132 <result> = ashr <ty> <op1>, <op2> ; yields ty:result
6133 <result> = ashr exact <ty> <op1>, <op2> ; yields ty:result
6138 The '``ashr``' instruction (arithmetic shift right) returns the first
6139 operand shifted to the right a specified number of bits with sign
6145 Both arguments to the '``ashr``' instruction must be the same
6146 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
6147 '``op2``' is treated as an unsigned value.
6152 This instruction always performs an arithmetic shift right operation,
6153 The most significant bits of the result will be filled with the sign bit
6154 of ``op1``. If ``op2`` is (statically or dynamically) equal to or larger
6155 than the number of bits in ``op1``, the result is undefined. If the
6156 arguments are vectors, each vector element of ``op1`` is shifted by the
6157 corresponding shift amount in ``op2``.
6159 If the ``exact`` keyword is present, the result value of the ``ashr`` is
6160 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
6166 .. code-block:: llvm
6168 <result> = ashr i32 4, 1 ; yields i32:result = 2
6169 <result> = ashr i32 4, 2 ; yields i32:result = 1
6170 <result> = ashr i8 4, 3 ; yields i8:result = 0
6171 <result> = ashr i8 -2, 1 ; yields i8:result = -1
6172 <result> = ashr i32 1, 32 ; undefined
6173 <result> = ashr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 3> ; yields: result=<2 x i32> < i32 -1, i32 0>
6175 '``and``' Instruction
6176 ^^^^^^^^^^^^^^^^^^^^^
6183 <result> = and <ty> <op1>, <op2> ; yields ty:result
6188 The '``and``' instruction returns the bitwise logical and of its two
6194 The two arguments to the '``and``' instruction must be
6195 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6196 arguments must have identical types.
6201 The truth table used for the '``and``' instruction is:
6218 .. code-block:: llvm
6220 <result> = and i32 4, %var ; yields i32:result = 4 & %var
6221 <result> = and i32 15, 40 ; yields i32:result = 8
6222 <result> = and i32 4, 8 ; yields i32:result = 0
6224 '``or``' Instruction
6225 ^^^^^^^^^^^^^^^^^^^^
6232 <result> = or <ty> <op1>, <op2> ; yields ty:result
6237 The '``or``' instruction returns the bitwise logical inclusive or of its
6243 The two arguments to the '``or``' instruction must be
6244 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6245 arguments must have identical types.
6250 The truth table used for the '``or``' instruction is:
6269 <result> = or i32 4, %var ; yields i32:result = 4 | %var
6270 <result> = or i32 15, 40 ; yields i32:result = 47
6271 <result> = or i32 4, 8 ; yields i32:result = 12
6273 '``xor``' Instruction
6274 ^^^^^^^^^^^^^^^^^^^^^
6281 <result> = xor <ty> <op1>, <op2> ; yields ty:result
6286 The '``xor``' instruction returns the bitwise logical exclusive or of
6287 its two operands. The ``xor`` is used to implement the "one's
6288 complement" operation, which is the "~" operator in C.
6293 The two arguments to the '``xor``' instruction must be
6294 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
6295 arguments must have identical types.
6300 The truth table used for the '``xor``' instruction is:
6317 .. code-block:: llvm
6319 <result> = xor i32 4, %var ; yields i32:result = 4 ^ %var
6320 <result> = xor i32 15, 40 ; yields i32:result = 39
6321 <result> = xor i32 4, 8 ; yields i32:result = 12
6322 <result> = xor i32 %V, -1 ; yields i32:result = ~%V
6327 LLVM supports several instructions to represent vector operations in a
6328 target-independent manner. These instructions cover the element-access
6329 and vector-specific operations needed to process vectors effectively.
6330 While LLVM does directly support these vector operations, many
6331 sophisticated algorithms will want to use target-specific intrinsics to
6332 take full advantage of a specific target.
6334 .. _i_extractelement:
6336 '``extractelement``' Instruction
6337 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6344 <result> = extractelement <n x <ty>> <val>, <ty2> <idx> ; yields <ty>
6349 The '``extractelement``' instruction extracts a single scalar element
6350 from a vector at a specified index.
6355 The first operand of an '``extractelement``' instruction is a value of
6356 :ref:`vector <t_vector>` type. The second operand is an index indicating
6357 the position from which to extract the element. The index may be a
6358 variable of any integer type.
6363 The result is a scalar of the same type as the element type of ``val``.
6364 Its value is the value at position ``idx`` of ``val``. If ``idx``
6365 exceeds the length of ``val``, the results are undefined.
6370 .. code-block:: llvm
6372 <result> = extractelement <4 x i32> %vec, i32 0 ; yields i32
6374 .. _i_insertelement:
6376 '``insertelement``' Instruction
6377 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6384 <result> = insertelement <n x <ty>> <val>, <ty> <elt>, <ty2> <idx> ; yields <n x <ty>>
6389 The '``insertelement``' instruction inserts a scalar element into a
6390 vector at a specified index.
6395 The first operand of an '``insertelement``' instruction is a value of
6396 :ref:`vector <t_vector>` type. The second operand is a scalar value whose
6397 type must equal the element type of the first operand. The third operand
6398 is an index indicating the position at which to insert the value. The
6399 index may be a variable of any integer type.
6404 The result is a vector of the same type as ``val``. Its element values
6405 are those of ``val`` except at position ``idx``, where it gets the value
6406 ``elt``. If ``idx`` exceeds the length of ``val``, the results are
6412 .. code-block:: llvm
6414 <result> = insertelement <4 x i32> %vec, i32 1, i32 0 ; yields <4 x i32>
6416 .. _i_shufflevector:
6418 '``shufflevector``' Instruction
6419 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6426 <result> = shufflevector <n x <ty>> <v1>, <n x <ty>> <v2>, <m x i32> <mask> ; yields <m x <ty>>
6431 The '``shufflevector``' instruction constructs a permutation of elements
6432 from two input vectors, returning a vector with the same element type as
6433 the input and length that is the same as the shuffle mask.
6438 The first two operands of a '``shufflevector``' instruction are vectors
6439 with the same type. The third argument is a shuffle mask whose element
6440 type is always 'i32'. The result of the instruction is a vector whose
6441 length is the same as the shuffle mask and whose element type is the
6442 same as the element type of the first two operands.
6444 The shuffle mask operand is required to be a constant vector with either
6445 constant integer or undef values.
6450 The elements of the two input vectors are numbered from left to right
6451 across both of the vectors. The shuffle mask operand specifies, for each
6452 element of the result vector, which element of the two input vectors the
6453 result element gets. The element selector may be undef (meaning "don't
6454 care") and the second operand may be undef if performing a shuffle from
6460 .. code-block:: llvm
6462 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
6463 <4 x i32> <i32 0, i32 4, i32 1, i32 5> ; yields <4 x i32>
6464 <result> = shufflevector <4 x i32> %v1, <4 x i32> undef,
6465 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32> - Identity shuffle.
6466 <result> = shufflevector <8 x i32> %v1, <8 x i32> undef,
6467 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32>
6468 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
6469 <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7 > ; yields <8 x i32>
6471 Aggregate Operations
6472 --------------------
6474 LLVM supports several instructions for working with
6475 :ref:`aggregate <t_aggregate>` values.
6479 '``extractvalue``' Instruction
6480 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6487 <result> = extractvalue <aggregate type> <val>, <idx>{, <idx>}*
6492 The '``extractvalue``' instruction extracts the value of a member field
6493 from an :ref:`aggregate <t_aggregate>` value.
6498 The first operand of an '``extractvalue``' instruction is a value of
6499 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The operands are
6500 constant indices to specify which value to extract in a similar manner
6501 as indices in a '``getelementptr``' instruction.
6503 The major differences to ``getelementptr`` indexing are:
6505 - Since the value being indexed is not a pointer, the first index is
6506 omitted and assumed to be zero.
6507 - At least one index must be specified.
6508 - Not only struct indices but also array indices must be in bounds.
6513 The result is the value at the position in the aggregate specified by
6519 .. code-block:: llvm
6521 <result> = extractvalue {i32, float} %agg, 0 ; yields i32
6525 '``insertvalue``' Instruction
6526 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6533 <result> = insertvalue <aggregate type> <val>, <ty> <elt>, <idx>{, <idx>}* ; yields <aggregate type>
6538 The '``insertvalue``' instruction inserts a value into a member field in
6539 an :ref:`aggregate <t_aggregate>` value.
6544 The first operand of an '``insertvalue``' instruction is a value of
6545 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The second operand is
6546 a first-class value to insert. The following operands are constant
6547 indices indicating the position at which to insert the value in a
6548 similar manner as indices in a '``extractvalue``' instruction. The value
6549 to insert must have the same type as the value identified by the
6555 The result is an aggregate of the same type as ``val``. Its value is
6556 that of ``val`` except that the value at the position specified by the
6557 indices is that of ``elt``.
6562 .. code-block:: llvm
6564 %agg1 = insertvalue {i32, float} undef, i32 1, 0 ; yields {i32 1, float undef}
6565 %agg2 = insertvalue {i32, float} %agg1, float %val, 1 ; yields {i32 1, float %val}
6566 %agg3 = insertvalue {i32, {float}} undef, float %val, 1, 0 ; yields {i32 undef, {float %val}}
6570 Memory Access and Addressing Operations
6571 ---------------------------------------
6573 A key design point of an SSA-based representation is how it represents
6574 memory. In LLVM, no memory locations are in SSA form, which makes things
6575 very simple. This section describes how to read, write, and allocate
6580 '``alloca``' Instruction
6581 ^^^^^^^^^^^^^^^^^^^^^^^^
6588 <result> = alloca [inalloca] <type> [, <ty> <NumElements>] [, align <alignment>] ; yields type*:result
6593 The '``alloca``' instruction allocates memory on the stack frame of the
6594 currently executing function, to be automatically released when this
6595 function returns to its caller. The object is always allocated in the
6596 generic address space (address space zero).
6601 The '``alloca``' instruction allocates ``sizeof(<type>)*NumElements``
6602 bytes of memory on the runtime stack, returning a pointer of the
6603 appropriate type to the program. If "NumElements" is specified, it is
6604 the number of elements allocated, otherwise "NumElements" is defaulted
6605 to be one. If a constant alignment is specified, the value result of the
6606 allocation is guaranteed to be aligned to at least that boundary. The
6607 alignment may not be greater than ``1 << 29``. If not specified, or if
6608 zero, the target can choose to align the allocation on any convenient
6609 boundary compatible with the type.
6611 '``type``' may be any sized type.
6616 Memory is allocated; a pointer is returned. The operation is undefined
6617 if there is insufficient stack space for the allocation. '``alloca``'d
6618 memory is automatically released when the function returns. The
6619 '``alloca``' instruction is commonly used to represent automatic
6620 variables that must have an address available. When the function returns
6621 (either with the ``ret`` or ``resume`` instructions), the memory is
6622 reclaimed. Allocating zero bytes is legal, but the result is undefined.
6623 The order in which memory is allocated (ie., which way the stack grows)
6629 .. code-block:: llvm
6631 %ptr = alloca i32 ; yields i32*:ptr
6632 %ptr = alloca i32, i32 4 ; yields i32*:ptr
6633 %ptr = alloca i32, i32 4, align 1024 ; yields i32*:ptr
6634 %ptr = alloca i32, align 1024 ; yields i32*:ptr
6638 '``load``' Instruction
6639 ^^^^^^^^^^^^^^^^^^^^^^
6646 <result> = load [volatile] <ty>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>][, !invariant.load !<index>][, !nonnull !<index>][, !dereferenceable !<index>][, !dereferenceable_or_null !<index>]
6647 <result> = load atomic [volatile] <ty>* <pointer> [singlethread] <ordering>, align <alignment>
6648 !<index> = !{ i32 1 }
6653 The '``load``' instruction is used to read from memory.
6658 The argument to the ``load`` instruction specifies the memory address
6659 from which to load. The type specified must be a :ref:`first
6660 class <t_firstclass>` type. If the ``load`` is marked as ``volatile``,
6661 then the optimizer is not allowed to modify the number or order of
6662 execution of this ``load`` with other :ref:`volatile
6663 operations <volatile>`.
6665 If the ``load`` is marked as ``atomic``, it takes an extra
6666 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
6667 ``release`` and ``acq_rel`` orderings are not valid on ``load``
6668 instructions. Atomic loads produce :ref:`defined <memmodel>` results
6669 when they may see multiple atomic stores. The type of the pointee must
6670 be an integer type whose bit width is a power of two greater than or
6671 equal to eight and less than or equal to a target-specific size limit.
6672 ``align`` must be explicitly specified on atomic loads, and the load has
6673 undefined behavior if the alignment is not set to a value which is at
6674 least the size in bytes of the pointee. ``!nontemporal`` does not have
6675 any defined semantics for atomic loads.
6677 The optional constant ``align`` argument specifies the alignment of the
6678 operation (that is, the alignment of the memory address). A value of 0
6679 or an omitted ``align`` argument means that the operation has the ABI
6680 alignment for the target. It is the responsibility of the code emitter
6681 to ensure that the alignment information is correct. Overestimating the
6682 alignment results in undefined behavior. Underestimating the alignment
6683 may produce less efficient code. An alignment of 1 is always safe. The
6684 maximum possible alignment is ``1 << 29``.
6686 The optional ``!nontemporal`` metadata must reference a single
6687 metadata name ``<index>`` corresponding to a metadata node with one
6688 ``i32`` entry of value 1. The existence of the ``!nontemporal``
6689 metadata on the instruction tells the optimizer and code generator
6690 that this load is not expected to be reused in the cache. The code
6691 generator may select special instructions to save cache bandwidth, such
6692 as the ``MOVNT`` instruction on x86.
6694 The optional ``!invariant.load`` metadata must reference a single
6695 metadata name ``<index>`` corresponding to a metadata node with no
6696 entries. The existence of the ``!invariant.load`` metadata on the
6697 instruction tells the optimizer and code generator that the address
6698 operand to this load points to memory which can be assumed unchanged.
6699 Being invariant does not imply that a location is dereferenceable,
6700 but it does imply that once the location is known dereferenceable
6701 its value is henceforth unchanging.
6703 The optional ``!nonnull`` metadata must reference a single
6704 metadata name ``<index>`` corresponding to a metadata node with no
6705 entries. The existence of the ``!nonnull`` metadata on the
6706 instruction tells the optimizer that the value loaded is known to
6707 never be null. This is analogous to the ''nonnull'' attribute
6708 on parameters and return values. This metadata can only be applied
6709 to loads of a pointer type.
6711 The optional ``!dereferenceable`` metadata must reference a single
6712 metadata name ``<index>`` corresponding to a metadata node with one ``i64``
6713 entry. The existence of the ``!dereferenceable`` metadata on the instruction
6714 tells the optimizer that the value loaded is known to be dereferenceable.
6715 The number of bytes known to be dereferenceable is specified by the integer
6716 value in the metadata node. This is analogous to the ''dereferenceable''
6717 attribute on parameters and return values. This metadata can only be applied
6718 to loads of a pointer type.
6720 The optional ``!dereferenceable_or_null`` metadata must reference a single
6721 metadata name ``<index>`` corresponding to a metadata node with one ``i64``
6722 entry. The existence of the ``!dereferenceable_or_null`` metadata on the
6723 instruction tells the optimizer that the value loaded is known to be either
6724 dereferenceable or null.
6725 The number of bytes known to be dereferenceable is specified by the integer
6726 value in the metadata node. This is analogous to the ''dereferenceable_or_null''
6727 attribute on parameters and return values. This metadata can only be applied
6728 to loads of a pointer type.
6733 The location of memory pointed to is loaded. If the value being loaded
6734 is of scalar type then the number of bytes read does not exceed the
6735 minimum number of bytes needed to hold all bits of the type. For
6736 example, loading an ``i24`` reads at most three bytes. When loading a
6737 value of a type like ``i20`` with a size that is not an integral number
6738 of bytes, the result is undefined if the value was not originally
6739 written using a store of the same type.
6744 .. code-block:: llvm
6746 %ptr = alloca i32 ; yields i32*:ptr
6747 store i32 3, i32* %ptr ; yields void
6748 %val = load i32, i32* %ptr ; yields i32:val = i32 3
6752 '``store``' Instruction
6753 ^^^^^^^^^^^^^^^^^^^^^^^
6760 store [volatile] <ty> <value>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>] ; yields void
6761 store atomic [volatile] <ty> <value>, <ty>* <pointer> [singlethread] <ordering>, align <alignment> ; yields void
6766 The '``store``' instruction is used to write to memory.
6771 There are two arguments to the ``store`` instruction: a value to store
6772 and an address at which to store it. The type of the ``<pointer>``
6773 operand must be a pointer to the :ref:`first class <t_firstclass>` type of
6774 the ``<value>`` operand. If the ``store`` is marked as ``volatile``,
6775 then the optimizer is not allowed to modify the number or order of
6776 execution of this ``store`` with other :ref:`volatile
6777 operations <volatile>`.
6779 If the ``store`` is marked as ``atomic``, it takes an extra
6780 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
6781 ``acquire`` and ``acq_rel`` orderings aren't valid on ``store``
6782 instructions. Atomic loads produce :ref:`defined <memmodel>` results
6783 when they may see multiple atomic stores. The type of the pointee must
6784 be an integer type whose bit width is a power of two greater than or
6785 equal to eight and less than or equal to a target-specific size limit.
6786 ``align`` must be explicitly specified on atomic stores, and the store
6787 has undefined behavior if the alignment is not set to a value which is
6788 at least the size in bytes of the pointee. ``!nontemporal`` does not
6789 have any defined semantics for atomic stores.
6791 The optional constant ``align`` argument specifies the alignment of the
6792 operation (that is, the alignment of the memory address). A value of 0
6793 or an omitted ``align`` argument means that the operation has the ABI
6794 alignment for the target. It is the responsibility of the code emitter
6795 to ensure that the alignment information is correct. Overestimating the
6796 alignment results in undefined behavior. Underestimating the
6797 alignment may produce less efficient code. An alignment of 1 is always
6798 safe. The maximum possible alignment is ``1 << 29``.
6800 The optional ``!nontemporal`` metadata must reference a single metadata
6801 name ``<index>`` corresponding to a metadata node with one ``i32`` entry of
6802 value 1. The existence of the ``!nontemporal`` metadata on the instruction
6803 tells the optimizer and code generator that this load is not expected to
6804 be reused in the cache. The code generator may select special
6805 instructions to save cache bandwidth, such as the MOVNT instruction on
6811 The contents of memory are updated to contain ``<value>`` at the
6812 location specified by the ``<pointer>`` operand. If ``<value>`` is
6813 of scalar type then the number of bytes written does not exceed the
6814 minimum number of bytes needed to hold all bits of the type. For
6815 example, storing an ``i24`` writes at most three bytes. When writing a
6816 value of a type like ``i20`` with a size that is not an integral number
6817 of bytes, it is unspecified what happens to the extra bits that do not
6818 belong to the type, but they will typically be overwritten.
6823 .. code-block:: llvm
6825 %ptr = alloca i32 ; yields i32*:ptr
6826 store i32 3, i32* %ptr ; yields void
6827 %val = load i32, i32* %ptr ; yields i32:val = i32 3
6831 '``fence``' Instruction
6832 ^^^^^^^^^^^^^^^^^^^^^^^
6839 fence [singlethread] <ordering> ; yields void
6844 The '``fence``' instruction is used to introduce happens-before edges
6850 '``fence``' instructions take an :ref:`ordering <ordering>` argument which
6851 defines what *synchronizes-with* edges they add. They can only be given
6852 ``acquire``, ``release``, ``acq_rel``, and ``seq_cst`` orderings.
