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 which 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.
83 #. Unnamed values are represented as an unsigned numeric value with
84 their prefix. For example, ``%12``, ``@2``, ``%44``.
85 #. Constants, which are described in the section Constants_ below.
87 LLVM requires that values start with a prefix for two reasons: Compilers
88 don't need to worry about name clashes with reserved words, and the set
89 of reserved words may be expanded in the future without penalty.
90 Additionally, unnamed identifiers allow a compiler to quickly come up
91 with a temporary variable without having to avoid symbol table
94 Reserved words in LLVM are very similar to reserved words in other
95 languages. There are keywords for different opcodes ('``add``',
96 '``bitcast``', '``ret``', etc...), for primitive type names ('``void``',
97 '``i32``', etc...), and others. These reserved words cannot conflict
98 with variable names, because none of them start with a prefix character
101 Here is an example of LLVM code to multiply the integer variable
108 %result = mul i32 %X, 8
110 After strength reduction:
114 %result = shl i32 %X, 3
120 %0 = add i32 %X, %X ; yields {i32}:%0
121 %1 = add i32 %0, %0 ; yields {i32}:%1
122 %result = add i32 %1, %1
124 This last way of multiplying ``%X`` by 8 illustrates several important
125 lexical features of LLVM:
127 #. Comments are delimited with a '``;``' and go until the end of line.
128 #. Unnamed temporaries are created when the result of a computation is
129 not assigned to a named value.
130 #. Unnamed temporaries are numbered sequentially (using a per-function
131 incrementing counter, starting with 0). Note that basic blocks are
132 included in this numbering. For example, if the entry basic block is not
133 given a label name, then it will get number 0.
135 It also shows a convention that we follow in this document. When
136 demonstrating instructions, we will follow an instruction with a comment
137 that defines the type and name of value produced.
145 LLVM programs are composed of ``Module``'s, each of which is a
146 translation unit of the input programs. Each module consists of
147 functions, global variables, and symbol table entries. Modules may be
148 combined together with the LLVM linker, which merges function (and
149 global variable) definitions, resolves forward declarations, and merges
150 symbol table entries. Here is an example of the "hello world" module:
154 ; Declare the string constant as a global constant.
155 @.str = private unnamed_addr constant [13 x i8] c"hello world\0A\00"
157 ; External declaration of the puts function
158 declare i32 @puts(i8* nocapture) nounwind
160 ; Definition of main function
161 define i32 @main() { ; i32()*
162 ; Convert [13 x i8]* to i8 *...
163 %cast210 = getelementptr [13 x i8]* @.str, i64 0, i64 0
165 ; Call puts function to write out the string to stdout.
166 call i32 @puts(i8* %cast210)
171 !1 = metadata !{i32 42}
174 This example is made up of a :ref:`global variable <globalvars>` named
175 "``.str``", an external declaration of the "``puts``" function, a
176 :ref:`function definition <functionstructure>` for "``main``" and
177 :ref:`named metadata <namedmetadatastructure>` "``foo``".
179 In general, a module is made up of a list of global values (where both
180 functions and global variables are global values). Global values are
181 represented by a pointer to a memory location (in this case, a pointer
182 to an array of char, and a pointer to a function), and have one of the
183 following :ref:`linkage types <linkage>`.
190 All Global Variables and Functions have one of the following types of
194 Global values with "``private``" linkage are only directly
195 accessible by objects in the current module. In particular, linking
196 code into a module with an private global value may cause the
197 private to be renamed as necessary to avoid collisions. Because the
198 symbol is private to the module, all references can be updated. This
199 doesn't show up in any symbol table in the object file.
201 Similar to private, but the value shows as a local symbol
202 (``STB_LOCAL`` in the case of ELF) in the object file. This
203 corresponds to the notion of the '``static``' keyword in C.
204 ``available_externally``
205 Globals with "``available_externally``" linkage are never emitted
206 into the object file corresponding to the LLVM module. They exist to
207 allow inlining and other optimizations to take place given knowledge
208 of the definition of the global, which is known to be somewhere
209 outside the module. Globals with ``available_externally`` linkage
210 are allowed to be discarded at will, and are otherwise the same as
211 ``linkonce_odr``. This linkage type is only allowed on definitions,
214 Globals with "``linkonce``" linkage are merged with other globals of
215 the same name when linkage occurs. This can be used to implement
216 some forms of inline functions, templates, or other code which must
217 be generated in each translation unit that uses it, but where the
218 body may be overridden with a more definitive definition later.
219 Unreferenced ``linkonce`` globals are allowed to be discarded. Note
220 that ``linkonce`` linkage does not actually allow the optimizer to
221 inline the body of this function into callers because it doesn't
222 know if this definition of the function is the definitive definition
223 within the program or whether it will be overridden by a stronger
224 definition. To enable inlining and other optimizations, use
225 "``linkonce_odr``" linkage.
227 "``weak``" linkage has the same merging semantics as ``linkonce``
228 linkage, except that unreferenced globals with ``weak`` linkage may
229 not be discarded. This is used for globals that are declared "weak"
232 "``common``" linkage is most similar to "``weak``" linkage, but they
233 are used for tentative definitions in C, such as "``int X;``" at
234 global scope. Symbols with "``common``" linkage are merged in the
235 same way as ``weak symbols``, and they may not be deleted if
236 unreferenced. ``common`` symbols may not have an explicit section,
237 must have a zero initializer, and may not be marked
238 ':ref:`constant <globalvars>`'. Functions and aliases may not have
241 .. _linkage_appending:
244 "``appending``" linkage may only be applied to global variables of
245 pointer to array type. When two global variables with appending
246 linkage are linked together, the two global arrays are appended
247 together. This is the LLVM, typesafe, equivalent of having the
248 system linker append together "sections" with identical names when
251 The semantics of this linkage follow the ELF object file model: the
252 symbol is weak until linked, if not linked, the symbol becomes null
253 instead of being an undefined reference.
254 ``linkonce_odr``, ``weak_odr``
255 Some languages allow differing globals to be merged, such as two
256 functions with different semantics. Other languages, such as
257 ``C++``, ensure that only equivalent globals are ever merged (the
258 "one definition rule" --- "ODR"). Such languages can use the
259 ``linkonce_odr`` and ``weak_odr`` linkage types to indicate that the
260 global will only be merged with equivalent globals. These linkage
261 types are otherwise the same as their non-``odr`` versions.
263 If none of the above identifiers are used, the global is externally
264 visible, meaning that it participates in linkage and can be used to
265 resolve external symbol references.
267 It is illegal for a function *declaration* to have any linkage type
268 other than ``external`` or ``extern_weak``.
275 LLVM :ref:`functions <functionstructure>`, :ref:`calls <i_call>` and
276 :ref:`invokes <i_invoke>` can all have an optional calling convention
277 specified for the call. The calling convention of any pair of dynamic
278 caller/callee must match, or the behavior of the program is undefined.
279 The following calling conventions are supported by LLVM, and more may be
282 "``ccc``" - The C calling convention
283 This calling convention (the default if no other calling convention
284 is specified) matches the target C calling conventions. This calling
285 convention supports varargs function calls and tolerates some
286 mismatch in the declared prototype and implemented declaration of
287 the function (as does normal C).
288 "``fastcc``" - The fast calling convention
289 This calling convention attempts to make calls as fast as possible
290 (e.g. by passing things in registers). This calling convention
291 allows the target to use whatever tricks it wants to produce fast
292 code for the target, without having to conform to an externally
293 specified ABI (Application Binary Interface). `Tail calls can only
294 be optimized when this, the GHC or the HiPE convention is
295 used. <CodeGenerator.html#id80>`_ This calling convention does not
296 support varargs and requires the prototype of all callees to exactly
297 match the prototype of the function definition.
298 "``coldcc``" - The cold calling convention
299 This calling convention attempts to make code in the caller as
300 efficient as possible under the assumption that the call is not
301 commonly executed. As such, these calls often preserve all registers
302 so that the call does not break any live ranges in the caller side.
303 This calling convention does not support varargs and requires the
304 prototype of all callees to exactly match the prototype of the
305 function definition. Furthermore the inliner doesn't consider such function
307 "``cc 10``" - GHC convention
308 This calling convention has been implemented specifically for use by
309 the `Glasgow Haskell Compiler (GHC) <http://www.haskell.org/ghc>`_.
310 It passes everything in registers, going to extremes to achieve this
311 by disabling callee save registers. This calling convention should
312 not be used lightly but only for specific situations such as an
313 alternative to the *register pinning* performance technique often
314 used when implementing functional programming languages. At the
315 moment only X86 supports this convention and it has the following
318 - On *X86-32* only supports up to 4 bit type parameters. No
319 floating point types are supported.
320 - On *X86-64* only supports up to 10 bit type parameters and 6
321 floating point parameters.
323 This calling convention supports `tail call
324 optimization <CodeGenerator.html#id80>`_ but requires both the
325 caller and callee are using it.
326 "``cc 11``" - The HiPE calling convention
327 This calling convention has been implemented specifically for use by
328 the `High-Performance Erlang
329 (HiPE) <http://www.it.uu.se/research/group/hipe/>`_ compiler, *the*
330 native code compiler of the `Ericsson's Open Source Erlang/OTP
331 system <http://www.erlang.org/download.shtml>`_. It uses more
332 registers for argument passing than the ordinary C calling
333 convention and defines no callee-saved registers. The calling
334 convention properly supports `tail call
335 optimization <CodeGenerator.html#id80>`_ but requires that both the
336 caller and the callee use it. It uses a *register pinning*
337 mechanism, similar to GHC's convention, for keeping frequently
338 accessed runtime components pinned to specific hardware registers.
339 At the moment only X86 supports this convention (both 32 and 64
341 "``webkit_jscc``" - WebKit's JavaScript calling convention
342 This calling convention has been implemented for `WebKit FTL JIT
343 <https://trac.webkit.org/wiki/FTLJIT>`_. It passes arguments on the
344 stack right to left (as cdecl does), and returns a value in the
345 platform's customary return register.
346 "``anyregcc``" - Dynamic calling convention for code patching
347 This is a special convention that supports patching an arbitrary code
348 sequence in place of a call site. This convention forces the call
349 arguments into registers but allows them to be dynamcially
350 allocated. This can currently only be used with calls to
351 llvm.experimental.patchpoint because only this intrinsic records
352 the location of its arguments in a side table. See :doc:`StackMaps`.
353 "``preserve_mostcc``" - The `PreserveMost` calling convention
354 This calling convention attempts to make the code in the caller as little
355 intrusive as possible. This calling convention behaves identical to the `C`
356 calling convention on how arguments and return values are passed, but it
357 uses a different set of caller/callee-saved registers. This alleviates the
358 burden of saving and recovering a large register set before and after the
359 call in the caller. If the arguments are passed in callee-saved registers,
360 then they will be preserved by the callee across the call. This doesn't
361 apply for values returned in callee-saved registers.
363 - On X86-64 the callee preserves all general purpose registers, except for
364 R11. R11 can be used as a scratch register. Floating-point registers
365 (XMMs/YMMs) are not preserved and need to be saved by the caller.
367 The idea behind this convention is to support calls to runtime functions
368 that have a hot path and a cold path. The hot path is usually a small piece
369 of code that doesn't many registers. The cold path might need to call out to
370 another function and therefore only needs to preserve the caller-saved
371 registers, which haven't already been saved by the caller. The
372 `PreserveMost` calling convention is very similar to the `cold` calling
373 convention in terms of caller/callee-saved registers, but they are used for
374 different types of function calls. `coldcc` is for function calls that are
375 rarely executed, whereas `preserve_mostcc` function calls are intended to be
376 on the hot path and definitely executed a lot. Furthermore `preserve_mostcc`
377 doesn't prevent the inliner from inlining the function call.
379 This calling convention will be used by a future version of the ObjectiveC
380 runtime and should therefore still be considered experimental at this time.
381 Although this convention was created to optimize certain runtime calls to
382 the ObjectiveC runtime, it is not limited to this runtime and might be used
383 by other runtimes in the future too. The current implementation only
384 supports X86-64, but the intention is to support more architectures in the
386 "``preserve_allcc``" - The `PreserveAll` calling convention
387 This calling convention attempts to make the code in the caller even less
388 intrusive than the `PreserveMost` calling convention. This calling
389 convention also behaves identical to the `C` calling convention on how
390 arguments and return values are passed, but it uses a different set of
391 caller/callee-saved registers. This removes the burden of saving and
392 recovering a large register set before and after the call in the caller. If
393 the arguments are passed in callee-saved registers, then they will be
394 preserved by the callee across the call. This doesn't apply for values
395 returned in callee-saved registers.
397 - On X86-64 the callee preserves all general purpose registers, except for
398 R11. R11 can be used as a scratch register. Furthermore it also preserves
399 all floating-point registers (XMMs/YMMs).
401 The idea behind this convention is to support calls to runtime functions
402 that don't need to call out to any other functions.
404 This calling convention, like the `PreserveMost` calling convention, will be
405 used by a future version of the ObjectiveC runtime and should be considered
406 experimental at this time.
407 "``cc <n>``" - Numbered convention
408 Any calling convention may be specified by number, allowing
409 target-specific calling conventions to be used. Target specific
410 calling conventions start at 64.
412 More calling conventions can be added/defined on an as-needed basis, to
413 support Pascal conventions or any other well-known target-independent
416 .. _visibilitystyles:
421 All Global Variables and Functions have one of the following visibility
424 "``default``" - Default style
425 On targets that use the ELF object file format, default visibility
426 means that the declaration is visible to other modules and, in
427 shared libraries, means that the declared entity may be overridden.
428 On Darwin, default visibility means that the declaration is visible
429 to other modules. Default visibility corresponds to "external
430 linkage" in the language.
431 "``hidden``" - Hidden style
432 Two declarations of an object with hidden visibility refer to the
433 same object if they are in the same shared object. Usually, hidden
434 visibility indicates that the symbol will not be placed into the
435 dynamic symbol table, so no other module (executable or shared
436 library) can reference it directly.
437 "``protected``" - Protected style
438 On ELF, protected visibility indicates that the symbol will be
439 placed in the dynamic symbol table, but that references within the
440 defining module will bind to the local symbol. That is, the symbol
441 cannot be overridden by another module.
448 All Global Variables, Functions and Aliases can have one of the following
452 "``dllimport``" causes the compiler to reference a function or variable via
453 a global pointer to a pointer that is set up by the DLL exporting the
454 symbol. On Microsoft Windows targets, the pointer name is formed by
455 combining ``__imp_`` and the function or variable name.
457 "``dllexport``" causes the compiler to provide a global pointer to a pointer
458 in a DLL, so that it can be referenced with the ``dllimport`` attribute. On
459 Microsoft Windows targets, the pointer name is formed by combining
460 ``__imp_`` and the function or variable name. Since this storage class
461 exists for defining a dll interface, the compiler, assembler and linker know
462 it is externally referenced and must refrain from deleting the symbol.
467 LLVM IR allows you to specify both "identified" and "literal" :ref:`structure
468 types <t_struct>`. Literal types are uniqued structurally, but identified types
469 are never uniqued. An :ref:`opaque structural type <t_opaque>` can also be used
470 to forward declare a type which is not yet available.
472 An example of a identified structure specification is:
476 %mytype = type { %mytype*, i32 }
478 Prior to the LLVM 3.0 release, identified types were structurally uniqued. Only
479 literal types are uniqued in recent versions of LLVM.
486 Global variables define regions of memory allocated at compilation time
489 Global variables definitions must be initialized, may have an explicit section
490 to be placed in, and may have an optional explicit alignment specified.
492 Global variables in other translation units can also be declared, in which
493 case they don't have an initializer.
495 A variable may be defined as ``thread_local``, which means that it will
496 not be shared by threads (each thread will have a separated copy of the
497 variable). Not all targets support thread-local variables. Optionally, a
498 TLS model may be specified:
501 For variables that are only used within the current shared library.
503 For variables in modules that will not be loaded dynamically.
505 For variables defined in the executable and only used within it.
507 The models correspond to the ELF TLS models; see `ELF Handling For
508 Thread-Local Storage <http://people.redhat.com/drepper/tls.pdf>`_ for
509 more information on under which circumstances the different models may
510 be used. The target may choose a different TLS model if the specified
511 model is not supported, or if a better choice of model can be made.
513 A variable may be defined as a global ``constant``, which indicates that
514 the contents of the variable will **never** be modified (enabling better
515 optimization, allowing the global data to be placed in the read-only
516 section of an executable, etc). Note that variables that need runtime
517 initialization cannot be marked ``constant`` as there is a store to the
520 LLVM explicitly allows *declarations* of global variables to be marked
521 constant, even if the final definition of the global is not. This
522 capability can be used to enable slightly better optimization of the
523 program, but requires the language definition to guarantee that
524 optimizations based on the 'constantness' are valid for the translation
525 units that do not include the definition.
527 As SSA values, global variables define pointer values that are in scope
528 (i.e. they dominate) all basic blocks in the program. Global variables
529 always define a pointer to their "content" type because they describe a
530 region of memory, and all memory objects in LLVM are accessed through
533 Global variables can be marked with ``unnamed_addr`` which indicates
534 that the address is not significant, only the content. Constants marked
535 like this can be merged with other constants if they have the same
536 initializer. Note that a constant with significant address *can* be
537 merged with a ``unnamed_addr`` constant, the result being a constant
538 whose address is significant.
540 A global variable may be declared to reside in a target-specific
541 numbered address space. For targets that support them, address spaces
542 may affect how optimizations are performed and/or what target
543 instructions are used to access the variable. The default address space
544 is zero. The address space qualifier must precede any other attributes.
546 LLVM allows an explicit section to be specified for globals. If the
547 target supports it, it will emit globals to the section specified.
549 By default, global initializers are optimized by assuming that global
550 variables defined within the module are not modified from their
551 initial values before the start of the global initializer. This is
552 true even for variables potentially accessible from outside the
553 module, including those with external linkage or appearing in
554 ``@llvm.used`` or dllexported variables. This assumption may be suppressed
555 by marking the variable with ``externally_initialized``.
557 An explicit alignment may be specified for a global, which must be a
558 power of 2. If not present, or if the alignment is set to zero, the
559 alignment of the global is set by the target to whatever it feels
560 convenient. If an explicit alignment is specified, the global is forced
561 to have exactly that alignment. Targets and optimizers are not allowed
562 to over-align the global if the global has an assigned section. In this
563 case, the extra alignment could be observable: for example, code could
564 assume that the globals are densely packed in their section and try to
565 iterate over them as an array, alignment padding would break this
568 Globals can also have a :ref:`DLL storage class <dllstorageclass>`.
572 [@<GlobalVarName> =] [Linkage] [Visibility] [DLLStorageClass] [ThreadLocal]
573 [AddrSpace] [unnamed_addr] [ExternallyInitialized]
574 <global | constant> <Type>
575 [, section "name"] [, align <Alignment>]
577 For example, the following defines a global in a numbered address space
578 with an initializer, section, and alignment:
582 @G = addrspace(5) constant float 1.0, section "foo", align 4
584 The following example just declares a global variable
588 @G = external global i32
590 The following example defines a thread-local global with the
591 ``initialexec`` TLS model:
595 @G = thread_local(initialexec) global i32 0, align 4
597 .. _functionstructure:
602 LLVM function definitions consist of the "``define``" keyword, an
603 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
604 style <visibility>`, an optional :ref:`DLL storage class <dllstorageclass>`,
605 an optional :ref:`calling convention <callingconv>`,
606 an optional ``unnamed_addr`` attribute, a return type, an optional
607 :ref:`parameter attribute <paramattrs>` for the return type, a function
608 name, a (possibly empty) argument list (each with optional :ref:`parameter
609 attributes <paramattrs>`), optional :ref:`function attributes <fnattrs>`,
610 an optional section, an optional alignment, an optional :ref:`garbage
611 collector name <gc>`, an optional :ref:`prefix <prefixdata>`, an opening
612 curly brace, a list of basic blocks, and a closing curly brace.
614 LLVM function declarations consist of the "``declare``" keyword, an
615 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
616 style <visibility>`, an optional :ref:`DLL storage class <dllstorageclass>`,
617 an optional :ref:`calling convention <callingconv>`,
618 an optional ``unnamed_addr`` attribute, a return type, an optional
619 :ref:`parameter attribute <paramattrs>` for the return type, a function
620 name, a possibly empty list of arguments, an optional alignment, an optional
621 :ref:`garbage collector name <gc>` and an optional :ref:`prefix <prefixdata>`.