6857 A fence A which has (at least) ``release`` ordering semantics
6858 *synchronizes with* a fence B with (at least) ``acquire`` ordering
6859 semantics if and only if there exist atomic operations X and Y, both
6860 operating on some atomic object M, such that A is sequenced before X, X
6861 modifies M (either directly or through some side effect of a sequence
6862 headed by X), Y is sequenced before B, and Y observes M. This provides a
6863 *happens-before* dependency between A and B. Rather than an explicit
6864 ``fence``, one (but not both) of the atomic operations X or Y might
6865 provide a ``release`` or ``acquire`` (resp.) ordering constraint and
6866 still *synchronize-with* the explicit ``fence`` and establish the
6867 *happens-before* edge.
6869 A ``fence`` which has ``seq_cst`` ordering, in addition to having both
6870 ``acquire`` and ``release`` semantics specified above, participates in
6871 the global program order of other ``seq_cst`` operations and/or fences.
6873 The optional ":ref:`singlethread <singlethread>`" argument specifies
6874 that the fence only synchronizes with other fences in the same thread.
6875 (This is useful for interacting with signal handlers.)
6880 .. code-block:: llvm
6882 fence acquire ; yields void
6883 fence singlethread seq_cst ; yields void
6887 '``cmpxchg``' Instruction
6888 ^^^^^^^^^^^^^^^^^^^^^^^^^
6895 cmpxchg [weak] [volatile] <ty>* <pointer>, <ty> <cmp>, <ty> <new> [singlethread] <success ordering> <failure ordering> ; yields { ty, i1 }
6900 The '``cmpxchg``' instruction is used to atomically modify memory. It
6901 loads a value in memory and compares it to a given value. If they are
6902 equal, it tries to store a new value into the memory.
6907 There are three arguments to the '``cmpxchg``' instruction: an address
6908 to operate on, a value to compare to the value currently be at that
6909 address, and a new value to place at that address if the compared values
6910 are equal. The type of '<cmp>' must be an integer type whose bit width
6911 is a power of two greater than or equal to eight and less than or equal
6912 to a target-specific size limit. '<cmp>' and '<new>' must have the same
6913 type, and the type of '<pointer>' must be a pointer to that type. If the
6914 ``cmpxchg`` is marked as ``volatile``, then the optimizer is not allowed
6915 to modify the number or order of execution of this ``cmpxchg`` with
6916 other :ref:`volatile operations <volatile>`.
6918 The success and failure :ref:`ordering <ordering>` arguments specify how this
6919 ``cmpxchg`` synchronizes with other atomic operations. Both ordering parameters
6920 must be at least ``monotonic``, the ordering constraint on failure must be no
6921 stronger than that on success, and the failure ordering cannot be either
6922 ``release`` or ``acq_rel``.
6924 The optional "``singlethread``" argument declares that the ``cmpxchg``
6925 is only atomic with respect to code (usually signal handlers) running in
6926 the same thread as the ``cmpxchg``. Otherwise the cmpxchg is atomic with
6927 respect to all other code in the system.
6929 The pointer passed into cmpxchg must have alignment greater than or
6930 equal to the size in memory of the operand.
6935 The contents of memory at the location specified by the '``<pointer>``' operand
6936 is read and compared to '``<cmp>``'; if the read value is the equal, the
6937 '``<new>``' is written. The original value at the location is returned, together
6938 with a flag indicating success (true) or failure (false).
6940 If the cmpxchg operation is marked as ``weak`` then a spurious failure is
6941 permitted: the operation may not write ``<new>`` even if the comparison
6944 If the cmpxchg operation is strong (the default), the i1 value is 1 if and only
6945 if the value loaded equals ``cmp``.
6947 A successful ``cmpxchg`` is a read-modify-write instruction for the purpose of
6948 identifying release sequences. A failed ``cmpxchg`` is equivalent to an atomic
6949 load with an ordering parameter determined the second ordering parameter.
6954 .. code-block:: llvm
6957 %orig = atomic load i32, i32* %ptr unordered ; yields i32
6961 %cmp = phi i32 [ %orig, %entry ], [%old, %loop]
6962 %squared = mul i32 %cmp, %cmp
6963 %val_success = cmpxchg i32* %ptr, i32 %cmp, i32 %squared acq_rel monotonic ; yields { i32, i1 }
6964 %value_loaded = extractvalue { i32, i1 } %val_success, 0
6965 %success = extractvalue { i32, i1 } %val_success, 1
6966 br i1 %success, label %done, label %loop
6973 '``atomicrmw``' Instruction
6974 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6981 atomicrmw [volatile] <operation> <ty>* <pointer>, <ty> <value> [singlethread] <ordering> ; yields ty
6986 The '``atomicrmw``' instruction is used to atomically modify memory.
6991 There are three arguments to the '``atomicrmw``' instruction: an
6992 operation to apply, an address whose value to modify, an argument to the
6993 operation. The operation must be one of the following keywords:
7007 The type of '<value>' must be an integer type whose bit width is a power
7008 of two greater than or equal to eight and less than or equal to a
7009 target-specific size limit. The type of the '``<pointer>``' operand must
7010 be a pointer to that type. If the ``atomicrmw`` is marked as
7011 ``volatile``, then the optimizer is not allowed to modify the number or
7012 order of execution of this ``atomicrmw`` with other :ref:`volatile
7013 operations <volatile>`.
7018 The contents of memory at the location specified by the '``<pointer>``'
7019 operand are atomically read, modified, and written back. The original
7020 value at the location is returned. The modification is specified by the
7023 - xchg: ``*ptr = val``
7024 - add: ``*ptr = *ptr + val``
7025 - sub: ``*ptr = *ptr - val``
7026 - and: ``*ptr = *ptr & val``
7027 - nand: ``*ptr = ~(*ptr & val)``
7028 - or: ``*ptr = *ptr | val``
7029 - xor: ``*ptr = *ptr ^ val``
7030 - max: ``*ptr = *ptr > val ? *ptr : val`` (using a signed comparison)
7031 - min: ``*ptr = *ptr < val ? *ptr : val`` (using a signed comparison)
7032 - umax: ``*ptr = *ptr > val ? *ptr : val`` (using an unsigned
7034 - umin: ``*ptr = *ptr < val ? *ptr : val`` (using an unsigned
7040 .. code-block:: llvm
7042 %old = atomicrmw add i32* %ptr, i32 1 acquire ; yields i32
7044 .. _i_getelementptr:
7046 '``getelementptr``' Instruction
7047 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7054 <result> = getelementptr <ty>, <ty>* <ptrval>{, <ty> <idx>}*
7055 <result> = getelementptr inbounds <ty>, <ty>* <ptrval>{, <ty> <idx>}*
7056 <result> = getelementptr <ty>, <ptr vector> <ptrval>, <vector index type> <idx>
7061 The '``getelementptr``' instruction is used to get the address of a
7062 subelement of an :ref:`aggregate <t_aggregate>` data structure. It performs
7063 address calculation only and does not access memory. The instruction can also
7064 be used to calculate a vector of such addresses.
7069 The first argument is always a type used as the basis for the calculations.
7070 The second argument is always a pointer or a vector of pointers, and is the
7071 base address to start from. The remaining arguments are indices
7072 that indicate which of the elements of the aggregate object are indexed.
7073 The interpretation of each index is dependent on the type being indexed
7074 into. The first index always indexes the pointer value given as the
7075 first argument, the second index indexes a value of the type pointed to
7076 (not necessarily the value directly pointed to, since the first index
7077 can be non-zero), etc. The first type indexed into must be a pointer
7078 value, subsequent types can be arrays, vectors, and structs. Note that
7079 subsequent types being indexed into can never be pointers, since that
7080 would require loading the pointer before continuing calculation.
7082 The type of each index argument depends on the type it is indexing into.
7083 When indexing into a (optionally packed) structure, only ``i32`` integer
7084 **constants** are allowed (when using a vector of indices they must all
7085 be the **same** ``i32`` integer constant). When indexing into an array,
7086 pointer or vector, integers of any width are allowed, and they are not
7087 required to be constant. These integers are treated as signed values
7090 For example, let's consider a C code fragment and how it gets compiled
7106 int *foo(struct ST *s) {
7107 return &s[1].Z.B[5][13];
7110 The LLVM code generated by Clang is:
7112 .. code-block:: llvm
7114 %struct.RT = type { i8, [10 x [20 x i32]], i8 }
7115 %struct.ST = type { i32, double, %struct.RT }
7117 define i32* @foo(%struct.ST* %s) nounwind uwtable readnone optsize ssp {
7119 %arrayidx = getelementptr inbounds %struct.ST, %struct.ST* %s, i64 1, i32 2, i32 1, i64 5, i64 13
7126 In the example above, the first index is indexing into the
7127 '``%struct.ST*``' type, which is a pointer, yielding a '``%struct.ST``'
7128 = '``{ i32, double, %struct.RT }``' type, a structure. The second index
7129 indexes into the third element of the structure, yielding a
7130 '``%struct.RT``' = '``{ i8 , [10 x [20 x i32]], i8 }``' type, another
7131 structure. The third index indexes into the second element of the
7132 structure, yielding a '``[10 x [20 x i32]]``' type, an array. The two
7133 dimensions of the array are subscripted into, yielding an '``i32``'
7134 type. The '``getelementptr``' instruction returns a pointer to this
7135 element, thus computing a value of '``i32*``' type.
7137 Note that it is perfectly legal to index partially through a structure,
7138 returning a pointer to an inner element. Because of this, the LLVM code
7139 for the given testcase is equivalent to:
7141 .. code-block:: llvm
7143 define i32* @foo(%struct.ST* %s) {
7144 %t1 = getelementptr %struct.ST, %struct.ST* %s, i32 1 ; yields %struct.ST*:%t1
7145 %t2 = getelementptr %struct.ST, %struct.ST* %t1, i32 0, i32 2 ; yields %struct.RT*:%t2
7146 %t3 = getelementptr %struct.RT, %struct.RT* %t2, i32 0, i32 1 ; yields [10 x [20 x i32]]*:%t3
7147 %t4 = getelementptr [10 x [20 x i32]], [10 x [20 x i32]]* %t3, i32 0, i32 5 ; yields [20 x i32]*:%t4
7148 %t5 = getelementptr [20 x i32], [20 x i32]* %t4, i32 0, i32 13 ; yields i32*:%t5
7152 If the ``inbounds`` keyword is present, the result value of the
7153 ``getelementptr`` is a :ref:`poison value <poisonvalues>` if the base
7154 pointer is not an *in bounds* address of an allocated object, or if any
7155 of the addresses that would be formed by successive addition of the
7156 offsets implied by the indices to the base address with infinitely
7157 precise signed arithmetic are not an *in bounds* address of that
7158 allocated object. The *in bounds* addresses for an allocated object are
7159 all the addresses that point into the object, plus the address one byte
7160 past the end. In cases where the base is a vector of pointers the
7161 ``inbounds`` keyword applies to each of the computations element-wise.
7163 If the ``inbounds`` keyword is not present, the offsets are added to the
7164 base address with silently-wrapping two's complement arithmetic. If the
7165 offsets have a different width from the pointer, they are sign-extended
7166 or truncated to the width of the pointer. The result value of the
7167 ``getelementptr`` may be outside the object pointed to by the base
7168 pointer. The result value may not necessarily be used to access memory
7169 though, even if it happens to point into allocated storage. See the
7170 :ref:`Pointer Aliasing Rules <pointeraliasing>` section for more
7173 The getelementptr instruction is often confusing. For some more insight
7174 into how it works, see :doc:`the getelementptr FAQ <GetElementPtr>`.
7179 .. code-block:: llvm
7181 ; yields [12 x i8]*:aptr
7182 %aptr = getelementptr {i32, [12 x i8]}, {i32, [12 x i8]}* %saptr, i64 0, i32 1
7184 %vptr = getelementptr {i32, <2 x i8>}, {i32, <2 x i8>}* %svptr, i64 0, i32 1, i32 1
7186 %eptr = getelementptr [12 x i8], [12 x i8]* %aptr, i64 0, i32 1
7188 %iptr = getelementptr [10 x i32], [10 x i32]* @arr, i16 0, i16 0
7193 The ``getelementptr`` returns a vector of pointers, instead of a single address,
7194 when one or more of its arguments is a vector. In such cases, all vector
7195 arguments should have the same number of elements, and every scalar argument
7196 will be effectively broadcast into a vector during address calculation.
7198 .. code-block:: llvm
7200 ; All arguments are vectors:
7201 ; A[i] = ptrs[i] + offsets[i]*sizeof(i8)
7202 %A = getelementptr i8, <4 x i8*> %ptrs, <4 x i64> %offsets
7204 ; Add the same scalar offset to each pointer of a vector:
7205 ; A[i] = ptrs[i] + offset*sizeof(i8)
7206 %A = getelementptr i8, <4 x i8*> %ptrs, i64 %offset
7208 ; Add distinct offsets to the same pointer:
7209 ; A[i] = ptr + offsets[i]*sizeof(i8)
7210 %A = getelementptr i8, i8* %ptr, <4 x i64> %offsets
7212 ; In all cases described above the type of the result is <4 x i8*>
7214 The two following instructions are equivalent:
7216 .. code-block:: llvm
7218 getelementptr %struct.ST, <4 x %struct.ST*> %s, <4 x i64> %ind1,
7219 <4 x i32> <i32 2, i32 2, i32 2, i32 2>,
7220 <4 x i32> <i32 1, i32 1, i32 1, i32 1>,
7222 <4 x i64> <i64 13, i64 13, i64 13, i64 13>
7224 getelementptr %struct.ST, <4 x %struct.ST*> %s, <4 x i64> %ind1,
7225 i32 2, i32 1, <4 x i32> %ind4, i64 13
7227 Let's look at the C code, where the vector version of ``getelementptr``
7232 // Let's assume that we vectorize the following loop:
7233 double *A, B; int *C;
7234 for (int i = 0; i < size; ++i) {
7238 .. code-block:: llvm
7240 ; get pointers for 8 elements from array B
7241 %ptrs = getelementptr double, double* %B, <8 x i32> %C
7242 ; load 8 elements from array B into A
7243 %A = call <8 x double> @llvm.masked.gather.v8f64(<8 x double*> %ptrs,
7244 i32 8, <8 x i1> %mask, <8 x double> %passthru)
7246 Conversion Operations
7247 ---------------------
7249 The instructions in this category are the conversion instructions
7250 (casting) which all take a single operand and a type. They perform
7251 various bit conversions on the operand.
7253 '``trunc .. to``' Instruction
7254 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7261 <result> = trunc <ty> <value> to <ty2> ; yields ty2
7266 The '``trunc``' instruction truncates its operand to the type ``ty2``.
7271 The '``trunc``' instruction takes a value to trunc, and a type to trunc
7272 it to. Both types must be of :ref:`integer <t_integer>` types, or vectors
7273 of the same number of integers. The bit size of the ``value`` must be
7274 larger than the bit size of the destination type, ``ty2``. Equal sized
7275 types are not allowed.
7280 The '``trunc``' instruction truncates the high order bits in ``value``
7281 and converts the remaining bits to ``ty2``. Since the source size must
7282 be larger than the destination size, ``trunc`` cannot be a *no-op cast*.
7283 It will always truncate bits.
7288 .. code-block:: llvm
7290 %X = trunc i32 257 to i8 ; yields i8:1
7291 %Y = trunc i32 123 to i1 ; yields i1:true
7292 %Z = trunc i32 122 to i1 ; yields i1:false
7293 %W = trunc <2 x i16> <i16 8, i16 7> to <2 x i8> ; yields <i8 8, i8 7>
7295 '``zext .. to``' Instruction
7296 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7303 <result> = zext <ty> <value> to <ty2> ; yields ty2
7308 The '``zext``' instruction zero extends its operand to type ``ty2``.
7313 The '``zext``' instruction takes a value to cast, and a type to cast it
7314 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
7315 the same number of integers. The bit size of the ``value`` must be
7316 smaller than the bit size of the destination type, ``ty2``.
7321 The ``zext`` fills the high order bits of the ``value`` with zero bits
7322 until it reaches the size of the destination type, ``ty2``.
7324 When zero extending from i1, the result will always be either 0 or 1.
7329 .. code-block:: llvm
7331 %X = zext i32 257 to i64 ; yields i64:257
7332 %Y = zext i1 true to i32 ; yields i32:1
7333 %Z = zext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
7335 '``sext .. to``' Instruction
7336 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7343 <result> = sext <ty> <value> to <ty2> ; yields ty2
7348 The '``sext``' sign extends ``value`` to the type ``ty2``.
7353 The '``sext``' instruction takes a value to cast, and a type to cast it
7354 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
7355 the same number of integers. The bit size of the ``value`` must be
7356 smaller than the bit size of the destination type, ``ty2``.
7361 The '``sext``' instruction performs a sign extension by copying the sign
7362 bit (highest order bit) of the ``value`` until it reaches the bit size
7363 of the type ``ty2``.
7365 When sign extending from i1, the extension always results in -1 or 0.
7370 .. code-block:: llvm
7372 %X = sext i8 -1 to i16 ; yields i16 :65535
7373 %Y = sext i1 true to i32 ; yields i32:-1
7374 %Z = sext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
7376 '``fptrunc .. to``' Instruction
7377 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7384 <result> = fptrunc <ty> <value> to <ty2> ; yields ty2
7389 The '``fptrunc``' instruction truncates ``value`` to type ``ty2``.
7394 The '``fptrunc``' instruction takes a :ref:`floating point <t_floating>`
7395 value to cast and a :ref:`floating point <t_floating>` type to cast it to.
7396 The size of ``value`` must be larger than the size of ``ty2``. This
7397 implies that ``fptrunc`` cannot be used to make a *no-op cast*.
7402 The '``fptrunc``' instruction truncates a ``value`` from a larger
7403 :ref:`floating point <t_floating>` type to a smaller :ref:`floating
7404 point <t_floating>` type. If the value cannot fit within the
7405 destination type, ``ty2``, then the results are undefined.
7410 .. code-block:: llvm
7412 %X = fptrunc double 123.0 to float ; yields float:123.0
7413 %Y = fptrunc double 1.0E+300 to float ; yields undefined
7415 '``fpext .. to``' Instruction
7416 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7423 <result> = fpext <ty> <value> to <ty2> ; yields ty2
7428 The '``fpext``' extends a floating point ``value`` to a larger floating
7434 The '``fpext``' instruction takes a :ref:`floating point <t_floating>`
7435 ``value`` to cast, and a :ref:`floating point <t_floating>` type to cast it
7436 to. The source type must be smaller than the destination type.
7441 The '``fpext``' instruction extends the ``value`` from a smaller
7442 :ref:`floating point <t_floating>` type to a larger :ref:`floating
7443 point <t_floating>` type. The ``fpext`` cannot be used to make a
7444 *no-op cast* because it always changes bits. Use ``bitcast`` to make a
7445 *no-op cast* for a floating point cast.
7450 .. code-block:: llvm
7452 %X = fpext float 3.125 to double ; yields double:3.125000e+00
7453 %Y = fpext double %X to fp128 ; yields fp128:0xL00000000000000004000900000000000
7455 '``fptoui .. to``' Instruction
7456 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7463 <result> = fptoui <ty> <value> to <ty2> ; yields ty2
7468 The '``fptoui``' converts a floating point ``value`` to its unsigned
7469 integer equivalent of type ``ty2``.