623 A function definition contains a list of basic blocks, forming the CFG (Control
624 Flow Graph) for the function. Each basic block may optionally start with a label
625 (giving the basic block a symbol table entry), contains a list of instructions,
626 and ends with a :ref:`terminator <terminators>` instruction (such as a branch or
627 function return). If an explicit label is not provided, a block is assigned an
628 implicit numbered label, using the next value from the same counter as used for
629 unnamed temporaries (:ref:`see above<identifiers>`). For example, if a function
630 entry block does not have an explicit label, it will be assigned label "%0",
631 then the first unnamed temporary in that block will be "%1", etc.
633 The first basic block in a function is special in two ways: it is
634 immediately executed on entrance to the function, and it is not allowed
635 to have predecessor basic blocks (i.e. there can not be any branches to
636 the entry block of a function). Because the block can have no
637 predecessors, it also cannot have any :ref:`PHI nodes <i_phi>`.
639 LLVM allows an explicit section to be specified for functions. If the
640 target supports it, it will emit functions to the section specified.
642 An explicit alignment may be specified for a function. If not present,
643 or if the alignment is set to zero, the alignment of the function is set
644 by the target to whatever it feels convenient. If an explicit alignment
645 is specified, the function is forced to have at least that much
646 alignment. All alignments must be a power of 2.
648 If the ``unnamed_addr`` attribute is given, the address is know to not
649 be significant and two identical functions can be merged.
653 define [linkage] [visibility] [DLLStorageClass]
655 <ResultType> @<FunctionName> ([argument list])
656 [unnamed_addr] [fn Attrs] [section "name"] [align N]
657 [gc] [prefix Constant] { ... }
664 Aliases act as "second name" for the aliasee value (which can be either
665 function, global variable, another alias or bitcast of global value).
666 Aliases may have an optional :ref:`linkage type <linkage>`, an optional
667 :ref:`visibility style <visibility>`, and an optional :ref:`DLL storage class
672 @<Name> = [Visibility] [DLLStorageClass] alias [Linkage] <AliaseeTy> @<Aliasee>
674 The linkage must be one of ``private``, ``internal``, ``linkonce``, ``weak``,
675 ``linkonce_odr``, ``weak_odr``, ``external``. Note that some system linkers
676 might not correctly handle dropping a weak symbol that is aliased by a non-weak
679 Alias that are not ``unnamed_addr`` are guaranteed to have the same address as
682 The aliasee must be a definition.
684 Aliases are not allowed to point to aliases with linkages that can be
685 overridden. Since they are only a second name, the possibility of the
686 intermediate alias being overridden cannot be represented in an object file.
688 .. _namedmetadatastructure:
693 Named metadata is a collection of metadata. :ref:`Metadata
694 nodes <metadata>` (but not metadata strings) are the only valid
695 operands for a named metadata.
699 ; Some unnamed metadata nodes, which are referenced by the named metadata.
700 !0 = metadata !{metadata !"zero"}
701 !1 = metadata !{metadata !"one"}
702 !2 = metadata !{metadata !"two"}
704 !name = !{!0, !1, !2}
711 The return type and each parameter of a function type may have a set of
712 *parameter attributes* associated with them. Parameter attributes are
713 used to communicate additional information about the result or
714 parameters of a function. Parameter attributes are considered to be part
715 of the function, not of the function type, so functions with different
716 parameter attributes can have the same function type.
718 Parameter attributes are simple keywords that follow the type specified.
719 If multiple parameter attributes are needed, they are space separated.
724 declare i32 @printf(i8* noalias nocapture, ...)
725 declare i32 @atoi(i8 zeroext)
726 declare signext i8 @returns_signed_char()
728 Note that any attributes for the function result (``nounwind``,
729 ``readonly``) come immediately after the argument list.
731 Currently, only the following parameter attributes are defined:
734 This indicates to the code generator that the parameter or return
735 value should be zero-extended to the extent required by the target's
736 ABI (which is usually 32-bits, but is 8-bits for a i1 on x86-64) by
737 the caller (for a parameter) or the callee (for a return value).
739 This indicates to the code generator that the parameter or return
740 value should be sign-extended to the extent required by the target's
741 ABI (which is usually 32-bits) by the caller (for a parameter) or
742 the callee (for a return value).
744 This indicates that this parameter or return value should be treated
745 in a special target-dependent fashion during while emitting code for
746 a function call or return (usually, by putting it in a register as
747 opposed to memory, though some targets use it to distinguish between
748 two different kinds of registers). Use of this attribute is
751 This indicates that the pointer parameter should really be passed by
752 value to the function. The attribute implies that a hidden copy of
753 the pointee is made between the caller and the callee, so the callee
754 is unable to modify the value in the caller. This attribute is only
755 valid on LLVM pointer arguments. It is generally used to pass
756 structs and arrays by value, but is also valid on pointers to
757 scalars. The copy is considered to belong to the caller not the
758 callee (for example, ``readonly`` functions should not write to
759 ``byval`` parameters). This is not a valid attribute for return
762 The byval attribute also supports specifying an alignment with the
763 align attribute. It indicates the alignment of the stack slot to
764 form and the known alignment of the pointer specified to the call
765 site. If the alignment is not specified, then the code generator
766 makes a target-specific assumption.
772 The ``inalloca`` argument attribute allows the caller to take the
773 address of outgoing stack arguments. An ``inalloca`` argument must
774 be a pointer to stack memory produced by an ``alloca`` instruction.
775 The alloca, or argument allocation, must also be tagged with the
776 inalloca keyword. Only the past argument may have the ``inalloca``
777 attribute, and that argument is guaranteed to be passed in memory.
779 An argument allocation may be used by a call at most once because
780 the call may deallocate it. The ``inalloca`` attribute cannot be
781 used in conjunction with other attributes that affect argument
782 storage, like ``inreg``, ``nest``, ``sret``, or ``byval``. The
783 ``inalloca`` attribute also disables LLVM's implicit lowering of
784 large aggregate return values, which means that frontend authors
785 must lower them with ``sret`` pointers.
787 When the call site is reached, the argument allocation must have
788 been the most recent stack allocation that is still live, or the
789 results are undefined. It is possible to allocate additional stack
790 space after an argument allocation and before its call site, but it
791 must be cleared off with :ref:`llvm.stackrestore
794 See :doc:`InAlloca` for more information on how to use this
798 This indicates that the pointer parameter specifies the address of a
799 structure that is the return value of the function in the source
800 program. This pointer must be guaranteed by the caller to be valid:
801 loads and stores to the structure may be assumed by the callee
802 not to trap and to be properly aligned. This may only be applied to
803 the first parameter. This is not a valid attribute for return
809 This indicates that pointer values :ref:`based <pointeraliasing>` on
810 the argument or return value do not alias pointer values which are
811 not *based* on it, ignoring certain "irrelevant" dependencies. For a
812 call to the parent function, dependencies between memory references
813 from before or after the call and from those during the call are
814 "irrelevant" to the ``noalias`` keyword for the arguments and return
815 value used in that call. The caller shares the responsibility with
816 the callee for ensuring that these requirements are met. For further
817 details, please see the discussion of the NoAlias response in :ref:`alias
818 analysis <Must, May, or No>`.
820 Note that this definition of ``noalias`` is intentionally similar
821 to the definition of ``restrict`` in C99 for function arguments,
822 though it is slightly weaker.
824 For function return values, C99's ``restrict`` is not meaningful,
825 while LLVM's ``noalias`` is.
827 This indicates that the callee does not make any copies of the
828 pointer that outlive the callee itself. This is not a valid
829 attribute for return values.
834 This indicates that the pointer parameter can be excised using the
835 :ref:`trampoline intrinsics <int_trampoline>`. This is not a valid
836 attribute for return values and can only be applied to one parameter.
839 This indicates that the function always returns the argument as its return
840 value. This is an optimization hint to the code generator when generating
841 the caller, allowing tail call optimization and omission of register saves
842 and restores in some cases; it is not checked or enforced when generating
843 the callee. The parameter and the function return type must be valid
844 operands for the :ref:`bitcast instruction <i_bitcast>`. This is not a
845 valid attribute for return values and can only be applied to one parameter.
849 Garbage Collector Names
850 -----------------------
852 Each function may specify a garbage collector name, which is simply a
857 define void @f() gc "name" { ... }
859 The compiler declares the supported values of *name*. Specifying a
860 collector which will cause the compiler to alter its output in order to
861 support the named garbage collection algorithm.
868 Prefix data is data associated with a function which the code generator
869 will emit immediately before the function body. The purpose of this feature
870 is to allow frontends to associate language-specific runtime metadata with
871 specific functions and make it available through the function pointer while
872 still allowing the function pointer to be called. To access the data for a
873 given function, a program may bitcast the function pointer to a pointer to
874 the constant's type. This implies that the IR symbol points to the start
877 To maintain the semantics of ordinary function calls, the prefix data must
878 have a particular format. Specifically, it must begin with a sequence of
879 bytes which decode to a sequence of machine instructions, valid for the
880 module's target, which transfer control to the point immediately succeeding
881 the prefix data, without performing any other visible action. This allows
882 the inliner and other passes to reason about the semantics of the function
883 definition without needing to reason about the prefix data. Obviously this
884 makes the format of the prefix data highly target dependent.
886 Prefix data is laid out as if it were an initializer for a global variable
887 of the prefix data's type. No padding is automatically placed between the
888 prefix data and the function body. If padding is required, it must be part
891 A trivial example of valid prefix data for the x86 architecture is ``i8 144``,
892 which encodes the ``nop`` instruction:
896 define void @f() prefix i8 144 { ... }
898 Generally prefix data can be formed by encoding a relative branch instruction
899 which skips the metadata, as in this example of valid prefix data for the
900 x86_64 architecture, where the first two bytes encode ``jmp .+10``:
904 %0 = type <{ i8, i8, i8* }>
906 define void @f() prefix %0 <{ i8 235, i8 8, i8* @md}> { ... }
908 A function may have prefix data but no body. This has similar semantics
909 to the ``available_externally`` linkage in that the data may be used by the
910 optimizers but will not be emitted in the object file.
917 Attribute groups are groups of attributes that are referenced by objects within
918 the IR. They are important for keeping ``.ll`` files readable, because a lot of
919 functions will use the same set of attributes. In the degenerative case of a
920 ``.ll`` file that corresponds to a single ``.c`` file, the single attribute
921 group will capture the important command line flags used to build that file.
923 An attribute group is a module-level object. To use an attribute group, an
924 object references the attribute group's ID (e.g. ``#37``). An object may refer
925 to more than one attribute group. In that situation, the attributes from the
926 different groups are merged.
928 Here is an example of attribute groups for a function that should always be
929 inlined, has a stack alignment of 4, and which shouldn't use SSE instructions:
933 ; Target-independent attributes:
934 attributes #0 = { alwaysinline alignstack=4 }
936 ; Target-dependent attributes:
937 attributes #1 = { "no-sse" }
939 ; Function @f has attributes: alwaysinline, alignstack=4, and "no-sse".
940 define void @f() #0 #1 { ... }
947 Function attributes are set to communicate additional information about
948 a function. Function attributes are considered to be part of the
949 function, not of the function type, so functions with different function
950 attributes can have the same function type.
952 Function attributes are simple keywords that follow the type specified.
953 If multiple attributes are needed, they are space separated. For
958 define void @f() noinline { ... }
959 define void @f() alwaysinline { ... }
960 define void @f() alwaysinline optsize { ... }
961 define void @f() optsize { ... }
964 This attribute indicates that, when emitting the prologue and
965 epilogue, the backend should forcibly align the stack pointer.
966 Specify the desired alignment, which must be a power of two, in
969 This attribute indicates that the inliner should attempt to inline
970 this function into callers whenever possible, ignoring any active
971 inlining size threshold for this caller.
973 This indicates that the callee function at a call site should be
974 recognized as a built-in function, even though the function's declaration
975 uses the ``nobuiltin`` attribute. This is only valid at call sites for
976 direct calls to functions which are declared with the ``nobuiltin``
979 This attribute indicates that this function is rarely called. When
980 computing edge weights, basic blocks post-dominated by a cold
981 function call are also considered to be cold; and, thus, given low
984 This attribute indicates that the source code contained a hint that
985 inlining this function is desirable (such as the "inline" keyword in
986 C/C++). It is just a hint; it imposes no requirements on the
989 This attribute suggests that optimization passes and code generator
990 passes make choices that keep the code size of this function as small
991 as possible and perform optimizations that may sacrifice runtime
992 performance in order to minimize the size of the generated code.
994 This attribute disables prologue / epilogue emission for the
995 function. This can have very system-specific consequences.
997 This indicates that the callee function at a call site is not recognized as
998 a built-in function. LLVM will retain the original call and not replace it
999 with equivalent code based on the semantics of the built-in function, unless
1000 the call site uses the ``builtin`` attribute. This is valid at call sites
1001 and on function declarations and definitions.
1003 This attribute indicates that calls to the function cannot be
1004 duplicated. A call to a ``noduplicate`` function may be moved
1005 within its parent function, but may not be duplicated within
1006 its parent function.
1008 A function containing a ``noduplicate`` call may still
1009 be an inlining candidate, provided that the call is not
1010 duplicated by inlining. That implies that the function has
1011 internal linkage and only has one call site, so the original
1012 call is dead after inlining.
1014 This attributes disables implicit floating point instructions.
1016 This attribute indicates that the inliner should never inline this
1017 function in any situation. This attribute may not be used together
1018 with the ``alwaysinline`` attribute.
1020 This attribute suppresses lazy symbol binding for the function. This
1021 may make calls to the function faster, at the cost of extra program
1022 startup time if the function is not called during program startup.
1024 This attribute indicates that the code generator should not use a
1025 red zone, even if the target-specific ABI normally permits it.
1027 This function attribute indicates that the function never returns
1028 normally. This produces undefined behavior at runtime if the
1029 function ever does dynamically return.
1031 This function attribute indicates that the function never returns
1032 with an unwind or exceptional control flow. If the function does
1033 unwind, its runtime behavior is undefined.
1035 This function attribute indicates that the function is not optimized
1036 by any optimization or code generator passes with the
1037 exception of interprocedural optimization passes.
1038 This attribute cannot be used together with the ``alwaysinline``
1039 attribute; this attribute is also incompatible
1040 with the ``minsize`` attribute and the ``optsize`` attribute.
1042 This attribute requires the ``noinline`` attribute to be specified on
1043 the function as well, so the function is never inlined into any caller.
1044 Only functions with the ``alwaysinline`` attribute are valid
1045 candidates for inlining into the body of this function.
1047 This attribute suggests that optimization passes and code generator
1048 passes make choices that keep the code size of this function low,
1049 and otherwise do optimizations specifically to reduce code size as
1050 long as they do not significantly impact runtime performance.
1052 On a function, this attribute indicates that the function computes its
1053 result (or decides to unwind an exception) based strictly on its arguments,
1054 without dereferencing any pointer arguments or otherwise accessing
1055 any mutable state (e.g. memory, control registers, etc) visible to
1056 caller functions. It does not write through any pointer arguments
1057 (including ``byval`` arguments) and never changes any state visible
1058 to callers. This means that it cannot unwind exceptions by calling
1059 the ``C++`` exception throwing methods.
1061 On an argument, this attribute indicates that the function does not
1062 dereference that pointer argument, even though it may read or write the
1063 memory that the pointer points to if accessed through other pointers.
1065 On a function, this attribute indicates that the function does not write
1066 through any pointer arguments (including ``byval`` arguments) or otherwise
1067 modify any state (e.g. memory, control registers, etc) visible to
1068 caller functions. It may dereference pointer arguments and read
1069 state that may be set in the caller. A readonly function always
1070 returns the same value (or unwinds an exception identically) when
1071 called with the same set of arguments and global state. It cannot
1072 unwind an exception by calling the ``C++`` exception throwing
1075 On an argument, this attribute indicates that the function does not write
1076 through this pointer argument, even though it may write to the memory that
1077 the pointer points to.
1079 This attribute indicates that this function can return twice. The C
1080 ``setjmp`` is an example of such a function. The compiler disables
1081 some optimizations (like tail calls) in the caller of these
1083 ``sanitize_address``
1084 This attribute indicates that AddressSanitizer checks
1085 (dynamic address safety analysis) are enabled for this function.
1087 This attribute indicates that MemorySanitizer checks (dynamic detection
1088 of accesses to uninitialized memory) are enabled for this function.
1090 This attribute indicates that ThreadSanitizer checks
1091 (dynamic thread safety analysis) are enabled for this function.
1093 This attribute indicates that the function should emit a stack
1094 smashing protector. It is in the form of a "canary" --- a random value
1095 placed on the stack before the local variables that's checked upon
1096 return from the function to see if it has been overwritten. A
1097 heuristic is used to determine if a function needs stack protectors
1098 or not. The heuristic used will enable protectors for functions with:
1100 - Character arrays larger than ``ssp-buffer-size`` (default 8).
1101 - Aggregates containing character arrays larger than ``ssp-buffer-size``.
1102 - Calls to alloca() with variable sizes or constant sizes greater than
1103 ``ssp-buffer-size``.
1105 Variables that are identified as requiring a protector will be arranged
1106 on the stack such that they are adjacent to the stack protector guard.
1108 If a function that has an ``ssp`` attribute is inlined into a
1109 function that doesn't have an ``ssp`` attribute, then the resulting
1110 function will have an ``ssp`` attribute.
1112 This attribute indicates that the function should *always* emit a
1113 stack smashing protector. This overrides the ``ssp`` function
1116 Variables that are identified as requiring a protector will be arranged
1117 on the stack such that they are adjacent to the stack protector guard.
1118 The specific layout rules are:
1120 #. Large arrays and structures containing large arrays
1121 (``>= ssp-buffer-size``) are closest to the stack protector.
1122 #. Small arrays and structures containing small arrays
1123 (``< ssp-buffer-size``) are 2nd closest to the protector.
1124 #. Variables that have had their address taken are 3rd closest to the
1127 If a function that has an ``sspreq`` attribute is inlined into a
1128 function that doesn't have an ``sspreq`` attribute or which has an
1129 ``ssp`` or ``sspstrong`` attribute, then the resulting function will have
1130 an ``sspreq`` attribute.
1132 This attribute indicates that the function should emit a stack smashing
1133 protector. This attribute causes a strong heuristic to be used when
1134 determining if a function needs stack protectors. The strong heuristic
1135 will enable protectors for functions with:
1137 - Arrays of any size and type
1138 - Aggregates containing an array of any size and type.
1139 - Calls to alloca().
1140 - Local variables that have had their address taken.
1142 Variables that are identified as requiring a protector will be arranged
1143 on the stack such that they are adjacent to the stack protector guard.
1144 The specific layout rules are:
1146 #. Large arrays and structures containing large arrays
1147 (``>= ssp-buffer-size``) are closest to the stack protector.
1148 #. Small arrays and structures containing small arrays
1149 (``< ssp-buffer-size``) are 2nd closest to the protector.
1150 #. Variables that have had their address taken are 3rd closest to the
1153 This overrides the ``ssp`` function attribute.
1155 If a function that has an ``sspstrong`` attribute is inlined into a
1156 function that doesn't have an ``sspstrong`` attribute, then the
1157 resulting function will have an ``sspstrong`` attribute.
1159 This attribute indicates that the ABI being targeted requires that
1160 an unwind table entry be produce for this function even if we can
1161 show that no exceptions passes by it. This is normally the case for
1162 the ELF x86-64 abi, but it can be disabled for some compilation
1167 Module-Level Inline Assembly
1168 ----------------------------
1170 Modules may contain "module-level inline asm" blocks, which corresponds
1171 to the GCC "file scope inline asm" blocks. These blocks are internally
1172 concatenated by LLVM and treated as a single unit, but may be separated
1173 in the ``.ll`` file if desired. The syntax is very simple:
1175 .. code-block:: llvm
1177 module asm "inline asm code goes here"
1178 module asm "more can go here"
1180 The strings can contain any character by escaping non-printable
1181 characters. The escape sequence used is simply "\\xx" where "xx" is the
1182 two digit hex code for the number.
1184 The inline asm code is simply printed to the machine code .s file when
1185 assembly code is generated.
1187 .. _langref_datalayout:
1192 A module may specify a target specific data layout string that specifies
1193 how data is to be laid out in memory. The syntax for the data layout is
1196 .. code-block:: llvm
1198 target datalayout = "layout specification"
1200 The *layout specification* consists of a list of specifications
1201 separated by the minus sign character ('-'). Each specification starts
1202 with a letter and may include other information after the letter to
1203 define some aspect of the data layout. The specifications accepted are
1207 Specifies that the target lays out data in big-endian form. That is,
1208 the bits with the most significance have the lowest address
1211 Specifies that the target lays out data in little-endian form. That
1212 is, the bits with the least significance have the lowest address
1215 Specifies the natural alignment of the stack in bits. Alignment
1216 promotion of stack variables is limited to the natural stack
1217 alignment to avoid dynamic stack realignment. The stack alignment
1218 must be a multiple of 8-bits. If omitted, the natural stack
1219 alignment defaults to "unspecified", which does not prevent any
1220 alignment promotions.
1221 ``p[n]:<size>:<abi>:<pref>``
1222 This specifies the *size* of a pointer and its ``<abi>`` and
1223 ``<pref>``\erred alignments for address space ``n``. All sizes are in
1224 bits. The address space, ``n`` is optional, and if not specified,
1225 denotes the default address space 0. The value of ``n`` must be
1226 in the range [1,2^23).
1227 ``i<size>:<abi>:<pref>``
1228 This specifies the alignment for an integer type of a given bit
1229 ``<size>``. The value of ``<size>`` must be in the range [1,2^23).
1230 ``v<size>:<abi>:<pref>``
1231 This specifies the alignment for a vector type of a given bit
1233 ``f<size>:<abi>:<pref>``
1234 This specifies the alignment for a floating point type of a given bit
1235 ``<size>``. Only values of ``<size>`` that are supported by the target
1236 will work. 32 (float) and 64 (double) are supported on all targets; 80
1237 or 128 (different flavors of long double) are also supported on some
1240 This specifies the alignment for an object of aggregate type.
1242 If present, specifies that llvm names are mangled in the output. The
1245 * ``e``: ELF mangling: Private symbols get a ``.L`` prefix.
1246 * ``m``: Mips mangling: Private symbols get a ``$`` prefix.
1247 * ``o``: Mach-O mangling: Private symbols get ``L`` prefix. Other
1248 symbols get a ``_`` prefix.
1249 * ``w``: Windows COFF prefix: Similar to Mach-O, but stdcall and fastcall
1250 functions also get a suffix based on the frame size.
1251 ``n<size1>:<size2>:<size3>...``
1252 This specifies a set of native integer widths for the target CPU in
1253 bits. For example, it might contain ``n32`` for 32-bit PowerPC,
1254 ``n32:64`` for PowerPC 64, or ``n8:16:32:64`` for X86-64. Elements of
1255 this set are considered to support most general arithmetic operations
1258 On every specification that takes a ``<abi>:<pref>``, specifying the
1259 ``<pref>`` alignment is optional. If omitted, the preceding ``:``
1260 should be omitted too and ``<pref>`` will be equal to ``<abi>``.
1262 When constructing the data layout for a given target, LLVM starts with a
1263 default set of specifications which are then (possibly) overridden by
1264 the specifications in the ``datalayout`` keyword. The default
1265 specifications are given in this list:
1267 - ``E`` - big endian
1268 - ``p:64:64:64`` - 64-bit pointers with 64-bit alignment.
1269 - ``p[n]:64:64:64`` - Other address spaces are assumed to be the
1270 same as the default address space.
1271 - ``S0`` - natural stack alignment is unspecified
1272 - ``i1:8:8`` - i1 is 8-bit (byte) aligned
1273 - ``i8:8:8`` - i8 is 8-bit (byte) aligned
1274 - ``i16:16:16`` - i16 is 16-bit aligned
1275 - ``i32:32:32`` - i32 is 32-bit aligned
1276 - ``i64:32:64`` - i64 has ABI alignment of 32-bits but preferred
1277 alignment of 64-bits
1278 - ``f16:16:16`` - half is 16-bit aligned
1279 - ``f32:32:32`` - float is 32-bit aligned
1280 - ``f64:64:64`` - double is 64-bit aligned
1281 - ``f128:128:128`` - quad is 128-bit aligned
1282 - ``v64:64:64`` - 64-bit vector is 64-bit aligned
1283 - ``v128:128:128`` - 128-bit vector is 128-bit aligned
1284 - ``a:0:64`` - aggregates are 64-bit aligned
1286 When LLVM is determining the alignment for a given type, it uses the
1289 #. If the type sought is an exact match for one of the specifications,
1290 that specification is used.
1291 #. If no match is found, and the type sought is an integer type, then
1292 the smallest integer type that is larger than the bitwidth of the
1293 sought type is used. If none of the specifications are larger than
1294 the bitwidth then the largest integer type is used. For example,
1295 given the default specifications above, the i7 type will use the
1296 alignment of i8 (next largest) while both i65 and i256 will use the
1297 alignment of i64 (largest specified).
1298 #. If no match is found, and the type sought is a vector type, then the
1299 largest vector type that is smaller than the sought vector type will
1300 be used as a fall back. This happens because <128 x double> can be
1301 implemented in terms of 64 <2 x double>, for example.
1303 The function of the data layout string may not be what you expect.
1304 Notably, this is not a specification from the frontend of what alignment
1305 the code generator should use.
1307 Instead, if specified, the target data layout is required to match what
1308 the ultimate *code generator* expects. This string is used by the
1309 mid-level optimizers to improve code, and this only works if it matches
1310 what the ultimate code generator uses. If you would like to generate IR
1311 that does not embed this target-specific detail into the IR, then you
1312 don't have to specify the string. This will disable some optimizations
1313 that require precise layout information, but this also prevents those
1314 optimizations from introducing target specificity into the IR.
1321 A module may specify a target triple string that describes the target
1322 host. The syntax for the target triple is simply:
1324 .. code-block:: llvm
1326 target triple = "x86_64-apple-macosx10.7.0"
1328 The *target triple* string consists of a series of identifiers delimited
1329 by the minus sign character ('-'). The canonical forms are:
1333 ARCHITECTURE-VENDOR-OPERATING_SYSTEM
1334 ARCHITECTURE-VENDOR-OPERATING_SYSTEM-ENVIRONMENT
1336 This information is passed along to the backend so that it generates
1337 code for the proper architecture. It's possible to override this on the
1338 command line with the ``-mtriple`` command line option.
1340 .. _pointeraliasing:
1342 Pointer Aliasing Rules
1343 ----------------------
1345 Any memory access must be done through a pointer value associated with
1346 an address range of the memory access, otherwise the behavior is
1347 undefined. Pointer values are associated with address ranges according
1348 to the following rules:
1350 - A pointer value is associated with the addresses associated with any
1351 value it is *based* on.
1352 - An address of a global variable is associated with the address range
1353 of the variable's storage.
1354 - The result value of an allocation instruction is associated with the
1355 address range of the allocated storage.
1356 - A null pointer in the default address-space is associated with no
1358 - An integer constant other than zero or a pointer value returned from
1359 a function not defined within LLVM may be associated with address
1360 ranges allocated through mechanisms other than those provided by
1361 LLVM. Such ranges shall not overlap with any ranges of addresses
1362 allocated by mechanisms provided by LLVM.
1364 A pointer value is *based* on another pointer value according to the
1367 - A pointer value formed from a ``getelementptr`` operation is *based*
1368 on the first operand of the ``getelementptr``.
1369 - The result value of a ``bitcast`` is *based* on the operand of the
1371 - A pointer value formed by an ``inttoptr`` is *based* on all pointer
1372 values that contribute (directly or indirectly) to the computation of
1373 the pointer's value.
1374 - The "*based* on" relationship is transitive.
1376 Note that this definition of *"based"* is intentionally similar to the
1377 definition of *"based"* in C99, though it is slightly weaker.
1379 LLVM IR does not associate types with memory. The result type of a
1380 ``load`` merely indicates the size and alignment of the memory from
1381 which to load, as well as the interpretation of the value. The first
1382 operand type of a ``store`` similarly only indicates the size and
1383 alignment of the store.
1385 Consequently, type-based alias analysis, aka TBAA, aka
1386 ``-fstrict-aliasing``, is not applicable to general unadorned LLVM IR.
1387 :ref:`Metadata <metadata>` may be used to encode additional information
1388 which specialized optimization passes may use to implement type-based
1393 Volatile Memory Accesses
1394 ------------------------
1396 Certain memory accesses, such as :ref:`load <i_load>`'s,
1397 :ref:`store <i_store>`'s, and :ref:`llvm.memcpy <int_memcpy>`'s may be
1398 marked ``volatile``. The optimizers must not change the number of
1399 volatile operations or change their order of execution relative to other
1400 volatile operations. The optimizers *may* change the order of volatile
1401 operations relative to non-volatile operations. This is not Java's
1402 "volatile" and has no cross-thread synchronization behavior.
1404 IR-level volatile loads and stores cannot safely be optimized into
1405 llvm.memcpy or llvm.memmove intrinsics even when those intrinsics are
1406 flagged volatile. Likewise, the backend should never split or merge
1407 target-legal volatile load/store instructions.
1409 .. admonition:: Rationale
1411 Platforms may rely on volatile loads and stores of natively supported
1412 data width to be executed as single instruction. For example, in C
1413 this holds for an l-value of volatile primitive type with native
1414 hardware support, but not necessarily for aggregate types. The
1415 frontend upholds these expectations, which are intentionally
1416 unspecified in the IR. The rules above ensure that IR transformation
1417 do not violate the frontend's contract with the language.
1421 Memory Model for Concurrent Operations
1422 --------------------------------------
1424 The LLVM IR does not define any way to start parallel threads of
1425 execution or to register signal handlers. Nonetheless, there are
1426 platform-specific ways to create them, and we define LLVM IR's behavior
1427 in their presence. This model is inspired by the C++0x memory model.
1429 For a more informal introduction to this model, see the :doc:`Atomics`.
1431 We define a *happens-before* partial order as the least partial order
1434 - Is a superset of single-thread program order, and
1435 - When a *synchronizes-with* ``b``, includes an edge from ``a`` to
1436 ``b``. *Synchronizes-with* pairs are introduced by platform-specific
1437 techniques, like pthread locks, thread creation, thread joining,
1438 etc., and by atomic instructions. (See also :ref:`Atomic Memory Ordering
1439 Constraints <ordering>`).
1441 Note that program order does not introduce *happens-before* edges
1442 between a thread and signals executing inside that thread.
1444 Every (defined) read operation (load instructions, memcpy, atomic
1445 loads/read-modify-writes, etc.) R reads a series of bytes written by
1446 (defined) write operations (store instructions, atomic
1447 stores/read-modify-writes, memcpy, etc.). For the purposes of this
1448 section, initialized globals are considered to have a write of the
1449 initializer which is atomic and happens before any other read or write
1450 of the memory in question. For each byte of a read R, R\ :sub:`byte`
1451 may see any write to the same byte, except:
1453 - If write\ :sub:`1` happens before write\ :sub:`2`, and
1454 write\ :sub:`2` happens before R\ :sub:`byte`, then
1455 R\ :sub:`byte` does not see write\ :sub:`1`.
1456 - If R\ :sub:`byte` happens before write\ :sub:`3`, then
1457 R\ :sub:`byte` does not see write\ :sub:`3`.
1459 Given that definition, R\ :sub:`byte` is defined as follows:
1461 - If R is volatile, the result is target-dependent. (Volatile is
1462 supposed to give guarantees which can support ``sig_atomic_t`` in
1463 C/C++, and may be used for accesses to addresses which do not behave
1464 like normal memory. It does not generally provide cross-thread
1466 - Otherwise, if there is no write to the same byte that happens before
1467 R\ :sub:`byte`, R\ :sub:`byte` returns ``undef`` for that byte.
1468 - Otherwise, if R\ :sub:`byte` may see exactly one write,
1469 R\ :sub:`byte` returns the value written by that write.
1470 - Otherwise, if R is atomic, and all the writes R\ :sub:`byte` may
1471 see are atomic, it chooses one of the values written. See the :ref:`Atomic
1472 Memory Ordering Constraints <ordering>` section for additional
1473 constraints on how the choice is made.
1474 - Otherwise R\ :sub:`byte` returns ``undef``.
1476 R returns the value composed of the series of bytes it read. This
1477 implies that some bytes within the value may be ``undef`` **without**
1478 the entire value being ``undef``. Note that this only defines the
1479 semantics of the operation; it doesn't mean that targets will emit more
1480 than one instruction to read the series of bytes.
1482 Note that in cases where none of the atomic intrinsics are used, this
1483 model places only one restriction on IR transformations on top of what
1484 is required for single-threaded execution: introducing a store to a byte
1485 which might not otherwise be stored is not allowed in general.
1486 (Specifically, in the case where another thread might write to and read
1487 from an address, introducing a store can change a load that may see
1488 exactly one write into a load that may see multiple writes.)
1492 Atomic Memory Ordering Constraints
1493 ----------------------------------
1495 Atomic instructions (:ref:`cmpxchg <i_cmpxchg>`,
1496 :ref:`atomicrmw <i_atomicrmw>`, :ref:`fence <i_fence>`,
1497 :ref:`atomic load <i_load>`, and :ref:`atomic store <i_store>`) take
1498 ordering parameters that determine which other atomic instructions on
1499 the same address they *synchronize with*. These semantics are borrowed
1500 from Java and C++0x, but are somewhat more colloquial. If these
1501 descriptions aren't precise enough, check those specs (see spec
1502 references in the :doc:`atomics guide <Atomics>`).
1503 :ref:`fence <i_fence>` instructions treat these orderings somewhat
1504 differently since they don't take an address. See that instruction's
1505 documentation for details.
1507 For a simpler introduction to the ordering constraints, see the
1511 The set of values that can be read is governed by the happens-before
1512 partial order. A value cannot be read unless some operation wrote
1513 it. This is intended to provide a guarantee strong enough to model
1514 Java's non-volatile shared variables. This ordering cannot be
1515 specified for read-modify-write operations; it is not strong enough
1516 to make them atomic in any interesting way.
1518 In addition to the guarantees of ``unordered``, there is a single
1519 total order for modifications by ``monotonic`` operations on each
1520 address. All modification orders must be compatible with the
1521 happens-before order. There is no guarantee that the modification
1522 orders can be combined to a global total order for the whole program
1523 (and this often will not be possible). The read in an atomic
1524 read-modify-write operation (:ref:`cmpxchg <i_cmpxchg>` and
1525 :ref:`atomicrmw <i_atomicrmw>`) reads the value in the modification
1526 order immediately before the value it writes. If one atomic read
1527 happens before another atomic read of the same address, the later
1528 read must see the same value or a later value in the address's
1529 modification order. This disallows reordering of ``monotonic`` (or
1530 stronger) operations on the same address. If an address is written
1531 ``monotonic``-ally by one thread, and other threads ``monotonic``-ally
1532 read that address repeatedly, the other threads must eventually see
1533 the write. This corresponds to the C++0x/C1x
1534 ``memory_order_relaxed``.
1536 In addition to the guarantees of ``monotonic``, a
1537 *synchronizes-with* edge may be formed with a ``release`` operation.
1538 This is intended to model C++'s ``memory_order_acquire``.
1540 In addition to the guarantees of ``monotonic``, if this operation
1541 writes a value which is subsequently read by an ``acquire``
1542 operation, it *synchronizes-with* that operation. (This isn't a
1543 complete description; see the C++0x definition of a release
1544 sequence.) This corresponds to the C++0x/C1x
1545 ``memory_order_release``.
1546 ``acq_rel`` (acquire+release)
1547 Acts as both an ``acquire`` and ``release`` operation on its
1548 address. This corresponds to the C++0x/C1x ``memory_order_acq_rel``.
1549 ``seq_cst`` (sequentially consistent)
1550 In addition to the guarantees of ``acq_rel`` (``acquire`` for an
1551 operation which only reads, ``release`` for an operation which only
1552 writes), there is a global total order on all
1553 sequentially-consistent operations on all addresses, which is
1554 consistent with the *happens-before* partial order and with the
1555 modification orders of all the affected addresses. Each
1556 sequentially-consistent read sees the last preceding write to the
1557 same address in this global order. This corresponds to the C++0x/C1x
1558 ``memory_order_seq_cst`` and Java volatile.
1562 If an atomic operation is marked ``singlethread``, it only *synchronizes
1563 with* or participates in modification and seq\_cst total orderings with
1564 other operations running in the same thread (for example, in signal
1572 LLVM IR floating-point binary ops (:ref:`fadd <i_fadd>`,
1573 :ref:`fsub <i_fsub>`, :ref:`fmul <i_fmul>`, :ref:`fdiv <i_fdiv>`,
1574 :ref:`frem <i_frem>`) have the following flags that can set to enable
1575 otherwise unsafe floating point operations
1578 No NaNs - Allow optimizations to assume the arguments and result are not
1579 NaN. Such optimizations are required to retain defined behavior over
1580 NaNs, but the value of the result is undefined.
1583 No Infs - Allow optimizations to assume the arguments and result are not
1584 +/-Inf. Such optimizations are required to retain defined behavior over
1585 +/-Inf, but the value of the result is undefined.
1588 No Signed Zeros - Allow optimizations to treat the sign of a zero
1589 argument or result as insignificant.
1592 Allow Reciprocal - Allow optimizations to use the reciprocal of an
1593 argument rather than perform division.
1596 Fast - Allow algebraically equivalent transformations that may
1597 dramatically change results in floating point (e.g. reassociate). This
1598 flag implies all the others.
1605 The LLVM type system is one of the most important features of the
1606 intermediate representation. Being typed enables a number of
1607 optimizations to be performed on the intermediate representation
1608 directly, without having to do extra analyses on the side before the
1609 transformation. A strong type system makes it easier to read the
1610 generated code and enables novel analyses and transformations that are
1611 not feasible to perform on normal three address code representations.
1621 The void type does not represent any value and has no size.
1639 The function type can be thought of as a function signature. It consists of a
1640 return type and a list of formal parameter types. The return type of a function
1641 type is a void type or first class type --- except for :ref:`label <t_label>`
1642 and :ref:`metadata <t_metadata>` types.
1648 <returntype> (<parameter list>)
1650 ...where '``<parameter list>``' is a comma-separated list of type
1651 specifiers. Optionally, the parameter list may include a type ``...``, which
1652 indicates that the function takes a variable number of arguments. Variable
1653 argument functions can access their arguments with the :ref:`variable argument
1654 handling intrinsic <int_varargs>` functions. '``<returntype>``' is any type
1655 except :ref:`label <t_label>` and :ref:`metadata <t_metadata>`.
1659 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1660 | ``i32 (i32)`` | function taking an ``i32``, returning an ``i32`` |
1661 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1662 | ``float (i16, i32 *) *`` | :ref:`Pointer <t_pointer>` to a function that takes an ``i16`` and a :ref:`pointer <t_pointer>` to ``i32``, returning ``float``. |
1663 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1664 | ``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. |
1665 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1666 | ``{i32, i32} (i32)`` | A function taking an ``i32``, returning a :ref:`structure <t_struct>` containing two ``i32`` values |
1667 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1674 The :ref:`first class <t_firstclass>` types are perhaps the most important.
1675 Values of these types are the only ones which can be produced by
1683 These are the types that are valid in registers from CodeGen's perspective.
1692 The integer type is a very simple type that simply specifies an
1693 arbitrary bit width for the integer type desired. Any bit width from 1
1694 bit to 2\ :sup:`23`\ -1 (about 8 million) can be specified.
1702 The number of bits the integer will occupy is specified by the ``N``
1708 +----------------+------------------------------------------------+
1709 | ``i1`` | a single-bit integer. |
1710 +----------------+------------------------------------------------+
1711 | ``i32`` | a 32-bit integer. |
1712 +----------------+------------------------------------------------+
1713 | ``i1942652`` | a really big integer of over 1 million bits. |
1714 +----------------+------------------------------------------------+
1718 Floating Point Types
1719 """"""""""""""""""""
1728 - 16-bit floating point value
1731 - 32-bit floating point value
1734 - 64-bit floating point value
1737 - 128-bit floating point value (112-bit mantissa)
1740 - 80-bit floating point value (X87)
1743 - 128-bit floating point value (two 64-bits)
1750 The x86_mmx type represents a value held in an MMX register on an x86
1751 machine. The operations allowed on it are quite limited: parameters and
1752 return values, load and store, and bitcast. User-specified MMX
1753 instructions are represented as intrinsic or asm calls with arguments
1754 and/or results of this type. There are no arrays, vectors or constants
1771 The pointer type is used to specify memory locations. Pointers are
1772 commonly used to reference objects in memory.
1774 Pointer types may have an optional address space attribute defining the
1775 numbered address space where the pointed-to object resides. The default
1776 address space is number zero. The semantics of non-zero address spaces
1777 are target-specific.
1779 Note that LLVM does not permit pointers to void (``void*``) nor does it
1780 permit pointers to labels (``label*``). Use ``i8*`` instead.