7474 The '``fptoui``' instruction takes a value to cast, which must be a
7475 scalar or vector :ref:`floating point <t_floating>` value, and a type to
7476 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
7477 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
7478 type with the same number of elements as ``ty``
7483 The '``fptoui``' instruction converts its :ref:`floating
7484 point <t_floating>` operand into the nearest (rounding towards zero)
7485 unsigned integer value. If the value cannot fit in ``ty2``, the results
7491 .. code-block:: llvm
7493 %X = fptoui double 123.0 to i32 ; yields i32:123
7494 %Y = fptoui float 1.0E+300 to i1 ; yields undefined:1
7495 %Z = fptoui float 1.04E+17 to i8 ; yields undefined:1
7497 '``fptosi .. to``' Instruction
7498 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7505 <result> = fptosi <ty> <value> to <ty2> ; yields ty2
7510 The '``fptosi``' instruction converts :ref:`floating point <t_floating>`
7511 ``value`` to type ``ty2``.
7516 The '``fptosi``' instruction takes a value to cast, which must be a
7517 scalar or vector :ref:`floating point <t_floating>` value, and a type to
7518 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
7519 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
7520 type with the same number of elements as ``ty``
7525 The '``fptosi``' instruction converts its :ref:`floating
7526 point <t_floating>` operand into the nearest (rounding towards zero)
7527 signed integer value. If the value cannot fit in ``ty2``, the results
7533 .. code-block:: llvm
7535 %X = fptosi double -123.0 to i32 ; yields i32:-123
7536 %Y = fptosi float 1.0E-247 to i1 ; yields undefined:1
7537 %Z = fptosi float 1.04E+17 to i8 ; yields undefined:1
7539 '``uitofp .. to``' Instruction
7540 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7547 <result> = uitofp <ty> <value> to <ty2> ; yields ty2
7552 The '``uitofp``' instruction regards ``value`` as an unsigned integer
7553 and converts that value to the ``ty2`` type.
7558 The '``uitofp``' instruction takes a value to cast, which must be a
7559 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
7560 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
7561 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
7562 type with the same number of elements as ``ty``
7567 The '``uitofp``' instruction interprets its operand as an unsigned
7568 integer quantity and converts it to the corresponding floating point
7569 value. If the value cannot fit in the floating point value, the results
7575 .. code-block:: llvm
7577 %X = uitofp i32 257 to float ; yields float:257.0
7578 %Y = uitofp i8 -1 to double ; yields double:255.0
7580 '``sitofp .. to``' Instruction
7581 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7588 <result> = sitofp <ty> <value> to <ty2> ; yields ty2
7593 The '``sitofp``' instruction regards ``value`` as a signed integer and
7594 converts that value to the ``ty2`` type.
7599 The '``sitofp``' instruction takes a value to cast, which must be a
7600 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
7601 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
7602 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
7603 type with the same number of elements as ``ty``
7608 The '``sitofp``' instruction interprets its operand as a signed integer
7609 quantity and converts it to the corresponding floating point value. If
7610 the value cannot fit in the floating point value, the results are
7616 .. code-block:: llvm
7618 %X = sitofp i32 257 to float ; yields float:257.0
7619 %Y = sitofp i8 -1 to double ; yields double:-1.0
7623 '``ptrtoint .. to``' Instruction
7624 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7631 <result> = ptrtoint <ty> <value> to <ty2> ; yields ty2
7636 The '``ptrtoint``' instruction converts the pointer or a vector of
7637 pointers ``value`` to the integer (or vector of integers) type ``ty2``.
7642 The '``ptrtoint``' instruction takes a ``value`` to cast, which must be
7643 a value of type :ref:`pointer <t_pointer>` or a vector of pointers, and a
7644 type to cast it to ``ty2``, which must be an :ref:`integer <t_integer>` or
7645 a vector of integers type.
7650 The '``ptrtoint``' instruction converts ``value`` to integer type
7651 ``ty2`` by interpreting the pointer value as an integer and either
7652 truncating or zero extending that value to the size of the integer type.
7653 If ``value`` is smaller than ``ty2`` then a zero extension is done. If
7654 ``value`` is larger than ``ty2`` then a truncation is done. If they are
7655 the same size, then nothing is done (*no-op cast*) other than a type
7661 .. code-block:: llvm
7663 %X = ptrtoint i32* %P to i8 ; yields truncation on 32-bit architecture
7664 %Y = ptrtoint i32* %P to i64 ; yields zero extension on 32-bit architecture
7665 %Z = ptrtoint <4 x i32*> %P to <4 x i64>; yields vector zero extension for a vector of addresses on 32-bit architecture
7669 '``inttoptr .. to``' Instruction
7670 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7677 <result> = inttoptr <ty> <value> to <ty2> ; yields ty2
7682 The '``inttoptr``' instruction converts an integer ``value`` to a
7683 pointer type, ``ty2``.
7688 The '``inttoptr``' instruction takes an :ref:`integer <t_integer>` value to
7689 cast, and a type to cast it to, which must be a :ref:`pointer <t_pointer>`
7695 The '``inttoptr``' instruction converts ``value`` to type ``ty2`` by
7696 applying either a zero extension or a truncation depending on the size
7697 of the integer ``value``. If ``value`` is larger than the size of a
7698 pointer then a truncation is done. If ``value`` is smaller than the size
7699 of a pointer then a zero extension is done. If they are the same size,
7700 nothing is done (*no-op cast*).
7705 .. code-block:: llvm
7707 %X = inttoptr i32 255 to i32* ; yields zero extension on 64-bit architecture
7708 %Y = inttoptr i32 255 to i32* ; yields no-op on 32-bit architecture
7709 %Z = inttoptr i64 0 to i32* ; yields truncation on 32-bit architecture
7710 %Z = inttoptr <4 x i32> %G to <4 x i8*>; yields truncation of vector G to four pointers
7714 '``bitcast .. to``' Instruction
7715 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7722 <result> = bitcast <ty> <value> to <ty2> ; yields ty2
7727 The '``bitcast``' instruction converts ``value`` to type ``ty2`` without
7733 The '``bitcast``' instruction takes a value to cast, which must be a
7734 non-aggregate first class value, and a type to cast it to, which must
7735 also be a non-aggregate :ref:`first class <t_firstclass>` type. The
7736 bit sizes of ``value`` and the destination type, ``ty2``, must be
7737 identical. If the source type is a pointer, the destination type must
7738 also be a pointer of the same size. This instruction supports bitwise
7739 conversion of vectors to integers and to vectors of other types (as
7740 long as they have the same size).
7745 The '``bitcast``' instruction converts ``value`` to type ``ty2``. It
7746 is always a *no-op cast* because no bits change with this
7747 conversion. The conversion is done as if the ``value`` had been stored
7748 to memory and read back as type ``ty2``. Pointer (or vector of
7749 pointers) types may only be converted to other pointer (or vector of
7750 pointers) types with the same address space through this instruction.
7751 To convert pointers to other types, use the :ref:`inttoptr <i_inttoptr>`
7752 or :ref:`ptrtoint <i_ptrtoint>` instructions first.
7757 .. code-block:: llvm
7759 %X = bitcast i8 255 to i8 ; yields i8 :-1
7760 %Y = bitcast i32* %x to sint* ; yields sint*:%x
7761 %Z = bitcast <2 x int> %V to i64; ; yields i64: %V
7762 %Z = bitcast <2 x i32*> %V to <2 x i64*> ; yields <2 x i64*>
7764 .. _i_addrspacecast:
7766 '``addrspacecast .. to``' Instruction
7767 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7774 <result> = addrspacecast <pty> <ptrval> to <pty2> ; yields pty2
7779 The '``addrspacecast``' instruction converts ``ptrval`` from ``pty`` in
7780 address space ``n`` to type ``pty2`` in address space ``m``.
7785 The '``addrspacecast``' instruction takes a pointer or vector of pointer value
7786 to cast and a pointer type to cast it to, which must have a different
7792 The '``addrspacecast``' instruction converts the pointer value
7793 ``ptrval`` to type ``pty2``. It can be a *no-op cast* or a complex
7794 value modification, depending on the target and the address space
7795 pair. Pointer conversions within the same address space must be
7796 performed with the ``bitcast`` instruction. Note that if the address space
7797 conversion is legal then both result and operand refer to the same memory
7803 .. code-block:: llvm
7805 %X = addrspacecast i32* %x to i32 addrspace(1)* ; yields i32 addrspace(1)*:%x
7806 %Y = addrspacecast i32 addrspace(1)* %y to i64 addrspace(2)* ; yields i64 addrspace(2)*:%y
7807 %Z = addrspacecast <4 x i32*> %z to <4 x float addrspace(3)*> ; yields <4 x float addrspace(3)*>:%z
7814 The instructions in this category are the "miscellaneous" instructions,
7815 which defy better classification.
7819 '``icmp``' Instruction
7820 ^^^^^^^^^^^^^^^^^^^^^^
7827 <result> = icmp <cond> <ty> <op1>, <op2> ; yields i1 or <N x i1>:result
7832 The '``icmp``' instruction returns a boolean value or a vector of
7833 boolean values based on comparison of its two integer, integer vector,
7834 pointer, or pointer vector operands.
7839 The '``icmp``' instruction takes three operands. The first operand is
7840 the condition code indicating the kind of comparison to perform. It is
7841 not a value, just a keyword. The possible condition code are:
7844 #. ``ne``: not equal
7845 #. ``ugt``: unsigned greater than
7846 #. ``uge``: unsigned greater or equal
7847 #. ``ult``: unsigned less than
7848 #. ``ule``: unsigned less or equal
7849 #. ``sgt``: signed greater than
7850 #. ``sge``: signed greater or equal
7851 #. ``slt``: signed less than
7852 #. ``sle``: signed less or equal
7854 The remaining two arguments must be :ref:`integer <t_integer>` or
7855 :ref:`pointer <t_pointer>` or integer :ref:`vector <t_vector>` typed. They
7856 must also be identical types.
7861 The '``icmp``' compares ``op1`` and ``op2`` according to the condition
7862 code given as ``cond``. The comparison performed always yields either an
7863 :ref:`i1 <t_integer>` or vector of ``i1`` result, as follows:
7865 #. ``eq``: yields ``true`` if the operands are equal, ``false``
7866 otherwise. No sign interpretation is necessary or performed.
7867 #. ``ne``: yields ``true`` if the operands are unequal, ``false``
7868 otherwise. No sign interpretation is necessary or performed.
7869 #. ``ugt``: interprets the operands as unsigned values and yields
7870 ``true`` if ``op1`` is greater than ``op2``.
7871 #. ``uge``: interprets the operands as unsigned values and yields
7872 ``true`` if ``op1`` is greater than or equal to ``op2``.
7873 #. ``ult``: interprets the operands as unsigned values and yields
7874 ``true`` if ``op1`` is less than ``op2``.
7875 #. ``ule``: interprets the operands as unsigned values and yields
7876 ``true`` if ``op1`` is less than or equal to ``op2``.
7877 #. ``sgt``: interprets the operands as signed values and yields ``true``
7878 if ``op1`` is greater than ``op2``.
7879 #. ``sge``: interprets the operands as signed values and yields ``true``
7880 if ``op1`` is greater than or equal to ``op2``.
7881 #. ``slt``: interprets the operands as signed values and yields ``true``
7882 if ``op1`` is less than ``op2``.
7883 #. ``sle``: interprets the operands as signed values and yields ``true``
7884 if ``op1`` is less than or equal to ``op2``.
7886 If the operands are :ref:`pointer <t_pointer>` typed, the pointer values
7887 are compared as if they were integers.
7889 If the operands are integer vectors, then they are compared element by
7890 element. The result is an ``i1`` vector with the same number of elements
7891 as the values being compared. Otherwise, the result is an ``i1``.
7896 .. code-block:: llvm
7898 <result> = icmp eq i32 4, 5 ; yields: result=false
7899 <result> = icmp ne float* %X, %X ; yields: result=false
7900 <result> = icmp ult i16 4, 5 ; yields: result=true
7901 <result> = icmp sgt i16 4, 5 ; yields: result=false
7902 <result> = icmp ule i16 -4, 5 ; yields: result=false
7903 <result> = icmp sge i16 4, 5 ; yields: result=false
7905 Note that the code generator does not yet support vector types with the
7906 ``icmp`` instruction.
7910 '``fcmp``' Instruction
7911 ^^^^^^^^^^^^^^^^^^^^^^
7918 <result> = fcmp [fast-math flags]* <cond> <ty> <op1>, <op2> ; yields i1 or <N x i1>:result
7923 The '``fcmp``' instruction returns a boolean value or vector of boolean
7924 values based on comparison of its operands.
7926 If the operands are floating point scalars, then the result type is a
7927 boolean (:ref:`i1 <t_integer>`).
7929 If the operands are floating point vectors, then the result type is a
7930 vector of boolean with the same number of elements as the operands being
7936 The '``fcmp``' instruction takes three operands. The first operand is
7937 the condition code indicating the kind of comparison to perform. It is
7938 not a value, just a keyword. The possible condition code are:
7940 #. ``false``: no comparison, always returns false
7941 #. ``oeq``: ordered and equal
7942 #. ``ogt``: ordered and greater than
7943 #. ``oge``: ordered and greater than or equal
7944 #. ``olt``: ordered and less than
7945 #. ``ole``: ordered and less than or equal
7946 #. ``one``: ordered and not equal
7947 #. ``ord``: ordered (no nans)
7948 #. ``ueq``: unordered or equal
7949 #. ``ugt``: unordered or greater than
7950 #. ``uge``: unordered or greater than or equal
7951 #. ``ult``: unordered or less than
7952 #. ``ule``: unordered or less than or equal
7953 #. ``une``: unordered or not equal
7954 #. ``uno``: unordered (either nans)
7955 #. ``true``: no comparison, always returns true
7957 *Ordered* means that neither operand is a QNAN while *unordered* means
7958 that either operand may be a QNAN.
7960 Each of ``val1`` and ``val2`` arguments must be either a :ref:`floating
7961 point <t_floating>` type or a :ref:`vector <t_vector>` of floating point
7962 type. They must have identical types.
7967 The '``fcmp``' instruction compares ``op1`` and ``op2`` according to the
7968 condition code given as ``cond``. If the operands are vectors, then the
7969 vectors are compared element by element. Each comparison performed
7970 always yields an :ref:`i1 <t_integer>` result, as follows:
7972 #. ``false``: always yields ``false``, regardless of operands.
7973 #. ``oeq``: yields ``true`` if both operands are not a QNAN and ``op1``
7974 is equal to ``op2``.
7975 #. ``ogt``: yields ``true`` if both operands are not a QNAN and ``op1``
7976 is greater than ``op2``.
7977 #. ``oge``: yields ``true`` if both operands are not a QNAN and ``op1``
7978 is greater than or equal to ``op2``.
7979 #. ``olt``: yields ``true`` if both operands are not a QNAN and ``op1``
7980 is less than ``op2``.
7981 #. ``ole``: yields ``true`` if both operands are not a QNAN and ``op1``
7982 is less than or equal to ``op2``.
7983 #. ``one``: yields ``true`` if both operands are not a QNAN and ``op1``
7984 is not equal to ``op2``.
7985 #. ``ord``: yields ``true`` if both operands are not a QNAN.
7986 #. ``ueq``: yields ``true`` if either operand is a QNAN or ``op1`` is
7988 #. ``ugt``: yields ``true`` if either operand is a QNAN or ``op1`` is
7989 greater than ``op2``.
7990 #. ``uge``: yields ``true`` if either operand is a QNAN or ``op1`` is
7991 greater than or equal to ``op2``.
7992 #. ``ult``: yields ``true`` if either operand is a QNAN or ``op1`` is
7994 #. ``ule``: yields ``true`` if either operand is a QNAN or ``op1`` is
7995 less than or equal to ``op2``.
7996 #. ``une``: yields ``true`` if either operand is a QNAN or ``op1`` is
7997 not equal to ``op2``.
7998 #. ``uno``: yields ``true`` if either operand is a QNAN.
7999 #. ``true``: always yields ``true``, regardless of operands.
8001 The ``fcmp`` instruction can also optionally take any number of
8002 :ref:`fast-math flags <fastmath>`, which are optimization hints to enable
8003 otherwise unsafe floating point optimizations.
8005 Any set of fast-math flags are legal on an ``fcmp`` instruction, but the
8006 only flags that have any effect on its semantics are those that allow
8007 assumptions to be made about the values of input arguments; namely
8008 ``nnan``, ``ninf``, and ``nsz``. See :ref:`fastmath` for more information.
8013 .. code-block:: llvm
8015 <result> = fcmp oeq float 4.0, 5.0 ; yields: result=false
8016 <result> = fcmp one float 4.0, 5.0 ; yields: result=true
8017 <result> = fcmp olt float 4.0, 5.0 ; yields: result=true
8018 <result> = fcmp ueq double 1.0, 2.0 ; yields: result=false
8020 Note that the code generator does not yet support vector types with the
8021 ``fcmp`` instruction.
8025 '``phi``' Instruction
8026 ^^^^^^^^^^^^^^^^^^^^^
8033 <result> = phi <ty> [ <val0>, <label0>], ...
8038 The '``phi``' instruction is used to implement the φ node in the SSA
8039 graph representing the function.
8044 The type of the incoming values is specified with the first type field.
8045 After this, the '``phi``' instruction takes a list of pairs as
8046 arguments, with one pair for each predecessor basic block of the current
8047 block. Only values of :ref:`first class <t_firstclass>` type may be used as
8048 the value arguments to the PHI node. Only labels may be used as the
8051 There must be no non-phi instructions between the start of a basic block
8052 and the PHI instructions: i.e. PHI instructions must be first in a basic
8055 For the purposes of the SSA form, the use of each incoming value is
8056 deemed to occur on the edge from the corresponding predecessor block to
8057 the current block (but after any definition of an '``invoke``'
8058 instruction's return value on the same edge).
8063 At runtime, the '``phi``' instruction logically takes on the value
8064 specified by the pair corresponding to the predecessor basic block that
8065 executed just prior to the current block.
8070 .. code-block:: llvm
8072 Loop: ; Infinite loop that counts from 0 on up...
8073 %indvar = phi i32 [ 0, %LoopHeader ], [ %nextindvar, %Loop ]
8074 %nextindvar = add i32 %indvar, 1
8079 '``select``' Instruction
8080 ^^^^^^^^^^^^^^^^^^^^^^^^
8087 <result> = select selty <cond>, <ty> <val1>, <ty> <val2> ; yields ty
8089 selty is either i1 or {<N x i1>}
8094 The '``select``' instruction is used to choose one value based on a
8095 condition, without IR-level branching.
8100 The '``select``' instruction requires an 'i1' value or a vector of 'i1'
8101 values indicating the condition, and two values of the same :ref:`first
8102 class <t_firstclass>` type.
8107 If the condition is an i1 and it evaluates to 1, the instruction returns
8108 the first value argument; otherwise, it returns the second value
8111 If the condition is a vector of i1, then the value arguments must be
8112 vectors of the same size, and the selection is done element by element.
8114 If the condition is an i1 and the value arguments are vectors of the
8115 same size, then an entire vector is selected.
8120 .. code-block:: llvm
8122 %X = select i1 true, i8 17, i8 42 ; yields i8:17
8126 '``call``' Instruction
8127 ^^^^^^^^^^^^^^^^^^^^^^
8134 <result> = [tail | musttail] call [cconv] [ret attrs] <ty> [<fnty>*] <fnptrval>(<function args>) [fn attrs]
8139 The '``call``' instruction represents a simple function call.
8144 This instruction requires several arguments:
8146 #. The optional ``tail`` and ``musttail`` markers indicate that the optimizers
8147 should perform tail call optimization. The ``tail`` marker is a hint that
8148 `can be ignored <CodeGenerator.html#sibcallopt>`_. The ``musttail`` marker
8149 means that the call must be tail call optimized in order for the program to
8150 be correct. The ``musttail`` marker provides these guarantees:
8152 #. The call will not cause unbounded stack growth if it is part of a
8153 recursive cycle in the call graph.
8154 #. Arguments with the :ref:`inalloca <attr_inalloca>` attribute are
8157 Both markers imply that the callee does not access allocas or varargs from
8158 the caller. Calls marked ``musttail`` must obey the following additional
8161 - The call must immediately precede a :ref:`ret <i_ret>` instruction,
8162 or a pointer bitcast followed by a ret instruction.
8163 - The ret instruction must return the (possibly bitcasted) value
8164 produced by the call or void.
8165 - The caller and callee prototypes must match. Pointer types of
8166 parameters or return types may differ in pointee type, but not
8168 - The calling conventions of the caller and callee must match.
8169 - All ABI-impacting function attributes, such as sret, byval, inreg,
8170 returned, and inalloca, must match.
8171 - The callee must be varargs iff the caller is varargs. Bitcasting a
8172 non-varargs function to the appropriate varargs type is legal so
8173 long as the non-varargs prefixes obey the other rules.
8175 Tail call optimization for calls marked ``tail`` is guaranteed to occur if
8176 the following conditions are met:
8178 - Caller and callee both have the calling convention ``fastcc``.
8179 - The call is in tail position (ret immediately follows call and ret
8180 uses value of call or is void).
8181 - Option ``-tailcallopt`` is enabled, or
8182 ``llvm::GuaranteedTailCallOpt`` is ``true``.
8183 - `Platform-specific constraints are
8184 met. <CodeGenerator.html#tailcallopt>`_
8186 #. The optional "cconv" marker indicates which :ref:`calling
8187 convention <callingconv>` the call should use. If none is
8188 specified, the call defaults to using C calling conventions. The
8189 calling convention of the call must match the calling convention of
8190 the target function, or else the behavior is undefined.
8191 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
8192 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
8194 #. '``ty``': the type of the call instruction itself which is also the
8195 type of the return value. Functions that return no value are marked
8197 #. '``fnty``': shall be the signature of the pointer to function value
8198 being invoked. The argument types must match the types implied by
8199 this signature. This type can be omitted if the function is not
8200 varargs and if the function type does not return a pointer to a
8202 #. '``fnptrval``': An LLVM value containing a pointer to a function to
8203 be invoked. In most cases, this is a direct function invocation, but
8204 indirect ``call``'s are just as possible, calling an arbitrary pointer
8206 #. '``function args``': argument list whose types match the function
8207 signature argument types and parameter attributes. All arguments must
8208 be of :ref:`first class <t_firstclass>` type. If the function signature
8209 indicates the function accepts a variable number of arguments, the
8210 extra arguments can be specified.
8211 #. The optional :ref:`function attributes <fnattrs>` list. Only
8212 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
8213 attributes are valid here.
8218 The '``call``' instruction is used to cause control flow to transfer to
8219 a specified function, with its incoming arguments bound to the specified
8220 values. Upon a '``ret``' instruction in the called function, control
8221 flow continues with the instruction after the function call, and the
8222 return value of the function is bound to the result argument.
8227 .. code-block:: llvm
8229 %retval = call i32 @test(i32 %argc)
8230 call i32 (i8*, ...)* @printf(i8* %msg, i32 12, i8 42) ; yields i32
8231 %X = tail call i32 @foo() ; yields i32
8232 %Y = tail call fastcc i32 @foo() ; yields i32
8233 call void %foo(i8 97 signext)
8235 %struct.A = type { i32, i8 }
8236 %r = call %struct.A @foo() ; yields { i32, i8 }
8237 %gr = extractvalue %struct.A %r, 0 ; yields i32
8238 %gr1 = extractvalue %struct.A %r, 1 ; yields i8
8239 %Z = call void @foo() noreturn ; indicates that %foo never returns normally
8240 %ZZ = call zeroext i32 @bar() ; Return value is %zero extended
8242 llvm treats calls to some functions with names and arguments that match
8243 the standard C99 library as being the C99 library functions, and may
8244 perform optimizations or generate code for them under that assumption.
8245 This is something we'd like to change in the future to provide better
8246 support for freestanding environments and non-C-based languages.
8250 '``va_arg``' Instruction
8251 ^^^^^^^^^^^^^^^^^^^^^^^^
8258 <resultval> = va_arg <va_list*> <arglist>, <argty>
8263 The '``va_arg``' instruction is used to access arguments passed through
8264 the "variable argument" area of a function call. It is used to implement
8265 the ``va_arg`` macro in C.
8270 This instruction takes a ``va_list*`` value and the type of the
8271 argument. It returns a value of the specified argument type and
8272 increments the ``va_list`` to point to the next argument. The actual
8273 type of ``va_list`` is target specific.
8278 The '``va_arg``' instruction loads an argument of the specified type
8279 from the specified ``va_list`` and causes the ``va_list`` to point to
8280 the next argument. For more information, see the variable argument
8281 handling :ref:`Intrinsic Functions <int_varargs>`.
8283 It is legal for this instruction to be called in a function which does
8284 not take a variable number of arguments, for example, the ``vfprintf``
8287 ``va_arg`` is an LLVM instruction instead of an :ref:`intrinsic
8288 function <intrinsics>` because it takes a type as an argument.
8293 See the :ref:`variable argument processing <int_varargs>` section.
8295 Note that the code generator does not yet fully support va\_arg on many
8296 targets. Also, it does not currently support va\_arg with aggregate
8297 types on any target.
8301 '``landingpad``' Instruction
8302 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8309 <resultval> = landingpad <resultty> <clause>+
8310 <resultval> = landingpad <resultty> cleanup <clause>*
8312 <clause> := catch <type> <value>
8313 <clause> := filter <array constant type> <array constant>
8318 The '``landingpad``' instruction is used by `LLVM's exception handling
8319 system <ExceptionHandling.html#overview>`_ to specify that a basic block
8320 is a landing pad --- one where the exception lands, and corresponds to the
8321 code found in the ``catch`` portion of a ``try``/``catch`` sequence. It
8322 defines values supplied by the :ref:`personality function <personalityfn>` upon
8323 re-entry to the function. The ``resultval`` has the type ``resultty``.
8329 ``cleanup`` flag indicates that the landing pad block is a cleanup.
8331 A ``clause`` begins with the clause type --- ``catch`` or ``filter`` --- and
8332 contains the global variable representing the "type" that may be caught
8333 or filtered respectively. Unlike the ``catch`` clause, the ``filter``
8334 clause takes an array constant as its argument. Use
8335 "``[0 x i8**] undef``" for a filter which cannot throw. The
8336 '``landingpad``' instruction must contain *at least* one ``clause`` or
8337 the ``cleanup`` flag.
8342 The '``landingpad``' instruction defines the values which are set by the
8343 :ref:`personality function <personalityfn>` upon re-entry to the function, and
8344 therefore the "result type" of the ``landingpad`` instruction. As with
8345 calling conventions, how the personality function results are
8346 represented in LLVM IR is target specific.
8348 The clauses are applied in order from top to bottom. If two
8349 ``landingpad`` instructions are merged together through inlining, the
8350 clauses from the calling function are appended to the list of clauses.
8351 When the call stack is being unwound due to an exception being thrown,
8352 the exception is compared against each ``clause`` in turn. If it doesn't
8353 match any of the clauses, and the ``cleanup`` flag is not set, then
8354 unwinding continues further up the call stack.
8356 The ``landingpad`` instruction has several restrictions:
8358 - A landing pad block is a basic block which is the unwind destination
8359 of an '``invoke``' instruction.
8360 - A landing pad block must have a '``landingpad``' instruction as its
8361 first non-PHI instruction.
8362 - There can be only one '``landingpad``' instruction within the landing
8364 - A basic block that is not a landing pad block may not include a
8365 '``landingpad``' instruction.
8370 .. code-block:: llvm
8372 ;; A landing pad which can catch an integer.
8373 %res = landingpad { i8*, i32 }
8375 ;; A landing pad that is a cleanup.
8376 %res = landingpad { i8*, i32 }
8378 ;; A landing pad which can catch an integer and can only throw a double.
8379 %res = landingpad { i8*, i32 }
8381 filter [1 x i8**] [@_ZTId]
8385 '``cleanuppad``' Instruction
8386 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8393 <resultval> = cleanuppad <resultty> [<args>*]
8398 The '``cleanuppad``' instruction is used by `LLVM's exception handling
8399 system <ExceptionHandling.html#overview>`_ to specify that a basic block
8400 is a cleanup block --- one where a personality routine attempts to
8401 transfer control to run cleanup actions.
8402 The ``args`` correspond to whatever additional
8403 information the :ref:`personality function <personalityfn>` requires to
8404 execute the cleanup.
8405 The ``resultval`` has the type ``resultty``.
8410 The instruction takes a list of arbitrary values which are interpreted
8411 by the :ref:`personality function <personalityfn>`.
8416 The '``cleanuppad``' instruction defines the values which are set by the
8417 :ref:`personality function <personalityfn>` upon re-entry to the function, and
8418 therefore the "result type" of the ``cleanuppad`` instruction. As with
8419 calling conventions, how the personality function results are
8420 represented in LLVM IR is target specific.
8422 When the call stack is being unwound due to an exception being thrown,
8423 the :ref:`personality function <personalityfn>` transfers control to the
8424 ``cleanuppad`` with the aid of the personality-specific arguments.
8426 The ``cleanuppad`` instruction has several restrictions:
8428 - A cleanup block is a basic block which is the unwind destination of
8429 an exceptional instruction.
8430 - A cleanup block must have a '``cleanuppad``' instruction as its
8431 first non-PHI instruction.
8432 - There can be only one '``cleanuppad``' instruction within the
8434 - A basic block that is not a cleanup block may not include a
8435 '``cleanuppad``' instruction.
8436 - It is undefined behavior for control to transfer from a ``cleanuppad`` to a
8437 ``catchret`` without first executing a ``cleanupret`` and a subsequent
8439 - It is undefined behavior for control to transfer from a ``cleanuppad`` to a
8440 ``ret`` without first executing a ``cleanupret``.
8445 .. code-block:: llvm
8447 %res = cleanuppad { i8*, i32 } [label %nextaction]
8454 LLVM supports the notion of an "intrinsic function". These functions
8455 have well known names and semantics and are required to follow certain
8456 restrictions. Overall, these intrinsics represent an extension mechanism
8457 for the LLVM language that does not require changing all of the
8458 transformations in LLVM when adding to the language (or the bitcode
8459 reader/writer, the parser, etc...).
8461 Intrinsic function names must all start with an "``llvm.``" prefix. This
8462 prefix is reserved in LLVM for intrinsic names; thus, function names may
8463 not begin with this prefix. Intrinsic functions must always be external
8464 functions: you cannot define the body of intrinsic functions. Intrinsic
8465 functions may only be used in call or invoke instructions: it is illegal
8466 to take the address of an intrinsic function. Additionally, because
8467 intrinsic functions are part of the LLVM language, it is required if any
8468 are added that they be documented here.
8470 Some intrinsic functions can be overloaded, i.e., the intrinsic
8471 represents a family of functions that perform the same operation but on
8472 different data types. Because LLVM can represent over 8 million
8473 different integer types, overloading is used commonly to allow an
8474 intrinsic function to operate on any integer type. One or more of the
8475 argument types or the result type can be overloaded to accept any
8476 integer type. Argument types may also be defined as exactly matching a
8477 previous argument's type or the result type. This allows an intrinsic
8478 function which accepts multiple arguments, but needs all of them to be
8479 of the same type, to only be overloaded with respect to a single
8480 argument or the result.
8482 Overloaded intrinsics will have the names of its overloaded argument
8483 types encoded into its function name, each preceded by a period. Only
8484 those types which are overloaded result in a name suffix. Arguments
8485 whose type is matched against another type do not. For example, the
8486 ``llvm.ctpop`` function can take an integer of any width and returns an
8487 integer of exactly the same integer width. This leads to a family of
8488 functions such as ``i8 @llvm.ctpop.i8(i8 %val)`` and
8489 ``i29 @llvm.ctpop.i29(i29 %val)``. Only one type, the return type, is
8490 overloaded, and only one type suffix is required. Because the argument's
8491 type is matched against the return type, it does not require its own
8494 To learn how to add an intrinsic function, please see the `Extending
8495 LLVM Guide <ExtendingLLVM.html>`_.
8499 Variable Argument Handling Intrinsics
8500 -------------------------------------
8502 Variable argument support is defined in LLVM with the
8503 :ref:`va_arg <i_va_arg>` instruction and these three intrinsic
8504 functions. These functions are related to the similarly named macros
8505 defined in the ``<stdarg.h>`` header file.
8507 All of these functions operate on arguments that use a target-specific
8508 value type "``va_list``". The LLVM assembly language reference manual
8509 does not define what this type is, so all transformations should be
8510 prepared to handle these functions regardless of the type used.
8512 This example shows how the :ref:`va_arg <i_va_arg>` instruction and the
8513 variable argument handling intrinsic functions are used.
8515 .. code-block:: llvm
8517 ; This struct is different for every platform. For most platforms,
8518 ; it is merely an i8*.
8519 %struct.va_list = type { i8* }
8521 ; For Unix x86_64 platforms, va_list is the following struct:
8522 ; %struct.va_list = type { i32, i32, i8*, i8* }
8524 define i32 @test(i32 %X, ...) {
8525 ; Initialize variable argument processing
8526 %ap = alloca %struct.va_list
8527 %ap2 = bitcast %struct.va_list* %ap to i8*
8528 call void @llvm.va_start(i8* %ap2)
8530 ; Read a single integer argument
8531 %tmp = va_arg i8* %ap2, i32
8533 ; Demonstrate usage of llvm.va_copy and llvm.va_end
8535 %aq2 = bitcast i8** %aq to i8*
8536 call void @llvm.va_copy(i8* %aq2, i8* %ap2)
8537 call void @llvm.va_end(i8* %aq2)
8539 ; Stop processing of arguments.
8540 call void @llvm.va_end(i8* %ap2)
8544 declare void @llvm.va_start(i8*)
8545 declare void @llvm.va_copy(i8*, i8*)
8546 declare void @llvm.va_end(i8*)
8550 '``llvm.va_start``' Intrinsic
8551 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8558 declare void @llvm.va_start(i8* <arglist>)
8563 The '``llvm.va_start``' intrinsic initializes ``*<arglist>`` for
8564 subsequent use by ``va_arg``.
8569 The argument is a pointer to a ``va_list`` element to initialize.
8574 The '``llvm.va_start``' intrinsic works just like the ``va_start`` macro
8575 available in C. In a target-dependent way, it initializes the
8576 ``va_list`` element to which the argument points, so that the next call
8577 to ``va_arg`` will produce the first variable argument passed to the
8578 function. Unlike the C ``va_start`` macro, this intrinsic does not need
8579 to know the last argument of the function as the compiler can figure
8582 '``llvm.va_end``' Intrinsic
8583 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8590 declare void @llvm.va_end(i8* <arglist>)
8595 The '``llvm.va_end``' intrinsic destroys ``*<arglist>``, which has been
8596 initialized previously with ``llvm.va_start`` or ``llvm.va_copy``.
8601 The argument is a pointer to a ``va_list`` to destroy.
8606 The '``llvm.va_end``' intrinsic works just like the ``va_end`` macro
8607 available in C. In a target-dependent way, it destroys the ``va_list``
8608 element to which the argument points. Calls to
8609 :ref:`llvm.va_start <int_va_start>` and
8610 :ref:`llvm.va_copy <int_va_copy>` must be matched exactly with calls to
8615 '``llvm.va_copy``' Intrinsic
8616 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8623 declare void @llvm.va_copy(i8* <destarglist>, i8* <srcarglist>)
8628 The '``llvm.va_copy``' intrinsic copies the current argument position
8629 from the source argument list to the destination argument list.
8634 The first argument is a pointer to a ``va_list`` element to initialize.
8635 The second argument is a pointer to a ``va_list`` element to copy from.
8640 The '``llvm.va_copy``' intrinsic works just like the ``va_copy`` macro
8641 available in C. In a target-dependent way, it copies the source
8642 ``va_list`` element into the destination ``va_list`` element. This
8643 intrinsic is necessary because the `` llvm.va_start`` intrinsic may be
8644 arbitrarily complex and require, for example, memory allocation.
8646 Accurate Garbage Collection Intrinsics
8647 --------------------------------------
8649 LLVM's support for `Accurate Garbage Collection <GarbageCollection.html>`_
8650 (GC) requires the frontend to generate code containing appropriate intrinsic
8651 calls and select an appropriate GC strategy which knows how to lower these
8652 intrinsics in a manner which is appropriate for the target collector.
8654 These intrinsics allow identification of :ref:`GC roots on the
8655 stack <int_gcroot>`, as well as garbage collector implementations that
8656 require :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers.
8657 Frontends for type-safe garbage collected languages should generate
8658 these intrinsics to make use of the LLVM garbage collectors. For more
8659 details, see `Garbage Collection with LLVM <GarbageCollection.html>`_.
8661 Experimental Statepoint Intrinsics
8662 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8664 LLVM provides an second experimental set of intrinsics for describing garbage
8665 collection safepoints in compiled code. These intrinsics are an alternative
8666 to the ``llvm.gcroot`` intrinsics, but are compatible with the ones for
8667 :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers. The
8668 differences in approach are covered in the `Garbage Collection with LLVM
8669 <GarbageCollection.html>`_ documentation. The intrinsics themselves are
8670 described in :doc:`Statepoints`.
8674 '``llvm.gcroot``' Intrinsic
8675 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8682 declare void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
8687 The '``llvm.gcroot``' intrinsic declares the existence of a GC root to
8688 the code generator, and allows some metadata to be associated with it.
8693 The first argument specifies the address of a stack object that contains
8694 the root pointer. The second pointer (which must be either a constant or
8695 a global value address) contains the meta-data to be associated with the
8701 At runtime, a call to this intrinsic stores a null pointer into the
8702 "ptrloc" location. At compile-time, the code generator generates
8703 information to allow the runtime to find the pointer at GC safe points.
8704 The '``llvm.gcroot``' intrinsic may only be used in a function which
8705 :ref:`specifies a GC algorithm <gc>`.
8709 '``llvm.gcread``' Intrinsic
8710 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8717 declare i8* @llvm.gcread(i8* %ObjPtr, i8** %Ptr)
8722 The '``llvm.gcread``' intrinsic identifies reads of references from heap
8723 locations, allowing garbage collector implementations that require read
8729 The second argument is the address to read from, which should be an
8730 address allocated from the garbage collector. The first object is a
8731 pointer to the start of the referenced object, if needed by the language
8732 runtime (otherwise null).
8737 The '``llvm.gcread``' intrinsic has the same semantics as a load
8738 instruction, but may be replaced with substantially more complex code by
8739 the garbage collector runtime, as needed. The '``llvm.gcread``'
8740 intrinsic may only be used in a function which :ref:`specifies a GC
8745 '``llvm.gcwrite``' Intrinsic
8746 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8753 declare void @llvm.gcwrite(i8* %P1, i8* %Obj, i8** %P2)
8758 The '``llvm.gcwrite``' intrinsic identifies writes of references to heap
8759 locations, allowing garbage collector implementations that require write
8760 barriers (such as generational or reference counting collectors).
8765 The first argument is the reference to store, the second is the start of
8766 the object to store it to, and the third is the address of the field of
8767 Obj to store to. If the runtime does not require a pointer to the
8768 object, Obj may be null.