1790 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1791 | ``[4 x i32]*`` | A :ref:`pointer <t_pointer>` to :ref:`array <t_array>` of four ``i32`` values. |
1792 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1793 | ``i32 (i32*) *`` | A :ref:`pointer <t_pointer>` to a :ref:`function <t_function>` that takes an ``i32*``, returning an ``i32``. |
1794 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1795 | ``i32 addrspace(5)*`` | A :ref:`pointer <t_pointer>` to an ``i32`` value that resides in address space #5. |
1796 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1805 A vector type is a simple derived type that represents a vector of
1806 elements. Vector types are used when multiple primitive data are
1807 operated in parallel using a single instruction (SIMD). A vector type
1808 requires a size (number of elements) and an underlying primitive data
1809 type. Vector types are considered :ref:`first class <t_firstclass>`.
1815 < <# elements> x <elementtype> >
1817 The number of elements is a constant integer value larger than 0;
1818 elementtype may be any integer or floating point type, or a pointer to
1819 these types. Vectors of size zero are not allowed.
1823 +-------------------+--------------------------------------------------+
1824 | ``<4 x i32>`` | Vector of 4 32-bit integer values. |
1825 +-------------------+--------------------------------------------------+
1826 | ``<8 x float>`` | Vector of 8 32-bit floating-point values. |
1827 +-------------------+--------------------------------------------------+
1828 | ``<2 x i64>`` | Vector of 2 64-bit integer values. |
1829 +-------------------+--------------------------------------------------+
1830 | ``<4 x i64*>`` | Vector of 4 pointers to 64-bit integer values. |
1831 +-------------------+--------------------------------------------------+
1840 The label type represents code labels.
1855 The metadata type represents embedded metadata. No derived types may be
1856 created from metadata except for :ref:`function <t_function>` arguments.
1869 Aggregate Types are a subset of derived types that can contain multiple
1870 member types. :ref:`Arrays <t_array>` and :ref:`structs <t_struct>` are
1871 aggregate types. :ref:`Vectors <t_vector>` are not considered to be
1881 The array type is a very simple derived type that arranges elements
1882 sequentially in memory. The array type requires a size (number of
1883 elements) and an underlying data type.
1889 [<# elements> x <elementtype>]
1891 The number of elements is a constant integer value; ``elementtype`` may
1892 be any type with a size.
1896 +------------------+--------------------------------------+
1897 | ``[40 x i32]`` | Array of 40 32-bit integer values. |
1898 +------------------+--------------------------------------+
1899 | ``[41 x i32]`` | Array of 41 32-bit integer values. |
1900 +------------------+--------------------------------------+
1901 | ``[4 x i8]`` | Array of 4 8-bit integer values. |
1902 +------------------+--------------------------------------+
1904 Here are some examples of multidimensional arrays:
1906 +-----------------------------+----------------------------------------------------------+
1907 | ``[3 x [4 x i32]]`` | 3x4 array of 32-bit integer values. |
1908 +-----------------------------+----------------------------------------------------------+
1909 | ``[12 x [10 x float]]`` | 12x10 array of single precision floating point values. |
1910 +-----------------------------+----------------------------------------------------------+
1911 | ``[2 x [3 x [4 x i16]]]`` | 2x3x4 array of 16-bit integer values. |
1912 +-----------------------------+----------------------------------------------------------+
1914 There is no restriction on indexing beyond the end of the array implied
1915 by a static type (though there are restrictions on indexing beyond the
1916 bounds of an allocated object in some cases). This means that
1917 single-dimension 'variable sized array' addressing can be implemented in
1918 LLVM with a zero length array type. An implementation of 'pascal style
1919 arrays' in LLVM could use the type "``{ i32, [0 x float]}``", for
1929 The structure type is used to represent a collection of data members
1930 together in memory. The elements of a structure may be any type that has
1933 Structures in memory are accessed using '``load``' and '``store``' by
1934 getting a pointer to a field with the '``getelementptr``' instruction.
1935 Structures in registers are accessed using the '``extractvalue``' and
1936 '``insertvalue``' instructions.
1938 Structures may optionally be "packed" structures, which indicate that
1939 the alignment of the struct is one byte, and that there is no padding
1940 between the elements. In non-packed structs, padding between field types
1941 is inserted as defined by the DataLayout string in the module, which is
1942 required to match what the underlying code generator expects.
1944 Structures can either be "literal" or "identified". A literal structure
1945 is defined inline with other types (e.g. ``{i32, i32}*``) whereas
1946 identified types are always defined at the top level with a name.
1947 Literal types are uniqued by their contents and can never be recursive
1948 or opaque since there is no way to write one. Identified types can be
1949 recursive, can be opaqued, and are never uniqued.
1955 %T1 = type { <type list> } ; Identified normal struct type
1956 %T2 = type <{ <type list> }> ; Identified packed struct type
1960 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1961 | ``{ i32, i32, i32 }`` | A triple of three ``i32`` values |
1962 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1963 | ``{ 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``. |
1964 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1965 | ``<{ i8, i32 }>`` | A packed struct known to be 5 bytes in size. |
1966 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1970 Opaque Structure Types
1971 """"""""""""""""""""""
1975 Opaque structure types are used to represent named structure types that
1976 do not have a body specified. This corresponds (for example) to the C
1977 notion of a forward declared structure.
1988 +--------------+-------------------+
1989 | ``opaque`` | An opaque type. |
1990 +--------------+-------------------+
1997 LLVM has several different basic types of constants. This section
1998 describes them all and their syntax.
2003 **Boolean constants**
2004 The two strings '``true``' and '``false``' are both valid constants
2006 **Integer constants**
2007 Standard integers (such as '4') are constants of the
2008 :ref:`integer <t_integer>` type. Negative numbers may be used with
2010 **Floating point constants**
2011 Floating point constants use standard decimal notation (e.g.
2012 123.421), exponential notation (e.g. 1.23421e+2), or a more precise
2013 hexadecimal notation (see below). The assembler requires the exact
2014 decimal value of a floating-point constant. For example, the
2015 assembler accepts 1.25 but rejects 1.3 because 1.3 is a repeating
2016 decimal in binary. Floating point constants must have a :ref:`floating
2017 point <t_floating>` type.
2018 **Null pointer constants**
2019 The identifier '``null``' is recognized as a null pointer constant
2020 and must be of :ref:`pointer type <t_pointer>`.
2022 The one non-intuitive notation for constants is the hexadecimal form of
2023 floating point constants. For example, the form
2024 '``double 0x432ff973cafa8000``' is equivalent to (but harder to read
2025 than) '``double 4.5e+15``'. The only time hexadecimal floating point
2026 constants are required (and the only time that they are generated by the
2027 disassembler) is when a floating point constant must be emitted but it
2028 cannot be represented as a decimal floating point number in a reasonable
2029 number of digits. For example, NaN's, infinities, and other special
2030 values are represented in their IEEE hexadecimal format so that assembly
2031 and disassembly do not cause any bits to change in the constants.
2033 When using the hexadecimal form, constants of types half, float, and
2034 double are represented using the 16-digit form shown above (which
2035 matches the IEEE754 representation for double); half and float values
2036 must, however, be exactly representable as IEEE 754 half and single
2037 precision, respectively. Hexadecimal format is always used for long
2038 double, and there are three forms of long double. The 80-bit format used
2039 by x86 is represented as ``0xK`` followed by 20 hexadecimal digits. The
2040 128-bit format used by PowerPC (two adjacent doubles) is represented by
2041 ``0xM`` followed by 32 hexadecimal digits. The IEEE 128-bit format is
2042 represented by ``0xL`` followed by 32 hexadecimal digits. Long doubles
2043 will only work if they match the long double format on your target.
2044 The IEEE 16-bit format (half precision) is represented by ``0xH``
2045 followed by 4 hexadecimal digits. All hexadecimal formats are big-endian
2046 (sign bit at the left).
2048 There are no constants of type x86_mmx.
2050 .. _complexconstants:
2055 Complex constants are a (potentially recursive) combination of simple
2056 constants and smaller complex constants.
2058 **Structure constants**
2059 Structure constants are represented with notation similar to
2060 structure type definitions (a comma separated list of elements,
2061 surrounded by braces (``{}``)). For example:
2062 "``{ i32 4, float 17.0, i32* @G }``", where "``@G``" is declared as
2063 "``@G = external global i32``". Structure constants must have
2064 :ref:`structure type <t_struct>`, and the number and types of elements
2065 must match those specified by the type.
2067 Array constants are represented with notation similar to array type
2068 definitions (a comma separated list of elements, surrounded by
2069 square brackets (``[]``)). For example:
2070 "``[ i32 42, i32 11, i32 74 ]``". Array constants must have
2071 :ref:`array type <t_array>`, and the number and types of elements must
2072 match those specified by the type.
2073 **Vector constants**
2074 Vector constants are represented with notation similar to vector
2075 type definitions (a comma separated list of elements, surrounded by
2076 less-than/greater-than's (``<>``)). For example:
2077 "``< i32 42, i32 11, i32 74, i32 100 >``". Vector constants
2078 must have :ref:`vector type <t_vector>`, and the number and types of
2079 elements must match those specified by the type.
2080 **Zero initialization**
2081 The string '``zeroinitializer``' can be used to zero initialize a
2082 value to zero of *any* type, including scalar and
2083 :ref:`aggregate <t_aggregate>` types. This is often used to avoid
2084 having to print large zero initializers (e.g. for large arrays) and
2085 is always exactly equivalent to using explicit zero initializers.
2087 A metadata node is a structure-like constant with :ref:`metadata
2088 type <t_metadata>`. For example:
2089 "``metadata !{ i32 0, metadata !"test" }``". Unlike other
2090 constants that are meant to be interpreted as part of the
2091 instruction stream, metadata is a place to attach additional
2092 information such as debug info.
2094 Global Variable and Function Addresses
2095 --------------------------------------
2097 The addresses of :ref:`global variables <globalvars>` and
2098 :ref:`functions <functionstructure>` are always implicitly valid
2099 (link-time) constants. These constants are explicitly referenced when
2100 the :ref:`identifier for the global <identifiers>` is used and always have
2101 :ref:`pointer <t_pointer>` type. For example, the following is a legal LLVM
2104 .. code-block:: llvm
2108 @Z = global [2 x i32*] [ i32* @X, i32* @Y ]
2115 The string '``undef``' can be used anywhere a constant is expected, and
2116 indicates that the user of the value may receive an unspecified
2117 bit-pattern. Undefined values may be of any type (other than '``label``'
2118 or '``void``') and be used anywhere a constant is permitted.
2120 Undefined values are useful because they indicate to the compiler that
2121 the program is well defined no matter what value is used. This gives the
2122 compiler more freedom to optimize. Here are some examples of
2123 (potentially surprising) transformations that are valid (in pseudo IR):
2125 .. code-block:: llvm
2135 This is safe because all of the output bits are affected by the undef
2136 bits. Any output bit can have a zero or one depending on the input bits.
2138 .. code-block:: llvm
2149 These logical operations have bits that are not always affected by the
2150 input. For example, if ``%X`` has a zero bit, then the output of the
2151 '``and``' operation will always be a zero for that bit, no matter what
2152 the corresponding bit from the '``undef``' is. As such, it is unsafe to
2153 optimize or assume that the result of the '``and``' is '``undef``'.
2154 However, it is safe to assume that all bits of the '``undef``' could be
2155 0, and optimize the '``and``' to 0. Likewise, it is safe to assume that
2156 all the bits of the '``undef``' operand to the '``or``' could be set,
2157 allowing the '``or``' to be folded to -1.
2159 .. code-block:: llvm
2161 %A = select undef, %X, %Y
2162 %B = select undef, 42, %Y
2163 %C = select %X, %Y, undef
2173 This set of examples shows that undefined '``select``' (and conditional
2174 branch) conditions can go *either way*, but they have to come from one
2175 of the two operands. In the ``%A`` example, if ``%X`` and ``%Y`` were
2176 both known to have a clear low bit, then ``%A`` would have to have a
2177 cleared low bit. However, in the ``%C`` example, the optimizer is
2178 allowed to assume that the '``undef``' operand could be the same as
2179 ``%Y``, allowing the whole '``select``' to be eliminated.
2181 .. code-block:: llvm
2183 %A = xor undef, undef
2200 This example points out that two '``undef``' operands are not
2201 necessarily the same. This can be surprising to people (and also matches
2202 C semantics) where they assume that "``X^X``" is always zero, even if
2203 ``X`` is undefined. This isn't true for a number of reasons, but the
2204 short answer is that an '``undef``' "variable" can arbitrarily change
2205 its value over its "live range". This is true because the variable
2206 doesn't actually *have a live range*. Instead, the value is logically
2207 read from arbitrary registers that happen to be around when needed, so
2208 the value is not necessarily consistent over time. In fact, ``%A`` and
2209 ``%C`` need to have the same semantics or the core LLVM "replace all
2210 uses with" concept would not hold.
2212 .. code-block:: llvm
2220 These examples show the crucial difference between an *undefined value*
2221 and *undefined behavior*. An undefined value (like '``undef``') is
2222 allowed to have an arbitrary bit-pattern. This means that the ``%A``
2223 operation can be constant folded to '``undef``', because the '``undef``'
2224 could be an SNaN, and ``fdiv`` is not (currently) defined on SNaN's.
2225 However, in the second example, we can make a more aggressive
2226 assumption: because the ``undef`` is allowed to be an arbitrary value,
2227 we are allowed to assume that it could be zero. Since a divide by zero
2228 has *undefined behavior*, we are allowed to assume that the operation
2229 does not execute at all. This allows us to delete the divide and all
2230 code after it. Because the undefined operation "can't happen", the
2231 optimizer can assume that it occurs in dead code.
2233 .. code-block:: llvm
2235 a: store undef -> %X
2236 b: store %X -> undef
2241 These examples reiterate the ``fdiv`` example: a store *of* an undefined
2242 value can be assumed to not have any effect; we can assume that the
2243 value is overwritten with bits that happen to match what was already
2244 there. However, a store *to* an undefined location could clobber
2245 arbitrary memory, therefore, it has undefined behavior.
2252 Poison values are similar to :ref:`undef values <undefvalues>`, however
2253 they also represent the fact that an instruction or constant expression
2254 which cannot evoke side effects has nevertheless detected a condition
2255 which results in undefined behavior.
2257 There is currently no way of representing a poison value in the IR; they
2258 only exist when produced by operations such as :ref:`add <i_add>` with
2261 Poison value behavior is defined in terms of value *dependence*:
2263 - Values other than :ref:`phi <i_phi>` nodes depend on their operands.
2264 - :ref:`Phi <i_phi>` nodes depend on the operand corresponding to
2265 their dynamic predecessor basic block.
2266 - Function arguments depend on the corresponding actual argument values
2267 in the dynamic callers of their functions.
2268 - :ref:`Call <i_call>` instructions depend on the :ref:`ret <i_ret>`
2269 instructions that dynamically transfer control back to them.
2270 - :ref:`Invoke <i_invoke>` instructions depend on the
2271 :ref:`ret <i_ret>`, :ref:`resume <i_resume>`, or exception-throwing
2272 call instructions that dynamically transfer control back to them.
2273 - Non-volatile loads and stores depend on the most recent stores to all
2274 of the referenced memory addresses, following the order in the IR
2275 (including loads and stores implied by intrinsics such as
2276 :ref:`@llvm.memcpy <int_memcpy>`.)
2277 - An instruction with externally visible side effects depends on the
2278 most recent preceding instruction with externally visible side
2279 effects, following the order in the IR. (This includes :ref:`volatile
2280 operations <volatile>`.)
2281 - An instruction *control-depends* on a :ref:`terminator
2282 instruction <terminators>` if the terminator instruction has
2283 multiple successors and the instruction is always executed when
2284 control transfers to one of the successors, and may not be executed
2285 when control is transferred to another.
2286 - Additionally, an instruction also *control-depends* on a terminator
2287 instruction if the set of instructions it otherwise depends on would
2288 be different if the terminator had transferred control to a different
2290 - Dependence is transitive.
2292 Poison Values have the same behavior as :ref:`undef values <undefvalues>`,
2293 with the additional affect that any instruction which has a *dependence*
2294 on a poison value has undefined behavior.
2296 Here are some examples:
2298 .. code-block:: llvm
2301 %poison = sub nuw i32 0, 1 ; Results in a poison value.
2302 %still_poison = and i32 %poison, 0 ; 0, but also poison.
2303 %poison_yet_again = getelementptr i32* @h, i32 %still_poison
2304 store i32 0, i32* %poison_yet_again ; memory at @h[0] is poisoned
2306 store i32 %poison, i32* @g ; Poison value stored to memory.
2307 %poison2 = load i32* @g ; Poison value loaded back from memory.
2309 store volatile i32 %poison, i32* @g ; External observation; undefined behavior.
2311 %narrowaddr = bitcast i32* @g to i16*
2312 %wideaddr = bitcast i32* @g to i64*
2313 %poison3 = load i16* %narrowaddr ; Returns a poison value.
2314 %poison4 = load i64* %wideaddr ; Returns a poison value.
2316 %cmp = icmp slt i32 %poison, 0 ; Returns a poison value.
2317 br i1 %cmp, label %true, label %end ; Branch to either destination.
2320 store volatile i32 0, i32* @g ; This is control-dependent on %cmp, so
2321 ; it has undefined behavior.
2325 %p = phi i32 [ 0, %entry ], [ 1, %true ]
2326 ; Both edges into this PHI are
2327 ; control-dependent on %cmp, so this
2328 ; always results in a poison value.
2330 store volatile i32 0, i32* @g ; This would depend on the store in %true
2331 ; if %cmp is true, or the store in %entry
2332 ; otherwise, so this is undefined behavior.
2334 br i1 %cmp, label %second_true, label %second_end
2335 ; The same branch again, but this time the
2336 ; true block doesn't have side effects.
2343 store volatile i32 0, i32* @g ; This time, the instruction always depends
2344 ; on the store in %end. Also, it is
2345 ; control-equivalent to %end, so this is
2346 ; well-defined (ignoring earlier undefined
2347 ; behavior in this example).
2351 Addresses of Basic Blocks
2352 -------------------------
2354 ``blockaddress(@function, %block)``
2356 The '``blockaddress``' constant computes the address of the specified
2357 basic block in the specified function, and always has an ``i8*`` type.
2358 Taking the address of the entry block is illegal.
2360 This value only has defined behavior when used as an operand to the
2361 ':ref:`indirectbr <i_indirectbr>`' instruction, or for comparisons
2362 against null. Pointer equality tests between labels addresses results in
2363 undefined behavior --- though, again, comparison against null is ok, and
2364 no label is equal to the null pointer. This may be passed around as an
2365 opaque pointer sized value as long as the bits are not inspected. This
2366 allows ``ptrtoint`` and arithmetic to be performed on these values so
2367 long as the original value is reconstituted before the ``indirectbr``
2370 Finally, some targets may provide defined semantics when using the value
2371 as the operand to an inline assembly, but that is target specific.
2375 Constant Expressions
2376 --------------------
2378 Constant expressions are used to allow expressions involving other
2379 constants to be used as constants. Constant expressions may be of any
2380 :ref:`first class <t_firstclass>` type and may involve any LLVM operation
2381 that does not have side effects (e.g. load and call are not supported).
2382 The following is the syntax for constant expressions:
2384 ``trunc (CST to TYPE)``
2385 Truncate a constant to another type. The bit size of CST must be
2386 larger than the bit size of TYPE. Both types must be integers.
2387 ``zext (CST to TYPE)``
2388 Zero extend a constant to another type. The bit size of CST must be
2389 smaller than the bit size of TYPE. Both types must be integers.
2390 ``sext (CST to TYPE)``
2391 Sign extend a constant to another type. The bit size of CST must be
2392 smaller than the bit size of TYPE. Both types must be integers.
2393 ``fptrunc (CST to TYPE)``
2394 Truncate a floating point constant to another floating point type.
2395 The size of CST must be larger than the size of TYPE. Both types
2396 must be floating point.
2397 ``fpext (CST to TYPE)``
2398 Floating point extend a constant to another type. The size of CST
2399 must be smaller or equal to the size of TYPE. Both types must be
2401 ``fptoui (CST to TYPE)``
2402 Convert a floating point constant to the corresponding unsigned
2403 integer constant. TYPE must be a scalar or vector integer type. CST
2404 must be of scalar or vector floating point type. Both CST and TYPE
2405 must be scalars, or vectors of the same number of elements. If the
2406 value won't fit in the integer type, the results are undefined.
2407 ``fptosi (CST to TYPE)``
2408 Convert a floating point constant to the corresponding signed
2409 integer constant. TYPE must be a scalar or vector integer type. CST
2410 must be of scalar or vector floating point type. Both CST and TYPE
2411 must be scalars, or vectors of the same number of elements. If the
2412 value won't fit in the integer type, the results are undefined.