8773 The '``llvm.gcwrite``' intrinsic has the same semantics as a store
8774 instruction, but may be replaced with substantially more complex code by
8775 the garbage collector runtime, as needed. The '``llvm.gcwrite``'
8776 intrinsic may only be used in a function which :ref:`specifies a GC
8779 Code Generator Intrinsics
8780 -------------------------
8782 These intrinsics are provided by LLVM to expose special features that
8783 may only be implemented with code generator support.
8785 '``llvm.returnaddress``' Intrinsic
8786 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8793 declare i8 *@llvm.returnaddress(i32 <level>)
8798 The '``llvm.returnaddress``' intrinsic attempts to compute a
8799 target-specific value indicating the return address of the current
8800 function or one of its callers.
8805 The argument to this intrinsic indicates which function to return the
8806 address for. Zero indicates the calling function, one indicates its
8807 caller, etc. The argument is **required** to be a constant integer
8813 The '``llvm.returnaddress``' intrinsic either returns a pointer
8814 indicating the return address of the specified call frame, or zero if it
8815 cannot be identified. The value returned by this intrinsic is likely to
8816 be incorrect or 0 for arguments other than zero, so it should only be
8817 used for debugging purposes.
8819 Note that calling this intrinsic does not prevent function inlining or
8820 other aggressive transformations, so the value returned may not be that
8821 of the obvious source-language caller.
8823 '``llvm.frameaddress``' Intrinsic
8824 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8831 declare i8* @llvm.frameaddress(i32 <level>)
8836 The '``llvm.frameaddress``' intrinsic attempts to return the
8837 target-specific frame pointer value for the specified stack frame.
8842 The argument to this intrinsic indicates which function to return the
8843 frame pointer for. Zero indicates the calling function, one indicates
8844 its caller, etc. The argument is **required** to be a constant integer
8850 The '``llvm.frameaddress``' intrinsic either returns a pointer
8851 indicating the frame address of the specified call frame, or zero if it
8852 cannot be identified. The value returned by this intrinsic is likely to
8853 be incorrect or 0 for arguments other than zero, so it should only be
8854 used for debugging purposes.
8856 Note that calling this intrinsic does not prevent function inlining or
8857 other aggressive transformations, so the value returned may not be that
8858 of the obvious source-language caller.
8860 '``llvm.localescape``' and '``llvm.localrecover``' Intrinsics
8861 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8868 declare void @llvm.localescape(...)
8869 declare i8* @llvm.localrecover(i8* %func, i8* %fp, i32 %idx)
8874 The '``llvm.localescape``' intrinsic escapes offsets of a collection of static
8875 allocas, and the '``llvm.localrecover``' intrinsic applies those offsets to a
8876 live frame pointer to recover the address of the allocation. The offset is
8877 computed during frame layout of the caller of ``llvm.localescape``.
8882 All arguments to '``llvm.localescape``' must be pointers to static allocas or
8883 casts of static allocas. Each function can only call '``llvm.localescape``'
8884 once, and it can only do so from the entry block.
8886 The ``func`` argument to '``llvm.localrecover``' must be a constant
8887 bitcasted pointer to a function defined in the current module. The code
8888 generator cannot determine the frame allocation offset of functions defined in
8891 The ``fp`` argument to '``llvm.localrecover``' must be a frame pointer of a
8892 call frame that is currently live. The return value of '``llvm.localaddress``'
8893 is one way to produce such a value, but various runtimes also expose a suitable
8894 pointer in platform-specific ways.
8896 The ``idx`` argument to '``llvm.localrecover``' indicates which alloca passed to
8897 '``llvm.localescape``' to recover. It is zero-indexed.
8902 These intrinsics allow a group of functions to share access to a set of local
8903 stack allocations of a one parent function. The parent function may call the
8904 '``llvm.localescape``' intrinsic once from the function entry block, and the
8905 child functions can use '``llvm.localrecover``' to access the escaped allocas.
8906 The '``llvm.localescape``' intrinsic blocks inlining, as inlining changes where
8907 the escaped allocas are allocated, which would break attempts to use
8908 '``llvm.localrecover``'.
8910 .. _int_read_register:
8911 .. _int_write_register:
8913 '``llvm.read_register``' and '``llvm.write_register``' Intrinsics
8914 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8921 declare i32 @llvm.read_register.i32(metadata)
8922 declare i64 @llvm.read_register.i64(metadata)
8923 declare void @llvm.write_register.i32(metadata, i32 @value)
8924 declare void @llvm.write_register.i64(metadata, i64 @value)
8930 The '``llvm.read_register``' and '``llvm.write_register``' intrinsics
8931 provides access to the named register. The register must be valid on
8932 the architecture being compiled to. The type needs to be compatible
8933 with the register being read.
8938 The '``llvm.read_register``' intrinsic returns the current value of the
8939 register, where possible. The '``llvm.write_register``' intrinsic sets
8940 the current value of the register, where possible.
8942 This is useful to implement named register global variables that need
8943 to always be mapped to a specific register, as is common practice on
8944 bare-metal programs including OS kernels.
8946 The compiler doesn't check for register availability or use of the used
8947 register in surrounding code, including inline assembly. Because of that,
8948 allocatable registers are not supported.
8950 Warning: So far it only works with the stack pointer on selected
8951 architectures (ARM, AArch64, PowerPC and x86_64). Significant amount of
8952 work is needed to support other registers and even more so, allocatable
8957 '``llvm.stacksave``' Intrinsic
8958 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8965 declare i8* @llvm.stacksave()
8970 The '``llvm.stacksave``' intrinsic is used to remember the current state
8971 of the function stack, for use with
8972 :ref:`llvm.stackrestore <int_stackrestore>`. This is useful for
8973 implementing language features like scoped automatic variable sized
8979 This intrinsic returns a opaque pointer value that can be passed to
8980 :ref:`llvm.stackrestore <int_stackrestore>`. When an
8981 ``llvm.stackrestore`` intrinsic is executed with a value saved from
8982 ``llvm.stacksave``, it effectively restores the state of the stack to
8983 the state it was in when the ``llvm.stacksave`` intrinsic executed. In
8984 practice, this pops any :ref:`alloca <i_alloca>` blocks from the stack that
8985 were allocated after the ``llvm.stacksave`` was executed.
8987 .. _int_stackrestore:
8989 '``llvm.stackrestore``' Intrinsic
8990 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8997 declare void @llvm.stackrestore(i8* %ptr)
9002 The '``llvm.stackrestore``' intrinsic is used to restore the state of
9003 the function stack to the state it was in when the corresponding
9004 :ref:`llvm.stacksave <int_stacksave>` intrinsic executed. This is
9005 useful for implementing language features like scoped automatic variable
9006 sized arrays in C99.
9011 See the description for :ref:`llvm.stacksave <int_stacksave>`.
9013 '``llvm.prefetch``' Intrinsic
9014 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9021 declare void @llvm.prefetch(i8* <address>, i32 <rw>, i32 <locality>, i32 <cache type>)
9026 The '``llvm.prefetch``' intrinsic is a hint to the code generator to
9027 insert a prefetch instruction if supported; otherwise, it is a noop.
9028 Prefetches have no effect on the behavior of the program but can change
9029 its performance characteristics.
9034 ``address`` is the address to be prefetched, ``rw`` is the specifier
9035 determining if the fetch should be for a read (0) or write (1), and
9036 ``locality`` is a temporal locality specifier ranging from (0) - no
9037 locality, to (3) - extremely local keep in cache. The ``cache type``
9038 specifies whether the prefetch is performed on the data (1) or
9039 instruction (0) cache. The ``rw``, ``locality`` and ``cache type``
9040 arguments must be constant integers.
9045 This intrinsic does not modify the behavior of the program. In
9046 particular, prefetches cannot trap and do not produce a value. On
9047 targets that support this intrinsic, the prefetch can provide hints to
9048 the processor cache for better performance.
9050 '``llvm.pcmarker``' Intrinsic
9051 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9058 declare void @llvm.pcmarker(i32 <id>)
9063 The '``llvm.pcmarker``' intrinsic is a method to export a Program
9064 Counter (PC) in a region of code to simulators and other tools. The
9065 method is target specific, but it is expected that the marker will use
9066 exported symbols to transmit the PC of the marker. The marker makes no
9067 guarantees that it will remain with any specific instruction after
9068 optimizations. It is possible that the presence of a marker will inhibit
9069 optimizations. The intended use is to be inserted after optimizations to
9070 allow correlations of simulation runs.
9075 ``id`` is a numerical id identifying the marker.
9080 This intrinsic does not modify the behavior of the program. Backends
9081 that do not support this intrinsic may ignore it.
9083 '``llvm.readcyclecounter``' Intrinsic
9084 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9091 declare i64 @llvm.readcyclecounter()
9096 The '``llvm.readcyclecounter``' intrinsic provides access to the cycle
9097 counter register (or similar low latency, high accuracy clocks) on those
9098 targets that support it. On X86, it should map to RDTSC. On Alpha, it
9099 should map to RPCC. As the backing counters overflow quickly (on the
9100 order of 9 seconds on alpha), this should only be used for small
9106 When directly supported, reading the cycle counter should not modify any
9107 memory. Implementations are allowed to either return a application
9108 specific value or a system wide value. On backends without support, this
9109 is lowered to a constant 0.
9111 Note that runtime support may be conditional on the privilege-level code is
9112 running at and the host platform.
9114 '``llvm.clear_cache``' Intrinsic
9115 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9122 declare void @llvm.clear_cache(i8*, i8*)
9127 The '``llvm.clear_cache``' intrinsic ensures visibility of modifications
9128 in the specified range to the execution unit of the processor. On
9129 targets with non-unified instruction and data cache, the implementation
9130 flushes the instruction cache.
9135 On platforms with coherent instruction and data caches (e.g. x86), this
9136 intrinsic is a nop. On platforms with non-coherent instruction and data
9137 cache (e.g. ARM, MIPS), the intrinsic is lowered either to appropriate
9138 instructions or a system call, if cache flushing requires special
9141 The default behavior is to emit a call to ``__clear_cache`` from the run
9144 This instrinsic does *not* empty the instruction pipeline. Modifications
9145 of the current function are outside the scope of the intrinsic.
9147 '``llvm.instrprof_increment``' Intrinsic
9148 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9155 declare void @llvm.instrprof_increment(i8* <name>, i64 <hash>,
9156 i32 <num-counters>, i32 <index>)
9161 The '``llvm.instrprof_increment``' intrinsic can be emitted by a
9162 frontend for use with instrumentation based profiling. These will be
9163 lowered by the ``-instrprof`` pass to generate execution counts of a
9169 The first argument is a pointer to a global variable containing the
9170 name of the entity being instrumented. This should generally be the
9171 (mangled) function name for a set of counters.
9173 The second argument is a hash value that can be used by the consumer
9174 of the profile data to detect changes to the instrumented source, and
9175 the third is the number of counters associated with ``name``. It is an
9176 error if ``hash`` or ``num-counters`` differ between two instances of
9177 ``instrprof_increment`` that refer to the same name.
9179 The last argument refers to which of the counters for ``name`` should
9180 be incremented. It should be a value between 0 and ``num-counters``.
9185 This intrinsic represents an increment of a profiling counter. It will
9186 cause the ``-instrprof`` pass to generate the appropriate data
9187 structures and the code to increment the appropriate value, in a
9188 format that can be written out by a compiler runtime and consumed via
9189 the ``llvm-profdata`` tool.
9191 Standard C Library Intrinsics
9192 -----------------------------
9194 LLVM provides intrinsics for a few important standard C library
9195 functions. These intrinsics allow source-language front-ends to pass
9196 information about the alignment of the pointer arguments to the code
9197 generator, providing opportunity for more efficient code generation.
9201 '``llvm.memcpy``' Intrinsic
9202 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9207 This is an overloaded intrinsic. You can use ``llvm.memcpy`` on any
9208 integer bit width and for different address spaces. Not all targets
9209 support all bit widths however.
9213 declare void @llvm.memcpy.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
9214 i32 <len>, i32 <align>, i1 <isvolatile>)
9215 declare void @llvm.memcpy.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
9216 i64 <len>, i32 <align>, i1 <isvolatile>)
9221 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
9222 source location to the destination location.
9224 Note that, unlike the standard libc function, the ``llvm.memcpy.*``
9225 intrinsics do not return a value, takes extra alignment/isvolatile
9226 arguments and the pointers can be in specified address spaces.
9231 The first argument is a pointer to the destination, the second is a
9232 pointer to the source. The third argument is an integer argument
9233 specifying the number of bytes to copy, the fourth argument is the
9234 alignment of the source and destination locations, and the fifth is a
9235 boolean indicating a volatile access.
9237 If the call to this intrinsic has an alignment value that is not 0 or 1,
9238 then the caller guarantees that both the source and destination pointers
9239 are aligned to that boundary.
9241 If the ``isvolatile`` parameter is ``true``, the ``llvm.memcpy`` call is
9242 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
9243 very cleanly specified and it is unwise to depend on it.
9248 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
9249 source location to the destination location, which are not allowed to
9250 overlap. It copies "len" bytes of memory over. If the argument is known
9251 to be aligned to some boundary, this can be specified as the fourth
9252 argument, otherwise it should be set to 0 or 1 (both meaning no alignment).
9254 '``llvm.memmove``' Intrinsic
9255 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9260 This is an overloaded intrinsic. You can use llvm.memmove on any integer
9261 bit width and for different address space. Not all targets support all
9266 declare void @llvm.memmove.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
9267 i32 <len>, i32 <align>, i1 <isvolatile>)
9268 declare void @llvm.memmove.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
9269 i64 <len>, i32 <align>, i1 <isvolatile>)
9274 The '``llvm.memmove.*``' intrinsics move a block of memory from the
9275 source location to the destination location. It is similar to the
9276 '``llvm.memcpy``' intrinsic but allows the two memory locations to
9279 Note that, unlike the standard libc function, the ``llvm.memmove.*``
9280 intrinsics do not return a value, takes extra alignment/isvolatile
9281 arguments and the pointers can be in specified address spaces.
9286 The first argument is a pointer to the destination, the second is a
9287 pointer to the source. The third argument is an integer argument
9288 specifying the number of bytes to copy, the fourth argument is the
9289 alignment of the source and destination locations, and the fifth is a
9290 boolean indicating a volatile access.
9292 If the call to this intrinsic has an alignment value that is not 0 or 1,
9293 then the caller guarantees that the source and destination pointers are
9294 aligned to that boundary.
9296 If the ``isvolatile`` parameter is ``true``, the ``llvm.memmove`` call
9297 is a :ref:`volatile operation <volatile>`. The detailed access behavior is
9298 not very cleanly specified and it is unwise to depend on it.
9303 The '``llvm.memmove.*``' intrinsics copy a block of memory from the
9304 source location to the destination location, which may overlap. It
9305 copies "len" bytes of memory over. If the argument is known to be
9306 aligned to some boundary, this can be specified as the fourth argument,
9307 otherwise it should be set to 0 or 1 (both meaning no alignment).
9309 '``llvm.memset.*``' Intrinsics
9310 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9315 This is an overloaded intrinsic. You can use llvm.memset on any integer
9316 bit width and for different address spaces. However, not all targets
9317 support all bit widths.
9321 declare void @llvm.memset.p0i8.i32(i8* <dest>, i8 <val>,
9322 i32 <len>, i32 <align>, i1 <isvolatile>)
9323 declare void @llvm.memset.p0i8.i64(i8* <dest>, i8 <val>,
9324 i64 <len>, i32 <align>, i1 <isvolatile>)
9329 The '``llvm.memset.*``' intrinsics fill a block of memory with a
9330 particular byte value.
9332 Note that, unlike the standard libc function, the ``llvm.memset``
9333 intrinsic does not return a value and takes extra alignment/volatile
9334 arguments. Also, the destination can be in an arbitrary address space.
9339 The first argument is a pointer to the destination to fill, the second
9340 is the byte value with which to fill it, the third argument is an
9341 integer argument specifying the number of bytes to fill, and the fourth
9342 argument is the known alignment of the destination location.
9344 If the call to this intrinsic has an alignment value that is not 0 or 1,
9345 then the caller guarantees that the destination pointer is aligned to
9348 If the ``isvolatile`` parameter is ``true``, the ``llvm.memset`` call is
9349 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
9350 very cleanly specified and it is unwise to depend on it.
9355 The '``llvm.memset.*``' intrinsics fill "len" bytes of memory starting
9356 at the destination location. If the argument is known to be aligned to
9357 some boundary, this can be specified as the fourth argument, otherwise
9358 it should be set to 0 or 1 (both meaning no alignment).
9360 '``llvm.sqrt.*``' Intrinsic
9361 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9366 This is an overloaded intrinsic. You can use ``llvm.sqrt`` on any
9367 floating point or vector of floating point type. Not all targets support
9372 declare float @llvm.sqrt.f32(float %Val)
9373 declare double @llvm.sqrt.f64(double %Val)
9374 declare x86_fp80 @llvm.sqrt.f80(x86_fp80 %Val)
9375 declare fp128 @llvm.sqrt.f128(fp128 %Val)
9376 declare ppc_fp128 @llvm.sqrt.ppcf128(ppc_fp128 %Val)
9381 The '``llvm.sqrt``' intrinsics return the sqrt of the specified operand,
9382 returning the same value as the libm '``sqrt``' functions would. Unlike
9383 ``sqrt`` in libm, however, ``llvm.sqrt`` has undefined behavior for
9384 negative numbers other than -0.0 (which allows for better optimization,
9385 because there is no need to worry about errno being set).
9386 ``llvm.sqrt(-0.0)`` is defined to return -0.0 like IEEE sqrt.
9391 The argument and return value are floating point numbers of the same
9397 This function returns the sqrt of the specified operand if it is a
9398 nonnegative floating point number.
9400 '``llvm.powi.*``' Intrinsic
9401 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9406 This is an overloaded intrinsic. You can use ``llvm.powi`` on any
9407 floating point or vector of floating point type. Not all targets support
9412 declare float @llvm.powi.f32(float %Val, i32 %power)
9413 declare double @llvm.powi.f64(double %Val, i32 %power)
9414 declare x86_fp80 @llvm.powi.f80(x86_fp80 %Val, i32 %power)
9415 declare fp128 @llvm.powi.f128(fp128 %Val, i32 %power)
9416 declare ppc_fp128 @llvm.powi.ppcf128(ppc_fp128 %Val, i32 %power)
9421 The '``llvm.powi.*``' intrinsics return the first operand raised to the
9422 specified (positive or negative) power. The order of evaluation of
9423 multiplications is not defined. When a vector of floating point type is
9424 used, the second argument remains a scalar integer value.
9429 The second argument is an integer power, and the first is a value to
9430 raise to that power.
9435 This function returns the first value raised to the second power with an
9436 unspecified sequence of rounding operations.
9438 '``llvm.sin.*``' Intrinsic
9439 ^^^^^^^^^^^^^^^^^^^^^^^^^^
9444 This is an overloaded intrinsic. You can use ``llvm.sin`` on any
9445 floating point or vector of floating point type. Not all targets support
9450 declare float @llvm.sin.f32(float %Val)
9451 declare double @llvm.sin.f64(double %Val)
9452 declare x86_fp80 @llvm.sin.f80(x86_fp80 %Val)
9453 declare fp128 @llvm.sin.f128(fp128 %Val)
9454 declare ppc_fp128 @llvm.sin.ppcf128(ppc_fp128 %Val)
9459 The '``llvm.sin.*``' intrinsics return the sine of the operand.
9464 The argument and return value are floating point numbers of the same
9470 This function returns the sine of the specified operand, returning the
9471 same values as the libm ``sin`` functions would, and handles error
9472 conditions in the same way.