2413 ``uitofp (CST to TYPE)``
2414 Convert an unsigned integer constant to the corresponding floating
2415 point constant. TYPE must be a scalar or vector floating point type.
2416 CST must be of scalar or vector integer type. Both CST and TYPE must
2417 be scalars, or vectors of the same number of elements. If the value
2418 won't fit in the floating point type, the results are undefined.
2419 ``sitofp (CST to TYPE)``
2420 Convert a signed integer constant to the corresponding floating
2421 point constant. TYPE must be a scalar or vector floating point type.
2422 CST must be of scalar or vector integer type. Both CST and TYPE must
2423 be scalars, or vectors of the same number of elements. If the value
2424 won't fit in the floating point type, the results are undefined.
2425 ``ptrtoint (CST to TYPE)``
2426 Convert a pointer typed constant to the corresponding integer
2427 constant. ``TYPE`` must be an integer type. ``CST`` must be of
2428 pointer type. The ``CST`` value is zero extended, truncated, or
2429 unchanged to make it fit in ``TYPE``.
2430 ``inttoptr (CST to TYPE)``
2431 Convert an integer constant to a pointer constant. TYPE must be a
2432 pointer type. CST must be of integer type. The CST value is zero
2433 extended, truncated, or unchanged to make it fit in a pointer size.
2434 This one is *really* dangerous!
2435 ``bitcast (CST to TYPE)``
2436 Convert a constant, CST, to another TYPE. The constraints of the
2437 operands are the same as those for the :ref:`bitcast
2438 instruction <i_bitcast>`.
2439 ``addrspacecast (CST to TYPE)``
2440 Convert a constant pointer or constant vector of pointer, CST, to another
2441 TYPE in a different address space. The constraints of the operands are the
2442 same as those for the :ref:`addrspacecast instruction <i_addrspacecast>`.
2443 ``getelementptr (CSTPTR, IDX0, IDX1, ...)``, ``getelementptr inbounds (CSTPTR, IDX0, IDX1, ...)``
2444 Perform the :ref:`getelementptr operation <i_getelementptr>` on
2445 constants. As with the :ref:`getelementptr <i_getelementptr>`
2446 instruction, the index list may have zero or more indexes, which are
2447 required to make sense for the type of "CSTPTR".
2448 ``select (COND, VAL1, VAL2)``
2449 Perform the :ref:`select operation <i_select>` on constants.
2450 ``icmp COND (VAL1, VAL2)``
2451 Performs the :ref:`icmp operation <i_icmp>` on constants.
2452 ``fcmp COND (VAL1, VAL2)``
2453 Performs the :ref:`fcmp operation <i_fcmp>` on constants.
2454 ``extractelement (VAL, IDX)``
2455 Perform the :ref:`extractelement operation <i_extractelement>` on
2457 ``insertelement (VAL, ELT, IDX)``
2458 Perform the :ref:`insertelement operation <i_insertelement>` on
2460 ``shufflevector (VEC1, VEC2, IDXMASK)``
2461 Perform the :ref:`shufflevector operation <i_shufflevector>` on
2463 ``extractvalue (VAL, IDX0, IDX1, ...)``
2464 Perform the :ref:`extractvalue operation <i_extractvalue>` on
2465 constants. The index list is interpreted in a similar manner as
2466 indices in a ':ref:`getelementptr <i_getelementptr>`' operation. At
2467 least one index value must be specified.
2468 ``insertvalue (VAL, ELT, IDX0, IDX1, ...)``
2469 Perform the :ref:`insertvalue operation <i_insertvalue>` on constants.
2470 The index list is interpreted in a similar manner as indices in a
2471 ':ref:`getelementptr <i_getelementptr>`' operation. At least one index
2472 value must be specified.
2473 ``OPCODE (LHS, RHS)``
2474 Perform the specified operation of the LHS and RHS constants. OPCODE
2475 may be any of the :ref:`binary <binaryops>` or :ref:`bitwise
2476 binary <bitwiseops>` operations. The constraints on operands are
2477 the same as those for the corresponding instruction (e.g. no bitwise
2478 operations on floating point values are allowed).
2485 Inline Assembler Expressions
2486 ----------------------------
2488 LLVM supports inline assembler expressions (as opposed to :ref:`Module-Level
2489 Inline Assembly <moduleasm>`) through the use of a special value. This
2490 value represents the inline assembler as a string (containing the
2491 instructions to emit), a list of operand constraints (stored as a
2492 string), a flag that indicates whether or not the inline asm expression
2493 has side effects, and a flag indicating whether the function containing
2494 the asm needs to align its stack conservatively. An example inline
2495 assembler expression is:
2497 .. code-block:: llvm
2499 i32 (i32) asm "bswap $0", "=r,r"
2501 Inline assembler expressions may **only** be used as the callee operand
2502 of a :ref:`call <i_call>` or an :ref:`invoke <i_invoke>` instruction.
2503 Thus, typically we have:
2505 .. code-block:: llvm
2507 %X = call i32 asm "bswap $0", "=r,r"(i32 %Y)
2509 Inline asms with side effects not visible in the constraint list must be
2510 marked as having side effects. This is done through the use of the
2511 '``sideeffect``' keyword, like so:
2513 .. code-block:: llvm
2515 call void asm sideeffect "eieio", ""()
2517 In some cases inline asms will contain code that will not work unless
2518 the stack is aligned in some way, such as calls or SSE instructions on
2519 x86, yet will not contain code that does that alignment within the asm.
2520 The compiler should make conservative assumptions about what the asm
2521 might contain and should generate its usual stack alignment code in the
2522 prologue if the '``alignstack``' keyword is present:
2524 .. code-block:: llvm
2526 call void asm alignstack "eieio", ""()
2528 Inline asms also support using non-standard assembly dialects. The
2529 assumed dialect is ATT. When the '``inteldialect``' keyword is present,
2530 the inline asm is using the Intel dialect. Currently, ATT and Intel are
2531 the only supported dialects. An example is:
2533 .. code-block:: llvm
2535 call void asm inteldialect "eieio", ""()
2537 If multiple keywords appear the '``sideeffect``' keyword must come
2538 first, the '``alignstack``' keyword second and the '``inteldialect``'
2544 The call instructions that wrap inline asm nodes may have a
2545 "``!srcloc``" MDNode attached to it that contains a list of constant
2546 integers. If present, the code generator will use the integer as the
2547 location cookie value when report errors through the ``LLVMContext``
2548 error reporting mechanisms. This allows a front-end to correlate backend
2549 errors that occur with inline asm back to the source code that produced
2552 .. code-block:: llvm
2554 call void asm sideeffect "something bad", ""(), !srcloc !42
2556 !42 = !{ i32 1234567 }
2558 It is up to the front-end to make sense of the magic numbers it places
2559 in the IR. If the MDNode contains multiple constants, the code generator
2560 will use the one that corresponds to the line of the asm that the error
2565 Metadata Nodes and Metadata Strings
2566 -----------------------------------
2568 LLVM IR allows metadata to be attached to instructions in the program
2569 that can convey extra information about the code to the optimizers and
2570 code generator. One example application of metadata is source-level
2571 debug information. There are two metadata primitives: strings and nodes.
2572 All metadata has the ``metadata`` type and is identified in syntax by a
2573 preceding exclamation point ('``!``').
2575 A metadata string is a string surrounded by double quotes. It can
2576 contain any character by escaping non-printable characters with
2577 "``\xx``" where "``xx``" is the two digit hex code. For example:
2580 Metadata nodes are represented with notation similar to structure
2581 constants (a comma separated list of elements, surrounded by braces and
2582 preceded by an exclamation point). Metadata nodes can have any values as
2583 their operand. For example:
2585 .. code-block:: llvm
2587 !{ metadata !"test\00", i32 10}
2589 A :ref:`named metadata <namedmetadatastructure>` is a collection of
2590 metadata nodes, which can be looked up in the module symbol table. For
2593 .. code-block:: llvm
2595 !foo = metadata !{!4, !3}
2597 Metadata can be used as function arguments. Here ``llvm.dbg.value``
2598 function is using two metadata arguments:
2600 .. code-block:: llvm
2602 call void @llvm.dbg.value(metadata !24, i64 0, metadata !25)
2604 Metadata can be attached with an instruction. Here metadata ``!21`` is
2605 attached to the ``add`` instruction using the ``!dbg`` identifier:
2607 .. code-block:: llvm
2609 %indvar.next = add i64 %indvar, 1, !dbg !21
2611 More information about specific metadata nodes recognized by the
2612 optimizers and code generator is found below.
2617 In LLVM IR, memory does not have types, so LLVM's own type system is not
2618 suitable for doing TBAA. Instead, metadata is added to the IR to
2619 describe a type system of a higher level language. This can be used to
2620 implement typical C/C++ TBAA, but it can also be used to implement
2621 custom alias analysis behavior for other languages.
2623 The current metadata format is very simple. TBAA metadata nodes have up
2624 to three fields, e.g.:
2626 .. code-block:: llvm
2628 !0 = metadata !{ metadata !"an example type tree" }
2629 !1 = metadata !{ metadata !"int", metadata !0 }
2630 !2 = metadata !{ metadata !"float", metadata !0 }
2631 !3 = metadata !{ metadata !"const float", metadata !2, i64 1 }
2633 The first field is an identity field. It can be any value, usually a
2634 metadata string, which uniquely identifies the type. The most important
2635 name in the tree is the name of the root node. Two trees with different
2636 root node names are entirely disjoint, even if they have leaves with
2639 The second field identifies the type's parent node in the tree, or is
2640 null or omitted for a root node. A type is considered to alias all of
2641 its descendants and all of its ancestors in the tree. Also, a type is
2642 considered to alias all types in other trees, so that bitcode produced
2643 from multiple front-ends is handled conservatively.
2645 If the third field is present, it's an integer which if equal to 1
2646 indicates that the type is "constant" (meaning
2647 ``pointsToConstantMemory`` should return true; see `other useful
2648 AliasAnalysis methods <AliasAnalysis.html#OtherItfs>`_).
2650 '``tbaa.struct``' Metadata
2651 ^^^^^^^^^^^^^^^^^^^^^^^^^^
2653 The :ref:`llvm.memcpy <int_memcpy>` is often used to implement
2654 aggregate assignment operations in C and similar languages, however it
2655 is defined to copy a contiguous region of memory, which is more than
2656 strictly necessary for aggregate types which contain holes due to
2657 padding. Also, it doesn't contain any TBAA information about the fields
2660 ``!tbaa.struct`` metadata can describe which memory subregions in a
2661 memcpy are padding and what the TBAA tags of the struct are.
2663 The current metadata format is very simple. ``!tbaa.struct`` metadata
2664 nodes are a list of operands which are in conceptual groups of three.
2665 For each group of three, the first operand gives the byte offset of a
2666 field in bytes, the second gives its size in bytes, and the third gives
2669 .. code-block:: llvm
2671 !4 = metadata !{ i64 0, i64 4, metadata !1, i64 8, i64 4, metadata !2 }
2673 This describes a struct with two fields. The first is at offset 0 bytes
2674 with size 4 bytes, and has tbaa tag !1. The second is at offset 8 bytes
2675 and has size 4 bytes and has tbaa tag !2.
2677 Note that the fields need not be contiguous. In this example, there is a
2678 4 byte gap between the two fields. This gap represents padding which
2679 does not carry useful data and need not be preserved.
2681 '``fpmath``' Metadata
2682 ^^^^^^^^^^^^^^^^^^^^^
2684 ``fpmath`` metadata may be attached to any instruction of floating point
2685 type. It can be used to express the maximum acceptable error in the
2686 result of that instruction, in ULPs, thus potentially allowing the
2687 compiler to use a more efficient but less accurate method of computing
2688 it. ULP is defined as follows:
2690 If ``x`` is a real number that lies between two finite consecutive
2691 floating-point numbers ``a`` and ``b``, without being equal to one
2692 of them, then ``ulp(x) = |b - a|``, otherwise ``ulp(x)`` is the
2693 distance between the two non-equal finite floating-point numbers
2694 nearest ``x``. Moreover, ``ulp(NaN)`` is ``NaN``.
2696 The metadata node shall consist of a single positive floating point
2697 number representing the maximum relative error, for example:
2699 .. code-block:: llvm
2701 !0 = metadata !{ float 2.5 } ; maximum acceptable inaccuracy is 2.5 ULPs
2703 '``range``' Metadata
2704 ^^^^^^^^^^^^^^^^^^^^
2706 ``range`` metadata may be attached only to loads of integer types. It
2707 expresses the possible ranges the loaded value is in. The ranges are
2708 represented with a flattened list of integers. The loaded value is known
2709 to be in the union of the ranges defined by each consecutive pair. Each
2710 pair has the following properties:
2712 - The type must match the type loaded by the instruction.
2713 - The pair ``a,b`` represents the range ``[a,b)``.
2714 - Both ``a`` and ``b`` are constants.
2715 - The range is allowed to wrap.
2716 - The range should not represent the full or empty set. That is,
2719 In addition, the pairs must be in signed order of the lower bound and
2720 they must be non-contiguous.
2724 .. code-block:: llvm
2726 %a = load i8* %x, align 1, !range !0 ; Can only be 0 or 1
2727 %b = load i8* %y, align 1, !range !1 ; Can only be 255 (-1), 0 or 1
2728 %c = load i8* %z, align 1, !range !2 ; Can only be 0, 1, 3, 4 or 5
2729 %d = load i8* %z, align 1, !range !3 ; Can only be -2, -1, 3, 4 or 5
2731 !0 = metadata !{ i8 0, i8 2 }
2732 !1 = metadata !{ i8 255, i8 2 }
2733 !2 = metadata !{ i8 0, i8 2, i8 3, i8 6 }
2734 !3 = metadata !{ i8 -2, i8 0, i8 3, i8 6 }
2739 It is sometimes useful to attach information to loop constructs. Currently,
2740 loop metadata is implemented as metadata attached to the branch instruction
2741 in the loop latch block. This type of metadata refer to a metadata node that is
2742 guaranteed to be separate for each loop. The loop identifier metadata is
2743 specified with the name ``llvm.loop``.
2745 The loop identifier metadata is implemented using a metadata that refers to
2746 itself to avoid merging it with any other identifier metadata, e.g.,
2747 during module linkage or function inlining. That is, each loop should refer
2748 to their own identification metadata even if they reside in separate functions.
2749 The following example contains loop identifier metadata for two separate loop
2752 .. code-block:: llvm
2754 !0 = metadata !{ metadata !0 }
2755 !1 = metadata !{ metadata !1 }
2757 The loop identifier metadata can be used to specify additional per-loop
2758 metadata. Any operands after the first operand can be treated as user-defined
2759 metadata. For example the ``llvm.vectorizer.unroll`` metadata is understood
2760 by the loop vectorizer to indicate how many times to unroll the loop:
2762 .. code-block:: llvm
2764 br i1 %exitcond, label %._crit_edge, label %.lr.ph, !llvm.loop !0
2766 !0 = metadata !{ metadata !0, metadata !1 }
2767 !1 = metadata !{ metadata !"llvm.vectorizer.unroll", i32 2 }
2772 Metadata types used to annotate memory accesses with information helpful
2773 for optimizations are prefixed with ``llvm.mem``.
2775 '``llvm.mem.parallel_loop_access``' Metadata
2776 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2778 For a loop to be parallel, in addition to using
2779 the ``llvm.loop`` metadata to mark the loop latch branch instruction,
2780 also all of the memory accessing instructions in the loop body need to be
2781 marked with the ``llvm.mem.parallel_loop_access`` metadata. If there
2782 is at least one memory accessing instruction not marked with the metadata,
2783 the loop must be considered a sequential loop. This causes parallel loops to be
2784 converted to sequential loops due to optimization passes that are unaware of
2785 the parallel semantics and that insert new memory instructions to the loop
2788 Example of a loop that is considered parallel due to its correct use of
2789 both ``llvm.loop`` and ``llvm.mem.parallel_loop_access``
2790 metadata types that refer to the same loop identifier metadata.
2792 .. code-block:: llvm
2796 %val0 = load i32* %arrayidx, !llvm.mem.parallel_loop_access !0
2798 store i32 %val0, i32* %arrayidx1, !llvm.mem.parallel_loop_access !0
2800 br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !0
2804 !0 = metadata !{ metadata !0 }
2806 It is also possible to have nested parallel loops. In that case the
2807 memory accesses refer to a list of loop identifier metadata nodes instead of
2808 the loop identifier metadata node directly:
2810 .. code-block:: llvm
2814 %val1 = load i32* %arrayidx3, !llvm.mem.parallel_loop_access !2
2816 br label %inner.for.body
2820 %val0 = load i32* %arrayidx1, !llvm.mem.parallel_loop_access !0
2822 store i32 %val0, i32* %arrayidx2, !llvm.mem.parallel_loop_access !0
2824 br i1 %exitcond, label %inner.for.end, label %inner.for.body, !llvm.loop !1
2828 store i32 %val1, i32* %arrayidx4, !llvm.mem.parallel_loop_access !2
2830 br i1 %exitcond, label %outer.for.end, label %outer.for.body, !llvm.loop !2
2832 outer.for.end: ; preds = %for.body
2834 !0 = metadata !{ metadata !1, metadata !2 } ; a list of loop identifiers
2835 !1 = metadata !{ metadata !1 } ; an identifier for the inner loop
2836 !2 = metadata !{ metadata !2 } ; an identifier for the outer loop
2838 '``llvm.vectorizer``'
2839 ^^^^^^^^^^^^^^^^^^^^^
2841 Metadata prefixed with ``llvm.vectorizer`` is used to control per-loop
2842 vectorization parameters such as vectorization factor and unroll factor.
2844 ``llvm.vectorizer`` metadata should be used in conjunction with ``llvm.loop``
2845 loop identification metadata.
2847 '``llvm.vectorizer.unroll``' Metadata
2848 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2850 This metadata instructs the loop vectorizer to unroll the specified
2851 loop exactly ``N`` times.
2853 The first operand is the string ``llvm.vectorizer.unroll`` and the second
2854 operand is an integer specifying the unroll factor. For example:
2856 .. code-block:: llvm
2858 !0 = metadata !{ metadata !"llvm.vectorizer.unroll", i32 4 }
2860 Note that setting ``llvm.vectorizer.unroll`` to 1 disables unrolling of the
2863 If ``llvm.vectorizer.unroll`` is set to 0 then the amount of unrolling will be
2864 determined automatically.
2866 '``llvm.vectorizer.width``' Metadata
2867 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2869 This metadata sets the target width of the vectorizer to ``N``. Without
2870 this metadata, the vectorizer will choose a width automatically.
2871 Regardless of this metadata, the vectorizer will only vectorize loops if
2872 it believes it is valid to do so.
2874 The first operand is the string ``llvm.vectorizer.width`` and the second
2875 operand is an integer specifying the width. For example:
2877 .. code-block:: llvm
2879 !0 = metadata !{ metadata !"llvm.vectorizer.width", i32 4 }
2881 Note that setting ``llvm.vectorizer.width`` to 1 disables vectorization of the
2884 If ``llvm.vectorizer.width`` is set to 0 then the width will be determined
2887 Module Flags Metadata
2888 =====================
2890 Information about the module as a whole is difficult to convey to LLVM's
2891 subsystems. The LLVM IR isn't sufficient to transmit this information.
2892 The ``llvm.module.flags`` named metadata exists in order to facilitate
2893 this. These flags are in the form of key / value pairs --- much like a
2894 dictionary --- making it easy for any subsystem who cares about a flag to
2897 The ``llvm.module.flags`` metadata contains a list of metadata triplets.
2898 Each triplet has the following form:
2900 - The first element is a *behavior* flag, which specifies the behavior
2901 when two (or more) modules are merged together, and it encounters two
2902 (or more) metadata with the same ID. The supported behaviors are
2904 - The second element is a metadata string that is a unique ID for the
2905 metadata. Each module may only have one flag entry for each unique ID (not
2906 including entries with the **Require** behavior).
2907 - The third element is the value of the flag.
2909 When two (or more) modules are merged together, the resulting
2910 ``llvm.module.flags`` metadata is the union of the modules' flags. That is, for
2911 each unique metadata ID string, there will be exactly one entry in the merged
2912 modules ``llvm.module.flags`` metadata table, and the value for that entry will
2913 be determined by the merge behavior flag, as described below. The only exception
2914 is that entries with the *Require* behavior are always preserved.
2916 The following behaviors are supported:
2927 Emits an error if two values disagree, otherwise the resulting value
2928 is that of the operands.
2932 Emits a warning if two values disagree. The result value will be the
2933 operand for the flag from the first module being linked.