9474 '``llvm.cos.*``' Intrinsic
9475 ^^^^^^^^^^^^^^^^^^^^^^^^^^
9480 This is an overloaded intrinsic. You can use ``llvm.cos`` on any
9481 floating point or vector of floating point type. Not all targets support
9486 declare float @llvm.cos.f32(float %Val)
9487 declare double @llvm.cos.f64(double %Val)
9488 declare x86_fp80 @llvm.cos.f80(x86_fp80 %Val)
9489 declare fp128 @llvm.cos.f128(fp128 %Val)
9490 declare ppc_fp128 @llvm.cos.ppcf128(ppc_fp128 %Val)
9495 The '``llvm.cos.*``' intrinsics return the cosine of the operand.
9500 The argument and return value are floating point numbers of the same
9506 This function returns the cosine of the specified operand, returning the
9507 same values as the libm ``cos`` functions would, and handles error
9508 conditions in the same way.
9510 '``llvm.pow.*``' Intrinsic
9511 ^^^^^^^^^^^^^^^^^^^^^^^^^^
9516 This is an overloaded intrinsic. You can use ``llvm.pow`` on any
9517 floating point or vector of floating point type. Not all targets support
9522 declare float @llvm.pow.f32(float %Val, float %Power)
9523 declare double @llvm.pow.f64(double %Val, double %Power)
9524 declare x86_fp80 @llvm.pow.f80(x86_fp80 %Val, x86_fp80 %Power)
9525 declare fp128 @llvm.pow.f128(fp128 %Val, fp128 %Power)
9526 declare ppc_fp128 @llvm.pow.ppcf128(ppc_fp128 %Val, ppc_fp128 Power)
9531 The '``llvm.pow.*``' intrinsics return the first operand raised to the
9532 specified (positive or negative) power.
9537 The second argument is a floating point power, and the first is a value
9538 to raise to that power.
9543 This function returns the first value raised to the second power,
9544 returning the same values as the libm ``pow`` functions would, and
9545 handles error conditions in the same way.
9547 '``llvm.exp.*``' Intrinsic
9548 ^^^^^^^^^^^^^^^^^^^^^^^^^^
9553 This is an overloaded intrinsic. You can use ``llvm.exp`` on any
9554 floating point or vector of floating point type. Not all targets support
9559 declare float @llvm.exp.f32(float %Val)
9560 declare double @llvm.exp.f64(double %Val)
9561 declare x86_fp80 @llvm.exp.f80(x86_fp80 %Val)
9562 declare fp128 @llvm.exp.f128(fp128 %Val)
9563 declare ppc_fp128 @llvm.exp.ppcf128(ppc_fp128 %Val)
9568 The '``llvm.exp.*``' intrinsics perform the exp function.
9573 The argument and return value are floating point numbers of the same
9579 This function returns the same values as the libm ``exp`` functions
9580 would, and handles error conditions in the same way.
9582 '``llvm.exp2.*``' Intrinsic
9583 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9588 This is an overloaded intrinsic. You can use ``llvm.exp2`` on any
9589 floating point or vector of floating point type. Not all targets support
9594 declare float @llvm.exp2.f32(float %Val)
9595 declare double @llvm.exp2.f64(double %Val)
9596 declare x86_fp80 @llvm.exp2.f80(x86_fp80 %Val)
9597 declare fp128 @llvm.exp2.f128(fp128 %Val)
9598 declare ppc_fp128 @llvm.exp2.ppcf128(ppc_fp128 %Val)
9603 The '``llvm.exp2.*``' intrinsics perform the exp2 function.
9608 The argument and return value are floating point numbers of the same
9614 This function returns the same values as the libm ``exp2`` functions
9615 would, and handles error conditions in the same way.
9617 '``llvm.log.*``' Intrinsic
9618 ^^^^^^^^^^^^^^^^^^^^^^^^^^
9623 This is an overloaded intrinsic. You can use ``llvm.log`` on any
9624 floating point or vector of floating point type. Not all targets support
9629 declare float @llvm.log.f32(float %Val)
9630 declare double @llvm.log.f64(double %Val)
9631 declare x86_fp80 @llvm.log.f80(x86_fp80 %Val)
9632 declare fp128 @llvm.log.f128(fp128 %Val)
9633 declare ppc_fp128 @llvm.log.ppcf128(ppc_fp128 %Val)
9638 The '``llvm.log.*``' intrinsics perform the log function.
9643 The argument and return value are floating point numbers of the same
9649 This function returns the same values as the libm ``log`` functions
9650 would, and handles error conditions in the same way.
9652 '``llvm.log10.*``' Intrinsic
9653 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9658 This is an overloaded intrinsic. You can use ``llvm.log10`` on any
9659 floating point or vector of floating point type. Not all targets support
9664 declare float @llvm.log10.f32(float %Val)
9665 declare double @llvm.log10.f64(double %Val)
9666 declare x86_fp80 @llvm.log10.f80(x86_fp80 %Val)
9667 declare fp128 @llvm.log10.f128(fp128 %Val)
9668 declare ppc_fp128 @llvm.log10.ppcf128(ppc_fp128 %Val)
9673 The '``llvm.log10.*``' intrinsics perform the log10 function.
9678 The argument and return value are floating point numbers of the same
9684 This function returns the same values as the libm ``log10`` functions
9685 would, and handles error conditions in the same way.
9687 '``llvm.log2.*``' Intrinsic
9688 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9693 This is an overloaded intrinsic. You can use ``llvm.log2`` on any
9694 floating point or vector of floating point type. Not all targets support
9699 declare float @llvm.log2.f32(float %Val)
9700 declare double @llvm.log2.f64(double %Val)
9701 declare x86_fp80 @llvm.log2.f80(x86_fp80 %Val)
9702 declare fp128 @llvm.log2.f128(fp128 %Val)
9703 declare ppc_fp128 @llvm.log2.ppcf128(ppc_fp128 %Val)
9708 The '``llvm.log2.*``' intrinsics perform the log2 function.
9713 The argument and return value are floating point numbers of the same
9719 This function returns the same values as the libm ``log2`` functions
9720 would, and handles error conditions in the same way.
9722 '``llvm.fma.*``' Intrinsic
9723 ^^^^^^^^^^^^^^^^^^^^^^^^^^
9728 This is an overloaded intrinsic. You can use ``llvm.fma`` on any
9729 floating point or vector of floating point type. Not all targets support
9734 declare float @llvm.fma.f32(float %a, float %b, float %c)
9735 declare double @llvm.fma.f64(double %a, double %b, double %c)
9736 declare x86_fp80 @llvm.fma.f80(x86_fp80 %a, x86_fp80 %b, x86_fp80 %c)
9737 declare fp128 @llvm.fma.f128(fp128 %a, fp128 %b, fp128 %c)
9738 declare ppc_fp128 @llvm.fma.ppcf128(ppc_fp128 %a, ppc_fp128 %b, ppc_fp128 %c)
9743 The '``llvm.fma.*``' intrinsics perform the fused multiply-add
9749 The argument and return value are floating point numbers of the same
9755 This function returns the same values as the libm ``fma`` functions
9756 would, and does not set errno.
9758 '``llvm.fabs.*``' Intrinsic
9759 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9764 This is an overloaded intrinsic. You can use ``llvm.fabs`` on any
9765 floating point or vector of floating point type. Not all targets support
9770 declare float @llvm.fabs.f32(float %Val)
9771 declare double @llvm.fabs.f64(double %Val)
9772 declare x86_fp80 @llvm.fabs.f80(x86_fp80 %Val)
9773 declare fp128 @llvm.fabs.f128(fp128 %Val)
9774 declare ppc_fp128 @llvm.fabs.ppcf128(ppc_fp128 %Val)
9779 The '``llvm.fabs.*``' intrinsics return the absolute value of the
9785 The argument and return value are floating point numbers of the same
9791 This function returns the same values as the libm ``fabs`` functions
9792 would, and handles error conditions in the same way.
9794 '``llvm.minnum.*``' Intrinsic
9795 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9800 This is an overloaded intrinsic. You can use ``llvm.minnum`` on any
9801 floating point or vector of floating point type. Not all targets support
9806 declare float @llvm.minnum.f32(float %Val0, float %Val1)
9807 declare double @llvm.minnum.f64(double %Val0, double %Val1)
9808 declare x86_fp80 @llvm.minnum.f80(x86_fp80 %Val0, x86_fp80 %Val1)
9809 declare fp128 @llvm.minnum.f128(fp128 %Val0, fp128 %Val1)
9810 declare ppc_fp128 @llvm.minnum.ppcf128(ppc_fp128 %Val0, ppc_fp128 %Val1)
9815 The '``llvm.minnum.*``' intrinsics return the minimum of the two
9822 The arguments and return value are floating point numbers of the same
9828 Follows the IEEE-754 semantics for minNum, which also match for libm's
9831 If either operand is a NaN, returns the other non-NaN operand. Returns
9832 NaN only if both operands are NaN. If the operands compare equal,
9833 returns a value that compares equal to both operands. This means that
9834 fmin(+/-0.0, +/-0.0) could return either -0.0 or 0.0.
9836 '``llvm.maxnum.*``' Intrinsic
9837 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9842 This is an overloaded intrinsic. You can use ``llvm.maxnum`` on any
9843 floating point or vector of floating point type. Not all targets support
9848 declare float @llvm.maxnum.f32(float %Val0, float %Val1l)
9849 declare double @llvm.maxnum.f64(double %Val0, double %Val1)
9850 declare x86_fp80 @llvm.maxnum.f80(x86_fp80 %Val0, x86_fp80 %Val1)
9851 declare fp128 @llvm.maxnum.f128(fp128 %Val0, fp128 %Val1)
9852 declare ppc_fp128 @llvm.maxnum.ppcf128(ppc_fp128 %Val0, ppc_fp128 %Val1)
9857 The '``llvm.maxnum.*``' intrinsics return the maximum of the two
9864 The arguments and return value are floating point numbers of the same
9869 Follows the IEEE-754 semantics for maxNum, which also match for libm's
9872 If either operand is a NaN, returns the other non-NaN operand. Returns
9873 NaN only if both operands are NaN. If the operands compare equal,
9874 returns a value that compares equal to both operands. This means that
9875 fmax(+/-0.0, +/-0.0) could return either -0.0 or 0.0.
9877 '``llvm.copysign.*``' Intrinsic
9878 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9883 This is an overloaded intrinsic. You can use ``llvm.copysign`` on any
9884 floating point or vector of floating point type. Not all targets support
9889 declare float @llvm.copysign.f32(float %Mag, float %Sgn)
9890 declare double @llvm.copysign.f64(double %Mag, double %Sgn)
9891 declare x86_fp80 @llvm.copysign.f80(x86_fp80 %Mag, x86_fp80 %Sgn)
9892 declare fp128 @llvm.copysign.f128(fp128 %Mag, fp128 %Sgn)
9893 declare ppc_fp128 @llvm.copysign.ppcf128(ppc_fp128 %Mag, ppc_fp128 %Sgn)
9898 The '``llvm.copysign.*``' intrinsics return a value with the magnitude of the
9899 first operand and the sign of the second operand.
9904 The arguments and return value are floating point numbers of the same
9910 This function returns the same values as the libm ``copysign``
9911 functions would, and handles error conditions in the same way.
9913 '``llvm.floor.*``' Intrinsic
9914 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9919 This is an overloaded intrinsic. You can use ``llvm.floor`` on any
9920 floating point or vector of floating point type. Not all targets support
9925 declare float @llvm.floor.f32(float %Val)
9926 declare double @llvm.floor.f64(double %Val)
9927 declare x86_fp80 @llvm.floor.f80(x86_fp80 %Val)
9928 declare fp128 @llvm.floor.f128(fp128 %Val)
9929 declare ppc_fp128 @llvm.floor.ppcf128(ppc_fp128 %Val)
9934 The '``llvm.floor.*``' intrinsics return the floor of the operand.
9939 The argument and return value are floating point numbers of the same
9945 This function returns the same values as the libm ``floor`` functions
9946 would, and handles error conditions in the same way.
9948 '``llvm.ceil.*``' Intrinsic
9949 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9954 This is an overloaded intrinsic. You can use ``llvm.ceil`` on any
9955 floating point or vector of floating point type. Not all targets support
9960 declare float @llvm.ceil.f32(float %Val)
9961 declare double @llvm.ceil.f64(double %Val)
9962 declare x86_fp80 @llvm.ceil.f80(x86_fp80 %Val)
9963 declare fp128 @llvm.ceil.f128(fp128 %Val)
9964 declare ppc_fp128 @llvm.ceil.ppcf128(ppc_fp128 %Val)
9969 The '``llvm.ceil.*``' intrinsics return the ceiling of the operand.
9974 The argument and return value are floating point numbers of the same
9980 This function returns the same values as the libm ``ceil`` functions
9981 would, and handles error conditions in the same way.
9983 '``llvm.trunc.*``' Intrinsic
9984 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9989 This is an overloaded intrinsic. You can use ``llvm.trunc`` on any
9990 floating point or vector of floating point type. Not all targets support
9995 declare float @llvm.trunc.f32(float %Val)
9996 declare double @llvm.trunc.f64(double %Val)
9997 declare x86_fp80 @llvm.trunc.f80(x86_fp80 %Val)
9998 declare fp128 @llvm.trunc.f128(fp128 %Val)
9999 declare ppc_fp128 @llvm.trunc.ppcf128(ppc_fp128 %Val)
10004 The '``llvm.trunc.*``' intrinsics returns the operand rounded to the
10005 nearest integer not larger in magnitude than the operand.
10010 The argument and return value are floating point numbers of the same
10016 This function returns the same values as the libm ``trunc`` functions
10017 would, and handles error conditions in the same way.
10019 '``llvm.rint.*``' Intrinsic
10020 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10025 This is an overloaded intrinsic. You can use ``llvm.rint`` on any
10026 floating point or vector of floating point type. Not all targets support
10031 declare float @llvm.rint.f32(float %Val)
10032 declare double @llvm.rint.f64(double %Val)
10033 declare x86_fp80 @llvm.rint.f80(x86_fp80 %Val)
10034 declare fp128 @llvm.rint.f128(fp128 %Val)
10035 declare ppc_fp128 @llvm.rint.ppcf128(ppc_fp128 %Val)
10040 The '``llvm.rint.*``' intrinsics returns the operand rounded to the
10041 nearest integer. It may raise an inexact floating-point exception if the
10042 operand isn't an integer.
10047 The argument and return value are floating point numbers of the same
10053 This function returns the same values as the libm ``rint`` functions
10054 would, and handles error conditions in the same way.
10056 '``llvm.nearbyint.*``' Intrinsic
10057 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10062 This is an overloaded intrinsic. You can use ``llvm.nearbyint`` on any
10063 floating point or vector of floating point type. Not all targets support
10068 declare float @llvm.nearbyint.f32(float %Val)
10069 declare double @llvm.nearbyint.f64(double %Val)
10070 declare x86_fp80 @llvm.nearbyint.f80(x86_fp80 %Val)
10071 declare fp128 @llvm.nearbyint.f128(fp128 %Val)
10072 declare ppc_fp128 @llvm.nearbyint.ppcf128(ppc_fp128 %Val)
10077 The '``llvm.nearbyint.*``' intrinsics returns the operand rounded to the
10083 The argument and return value are floating point numbers of the same
10089 This function returns the same values as the libm ``nearbyint``
10090 functions would, and handles error conditions in the same way.
10092 '``llvm.round.*``' Intrinsic
10093 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10098 This is an overloaded intrinsic. You can use ``llvm.round`` on any
10099 floating point or vector of floating point type. Not all targets support
10104 declare float @llvm.round.f32(float %Val)
10105 declare double @llvm.round.f64(double %Val)
10106 declare x86_fp80 @llvm.round.f80(x86_fp80 %Val)
10107 declare fp128 @llvm.round.f128(fp128 %Val)
10108 declare ppc_fp128 @llvm.round.ppcf128(ppc_fp128 %Val)
10113 The '``llvm.round.*``' intrinsics returns the operand rounded to the
10119 The argument and return value are floating point numbers of the same
10125 This function returns the same values as the libm ``round``
10126 functions would, and handles error conditions in the same way.
10128 Bit Manipulation Intrinsics
10129 ---------------------------
10131 LLVM provides intrinsics for a few important bit manipulation
10132 operations. These allow efficient code generation for some algorithms.
10134 '``llvm.bswap.*``' Intrinsics
10135 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10140 This is an overloaded intrinsic function. You can use bswap on any
10141 integer type that is an even number of bytes (i.e. BitWidth % 16 == 0).
10145 declare i16 @llvm.bswap.i16(i16 <id>)
10146 declare i32 @llvm.bswap.i32(i32 <id>)
10147 declare i64 @llvm.bswap.i64(i64 <id>)
10152 The '``llvm.bswap``' family of intrinsics is used to byte swap integer
10153 values with an even number of bytes (positive multiple of 16 bits).
10154 These are useful for performing operations on data that is not in the
10155 target's native byte order.
10160 The ``llvm.bswap.i16`` intrinsic returns an i16 value that has the high
10161 and low byte of the input i16 swapped. Similarly, the ``llvm.bswap.i32``
10162 intrinsic returns an i32 value that has the four bytes of the input i32
10163 swapped, so that if the input bytes are numbered 0, 1, 2, 3 then the
10164 returned i32 will have its bytes in 3, 2, 1, 0 order. The
10165 ``llvm.bswap.i48``, ``llvm.bswap.i64`` and other intrinsics extend this
10166 concept to additional even-byte lengths (6 bytes, 8 bytes and more,
10169 '``llvm.ctpop.*``' Intrinsic
10170 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10175 This is an overloaded intrinsic. You can use llvm.ctpop on any integer
10176 bit width, or on any vector with integer elements. Not all targets
10177 support all bit widths or vector types, however.
10181 declare i8 @llvm.ctpop.i8(i8 <src>)
10182 declare i16 @llvm.ctpop.i16(i16 <src>)
10183 declare i32 @llvm.ctpop.i32(i32 <src>)
10184 declare i64 @llvm.ctpop.i64(i64 <src>)
10185 declare i256 @llvm.ctpop.i256(i256 <src>)
10186 declare <2 x i32> @llvm.ctpop.v2i32(<2 x i32> <src>)
10191 The '``llvm.ctpop``' family of intrinsics counts the number of bits set
10197 The only argument is the value to be counted. The argument may be of any
10198 integer type, or a vector with integer elements. The return type must
10199 match the argument type.
10204 The '``llvm.ctpop``' intrinsic counts the 1's in a variable, or within
10205 each element of a vector.
10207 '``llvm.ctlz.*``' Intrinsic
10208 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10213 This is an overloaded intrinsic. You can use ``llvm.ctlz`` on any
10214 integer bit width, or any vector whose elements are integers. Not all
10215 targets support all bit widths or vector types, however.
10219 declare i8 @llvm.ctlz.i8 (i8 <src>, i1 <is_zero_undef>)
10220 declare i16 @llvm.ctlz.i16 (i16 <src>, i1 <is_zero_undef>)
10221 declare i32 @llvm.ctlz.i32 (i32 <src>, i1 <is_zero_undef>)
10222 declare i64 @llvm.ctlz.i64 (i64 <src>, i1 <is_zero_undef>)
10223 declare i256 @llvm.ctlz.i256(i256 <src>, i1 <is_zero_undef>)
10224 declase <2 x i32> @llvm.ctlz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
10229 The '``llvm.ctlz``' family of intrinsic functions counts the number of
10230 leading zeros in a variable.
10235 The first argument is the value to be counted. This argument may be of
10236 any integer type, or a vector with integer element type. The return
10237 type must match the first argument type.
10239 The second argument must be a constant and is a flag to indicate whether
10240 the intrinsic should ensure that a zero as the first argument produces a
10241 defined result. Historically some architectures did not provide a
10242 defined result for zero values as efficiently, and many algorithms are
10243 now predicated on avoiding zero-value inputs.
10248 The '``llvm.ctlz``' intrinsic counts the leading (most significant)
10249 zeros in a variable, or within each element of the vector. If
10250 ``src == 0`` then the result is the size in bits of the type of ``src``
10251 if ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
10252 ``llvm.ctlz(i32 2) = 30``.
10254 '``llvm.cttz.*``' Intrinsic
10255 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
10260 This is an overloaded intrinsic. You can use ``llvm.cttz`` on any
10261 integer bit width, or any vector of integer elements. Not all targets
10262 support all bit widths or vector types, however.
10266 declare i8 @llvm.cttz.i8 (i8 <src>, i1 <is_zero_undef>)
10267 declare i16 @llvm.cttz.i16 (i16 <src>, i1 <is_zero_undef>)
10268 declare i32 @llvm.cttz.i32 (i32 <src>, i1 <is_zero_undef>)
10269 declare i64 @llvm.cttz.i64 (i64 <src>, i1 <is_zero_undef>)
10270 declare i256 @llvm.cttz.i256(i256 <src>, i1 <is_zero_undef>)
10271 declase <2 x i32> @llvm.cttz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
10276 The '``llvm.cttz``' family of intrinsic functions counts the number of
10282 The first argument is the value to be counted. This argument may be of
10283 any integer type, or a vector with integer element type. The return
10284 type must match the first argument type.
10286 The second argument must be a constant and is a flag to indicate whether
10287 the intrinsic should ensure that a zero as the first argument produces a
10288 defined result. Historically some architectures did not provide a
10289 defined result for zero values as efficiently, and many algorithms are
10290 now predicated on avoiding zero-value inputs.
10295 The '``llvm.cttz``' intrinsic counts the trailing (least significant)
10296 zeros in a variable, or within each element of a vector. If ``src == 0``
10297 then the result is the size in bits of the type of ``src`` if
10298 ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
10299 ``llvm.cttz(2) = 1``.
10303 Arithmetic with Overflow Intrinsics
10304 -----------------------------------
10306 LLVM provides intrinsics for some arithmetic with overflow operations.
10308 '``llvm.sadd.with.overflow.*``' Intrinsics
10309 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10314 This is an overloaded intrinsic. You can use ``llvm.sadd.with.overflow``
10315 on any integer bit width.
10319 declare {i16, i1} @llvm.sadd.with.overflow.i16(i16 %a, i16 %b)
10320 declare {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
10321 declare {i64, i1} @llvm.sadd.with.overflow.i64(i64 %a, i64 %b)
10326 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
10327 a signed addition of the two arguments, and indicate whether an overflow
10328 occurred during the signed summation.
10333 The arguments (%a and %b) and the first element of the result structure
10334 may be of integer types of any bit width, but they must have the same
10335 bit width. The second element of the result structure must be of type
10336 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
10342 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
10343 a signed addition of the two variables. They return a structure --- the
10344 first element of which is the signed summation, and the second element
10345 of which is a bit specifying if the signed summation resulted in an
10351 .. code-block:: llvm
10353 %res = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
10354 %sum = extractvalue {i32, i1} %res, 0
10355 %obit = extractvalue {i32, i1} %res, 1
10356 br i1 %obit, label %overflow, label %normal
10358 '``llvm.uadd.with.overflow.*``' Intrinsics
10359 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10364 This is an overloaded intrinsic. You can use ``llvm.uadd.with.overflow``
10365 on any integer bit width.
10369 declare {i16, i1} @llvm.uadd.with.overflow.i16(i16 %a, i16 %b)
10370 declare {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
10371 declare {i64, i1} @llvm.uadd.with.overflow.i64(i64 %a, i64 %b)
10376 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
10377 an unsigned addition of the two arguments, and indicate whether a carry
10378 occurred during the unsigned summation.
10383 The arguments (%a and %b) and the first element of the result structure
10384 may be of integer types of any bit width, but they must have the same
10385 bit width. The second element of the result structure must be of type
10386 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
10392 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
10393 an unsigned addition of the two arguments. They return a structure --- the
10394 first element of which is the sum, and the second element of which is a
10395 bit specifying if the unsigned summation resulted in a carry.
10400 .. code-block:: llvm
10402 %res = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
10403 %sum = extractvalue {i32, i1} %res, 0
10404 %obit = extractvalue {i32, i1} %res, 1
10405 br i1 %obit, label %carry, label %normal
10407 '``llvm.ssub.with.overflow.*``' Intrinsics
10408 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10413 This is an overloaded intrinsic. You can use ``llvm.ssub.with.overflow``
10414 on any integer bit width.
10418 declare {i16, i1} @llvm.ssub.with.overflow.i16(i16 %a, i16 %b)
10419 declare {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
10420 declare {i64, i1} @llvm.ssub.with.overflow.i64(i64 %a, i64 %b)
10425 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
10426 a signed subtraction of the two arguments, and indicate whether an
10427 overflow occurred during the signed subtraction.
10432 The arguments (%a and %b) and the first element of the result structure
10433 may be of integer types of any bit width, but they must have the same
10434 bit width. The second element of the result structure must be of type
10435 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
10441 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
10442 a signed subtraction of the two arguments. They return a structure --- the
10443 first element of which is the subtraction, and the second element of
10444 which is a bit specifying if the signed subtraction resulted in an
10450 .. code-block:: llvm
10452 %res = call {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
10453 %sum = extractvalue {i32, i1} %res, 0
10454 %obit = extractvalue {i32, i1} %res, 1
10455 br i1 %obit, label %overflow, label %normal
10457 '``llvm.usub.with.overflow.*``' Intrinsics
10458 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10463 This is an overloaded intrinsic. You can use ``llvm.usub.with.overflow``
10464 on any integer bit width.
10468 declare {i16, i1} @llvm.usub.with.overflow.i16(i16 %a, i16 %b)
10469 declare {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
10470 declare {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
10475 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
10476 an unsigned subtraction of the two arguments, and indicate whether an
10477 overflow occurred during the unsigned subtraction.
10482 The arguments (%a and %b) and the first element of the result structure
10483 may be of integer types of any bit width, but they must have the same
10484 bit width. The second element of the result structure must be of type
10485 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
10491 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
10492 an unsigned subtraction of the two arguments. They return a structure ---
10493 the first element of which is the subtraction, and the second element of
10494 which is a bit specifying if the unsigned subtraction resulted in an
10500 .. code-block:: llvm
10502 %res = call {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
10503 %sum = extractvalue {i32, i1} %res, 0
10504 %obit = extractvalue {i32, i1} %res, 1
10505 br i1 %obit, label %overflow, label %normal
10507 '``llvm.smul.with.overflow.*``' Intrinsics
10508 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10513 This is an overloaded intrinsic. You can use ``llvm.smul.with.overflow``
10514 on any integer bit width.
10518 declare {i16, i1} @llvm.smul.with.overflow.i16(i16 %a, i16 %b)
10519 declare {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
10520 declare {i64, i1} @llvm.smul.with.overflow.i64(i64 %a, i64 %b)
10525 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
10526 a signed multiplication of the two arguments, and indicate whether an
10527 overflow occurred during the signed multiplication.
10532 The arguments (%a and %b) and the first element of the result structure
10533 may be of integer types of any bit width, but they must have the same
10534 bit width. The second element of the result structure must be of type
10535 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
10541 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
10542 a signed multiplication of the two arguments. They return a structure ---
10543 the first element of which is the multiplication, and the second element
10544 of which is a bit specifying if the signed multiplication resulted in an
10550 .. code-block:: llvm
10552 %res = call {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
10553 %sum = extractvalue {i32, i1} %res, 0
10554 %obit = extractvalue {i32, i1} %res, 1
10555 br i1 %obit, label %overflow, label %normal
10557 '``llvm.umul.with.overflow.*``' Intrinsics
10558 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10563 This is an overloaded intrinsic. You can use ``llvm.umul.with.overflow``
10564 on any integer bit width.
10568 declare {i16, i1} @llvm.umul.with.overflow.i16(i16 %a, i16 %b)
10569 declare {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
10570 declare {i64, i1} @llvm.umul.with.overflow.i64(i64 %a, i64 %b)
10575 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
10576 a unsigned multiplication of the two arguments, and indicate whether an
10577 overflow occurred during the unsigned multiplication.
10582 The arguments (%a and %b) and the first element of the result structure
10583 may be of integer types of any bit width, but they must have the same
10584 bit width. The second element of the result structure must be of type
10585 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
10591 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
10592 an unsigned multiplication of the two arguments. They return a structure ---
10593 the first element of which is the multiplication, and the second
10594 element of which is a bit specifying if the unsigned multiplication
10595 resulted in an overflow.
10600 .. code-block:: llvm
10602 %res = call {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
10603 %sum = extractvalue {i32, i1} %res, 0
10604 %obit = extractvalue {i32, i1} %res, 1
10605 br i1 %obit, label %overflow, label %normal
10607 Specialised Arithmetic Intrinsics
10608 ---------------------------------
10610 '``llvm.canonicalize.*``' Intrinsic
10611 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10618 declare float @llvm.canonicalize.f32(float %a)
10619 declare double @llvm.canonicalize.f64(double %b)
10624 The '``llvm.canonicalize.*``' intrinsic returns the platform specific canonical
10625 encoding of a floating point number. This canonicalization is useful for
10626 implementing certain numeric primitives such as frexp. The canonical encoding is
10627 defined by IEEE-754-2008 to be:
10631 2.1.8 canonical encoding: The preferred encoding of a floating-point
10632 representation in a format. Applied to declets, significands of finite
10633 numbers, infinities, and NaNs, especially in decimal formats.
10635 This operation can also be considered equivalent to the IEEE-754-2008
10636 conversion of a floating-point value to the same format. NaNs are handled
10637 according to section 6.2.
10639 Examples of non-canonical encodings:
10641 - x87 pseudo denormals, pseudo NaNs, pseudo Infinity, Unnormals. These are
10642 converted to a canonical representation per hardware-specific protocol.
10643 - Many normal decimal floating point numbers have non-canonical alternative
10645 - Some machines, like GPUs or ARMv7 NEON, do not support subnormal values.
10646 These are treated as non-canonical encodings of zero and with be flushed to
10647 a zero of the same sign by this operation.
10649 Note that per IEEE-754-2008 6.2, systems that support signaling NaNs with
10650 default exception handling must signal an invalid exception, and produce a
10653 This function should always be implementable as multiplication by 1.0, provided
10654 that the compiler does not constant fold the operation. Likewise, division by
10655 1.0 and ``llvm.minnum(x, x)`` are possible implementations. Addition with
10656 -0.0 is also sufficient provided that the rounding mode is not -Infinity.
10658 ``@llvm.canonicalize`` must preserve the equality relation. That is:
10660 - ``(@llvm.canonicalize(x) == x)`` is equivalent to ``(x == x)``
10661 - ``(@llvm.canonicalize(x) == @llvm.canonicalize(y))`` is equivalent to
10664 Additionally, the sign of zero must be conserved:
10665 ``@llvm.canonicalize(-0.0) = -0.0`` and ``@llvm.canonicalize(+0.0) = +0.0``
10667 The payload bits of a NaN must be conserved, with two exceptions.
10668 First, environments which use only a single canonical representation of NaN
10669 must perform said canonicalization. Second, SNaNs must be quieted per the
10672 The canonicalization operation may be optimized away if:
10674 - The input is known to be canonical. For example, it was produced by a
10675 floating-point operation that is required by the standard to be canonical.
10676 - The result is consumed only by (or fused with) other floating-point
10677 operations. That is, the bits of the floating point value are not examined.
10679 '``llvm.fmuladd.*``' Intrinsic
10680 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10687 declare float @llvm.fmuladd.f32(float %a, float %b, float %c)
10688 declare double @llvm.fmuladd.f64(double %a, double %b, double %c)
10693 The '``llvm.fmuladd.*``' intrinsic functions represent multiply-add
10694 expressions that can be fused if the code generator determines that (a) the
10695 target instruction set has support for a fused operation, and (b) that the
10696 fused operation is more efficient than the equivalent, separate pair of mul
10697 and add instructions.
10702 The '``llvm.fmuladd.*``' intrinsics each take three arguments: two
10703 multiplicands, a and b, and an addend c.
10712 %0 = call float @llvm.fmuladd.f32(%a, %b, %c)
10714 is equivalent to the expression a \* b + c, except that rounding will
10715 not be performed between the multiplication and addition steps if the
10716 code generator fuses the operations. Fusion is not guaranteed, even if
10717 the target platform supports it. If a fused multiply-add is required the
10718 corresponding llvm.fma.\* intrinsic function should be used
10719 instead. This never sets errno, just as '``llvm.fma.*``'.
10724 .. code-block:: llvm
10726 %r2 = call float @llvm.fmuladd.f32(float %a, float %b, float %c) ; yields float:r2 = (a * b) + c
10729 '``llvm.uabsdiff.*``' and '``llvm.sabsdiff.*``' Intrinsics
10730 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10734 This is an overloaded intrinsic. The loaded data is a vector of any integer bit width.
10736 .. code-block:: llvm
10738 declare <4 x integer> @llvm.uabsdiff.v4i32(<4 x integer> %a, <4 x integer> %b)
10744 The ``llvm.uabsdiff`` intrinsic returns a vector result of the absolute difference of the two operands,
10745 treating them both as unsigned integers.
10747 The ``llvm.sabsdiff`` intrinsic returns a vector result of the absolute difference of the two operands,
10748 treating them both as signed integers.
10752 These intrinsics are primarily used during the code generation stage of compilation.
10753 They are generated by compiler passes such as the Loop and SLP vectorizers.it is not
10754 recommended for users to create them manually.
10759 Both intrinsics take two integer of the same bitwidth.
10766 call <4 x i32> @llvm.uabsdiff.v4i32(<4 x i32> %a, <4 x i32> %b)
10770 %sub = sub <4 x i32> %a, %b
10771 %ispos = icmp ugt <4 x i32> %sub, <i32 -1, i32 -1, i32 -1, i32 -1>
10772 %neg = sub <4 x i32> zeroinitializer, %sub
10773 %1 = select <4 x i1> %ispos, <4 x i32> %sub, <4 x i32> %neg
10775 Similarly the expression::
10777 call <4 x i32> @llvm.sabsdiff.v4i32(<4 x i32> %a, <4 x i32> %b)
10781 %sub = sub nsw <4 x i32> %a, %b
10782 %ispos = icmp sgt <4 x i32> %sub, <i32 -1, i32 -1, i32 -1, i32 -1>
10783 %neg = sub nsw <4 x i32> zeroinitializer, %sub
10784 %1 = select <4 x i1> %ispos, <4 x i32> %sub, <4 x i32> %neg
10787 Half Precision Floating Point Intrinsics
10788 ----------------------------------------
10790 For most target platforms, half precision floating point is a
10791 storage-only format. This means that it is a dense encoding (in memory)
10792 but does not support computation in the format.
10794 This means that code must first load the half-precision floating point
10795 value as an i16, then convert it to float with
10796 :ref:`llvm.convert.from.fp16 <int_convert_from_fp16>`. Computation can
10797 then be performed on the float value (including extending to double
10798 etc). To store the value back to memory, it is first converted to float
10799 if needed, then converted to i16 with
10800 :ref:`llvm.convert.to.fp16 <int_convert_to_fp16>`, then storing as an
10803 .. _int_convert_to_fp16:
10805 '``llvm.convert.to.fp16``' Intrinsic
10806 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10813 declare i16 @llvm.convert.to.fp16.f32(float %a)
10814 declare i16 @llvm.convert.to.fp16.f64(double %a)
10819 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion from a
10820 conventional floating point type to half precision floating point format.
10825 The intrinsic function contains single argument - the value to be
10831 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion from a
10832 conventional floating point format to half precision floating point format. The
10833 return value is an ``i16`` which contains the converted number.
10838 .. code-block:: llvm
10840 %res = call i16 @llvm.convert.to.fp16.f32(float %a)
10841 store i16 %res, i16* @x, align 2
10843 .. _int_convert_from_fp16:
10845 '``llvm.convert.from.fp16``' Intrinsic
10846 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10853 declare float @llvm.convert.from.fp16.f32(i16 %a)
10854 declare double @llvm.convert.from.fp16.f64(i16 %a)
10859 The '``llvm.convert.from.fp16``' intrinsic function performs a
10860 conversion from half precision floating point format to single precision
10861 floating point format.
10866 The intrinsic function contains single argument - the value to be
10872 The '``llvm.convert.from.fp16``' intrinsic function performs a
10873 conversion from half single precision floating point format to single
10874 precision floating point format. The input half-float value is
10875 represented by an ``i16`` value.
10880 .. code-block:: llvm
10882 %a = load i16, i16* @x, align 2
10883 %res = call float @llvm.convert.from.fp16(i16 %a)
10885 .. _dbg_intrinsics:
10887 Debugger Intrinsics
10888 -------------------
10890 The LLVM debugger intrinsics (which all start with ``llvm.dbg.``
10891 prefix), are described in the `LLVM Source Level
10892 Debugging <SourceLevelDebugging.html#format_common_intrinsics>`_
10895 Exception Handling Intrinsics
10896 -----------------------------
10898 The LLVM exception handling intrinsics (which all start with
10899 ``llvm.eh.`` prefix), are described in the `LLVM Exception
10900 Handling <ExceptionHandling.html#format_common_intrinsics>`_ document.
10902 .. _int_trampoline:
10904 Trampoline Intrinsics
10905 ---------------------
10907 These intrinsics make it possible to excise one parameter, marked with
10908 the :ref:`nest <nest>` attribute, from a function. The result is a
10909 callable function pointer lacking the nest parameter - the caller does
10910 not need to provide a value for it. Instead, the value to use is stored
10911 in advance in a "trampoline", a block of memory usually allocated on the
10912 stack, which also contains code to splice the nest value into the
10913 argument list. This is used to implement the GCC nested function address
10916 For example, if the function is ``i32 f(i8* nest %c, i32 %x, i32 %y)``
10917 then the resulting function pointer has signature ``i32 (i32, i32)*``.
10918 It can be created as follows:
10920 .. code-block:: llvm
10922 %tramp = alloca [10 x i8], align 4 ; size and alignment only correct for X86
10923 %tramp1 = getelementptr [10 x i8], [10 x i8]* %tramp, i32 0, i32 0
10924 call i8* @llvm.init.trampoline(i8* %tramp1, i8* bitcast (i32 (i8*, i32, i32)* @f to i8*), i8* %nval)
10925 %p = call i8* @llvm.adjust.trampoline(i8* %tramp1)
10926 %fp = bitcast i8* %p to i32 (i32, i32)*
10928 The call ``%val = call i32 %fp(i32 %x, i32 %y)`` is then equivalent to
10929 ``%val = call i32 %f(i8* %nval, i32 %x, i32 %y)``.
10933 '``llvm.init.trampoline``' Intrinsic
10934 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10941 declare void @llvm.init.trampoline(i8* <tramp>, i8* <func>, i8* <nval>)
10946 This fills the memory pointed to by ``tramp`` with executable code,
10947 turning it into a trampoline.
10952 The ``llvm.init.trampoline`` intrinsic takes three arguments, all
10953 pointers. The ``tramp`` argument must point to a sufficiently large and
10954 sufficiently aligned block of memory; this memory is written to by the
10955 intrinsic. Note that the size and the alignment are target-specific -
10956 LLVM currently provides no portable way of determining them, so a
10957 front-end that generates this intrinsic needs to have some
10958 target-specific knowledge. The ``func`` argument must hold a function
10959 bitcast to an ``i8*``.