2937 Adds a requirement that another module flag be present and have a
2938 specified value after linking is performed. The value must be a
2939 metadata pair, where the first element of the pair is the ID of the
2940 module flag to be restricted, and the second element of the pair is
2941 the value the module flag should be restricted to. This behavior can
2942 be used to restrict the allowable results (via triggering of an
2943 error) of linking IDs with the **Override** behavior.
2947 Uses the specified value, regardless of the behavior or value of the
2948 other module. If both modules specify **Override**, but the values
2949 differ, an error will be emitted.
2953 Appends the two values, which are required to be metadata nodes.
2957 Appends the two values, which are required to be metadata
2958 nodes. However, duplicate entries in the second list are dropped
2959 during the append operation.
2961 It is an error for a particular unique flag ID to have multiple behaviors,
2962 except in the case of **Require** (which adds restrictions on another metadata
2963 value) or **Override**.
2965 An example of module flags:
2967 .. code-block:: llvm
2969 !0 = metadata !{ i32 1, metadata !"foo", i32 1 }
2970 !1 = metadata !{ i32 4, metadata !"bar", i32 37 }
2971 !2 = metadata !{ i32 2, metadata !"qux", i32 42 }
2972 !3 = metadata !{ i32 3, metadata !"qux",
2974 metadata !"foo", i32 1
2977 !llvm.module.flags = !{ !0, !1, !2, !3 }
2979 - Metadata ``!0`` has the ID ``!"foo"`` and the value '1'. The behavior
2980 if two or more ``!"foo"`` flags are seen is to emit an error if their
2981 values are not equal.
2983 - Metadata ``!1`` has the ID ``!"bar"`` and the value '37'. The
2984 behavior if two or more ``!"bar"`` flags are seen is to use the value
2987 - Metadata ``!2`` has the ID ``!"qux"`` and the value '42'. The
2988 behavior if two or more ``!"qux"`` flags are seen is to emit a
2989 warning if their values are not equal.
2991 - Metadata ``!3`` has the ID ``!"qux"`` and the value:
2995 metadata !{ metadata !"foo", i32 1 }
2997 The behavior is to emit an error if the ``llvm.module.flags`` does not
2998 contain a flag with the ID ``!"foo"`` that has the value '1' after linking is
3001 Objective-C Garbage Collection Module Flags Metadata
3002 ----------------------------------------------------
3004 On the Mach-O platform, Objective-C stores metadata about garbage
3005 collection in a special section called "image info". The metadata
3006 consists of a version number and a bitmask specifying what types of
3007 garbage collection are supported (if any) by the file. If two or more
3008 modules are linked together their garbage collection metadata needs to
3009 be merged rather than appended together.
3011 The Objective-C garbage collection module flags metadata consists of the
3012 following key-value pairs:
3021 * - ``Objective-C Version``
3022 - **[Required]** --- The Objective-C ABI version. Valid values are 1 and 2.
3024 * - ``Objective-C Image Info Version``
3025 - **[Required]** --- The version of the image info section. Currently
3028 * - ``Objective-C Image Info Section``
3029 - **[Required]** --- The section to place the metadata. Valid values are
3030 ``"__OBJC, __image_info, regular"`` for Objective-C ABI version 1, and
3031 ``"__DATA,__objc_imageinfo, regular, no_dead_strip"`` for
3032 Objective-C ABI version 2.
3034 * - ``Objective-C Garbage Collection``
3035 - **[Required]** --- Specifies whether garbage collection is supported or
3036 not. Valid values are 0, for no garbage collection, and 2, for garbage
3037 collection supported.
3039 * - ``Objective-C GC Only``
3040 - **[Optional]** --- Specifies that only garbage collection is supported.
3041 If present, its value must be 6. This flag requires that the
3042 ``Objective-C Garbage Collection`` flag have the value 2.
3044 Some important flag interactions:
3046 - If a module with ``Objective-C Garbage Collection`` set to 0 is
3047 merged with a module with ``Objective-C Garbage Collection`` set to
3048 2, then the resulting module has the
3049 ``Objective-C Garbage Collection`` flag set to 0.
3050 - A module with ``Objective-C Garbage Collection`` set to 0 cannot be
3051 merged with a module with ``Objective-C GC Only`` set to 6.
3053 Automatic Linker Flags Module Flags Metadata
3054 --------------------------------------------
3056 Some targets support embedding flags to the linker inside individual object
3057 files. Typically this is used in conjunction with language extensions which
3058 allow source files to explicitly declare the libraries they depend on, and have
3059 these automatically be transmitted to the linker via object files.
3061 These flags are encoded in the IR using metadata in the module flags section,
3062 using the ``Linker Options`` key. The merge behavior for this flag is required
3063 to be ``AppendUnique``, and the value for the key is expected to be a metadata
3064 node which should be a list of other metadata nodes, each of which should be a
3065 list of metadata strings defining linker options.
3067 For example, the following metadata section specifies two separate sets of
3068 linker options, presumably to link against ``libz`` and the ``Cocoa``
3071 !0 = metadata !{ i32 6, metadata !"Linker Options",
3073 metadata !{ metadata !"-lz" },
3074 metadata !{ metadata !"-framework", metadata !"Cocoa" } } }
3075 !llvm.module.flags = !{ !0 }
3077 The metadata encoding as lists of lists of options, as opposed to a collapsed
3078 list of options, is chosen so that the IR encoding can use multiple option
3079 strings to specify e.g., a single library, while still having that specifier be
3080 preserved as an atomic element that can be recognized by a target specific
3081 assembly writer or object file emitter.
3083 Each individual option is required to be either a valid option for the target's
3084 linker, or an option that is reserved by the target specific assembly writer or
3085 object file emitter. No other aspect of these options is defined by the IR.
3087 .. _intrinsicglobalvariables:
3089 Intrinsic Global Variables
3090 ==========================
3092 LLVM has a number of "magic" global variables that contain data that
3093 affect code generation or other IR semantics. These are documented here.
3094 All globals of this sort should have a section specified as
3095 "``llvm.metadata``". This section and all globals that start with
3096 "``llvm.``" are reserved for use by LLVM.
3100 The '``llvm.used``' Global Variable
3101 -----------------------------------
3103 The ``@llvm.used`` global is an array which has
3104 :ref:`appending linkage <linkage_appending>`. This array contains a list of
3105 pointers to named global variables, functions and aliases which may optionally
3106 have a pointer cast formed of bitcast or getelementptr. For example, a legal
3109 .. code-block:: llvm
3114 @llvm.used = appending global [2 x i8*] [
3116 i8* bitcast (i32* @Y to i8*)
3117 ], section "llvm.metadata"
3119 If a symbol appears in the ``@llvm.used`` list, then the compiler, assembler,
3120 and linker are required to treat the symbol as if there is a reference to the
3121 symbol that it cannot see (which is why they have to be named). For example, if
3122 a variable has internal linkage and no references other than that from the
3123 ``@llvm.used`` list, it cannot be deleted. This is commonly used to represent
3124 references from inline asms and other things the compiler cannot "see", and
3125 corresponds to "``attribute((used))``" in GNU C.
3127 On some targets, the code generator must emit a directive to the
3128 assembler or object file to prevent the assembler and linker from
3129 molesting the symbol.
3131 .. _gv_llvmcompilerused:
3133 The '``llvm.compiler.used``' Global Variable
3134 --------------------------------------------
3136 The ``@llvm.compiler.used`` directive is the same as the ``@llvm.used``
3137 directive, except that it only prevents the compiler from touching the
3138 symbol. On targets that support it, this allows an intelligent linker to
3139 optimize references to the symbol without being impeded as it would be
3142 This is a rare construct that should only be used in rare circumstances,
3143 and should not be exposed to source languages.
3145 .. _gv_llvmglobalctors:
3147 The '``llvm.global_ctors``' Global Variable
3148 -------------------------------------------
3150 .. code-block:: llvm
3152 %0 = type { i32, void ()* }
3153 @llvm.global_ctors = appending global [1 x %0] [%0 { i32 65535, void ()* @ctor }]
3155 The ``@llvm.global_ctors`` array contains a list of constructor
3156 functions and associated priorities. The functions referenced by this
3157 array will be called in ascending order of priority (i.e. lowest first)
3158 when the module is loaded. The order of functions with the same priority
3161 .. _llvmglobaldtors:
3163 The '``llvm.global_dtors``' Global Variable
3164 -------------------------------------------
3166 .. code-block:: llvm
3168 %0 = type { i32, void ()* }
3169 @llvm.global_dtors = appending global [1 x %0] [%0 { i32 65535, void ()* @dtor }]
3171 The ``@llvm.global_dtors`` array contains a list of destructor functions
3172 and associated priorities. The functions referenced by this array will
3173 be called in descending order of priority (i.e. highest first) when the
3174 module is loaded. The order of functions with the same priority is not
3177 Instruction Reference
3178 =====================
3180 The LLVM instruction set consists of several different classifications
3181 of instructions: :ref:`terminator instructions <terminators>`, :ref:`binary
3182 instructions <binaryops>`, :ref:`bitwise binary
3183 instructions <bitwiseops>`, :ref:`memory instructions <memoryops>`, and
3184 :ref:`other instructions <otherops>`.
3188 Terminator Instructions
3189 -----------------------
3191 As mentioned :ref:`previously <functionstructure>`, every basic block in a
3192 program ends with a "Terminator" instruction, which indicates which
3193 block should be executed after the current block is finished. These
3194 terminator instructions typically yield a '``void``' value: they produce
3195 control flow, not values (the one exception being the
3196 ':ref:`invoke <i_invoke>`' instruction).
3198 The terminator instructions are: ':ref:`ret <i_ret>`',
3199 ':ref:`br <i_br>`', ':ref:`switch <i_switch>`',
3200 ':ref:`indirectbr <i_indirectbr>`', ':ref:`invoke <i_invoke>`',
3201 ':ref:`resume <i_resume>`', and ':ref:`unreachable <i_unreachable>`'.
3205 '``ret``' Instruction
3206 ^^^^^^^^^^^^^^^^^^^^^
3213 ret <type> <value> ; Return a value from a non-void function
3214 ret void ; Return from void function
3219 The '``ret``' instruction is used to return control flow (and optionally
3220 a value) from a function back to the caller.
3222 There are two forms of the '``ret``' instruction: one that returns a
3223 value and then causes control flow, and one that just causes control
3229 The '``ret``' instruction optionally accepts a single argument, the
3230 return value. The type of the return value must be a ':ref:`first
3231 class <t_firstclass>`' type.
3233 A function is not :ref:`well formed <wellformed>` if it it has a non-void
3234 return type and contains a '``ret``' instruction with no return value or
3235 a return value with a type that does not match its type, or if it has a
3236 void return type and contains a '``ret``' instruction with a return
3242 When the '``ret``' instruction is executed, control flow returns back to
3243 the calling function's context. If the caller is a
3244 ":ref:`call <i_call>`" instruction, execution continues at the
3245 instruction after the call. If the caller was an
3246 ":ref:`invoke <i_invoke>`" instruction, execution continues at the
3247 beginning of the "normal" destination block. If the instruction returns
3248 a value, that value shall set the call or invoke instruction's return
3254 .. code-block:: llvm
3256 ret i32 5 ; Return an integer value of 5
3257 ret void ; Return from a void function
3258 ret { i32, i8 } { i32 4, i8 2 } ; Return a struct of values 4 and 2
3262 '``br``' Instruction
3263 ^^^^^^^^^^^^^^^^^^^^
3270 br i1 <cond>, label <iftrue>, label <iffalse>
3271 br label <dest> ; Unconditional branch
3276 The '``br``' instruction is used to cause control flow to transfer to a
3277 different basic block in the current function. There are two forms of
3278 this instruction, corresponding to a conditional branch and an
3279 unconditional branch.
3284 The conditional branch form of the '``br``' instruction takes a single
3285 '``i1``' value and two '``label``' values. The unconditional form of the
3286 '``br``' instruction takes a single '``label``' value as a target.
3291 Upon execution of a conditional '``br``' instruction, the '``i1``'
3292 argument is evaluated. If the value is ``true``, control flows to the
3293 '``iftrue``' ``label`` argument. If "cond" is ``false``, control flows
3294 to the '``iffalse``' ``label`` argument.
3299 .. code-block:: llvm
3302 %cond = icmp eq i32 %a, %b
3303 br i1 %cond, label %IfEqual, label %IfUnequal
3311 '``switch``' Instruction
3312 ^^^^^^^^^^^^^^^^^^^^^^^^
3319 switch <intty> <value>, label <defaultdest> [ <intty> <val>, label <dest> ... ]
3324 The '``switch``' instruction is used to transfer control flow to one of
3325 several different places. It is a generalization of the '``br``'
3326 instruction, allowing a branch to occur to one of many possible
3332 The '``switch``' instruction uses three parameters: an integer
3333 comparison value '``value``', a default '``label``' destination, and an
3334 array of pairs of comparison value constants and '``label``'s. The table
3335 is not allowed to contain duplicate constant entries.
3340 The ``switch`` instruction specifies a table of values and destinations.
3341 When the '``switch``' instruction is executed, this table is searched
3342 for the given value. If the value is found, control flow is transferred
3343 to the corresponding destination; otherwise, control flow is transferred
3344 to the default destination.
3349 Depending on properties of the target machine and the particular
3350 ``switch`` instruction, this instruction may be code generated in
3351 different ways. For example, it could be generated as a series of
3352 chained conditional branches or with a lookup table.
3357 .. code-block:: llvm
3359 ; Emulate a conditional br instruction
3360 %Val = zext i1 %value to i32
3361 switch i32 %Val, label %truedest [ i32 0, label %falsedest ]
3363 ; Emulate an unconditional br instruction
3364 switch i32 0, label %dest [ ]
3366 ; Implement a jump table:
3367 switch i32 %val, label %otherwise [ i32 0, label %onzero
3369 i32 2, label %ontwo ]
3373 '``indirectbr``' Instruction
3374 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3381 indirectbr <somety>* <address>, [ label <dest1>, label <dest2>, ... ]
3386 The '``indirectbr``' instruction implements an indirect branch to a
3387 label within the current function, whose address is specified by
3388 "``address``". Address must be derived from a
3389 :ref:`blockaddress <blockaddress>` constant.
3394 The '``address``' argument is the address of the label to jump to. The
3395 rest of the arguments indicate the full set of possible destinations
3396 that the address may point to. Blocks are allowed to occur multiple
3397 times in the destination list, though this isn't particularly useful.
3399 This destination list is required so that dataflow analysis has an
3400 accurate understanding of the CFG.
3405 Control transfers to the block specified in the address argument. All
3406 possible destination blocks must be listed in the label list, otherwise
3407 this instruction has undefined behavior. This implies that jumps to
3408 labels defined in other functions have undefined behavior as well.
3413 This is typically implemented with a jump through a register.
3418 .. code-block:: llvm
3420 indirectbr i8* %Addr, [ label %bb1, label %bb2, label %bb3 ]
3424 '``invoke``' Instruction
3425 ^^^^^^^^^^^^^^^^^^^^^^^^
3432 <result> = invoke [cconv] [ret attrs] <ptr to function ty> <function ptr val>(<function args>) [fn attrs]
3433 to label <normal label> unwind label <exception label>
3438 The '``invoke``' instruction causes control to transfer to a specified
3439 function, with the possibility of control flow transfer to either the
3440 '``normal``' label or the '``exception``' label. If the callee function
3441 returns with the "``ret``" instruction, control flow will return to the
3442 "normal" label. If the callee (or any indirect callees) returns via the
3443 ":ref:`resume <i_resume>`" instruction or other exception handling
3444 mechanism, control is interrupted and continued at the dynamically
3445 nearest "exception" label.
3447 The '``exception``' label is a `landing
3448 pad <ExceptionHandling.html#overview>`_ for the exception. As such,
3449 '``exception``' label is required to have the
3450 ":ref:`landingpad <i_landingpad>`" instruction, which contains the
3451 information about the behavior of the program after unwinding happens,
3452 as its first non-PHI instruction. The restrictions on the
3453 "``landingpad``" instruction's tightly couples it to the "``invoke``"
3454 instruction, so that the important information contained within the
3455 "``landingpad``" instruction can't be lost through normal code motion.
3460 This instruction requires several arguments:
3462 #. The optional "cconv" marker indicates which :ref:`calling
3463 convention <callingconv>` the call should use. If none is
3464 specified, the call defaults to using C calling conventions.
3465 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
3466 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
3468 #. '``ptr to function ty``': shall be the signature of the pointer to
3469 function value being invoked. In most cases, this is a direct
3470 function invocation, but indirect ``invoke``'s are just as possible,
3471 branching off an arbitrary pointer to function value.
3472 #. '``function ptr val``': An LLVM value containing a pointer to a
3473 function to be invoked.
3474 #. '``function args``': argument list whose types match the function
3475 signature argument types and parameter attributes. All arguments must
3476 be of :ref:`first class <t_firstclass>` type. If the function signature
3477 indicates the function accepts a variable number of arguments, the
3478 extra arguments can be specified.
3479 #. '``normal label``': the label reached when the called function
3480 executes a '``ret``' instruction.
3481 #. '``exception label``': the label reached when a callee returns via
3482 the :ref:`resume <i_resume>` instruction or other exception handling
3484 #. The optional :ref:`function attributes <fnattrs>` list. Only
3485 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
3486 attributes are valid here.
3491 This instruction is designed to operate as a standard '``call``'
3492 instruction in most regards. The primary difference is that it
3493 establishes an association with a label, which is used by the runtime
3494 library to unwind the stack.
3496 This instruction is used in languages with destructors to ensure that
3497 proper cleanup is performed in the case of either a ``longjmp`` or a
3498 thrown exception. Additionally, this is important for implementation of
3499 '``catch``' clauses in high-level languages that support them.
3501 For the purposes of the SSA form, the definition of the value returned
3502 by the '``invoke``' instruction is deemed to occur on the edge from the
3503 current block to the "normal" label. If the callee unwinds then no
3504 return value is available.
3509 .. code-block:: llvm
3511 %retval = invoke i32 @Test(i32 15) to label %Continue
3512 unwind label %TestCleanup ; {i32}:retval set
3513 %retval = invoke coldcc i32 %Testfnptr(i32 15) to label %Continue
3514 unwind label %TestCleanup ; {i32}:retval set
3518 '``resume``' Instruction
3519 ^^^^^^^^^^^^^^^^^^^^^^^^
3526 resume <type> <value>
3531 The '``resume``' instruction is a terminator instruction that has no
3537 The '``resume``' instruction requires one argument, which must have the
3538 same type as the result of any '``landingpad``' instruction in the same
3544 The '``resume``' instruction resumes propagation of an existing
3545 (in-flight) exception whose unwinding was interrupted with a
3546 :ref:`landingpad <i_landingpad>` instruction.
3551 .. code-block:: llvm
3553 resume { i8*, i32 } %exn
3557 '``unreachable``' Instruction
3558 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3570 The '``unreachable``' instruction has no defined semantics. This
3571 instruction is used to inform the optimizer that a particular portion of
3572 the code is not reachable. This can be used to indicate that the code
3573 after a no-return function cannot be reached, and other facts.
3578 The '``unreachable``' instruction has no defined semantics.
3585 Binary operators are used to do most of the computation in a program.
3586 They require two operands of the same type, execute an operation on
3587 them, and produce a single value. The operands might represent multiple
3588 data, as is the case with the :ref:`vector <t_vector>` data type. The
3589 result value has the same type as its operands.
3591 There are several different binary operators:
3595 '``add``' Instruction
3596 ^^^^^^^^^^^^^^^^^^^^^
3603 <result> = add <ty> <op1>, <op2> ; yields {ty}:result
3604 <result> = add nuw <ty> <op1>, <op2> ; yields {ty}:result
3605 <result> = add nsw <ty> <op1>, <op2> ; yields {ty}:result
3606 <result> = add nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3611 The '``add``' instruction returns the sum of its two operands.
3616 The two arguments to the '``add``' instruction must be
3617 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3618 arguments must have identical types.
3623 The value produced is the integer sum of the two operands.
3625 If the sum has unsigned overflow, the result returned is the
3626 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
3629 Because LLVM integers use a two's complement representation, this
3630 instruction is appropriate for both signed and unsigned integers.
3632 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3633 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3634 result value of the ``add`` is a :ref:`poison value <poisonvalues>` if
3635 unsigned and/or signed overflow, respectively, occurs.
3640 .. code-block:: llvm
3642 <result> = add i32 4, %var ; yields {i32}:result = 4 + %var
3646 '``fadd``' Instruction
3647 ^^^^^^^^^^^^^^^^^^^^^^
3654 <result> = fadd [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3659 The '``fadd``' instruction returns the sum of its two operands.
3664 The two arguments to the '``fadd``' instruction must be :ref:`floating
3665 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3666 Both arguments must have identical types.
3671 The value produced is the floating point sum of the two operands. This
3672 instruction can also take any number of :ref:`fast-math flags <fastmath>`,
3673 which are optimization hints to enable otherwise unsafe floating point
3679 .. code-block:: llvm
3681 <result> = fadd float 4.0, %var ; yields {float}:result = 4.0 + %var
3683 '``sub``' Instruction
3684 ^^^^^^^^^^^^^^^^^^^^^
3691 <result> = sub <ty> <op1>, <op2> ; yields {ty}:result
3692 <result> = sub nuw <ty> <op1>, <op2> ; yields {ty}:result
3693 <result> = sub nsw <ty> <op1>, <op2> ; yields {ty}:result
3694 <result> = sub nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3699 The '``sub``' instruction returns the difference of its two operands.
3701 Note that the '``sub``' instruction is used to represent the '``neg``'
3702 instruction present in most other intermediate representations.
3707 The two arguments to the '``sub``' instruction must be
3708 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3709 arguments must have identical types.
3714 The value produced is the integer difference of the two operands.
3716 If the difference has unsigned overflow, the result returned is the
3717 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
3720 Because LLVM integers use a two's complement representation, this
3721 instruction is appropriate for both signed and unsigned integers.
3723 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3724 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3725 result value of the ``sub`` is a :ref:`poison value <poisonvalues>` if
3726 unsigned and/or signed overflow, respectively, occurs.
3731 .. code-block:: llvm
3733 <result> = sub i32 4, %var ; yields {i32}:result = 4 - %var
3734 <result> = sub i32 0, %val ; yields {i32}:result = -%var
3738 '``fsub``' Instruction
3739 ^^^^^^^^^^^^^^^^^^^^^^
3746 <result> = fsub [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3751 The '``fsub``' instruction returns the difference of its two operands.
3753 Note that the '``fsub``' instruction is used to represent the '``fneg``'
3754 instruction present in most other intermediate representations.
3759 The two arguments to the '``fsub``' instruction must be :ref:`floating
3760 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3761 Both arguments must have identical types.
3766 The value produced is the floating point difference of the two operands.
3767 This instruction can also take any number of :ref:`fast-math
3768 flags <fastmath>`, which are optimization hints to enable otherwise
3769 unsafe floating point optimizations:
3774 .. code-block:: llvm
3776 <result> = fsub float 4.0, %var ; yields {float}:result = 4.0 - %var
3777 <result> = fsub float -0.0, %val ; yields {float}:result = -%var
3779 '``mul``' Instruction
3780 ^^^^^^^^^^^^^^^^^^^^^
3787 <result> = mul <ty> <op1>, <op2> ; yields {ty}:result
3788 <result> = mul nuw <ty> <op1>, <op2> ; yields {ty}:result
3789 <result> = mul nsw <ty> <op1>, <op2> ; yields {ty}:result
3790 <result> = mul nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3795 The '``mul``' instruction returns the product of its two operands.
3800 The two arguments to the '``mul``' instruction must be
3801 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3802 arguments must have identical types.
3807 The value produced is the integer product of the two operands.
3809 If the result of the multiplication has unsigned overflow, the result
3810 returned is the mathematical result modulo 2\ :sup:`n`\ , where n is the
3811 bit width of the result.
3813 Because LLVM integers use a two's complement representation, and the
3814 result is the same width as the operands, this instruction returns the
3815 correct result for both signed and unsigned integers. If a full product
3816 (e.g. ``i32`` * ``i32`` -> ``i64``) is needed, the operands should be
3817 sign-extended or zero-extended as appropriate to the width of the full
3820 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3821 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3822 result value of the ``mul`` is a :ref:`poison value <poisonvalues>` if
3823 unsigned and/or signed overflow, respectively, occurs.
3828 .. code-block:: llvm
3830 <result> = mul i32 4, %var ; yields {i32}:result = 4 * %var
3834 '``fmul``' Instruction
3835 ^^^^^^^^^^^^^^^^^^^^^^
3842 <result> = fmul [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3847 The '``fmul``' instruction returns the product of its two operands.
3852 The two arguments to the '``fmul``' instruction must be :ref:`floating
3853 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3854 Both arguments must have identical types.
3859 The value produced is the floating point product of the two operands.
3860 This instruction can also take any number of :ref:`fast-math
3861 flags <fastmath>`, which are optimization hints to enable otherwise
3862 unsafe floating point optimizations:
3867 .. code-block:: llvm
3869 <result> = fmul float 4.0, %var ; yields {float}:result = 4.0 * %var
3871 '``udiv``' Instruction
3872 ^^^^^^^^^^^^^^^^^^^^^^
3879 <result> = udiv <ty> <op1>, <op2> ; yields {ty}:result
3880 <result> = udiv exact <ty> <op1>, <op2> ; yields {ty}:result
3885 The '``udiv``' instruction returns the quotient of its two operands.
3890 The two arguments to the '``udiv``' instruction must be
3891 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3892 arguments must have identical types.
3897 The value produced is the unsigned integer quotient of the two operands.
3899 Note that unsigned integer division and signed integer division are
3900 distinct operations; for signed integer division, use '``sdiv``'.
3902 Division by zero leads to undefined behavior.
3904 If the ``exact`` keyword is present, the result value of the ``udiv`` is
3905 a :ref:`poison value <poisonvalues>` if %op1 is not a multiple of %op2 (as
3906 such, "((a udiv exact b) mul b) == a").
3911 .. code-block:: llvm
3913 <result> = udiv i32 4, %var ; yields {i32}:result = 4 / %var
3915 '``sdiv``' Instruction
3916 ^^^^^^^^^^^^^^^^^^^^^^
3923 <result> = sdiv <ty> <op1>, <op2> ; yields {ty}:result
3924 <result> = sdiv exact <ty> <op1>, <op2> ; yields {ty}:result
3929 The '``sdiv``' instruction returns the quotient of its two operands.
3934 The two arguments to the '``sdiv``' instruction must be
3935 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3936 arguments must have identical types.
3941 The value produced is the signed integer quotient of the two operands
3942 rounded towards zero.
3944 Note that signed integer division and unsigned integer division are
3945 distinct operations; for unsigned integer division, use '``udiv``'.
3947 Division by zero leads to undefined behavior. Overflow also leads to
3948 undefined behavior; this is a rare case, but can occur, for example, by
3949 doing a 32-bit division of -2147483648 by -1.
3951 If the ``exact`` keyword is present, the result value of the ``sdiv`` is
3952 a :ref:`poison value <poisonvalues>` if the result would be rounded.
3957 .. code-block:: llvm
3959 <result> = sdiv i32 4, %var ; yields {i32}:result = 4 / %var
3963 '``fdiv``' Instruction
3964 ^^^^^^^^^^^^^^^^^^^^^^
3971 <result> = fdiv [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3976 The '``fdiv``' instruction returns the quotient of its two operands.
3981 The two arguments to the '``fdiv``' instruction must be :ref:`floating
3982 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3983 Both arguments must have identical types.
3988 The value produced is the floating point quotient of the two operands.
3989 This instruction can also take any number of :ref:`fast-math
3990 flags <fastmath>`, which are optimization hints to enable otherwise
3991 unsafe floating point optimizations:
3996 .. code-block:: llvm
3998 <result> = fdiv float 4.0, %var ; yields {float}:result = 4.0 / %var
4000 '``urem``' Instruction
4001 ^^^^^^^^^^^^^^^^^^^^^^
4008 <result> = urem <ty> <op1>, <op2> ; yields {ty}:result
4013 The '``urem``' instruction returns the remainder from the unsigned
4014 division of its two arguments.
4019 The two arguments to the '``urem``' instruction must be
4020 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4021 arguments must have identical types.
4026 This instruction returns the unsigned integer *remainder* of a division.
4027 This instruction always performs an unsigned division to get the
4030 Note that unsigned integer remainder and signed integer remainder are
4031 distinct operations; for signed integer remainder, use '``srem``'.
4033 Taking the remainder of a division by zero leads to undefined behavior.
4038 .. code-block:: llvm
4040 <result> = urem i32 4, %var ; yields {i32}:result = 4 % %var
4042 '``srem``' Instruction
4043 ^^^^^^^^^^^^^^^^^^^^^^
4050 <result> = srem <ty> <op1>, <op2> ; yields {ty}:result
4055 The '``srem``' instruction returns the remainder from the signed
4056 division of its two operands. This instruction can also take
4057 :ref:`vector <t_vector>` versions of the values in which case the elements
4063 The two arguments to the '``srem``' instruction must be
4064 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4065 arguments must have identical types.
4070 This instruction returns the *remainder* of a division (where the result
4071 is either zero or has the same sign as the dividend, ``op1``), not the
4072 *modulo* operator (where the result is either zero or has the same sign
4073 as the divisor, ``op2``) of a value. For more information about the
4074 difference, see `The Math
4075 Forum <http://mathforum.org/dr.math/problems/anne.4.28.99.html>`_. For a
4076 table of how this is implemented in various languages, please see
4078 operation <http://en.wikipedia.org/wiki/Modulo_operation>`_.
4080 Note that signed integer remainder and unsigned integer remainder are
4081 distinct operations; for unsigned integer remainder, use '``urem``'.
4083 Taking the remainder of a division by zero leads to undefined behavior.
4084 Overflow also leads to undefined behavior; this is a rare case, but can
4085 occur, for example, by taking the remainder of a 32-bit division of
4086 -2147483648 by -1. (The remainder doesn't actually overflow, but this
4087 rule lets srem be implemented using instructions that return both the
4088 result of the division and the remainder.)
4093 .. code-block:: llvm
4095 <result> = srem i32 4, %var ; yields {i32}:result = 4 % %var
4099 '``frem``' Instruction
4100 ^^^^^^^^^^^^^^^^^^^^^^
4107 <result> = frem [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
4112 The '``frem``' instruction returns the remainder from the division of
4118 The two arguments to the '``frem``' instruction must be :ref:`floating
4119 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4120 Both arguments must have identical types.
4125 This instruction returns the *remainder* of a division. The remainder
4126 has the same sign as the dividend. This instruction can also take any
4127 number of :ref:`fast-math flags <fastmath>`, which are optimization hints
4128 to enable otherwise unsafe floating point optimizations:
4133 .. code-block:: llvm
4135 <result> = frem float 4.0, %var ; yields {float}:result = 4.0 % %var
4139 Bitwise Binary Operations
4140 -------------------------
4142 Bitwise binary operators are used to do various forms of bit-twiddling
4143 in a program. They are generally very efficient instructions and can
4144 commonly be strength reduced from other instructions. They require two
4145 operands of the same type, execute an operation on them, and produce a
4146 single value. The resulting value is the same type as its operands.
4148 '``shl``' Instruction
4149 ^^^^^^^^^^^^^^^^^^^^^
4156 <result> = shl <ty> <op1>, <op2> ; yields {ty}:result
4157 <result> = shl nuw <ty> <op1>, <op2> ; yields {ty}:result
4158 <result> = shl nsw <ty> <op1>, <op2> ; yields {ty}:result
4159 <result> = shl nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
4164 The '``shl``' instruction returns the first operand shifted to the left
4165 a specified number of bits.
4170 Both arguments to the '``shl``' instruction must be the same
4171 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4172 '``op2``' is treated as an unsigned value.
4177 The value produced is ``op1`` \* 2\ :sup:`op2` mod 2\ :sup:`n`,
4178 where ``n`` is the width of the result. If ``op2`` is (statically or
4179 dynamically) negative or equal to or larger than the number of bits in
4180 ``op1``, the result is undefined. If the arguments are vectors, each
4181 vector element of ``op1`` is shifted by the corresponding shift amount
4184 If the ``nuw`` keyword is present, then the shift produces a :ref:`poison
4185 value <poisonvalues>` if it shifts out any non-zero bits. If the
4186 ``nsw`` keyword is present, then the shift produces a :ref:`poison
4187 value <poisonvalues>` if it shifts out any bits that disagree with the
4188 resultant sign bit. As such, NUW/NSW have the same semantics as they
4189 would if the shift were expressed as a mul instruction with the same
4190 nsw/nuw bits in (mul %op1, (shl 1, %op2)).
4195 .. code-block:: llvm
4197 <result> = shl i32 4, %var ; yields {i32}: 4 << %var
4198 <result> = shl i32 4, 2 ; yields {i32}: 16
4199 <result> = shl i32 1, 10 ; yields {i32}: 1024
4200 <result> = shl i32 1, 32 ; undefined
4201 <result> = shl <2 x i32> < i32 1, i32 1>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 2, i32 4>
4203 '``lshr``' Instruction
4204 ^^^^^^^^^^^^^^^^^^^^^^
4211 <result> = lshr <ty> <op1>, <op2> ; yields {ty}:result
4212 <result> = lshr exact <ty> <op1>, <op2> ; yields {ty}:result
4217 The '``lshr``' instruction (logical shift right) returns the first
4218 operand shifted to the right a specified number of bits with zero fill.
4223 Both arguments to the '``lshr``' instruction must be the same
4224 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4225 '``op2``' is treated as an unsigned value.
4230 This instruction always performs a logical shift right operation. The
4231 most significant bits of the result will be filled with zero bits after
4232 the shift. If ``op2`` is (statically or dynamically) equal to or larger
4233 than the number of bits in ``op1``, the result is undefined. If the
4234 arguments are vectors, each vector element of ``op1`` is shifted by the
4235 corresponding shift amount in ``op2``.
4237 If the ``exact`` keyword is present, the result value of the ``lshr`` is
4238 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
4244 .. code-block:: llvm
4246 <result> = lshr i32 4, 1 ; yields {i32}:result = 2
4247 <result> = lshr i32 4, 2 ; yields {i32}:result = 1
4248 <result> = lshr i8 4, 3 ; yields {i8}:result = 0
4249 <result> = lshr i8 -2, 1 ; yields {i8}:result = 0x7F
4250 <result> = lshr i32 1, 32 ; undefined
4251 <result> = lshr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 0x7FFFFFFF, i32 1>
4253 '``ashr``' Instruction
4254 ^^^^^^^^^^^^^^^^^^^^^^
4261 <result> = ashr <ty> <op1>, <op2> ; yields {ty}:result
4262 <result> = ashr exact <ty> <op1>, <op2> ; yields {ty}:result
4267 The '``ashr``' instruction (arithmetic shift right) returns the first
4268 operand shifted to the right a specified number of bits with sign
4274 Both arguments to the '``ashr``' instruction must be the same
4275 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4276 '``op2``' is treated as an unsigned value.
4281 This instruction always performs an arithmetic shift right operation,
4282 The most significant bits of the result will be filled with the sign bit
4283 of ``op1``. If ``op2`` is (statically or dynamically) equal to or larger
4284 than the number of bits in ``op1``, the result is undefined. If the
4285 arguments are vectors, each vector element of ``op1`` is shifted by the
4286 corresponding shift amount in ``op2``.
4288 If the ``exact`` keyword is present, the result value of the ``ashr`` is
4289 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
4295 .. code-block:: llvm
4297 <result> = ashr i32 4, 1 ; yields {i32}:result = 2
4298 <result> = ashr i32 4, 2 ; yields {i32}:result = 1
4299 <result> = ashr i8 4, 3 ; yields {i8}:result = 0
4300 <result> = ashr i8 -2, 1 ; yields {i8}:result = -1
4301 <result> = ashr i32 1, 32 ; undefined
4302 <result> = ashr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 3> ; yields: result=<2 x i32> < i32 -1, i32 0>
4304 '``and``' Instruction
4305 ^^^^^^^^^^^^^^^^^^^^^
4312 <result> = and <ty> <op1>, <op2> ; yields {ty}:result
4317 The '``and``' instruction returns the bitwise logical and of its two
4323 The two arguments to the '``and``' instruction must be
4324 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4325 arguments must have identical types.
4330 The truth table used for the '``and``' instruction is:
4347 .. code-block:: llvm
4349 <result> = and i32 4, %var ; yields {i32}:result = 4 & %var
4350 <result> = and i32 15, 40 ; yields {i32}:result = 8
4351 <result> = and i32 4, 8 ; yields {i32}:result = 0
4353 '``or``' Instruction
4354 ^^^^^^^^^^^^^^^^^^^^
4361 <result> = or <ty> <op1>, <op2> ; yields {ty}:result
4366 The '``or``' instruction returns the bitwise logical inclusive or of its
4372 The two arguments to the '``or``' instruction must be
4373 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4374 arguments must have identical types.
4379 The truth table used for the '``or``' instruction is:
4398 <result> = or i32 4, %var ; yields {i32}:result = 4 | %var
4399 <result> = or i32 15, 40 ; yields {i32}:result = 47
4400 <result> = or i32 4, 8 ; yields {i32}:result = 12
4402 '``xor``' Instruction
4403 ^^^^^^^^^^^^^^^^^^^^^
4410 <result> = xor <ty> <op1>, <op2> ; yields {ty}:result
4415 The '``xor``' instruction returns the bitwise logical exclusive or of
4416 its two operands. The ``xor`` is used to implement the "one's
4417 complement" operation, which is the "~" operator in C.
4422 The two arguments to the '``xor``' instruction must be
4423 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4424 arguments must have identical types.
4429 The truth table used for the '``xor``' instruction is:
4446 .. code-block:: llvm
4448 <result> = xor i32 4, %var ; yields {i32}:result = 4 ^ %var
4449 <result> = xor i32 15, 40 ; yields {i32}:result = 39
4450 <result> = xor i32 4, 8 ; yields {i32}:result = 12
4451 <result> = xor i32 %V, -1 ; yields {i32}:result = ~%V
4456 LLVM supports several instructions to represent vector operations in a
4457 target-independent manner. These instructions cover the element-access
4458 and vector-specific operations needed to process vectors effectively.
4459 While LLVM does directly support these vector operations, many
4460 sophisticated algorithms will want to use target-specific intrinsics to
4461 take full advantage of a specific target.
4463 .. _i_extractelement:
4465 '``extractelement``' Instruction
4466 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4473 <result> = extractelement <n x <ty>> <val>, <ty2> <idx> ; yields <ty>
4478 The '``extractelement``' instruction extracts a single scalar element
4479 from a vector at a specified index.
4484 The first operand of an '``extractelement``' instruction is a value of
4485 :ref:`vector <t_vector>` type. The second operand is an index indicating
4486 the position from which to extract the element. The index may be a
4487 variable of any integer type.
4492 The result is a scalar of the same type as the element type of ``val``.
4493 Its value is the value at position ``idx`` of ``val``. If ``idx``
4494 exceeds the length of ``val``, the results are undefined.
4499 .. code-block:: llvm
4501 <result> = extractelement <4 x i32> %vec, i32 0 ; yields i32
4503 .. _i_insertelement:
4505 '``insertelement``' Instruction
4506 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4513 <result> = insertelement <n x <ty>> <val>, <ty> <elt>, <ty2> <idx> ; yields <n x <ty>>
4518 The '``insertelement``' instruction inserts a scalar element into a
4519 vector at a specified index.
4524 The first operand of an '``insertelement``' instruction is a value of
4525 :ref:`vector <t_vector>` type. The second operand is a scalar value whose
4526 type must equal the element type of the first operand. The third operand
4527 is an index indicating the position at which to insert the value. The
4528 index may be a variable of any integer type.
4533 The result is a vector of the same type as ``val``. Its element values
4534 are those of ``val`` except at position ``idx``, where it gets the value
4535 ``elt``. If ``idx`` exceeds the length of ``val``, the results are
4541 .. code-block:: llvm
4543 <result> = insertelement <4 x i32> %vec, i32 1, i32 0 ; yields <4 x i32>
4545 .. _i_shufflevector:
4547 '``shufflevector``' Instruction
4548 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4555 <result> = shufflevector <n x <ty>> <v1>, <n x <ty>> <v2>, <m x i32> <mask> ; yields <m x <ty>>
4560 The '``shufflevector``' instruction constructs a permutation of elements
4561 from two input vectors, returning a vector with the same element type as
4562 the input and length that is the same as the shuffle mask.
4567 The first two operands of a '``shufflevector``' instruction are vectors
4568 with the same type. The third argument is a shuffle mask whose element
4569 type is always 'i32'. The result of the instruction is a vector whose
4570 length is the same as the shuffle mask and whose element type is the
4571 same as the element type of the first two operands.
4573 The shuffle mask operand is required to be a constant vector with either
4574 constant integer or undef values.