10964 The block of memory pointed to by ``tramp`` is filled with target
10965 dependent code, turning it into a function. Then ``tramp`` needs to be
10966 passed to :ref:`llvm.adjust.trampoline <int_at>` to get a pointer which can
10967 be :ref:`bitcast (to a new function) and called <int_trampoline>`. The new
10968 function's signature is the same as that of ``func`` with any arguments
10969 marked with the ``nest`` attribute removed. At most one such ``nest``
10970 argument is allowed, and it must be of pointer type. Calling the new
10971 function is equivalent to calling ``func`` with the same argument list,
10972 but with ``nval`` used for the missing ``nest`` argument. If, after
10973 calling ``llvm.init.trampoline``, the memory pointed to by ``tramp`` is
10974 modified, then the effect of any later call to the returned function
10975 pointer is undefined.
10979 '``llvm.adjust.trampoline``' Intrinsic
10980 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
10987 declare i8* @llvm.adjust.trampoline(i8* <tramp>)
10992 This performs any required machine-specific adjustment to the address of
10993 a trampoline (passed as ``tramp``).
10998 ``tramp`` must point to a block of memory which already has trampoline
10999 code filled in by a previous call to
11000 :ref:`llvm.init.trampoline <int_it>`.
11005 On some architectures the address of the code to be executed needs to be
11006 different than the address where the trampoline is actually stored. This
11007 intrinsic returns the executable address corresponding to ``tramp``
11008 after performing the required machine specific adjustments. The pointer
11009 returned can then be :ref:`bitcast and executed <int_trampoline>`.
11011 .. _int_mload_mstore:
11013 Masked Vector Load and Store Intrinsics
11014 ---------------------------------------
11016 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.
11020 '``llvm.masked.load.*``' Intrinsics
11021 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11025 This is an overloaded intrinsic. The loaded data is a vector of any integer or floating point data type.
11029 declare <16 x float> @llvm.masked.load.v16f32 (<16 x float>* <ptr>, i32 <alignment>, <16 x i1> <mask>, <16 x float> <passthru>)
11030 declare <2 x double> @llvm.masked.load.v2f64 (<2 x double>* <ptr>, i32 <alignment>, <2 x i1> <mask>, <2 x double> <passthru>)
11035 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.
11041 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.
11047 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.
11048 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.
11053 %res = call <16 x float> @llvm.masked.load.v16f32 (<16 x float>* %ptr, i32 4, <16 x i1>%mask, <16 x float> %passthru)
11055 ;; The result of the two following instructions is identical aside from potential memory access exception
11056 %loadlal = load <16 x float>, <16 x float>* %ptr, align 4
11057 %res = select <16 x i1> %mask, <16 x float> %loadlal, <16 x float> %passthru
11061 '``llvm.masked.store.*``' Intrinsics
11062 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11066 This is an overloaded intrinsic. The data stored in memory is a vector of any integer or floating point data type.
11070 declare void @llvm.masked.store.v8i32 (<8 x i32> <value>, <8 x i32> * <ptr>, i32 <alignment>, <8 x i1> <mask>)
11071 declare void @llvm.masked.store.v16f32(<16 x i32> <value>, <16 x i32>* <ptr>, i32 <alignment>, <16 x i1> <mask>)
11076 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.
11081 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.
11087 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.
11088 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.
11092 call void @llvm.masked.store.v16f32(<16 x float> %value, <16 x float>* %ptr, i32 4, <16 x i1> %mask)
11094 ;; The result of the following instructions is identical aside from potential data races and memory access exceptions
11095 %oldval = load <16 x float>, <16 x float>* %ptr, align 4
11096 %res = select <16 x i1> %mask, <16 x float> %value, <16 x float> %oldval
11097 store <16 x float> %res, <16 x float>* %ptr, align 4
11100 Masked Vector Gather and Scatter Intrinsics
11101 -------------------------------------------
11103 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.
11107 '``llvm.masked.gather.*``' Intrinsics
11108 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11112 This is an overloaded intrinsic. The loaded data are multiple scalar values of any integer or floating point data type gathered together into one vector.
11116 declare <16 x float> @llvm.masked.gather.v16f32 (<16 x float*> <ptrs>, i32 <alignment>, <16 x i1> <mask>, <16 x float> <passthru>)
11117 declare <2 x double> @llvm.masked.gather.v2f64 (<2 x double*> <ptrs>, i32 <alignment>, <2 x i1> <mask>, <2 x double> <passthru>)
11122 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.
11128 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.
11134 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.
11135 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.
11140 %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>)
11142 ;; The gather with all-true mask is equivalent to the following instruction sequence
11143 %ptr0 = extractelement <4 x double*> %ptrs, i32 0
11144 %ptr1 = extractelement <4 x double*> %ptrs, i32 1
11145 %ptr2 = extractelement <4 x double*> %ptrs, i32 2
11146 %ptr3 = extractelement <4 x double*> %ptrs, i32 3
11148 %val0 = load double, double* %ptr0, align 8
11149 %val1 = load double, double* %ptr1, align 8
11150 %val2 = load double, double* %ptr2, align 8
11151 %val3 = load double, double* %ptr3, align 8
11153 %vec0 = insertelement <4 x double>undef, %val0, 0
11154 %vec01 = insertelement <4 x double>%vec0, %val1, 1
11155 %vec012 = insertelement <4 x double>%vec01, %val2, 2
11156 %vec0123 = insertelement <4 x double>%vec012, %val3, 3
11160 '``llvm.masked.scatter.*``' Intrinsics
11161 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11165 This is an overloaded intrinsic. The data stored in memory is a vector of any integer or floating point data type. Each vector element is stored in an arbitrary memory addresses. Scatter with overlapping addresses is guaranteed to be ordered from least-significant to most-significant element.
11169 declare void @llvm.masked.scatter.v8i32 (<8 x i32> <value>, <8 x i32*> <ptrs>, i32 <alignment>, <8 x i1> <mask>)
11170 declare void @llvm.masked.scatter.v16f32(<16 x i32> <value>, <16 x i32*> <ptrs>, i32 <alignment>, <16 x i1> <mask>)
11175 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.
11180 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.
11186 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 divergency. Other targets may support this intrinsic differently, for example by lowering it into a sequence of branches that guard scalar store operations.
11190 ;; This instruction unconditionaly stores data vector in multiple addresses
11191 call @llvm.masked.scatter.v8i32 (<8 x i32> %value, <8 x i32*> %ptrs, i32 4, <8 x i1> <true, true, .. true>)
11193 ;; It is equivalent to a list of scalar stores
11194 %val0 = extractelement <8 x i32> %value, i32 0
11195 %val1 = extractelement <8 x i32> %value, i32 1
11197 %val7 = extractelement <8 x i32> %value, i32 7
11198 %ptr0 = extractelement <8 x i32*> %ptrs, i32 0
11199 %ptr1 = extractelement <8 x i32*> %ptrs, i32 1
11201 %ptr7 = extractelement <8 x i32*> %ptrs, i32 7
11202 ;; Note: the order of the following stores is important when they overlap:
11203 store i32 %val0, i32* %ptr0, align 4
11204 store i32 %val1, i32* %ptr1, align 4
11206 store i32 %val7, i32* %ptr7, align 4
11212 This class of intrinsics provides information about the lifetime of
11213 memory objects and ranges where variables are immutable.
11217 '``llvm.lifetime.start``' Intrinsic
11218 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11225 declare void @llvm.lifetime.start(i64 <size>, i8* nocapture <ptr>)
11230 The '``llvm.lifetime.start``' intrinsic specifies the start of a memory
11236 The first argument is a constant integer representing the size of the
11237 object, or -1 if it is variable sized. The second argument is a pointer
11243 This intrinsic indicates that before this point in the code, the value
11244 of the memory pointed to by ``ptr`` is dead. This means that it is known
11245 to never be used and has an undefined value. A load from the pointer
11246 that precedes this intrinsic can be replaced with ``'undef'``.
11250 '``llvm.lifetime.end``' Intrinsic
11251 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11258 declare void @llvm.lifetime.end(i64 <size>, i8* nocapture <ptr>)
11263 The '``llvm.lifetime.end``' intrinsic specifies the end of a memory
11269 The first argument is a constant integer representing the size of the
11270 object, or -1 if it is variable sized. The second argument is a pointer
11276 This intrinsic indicates that after this point in the code, the value of
11277 the memory pointed to by ``ptr`` is dead. This means that it is known to
11278 never be used and has an undefined value. Any stores into the memory
11279 object following this intrinsic may be removed as dead.
11281 '``llvm.invariant.start``' Intrinsic
11282 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11289 declare {}* @llvm.invariant.start(i64 <size>, i8* nocapture <ptr>)
11294 The '``llvm.invariant.start``' intrinsic specifies that the contents of
11295 a memory object will not change.
11300 The first argument is a constant integer representing the size of the
11301 object, or -1 if it is variable sized. The second argument is a pointer
11307 This intrinsic indicates that until an ``llvm.invariant.end`` that uses
11308 the return value, the referenced memory location is constant and
11311 '``llvm.invariant.end``' Intrinsic
11312 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11319 declare void @llvm.invariant.end({}* <start>, i64 <size>, i8* nocapture <ptr>)
11324 The '``llvm.invariant.end``' intrinsic specifies that the contents of a
11325 memory object are mutable.
11330 The first argument is the matching ``llvm.invariant.start`` intrinsic.
11331 The second argument is a constant integer representing the size of the
11332 object, or -1 if it is variable sized and the third argument is a
11333 pointer to the object.
11338 This intrinsic indicates that the memory is mutable again.
11343 This class of intrinsics is designed to be generic and has no specific
11346 '``llvm.var.annotation``' Intrinsic
11347 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11354 declare void @llvm.var.annotation(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
11359 The '``llvm.var.annotation``' intrinsic.
11364 The first argument is a pointer to a value, the second is a pointer to a
11365 global string, the third is a pointer to a global string which is the
11366 source file name, and the last argument is the line number.
11371 This intrinsic allows annotation of local variables with arbitrary
11372 strings. This can be useful for special purpose optimizations that want
11373 to look for these annotations. These have no other defined use; they are
11374 ignored by code generation and optimization.
11376 '``llvm.ptr.annotation.*``' Intrinsic
11377 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11382 This is an overloaded intrinsic. You can use '``llvm.ptr.annotation``' on a
11383 pointer to an integer of any width. *NOTE* you must specify an address space for
11384 the pointer. The identifier for the default address space is the integer
11389 declare i8* @llvm.ptr.annotation.p<address space>i8(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
11390 declare i16* @llvm.ptr.annotation.p<address space>i16(i16* <val>, i8* <str>, i8* <str>, i32 <int>)
11391 declare i32* @llvm.ptr.annotation.p<address space>i32(i32* <val>, i8* <str>, i8* <str>, i32 <int>)
11392 declare i64* @llvm.ptr.annotation.p<address space>i64(i64* <val>, i8* <str>, i8* <str>, i32 <int>)
11393 declare i256* @llvm.ptr.annotation.p<address space>i256(i256* <val>, i8* <str>, i8* <str>, i32 <int>)
11398 The '``llvm.ptr.annotation``' intrinsic.
11403 The first argument is a pointer to an integer value of arbitrary bitwidth
11404 (result of some expression), the second is a pointer to a global string, the
11405 third is a pointer to a global string which is the source file name, and the
11406 last argument is the line number. It returns the value of the first argument.
11411 This intrinsic allows annotation of a pointer to an integer with arbitrary
11412 strings. This can be useful for special purpose optimizations that want to look
11413 for these annotations. These have no other defined use; they are ignored by code
11414 generation and optimization.
11416 '``llvm.annotation.*``' Intrinsic
11417 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11422 This is an overloaded intrinsic. You can use '``llvm.annotation``' on
11423 any integer bit width.
11427 declare i8 @llvm.annotation.i8(i8 <val>, i8* <str>, i8* <str>, i32 <int>)
11428 declare i16 @llvm.annotation.i16(i16 <val>, i8* <str>, i8* <str>, i32 <int>)
11429 declare i32 @llvm.annotation.i32(i32 <val>, i8* <str>, i8* <str>, i32 <int>)
11430 declare i64 @llvm.annotation.i64(i64 <val>, i8* <str>, i8* <str>, i32 <int>)
11431 declare i256 @llvm.annotation.i256(i256 <val>, i8* <str>, i8* <str>, i32 <int>)
11436 The '``llvm.annotation``' intrinsic.
11441 The first argument is an integer value (result of some expression), the
11442 second is a pointer to a global string, the third is a pointer to a
11443 global string which is the source file name, and the last argument is
11444 the line number. It returns the value of the first argument.
11449 This intrinsic allows annotations to be put on arbitrary expressions
11450 with arbitrary strings. This can be useful for special purpose
11451 optimizations that want to look for these annotations. These have no
11452 other defined use; they are ignored by code generation and optimization.
11454 '``llvm.trap``' Intrinsic
11455 ^^^^^^^^^^^^^^^^^^^^^^^^^
11462 declare void @llvm.trap() noreturn nounwind
11467 The '``llvm.trap``' intrinsic.
11477 This intrinsic is lowered to the target dependent trap instruction. If
11478 the target does not have a trap instruction, this intrinsic will be
11479 lowered to a call of the ``abort()`` function.
11481 '``llvm.debugtrap``' Intrinsic
11482 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11489 declare void @llvm.debugtrap() nounwind
11494 The '``llvm.debugtrap``' intrinsic.
11504 This intrinsic is lowered to code which is intended to cause an
11505 execution trap with the intention of requesting the attention of a
11508 '``llvm.stackprotector``' Intrinsic
11509 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11516 declare void @llvm.stackprotector(i8* <guard>, i8** <slot>)
11521 The ``llvm.stackprotector`` intrinsic takes the ``guard`` and stores it
11522 onto the stack at ``slot``. The stack slot is adjusted to ensure that it
11523 is placed on the stack before local variables.
11528 The ``llvm.stackprotector`` intrinsic requires two pointer arguments.
11529 The first argument is the value loaded from the stack guard
11530 ``@__stack_chk_guard``. The second variable is an ``alloca`` that has
11531 enough space to hold the value of the guard.
11536 This intrinsic causes the prologue/epilogue inserter to force the position of
11537 the ``AllocaInst`` stack slot to be before local variables on the stack. This is
11538 to ensure that if a local variable on the stack is overwritten, it will destroy
11539 the value of the guard. When the function exits, the guard on the stack is
11540 checked against the original guard by ``llvm.stackprotectorcheck``. If they are
11541 different, then ``llvm.stackprotectorcheck`` causes the program to abort by
11542 calling the ``__stack_chk_fail()`` function.
11544 '``llvm.stackprotectorcheck``' Intrinsic
11545 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11552 declare void @llvm.stackprotectorcheck(i8** <guard>)
11557 The ``llvm.stackprotectorcheck`` intrinsic compares ``guard`` against an already
11558 created stack protector and if they are not equal calls the
11559 ``__stack_chk_fail()`` function.
11564 The ``llvm.stackprotectorcheck`` intrinsic requires one pointer argument, the
11565 the variable ``@__stack_chk_guard``.
11570 This intrinsic is provided to perform the stack protector check by comparing
11571 ``guard`` with the stack slot created by ``llvm.stackprotector`` and if the
11572 values do not match call the ``__stack_chk_fail()`` function.
11574 The reason to provide this as an IR level intrinsic instead of implementing it
11575 via other IR operations is that in order to perform this operation at the IR
11576 level without an intrinsic, one would need to create additional basic blocks to
11577 handle the success/failure cases. This makes it difficult to stop the stack
11578 protector check from disrupting sibling tail calls in Codegen. With this
11579 intrinsic, we are able to generate the stack protector basic blocks late in
11580 codegen after the tail call decision has occurred.
11582 '``llvm.objectsize``' Intrinsic
11583 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11590 declare i32 @llvm.objectsize.i32(i8* <object>, i1 <min>)
11591 declare i64 @llvm.objectsize.i64(i8* <object>, i1 <min>)
11596 The ``llvm.objectsize`` intrinsic is designed to provide information to
11597 the optimizers to determine at compile time whether a) an operation
11598 (like memcpy) will overflow a buffer that corresponds to an object, or
11599 b) that a runtime check for overflow isn't necessary. An object in this
11600 context means an allocation of a specific class, structure, array, or
11606 The ``llvm.objectsize`` intrinsic takes two arguments. The first
11607 argument is a pointer to or into the ``object``. The second argument is
11608 a boolean and determines whether ``llvm.objectsize`` returns 0 (if true)
11609 or -1 (if false) when the object size is unknown. The second argument
11610 only accepts constants.
11615 The ``llvm.objectsize`` intrinsic is lowered to a constant representing
11616 the size of the object concerned. If the size cannot be determined at
11617 compile time, ``llvm.objectsize`` returns ``i32/i64 -1 or 0`` (depending
11618 on the ``min`` argument).
11620 '``llvm.expect``' Intrinsic
11621 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
11626 This is an overloaded intrinsic. You can use ``llvm.expect`` on any
11631 declare i1 @llvm.expect.i1(i1 <val>, i1 <expected_val>)
11632 declare i32 @llvm.expect.i32(i32 <val>, i32 <expected_val>)
11633 declare i64 @llvm.expect.i64(i64 <val>, i64 <expected_val>)
11638 The ``llvm.expect`` intrinsic provides information about expected (the
11639 most probable) value of ``val``, which can be used by optimizers.
11644 The ``llvm.expect`` intrinsic takes two arguments. The first argument is
11645 a value. The second argument is an expected value, this needs to be a
11646 constant value, variables are not allowed.
11651 This intrinsic is lowered to the ``val``.
11655 '``llvm.assume``' Intrinsic
11656 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11663 declare void @llvm.assume(i1 %cond)
11668 The ``llvm.assume`` allows the optimizer to assume that the provided
11669 condition is true. This information can then be used in simplifying other parts
11675 The condition which the optimizer may assume is always true.
11680 The intrinsic allows the optimizer to assume that the provided condition is
11681 always true whenever the control flow reaches the intrinsic call. No code is
11682 generated for this intrinsic, and instructions that contribute only to the
11683 provided condition are not used for code generation. If the condition is
11684 violated during execution, the behavior is undefined.
11686 Note that the optimizer might limit the transformations performed on values
11687 used by the ``llvm.assume`` intrinsic in order to preserve the instructions
11688 only used to form the intrinsic's input argument. This might prove undesirable
11689 if the extra information provided by the ``llvm.assume`` intrinsic does not cause
11690 sufficient overall improvement in code quality. For this reason,
11691 ``llvm.assume`` should not be used to document basic mathematical invariants
11692 that the optimizer can otherwise deduce or facts that are of little use to the
11697 '``llvm.bitset.test``' Intrinsic
11698 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11705 declare i1 @llvm.bitset.test(i8* %ptr, metadata %bitset) nounwind readnone
11711 The first argument is a pointer to be tested. The second argument is a
11712 metadata string containing the name of a :doc:`bitset <BitSets>`.
11717 The ``llvm.bitset.test`` intrinsic tests whether the given pointer is a
11718 member of the given bitset.
11720 '``llvm.donothing``' Intrinsic
11721 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11728 declare void @llvm.donothing() nounwind readnone
11733 The ``llvm.donothing`` intrinsic doesn't perform any operation. It's one of only
11734 two intrinsics (besides ``llvm.experimental.patchpoint``) that can be called
11735 with an invoke instruction.
11745 This intrinsic does nothing, and it's removed by optimizers and ignored
11748 Stack Map Intrinsics
11749 --------------------
11751 LLVM provides experimental intrinsics to support runtime patching
11752 mechanisms commonly desired in dynamic language JITs. These intrinsics
11753 are described in :doc:`StackMaps`.