4579 The elements of the two input vectors are numbered from left to right
4580 across both of the vectors. The shuffle mask operand specifies, for each
4581 element of the result vector, which element of the two input vectors the
4582 result element gets. The element selector may be undef (meaning "don't
4583 care") and the second operand may be undef if performing a shuffle from
4589 .. code-block:: llvm
4591 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4592 <4 x i32> <i32 0, i32 4, i32 1, i32 5> ; yields <4 x i32>
4593 <result> = shufflevector <4 x i32> %v1, <4 x i32> undef,
4594 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32> - Identity shuffle.
4595 <result> = shufflevector <8 x i32> %v1, <8 x i32> undef,
4596 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32>
4597 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4598 <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7 > ; yields <8 x i32>
4600 Aggregate Operations
4601 --------------------
4603 LLVM supports several instructions for working with
4604 :ref:`aggregate <t_aggregate>` values.
4608 '``extractvalue``' Instruction
4609 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4616 <result> = extractvalue <aggregate type> <val>, <idx>{, <idx>}*
4621 The '``extractvalue``' instruction extracts the value of a member field
4622 from an :ref:`aggregate <t_aggregate>` value.
4627 The first operand of an '``extractvalue``' instruction is a value of
4628 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The operands are
4629 constant indices to specify which value to extract in a similar manner
4630 as indices in a '``getelementptr``' instruction.
4632 The major differences to ``getelementptr`` indexing are:
4634 - Since the value being indexed is not a pointer, the first index is
4635 omitted and assumed to be zero.
4636 - At least one index must be specified.
4637 - Not only struct indices but also array indices must be in bounds.
4642 The result is the value at the position in the aggregate specified by
4648 .. code-block:: llvm
4650 <result> = extractvalue {i32, float} %agg, 0 ; yields i32
4654 '``insertvalue``' Instruction
4655 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4662 <result> = insertvalue <aggregate type> <val>, <ty> <elt>, <idx>{, <idx>}* ; yields <aggregate type>
4667 The '``insertvalue``' instruction inserts a value into a member field in
4668 an :ref:`aggregate <t_aggregate>` value.
4673 The first operand of an '``insertvalue``' instruction is a value of
4674 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The second operand is
4675 a first-class value to insert. The following operands are constant
4676 indices indicating the position at which to insert the value in a
4677 similar manner as indices in a '``extractvalue``' instruction. The value
4678 to insert must have the same type as the value identified by the
4684 The result is an aggregate of the same type as ``val``. Its value is
4685 that of ``val`` except that the value at the position specified by the
4686 indices is that of ``elt``.
4691 .. code-block:: llvm
4693 %agg1 = insertvalue {i32, float} undef, i32 1, 0 ; yields {i32 1, float undef}
4694 %agg2 = insertvalue {i32, float} %agg1, float %val, 1 ; yields {i32 1, float %val}
4695 %agg3 = insertvalue {i32, {float}} %agg1, float %val, 1, 0 ; yields {i32 1, float %val}
4699 Memory Access and Addressing Operations
4700 ---------------------------------------
4702 A key design point of an SSA-based representation is how it represents
4703 memory. In LLVM, no memory locations are in SSA form, which makes things
4704 very simple. This section describes how to read, write, and allocate
4709 '``alloca``' Instruction
4710 ^^^^^^^^^^^^^^^^^^^^^^^^
4717 <result> = alloca [inalloca] <type> [, <ty> <NumElements>] [, align <alignment>] ; yields {type*}:result
4722 The '``alloca``' instruction allocates memory on the stack frame of the
4723 currently executing function, to be automatically released when this
4724 function returns to its caller. The object is always allocated in the
4725 generic address space (address space zero).
4730 The '``alloca``' instruction allocates ``sizeof(<type>)*NumElements``
4731 bytes of memory on the runtime stack, returning a pointer of the
4732 appropriate type to the program. If "NumElements" is specified, it is
4733 the number of elements allocated, otherwise "NumElements" is defaulted
4734 to be one. If a constant alignment is specified, the value result of the
4735 allocation is guaranteed to be aligned to at least that boundary. If not
4736 specified, or if zero, the target can choose to align the allocation on
4737 any convenient boundary compatible with the type.
4739 '``type``' may be any sized type.
4744 Memory is allocated; a pointer is returned. The operation is undefined
4745 if there is insufficient stack space for the allocation. '``alloca``'d
4746 memory is automatically released when the function returns. The
4747 '``alloca``' instruction is commonly used to represent automatic
4748 variables that must have an address available. When the function returns
4749 (either with the ``ret`` or ``resume`` instructions), the memory is
4750 reclaimed. Allocating zero bytes is legal, but the result is undefined.
4751 The order in which memory is allocated (ie., which way the stack grows)
4757 .. code-block:: llvm
4759 %ptr = alloca i32 ; yields {i32*}:ptr
4760 %ptr = alloca i32, i32 4 ; yields {i32*}:ptr
4761 %ptr = alloca i32, i32 4, align 1024 ; yields {i32*}:ptr
4762 %ptr = alloca i32, align 1024 ; yields {i32*}:ptr
4766 '``load``' Instruction
4767 ^^^^^^^^^^^^^^^^^^^^^^
4774 <result> = load [volatile] <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>][, !invariant.load !<index>]
4775 <result> = load atomic [volatile] <ty>* <pointer> [singlethread] <ordering>, align <alignment>
4776 !<index> = !{ i32 1 }
4781 The '``load``' instruction is used to read from memory.
4786 The argument to the ``load`` instruction specifies the memory address
4787 from which to load. The pointer must point to a :ref:`first
4788 class <t_firstclass>` type. If the ``load`` is marked as ``volatile``,
4789 then the optimizer is not allowed to modify the number or order of
4790 execution of this ``load`` with other :ref:`volatile
4791 operations <volatile>`.
4793 If the ``load`` is marked as ``atomic``, it takes an extra
4794 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
4795 ``release`` and ``acq_rel`` orderings are not valid on ``load``
4796 instructions. Atomic loads produce :ref:`defined <memmodel>` results
4797 when they may see multiple atomic stores. The type of the pointee must
4798 be an integer type whose bit width is a power of two greater than or
4799 equal to eight and less than or equal to a target-specific size limit.
4800 ``align`` must be explicitly specified on atomic loads, and the load has
4801 undefined behavior if the alignment is not set to a value which is at
4802 least the size in bytes of the pointee. ``!nontemporal`` does not have
4803 any defined semantics for atomic loads.
4805 The optional constant ``align`` argument specifies the alignment of the
4806 operation (that is, the alignment of the memory address). A value of 0
4807 or an omitted ``align`` argument means that the operation has the ABI
4808 alignment for the target. It is the responsibility of the code emitter
4809 to ensure that the alignment information is correct. Overestimating the
4810 alignment results in undefined behavior. Underestimating the alignment
4811 may produce less efficient code. An alignment of 1 is always safe.
4813 The optional ``!nontemporal`` metadata must reference a single
4814 metadata name ``<index>`` corresponding to a metadata node with one
4815 ``i32`` entry of value 1. The existence of the ``!nontemporal``
4816 metadata on the instruction tells the optimizer and code generator
4817 that this load is not expected to be reused in the cache. The code
4818 generator may select special instructions to save cache bandwidth, such
4819 as the ``MOVNT`` instruction on x86.
4821 The optional ``!invariant.load`` metadata must reference a single
4822 metadata name ``<index>`` corresponding to a metadata node with no
4823 entries. The existence of the ``!invariant.load`` metadata on the
4824 instruction tells the optimizer and code generator that this load
4825 address points to memory which does not change value during program
4826 execution. The optimizer may then move this load around, for example, by
4827 hoisting it out of loops using loop invariant code motion.
4832 The location of memory pointed to is loaded. If the value being loaded
4833 is of scalar type then the number of bytes read does not exceed the
4834 minimum number of bytes needed to hold all bits of the type. For
4835 example, loading an ``i24`` reads at most three bytes. When loading a
4836 value of a type like ``i20`` with a size that is not an integral number
4837 of bytes, the result is undefined if the value was not originally
4838 written using a store of the same type.
4843 .. code-block:: llvm
4845 %ptr = alloca i32 ; yields {i32*}:ptr
4846 store i32 3, i32* %ptr ; yields {void}
4847 %val = load i32* %ptr ; yields {i32}:val = i32 3
4851 '``store``' Instruction
4852 ^^^^^^^^^^^^^^^^^^^^^^^
4859 store [volatile] <ty> <value>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>] ; yields {void}
4860 store atomic [volatile] <ty> <value>, <ty>* <pointer> [singlethread] <ordering>, align <alignment> ; yields {void}
4865 The '``store``' instruction is used to write to memory.
4870 There are two arguments to the ``store`` instruction: a value to store
4871 and an address at which to store it. The type of the ``<pointer>``
4872 operand must be a pointer to the :ref:`first class <t_firstclass>` type of
4873 the ``<value>`` operand. If the ``store`` is marked as ``volatile``,
4874 then the optimizer is not allowed to modify the number or order of
4875 execution of this ``store`` with other :ref:`volatile
4876 operations <volatile>`.
4878 If the ``store`` is marked as ``atomic``, it takes an extra
4879 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
4880 ``acquire`` and ``acq_rel`` orderings aren't valid on ``store``
4881 instructions. Atomic loads produce :ref:`defined <memmodel>` results
4882 when they may see multiple atomic stores. The type of the pointee must
4883 be an integer type whose bit width is a power of two greater than or
4884 equal to eight and less than or equal to a target-specific size limit.
4885 ``align`` must be explicitly specified on atomic stores, and the store
4886 has undefined behavior if the alignment is not set to a value which is
4887 at least the size in bytes of the pointee. ``!nontemporal`` does not
4888 have any defined semantics for atomic stores.
4890 The optional constant ``align`` argument specifies the alignment of the
4891 operation (that is, the alignment of the memory address). A value of 0
4892 or an omitted ``align`` argument means that the operation has the ABI
4893 alignment for the target. It is the responsibility of the code emitter
4894 to ensure that the alignment information is correct. Overestimating the
4895 alignment results in undefined behavior. Underestimating the
4896 alignment may produce less efficient code. An alignment of 1 is always
4899 The optional ``!nontemporal`` metadata must reference a single metadata
4900 name ``<index>`` corresponding to a metadata node with one ``i32`` entry of
4901 value 1. The existence of the ``!nontemporal`` metadata on the instruction
4902 tells the optimizer and code generator that this load is not expected to
4903 be reused in the cache. The code generator may select special
4904 instructions to save cache bandwidth, such as the MOVNT instruction on
4910 The contents of memory are updated to contain ``<value>`` at the
4911 location specified by the ``<pointer>`` operand. If ``<value>`` is
4912 of scalar type then the number of bytes written does not exceed the
4913 minimum number of bytes needed to hold all bits of the type. For
4914 example, storing an ``i24`` writes at most three bytes. When writing a
4915 value of a type like ``i20`` with a size that is not an integral number
4916 of bytes, it is unspecified what happens to the extra bits that do not
4917 belong to the type, but they will typically be overwritten.
4922 .. code-block:: llvm
4924 %ptr = alloca i32 ; yields {i32*}:ptr
4925 store i32 3, i32* %ptr ; yields {void}
4926 %val = load i32* %ptr ; yields {i32}:val = i32 3
4930 '``fence``' Instruction
4931 ^^^^^^^^^^^^^^^^^^^^^^^
4938 fence [singlethread] <ordering> ; yields {void}
4943 The '``fence``' instruction is used to introduce happens-before edges
4949 '``fence``' instructions take an :ref:`ordering <ordering>` argument which
4950 defines what *synchronizes-with* edges they add. They can only be given
4951 ``acquire``, ``release``, ``acq_rel``, and ``seq_cst`` orderings.
4956 A fence A which has (at least) ``release`` ordering semantics
4957 *synchronizes with* a fence B with (at least) ``acquire`` ordering
4958 semantics if and only if there exist atomic operations X and Y, both
4959 operating on some atomic object M, such that A is sequenced before X, X
4960 modifies M (either directly or through some side effect of a sequence
4961 headed by X), Y is sequenced before B, and Y observes M. This provides a
4962 *happens-before* dependency between A and B. Rather than an explicit
4963 ``fence``, one (but not both) of the atomic operations X or Y might
4964 provide a ``release`` or ``acquire`` (resp.) ordering constraint and
4965 still *synchronize-with* the explicit ``fence`` and establish the
4966 *happens-before* edge.
4968 A ``fence`` which has ``seq_cst`` ordering, in addition to having both
4969 ``acquire`` and ``release`` semantics specified above, participates in
4970 the global program order of other ``seq_cst`` operations and/or fences.
4972 The optional ":ref:`singlethread <singlethread>`" argument specifies
4973 that the fence only synchronizes with other fences in the same thread.
4974 (This is useful for interacting with signal handlers.)
4979 .. code-block:: llvm
4981 fence acquire ; yields {void}
4982 fence singlethread seq_cst ; yields {void}
4986 '``cmpxchg``' Instruction
4987 ^^^^^^^^^^^^^^^^^^^^^^^^^
4994 cmpxchg [volatile] <ty>* <pointer>, <ty> <cmp>, <ty> <new> [singlethread] <success ordering> <failure ordering> ; yields {ty}
4999 The '``cmpxchg``' instruction is used to atomically modify memory. It
5000 loads a value in memory and compares it to a given value. If they are
5001 equal, it stores a new value into the memory.
5006 There are three arguments to the '``cmpxchg``' instruction: an address
5007 to operate on, a value to compare to the value currently be at that
5008 address, and a new value to place at that address if the compared values
5009 are equal. The type of '<cmp>' must be an integer type whose bit width
5010 is a power of two greater than or equal to eight and less than or equal
5011 to a target-specific size limit. '<cmp>' and '<new>' must have the same
5012 type, and the type of '<pointer>' must be a pointer to that type. If the
5013 ``cmpxchg`` is marked as ``volatile``, then the optimizer is not allowed
5014 to modify the number or order of execution of this ``cmpxchg`` with
5015 other :ref:`volatile operations <volatile>`.
5017 The success and failure :ref:`ordering <ordering>` arguments specify how this
5018 ``cmpxchg`` synchronizes with other atomic operations. The both ordering
5019 parameters must be at least ``monotonic``, the ordering constraint on failure
5020 must be no stronger than that on success, and the failure ordering cannot be
5021 either ``release`` or ``acq_rel``.
5023 The optional "``singlethread``" argument declares that the ``cmpxchg``
5024 is only atomic with respect to code (usually signal handlers) running in
5025 the same thread as the ``cmpxchg``. Otherwise the cmpxchg is atomic with
5026 respect to all other code in the system.
5028 The pointer passed into cmpxchg must have alignment greater than or
5029 equal to the size in memory of the operand.
5034 The contents of memory at the location specified by the '``<pointer>``'
5035 operand is read and compared to '``<cmp>``'; if the read value is the
5036 equal, '``<new>``' is written. The original value at the location is
5039 A successful ``cmpxchg`` is a read-modify-write instruction for the purpose of
5040 identifying release sequences. A failed ``cmpxchg`` is equivalent to an atomic
5041 load with an ordering parameter determined the second ordering parameter.
5046 .. code-block:: llvm
5049 %orig = atomic load i32* %ptr unordered ; yields {i32}
5053 %cmp = phi i32 [ %orig, %entry ], [%old, %loop]
5054 %squared = mul i32 %cmp, %cmp
5055 %old = cmpxchg i32* %ptr, i32 %cmp, i32 %squared acq_rel monotonic ; yields {i32}
5056 %success = icmp eq i32 %cmp, %old
5057 br i1 %success, label %done, label %loop
5064 '``atomicrmw``' Instruction
5065 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
5072 atomicrmw [volatile] <operation> <ty>* <pointer>, <ty> <value> [singlethread] <ordering> ; yields {ty}
5077 The '``atomicrmw``' instruction is used to atomically modify memory.
5082 There are three arguments to the '``atomicrmw``' instruction: an
5083 operation to apply, an address whose value to modify, an argument to the
5084 operation. The operation must be one of the following keywords:
5098 The type of '<value>' must be an integer type whose bit width is a power
5099 of two greater than or equal to eight and less than or equal to a
5100 target-specific size limit. The type of the '``<pointer>``' operand must
5101 be a pointer to that type. If the ``atomicrmw`` is marked as
5102 ``volatile``, then the optimizer is not allowed to modify the number or
5103 order of execution of this ``atomicrmw`` with other :ref:`volatile
5104 operations <volatile>`.
5109 The contents of memory at the location specified by the '``<pointer>``'
5110 operand are atomically read, modified, and written back. The original
5111 value at the location is returned. The modification is specified by the
5114 - xchg: ``*ptr = val``
5115 - add: ``*ptr = *ptr + val``
5116 - sub: ``*ptr = *ptr - val``
5117 - and: ``*ptr = *ptr & val``
5118 - nand: ``*ptr = ~(*ptr & val)``
5119 - or: ``*ptr = *ptr | val``
5120 - xor: ``*ptr = *ptr ^ val``
5121 - max: ``*ptr = *ptr > val ? *ptr : val`` (using a signed comparison)
5122 - min: ``*ptr = *ptr < val ? *ptr : val`` (using a signed comparison)
5123 - umax: ``*ptr = *ptr > val ? *ptr : val`` (using an unsigned
5125 - umin: ``*ptr = *ptr < val ? *ptr : val`` (using an unsigned
5131 .. code-block:: llvm
5133 %old = atomicrmw add i32* %ptr, i32 1 acquire ; yields {i32}
5135 .. _i_getelementptr:
5137 '``getelementptr``' Instruction
5138 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5145 <result> = getelementptr <pty>* <ptrval>{, <ty> <idx>}*
5146 <result> = getelementptr inbounds <pty>* <ptrval>{, <ty> <idx>}*
5147 <result> = getelementptr <ptr vector> ptrval, <vector index type> idx
5152 The '``getelementptr``' instruction is used to get the address of a
5153 subelement of an :ref:`aggregate <t_aggregate>` data structure. It performs
5154 address calculation only and does not access memory.
5159 The first argument is always a pointer or a vector of pointers, and
5160 forms the basis of the calculation. The remaining arguments are indices
5161 that indicate which of the elements of the aggregate object are indexed.
5162 The interpretation of each index is dependent on the type being indexed
5163 into. The first index always indexes the pointer value given as the
5164 first argument, the second index indexes a value of the type pointed to
5165 (not necessarily the value directly pointed to, since the first index
5166 can be non-zero), etc. The first type indexed into must be a pointer
5167 value, subsequent types can be arrays, vectors, and structs. Note that
5168 subsequent types being indexed into can never be pointers, since that
5169 would require loading the pointer before continuing calculation.
5171 The type of each index argument depends on the type it is indexing into.
5172 When indexing into a (optionally packed) structure, only ``i32`` integer
5173 **constants** are allowed (when using a vector of indices they must all
5174 be the **same** ``i32`` integer constant). When indexing into an array,
5175 pointer or vector, integers of any width are allowed, and they are not
5176 required to be constant. These integers are treated as signed values
5179 For example, let's consider a C code fragment and how it gets compiled
5195 int *foo(struct ST *s) {
5196 return &s[1].Z.B[5][13];
5199 The LLVM code generated by Clang is:
5201 .. code-block:: llvm
5203 %struct.RT = type { i8, [10 x [20 x i32]], i8 }
5204 %struct.ST = type { i32, double, %struct.RT }
5206 define i32* @foo(%struct.ST* %s) nounwind uwtable readnone optsize ssp {
5208 %arrayidx = getelementptr inbounds %struct.ST* %s, i64 1, i32 2, i32 1, i64 5, i64 13
5215 In the example above, the first index is indexing into the
5216 '``%struct.ST*``' type, which is a pointer, yielding a '``%struct.ST``'
5217 = '``{ i32, double, %struct.RT }``' type, a structure. The second index
5218 indexes into the third element of the structure, yielding a
5219 '``%struct.RT``' = '``{ i8 , [10 x [20 x i32]], i8 }``' type, another
5220 structure. The third index indexes into the second element of the
5221 structure, yielding a '``[10 x [20 x i32]]``' type, an array. The two
5222 dimensions of the array are subscripted into, yielding an '``i32``'
5223 type. The '``getelementptr``' instruction returns a pointer to this
5224 element, thus computing a value of '``i32*``' type.
5226 Note that it is perfectly legal to index partially through a structure,
5227 returning a pointer to an inner element. Because of this, the LLVM code
5228 for the given testcase is equivalent to:
5230 .. code-block:: llvm
5232 define i32* @foo(%struct.ST* %s) {
5233 %t1 = getelementptr %struct.ST* %s, i32 1 ; yields %struct.ST*:%t1
5234 %t2 = getelementptr %struct.ST* %t1, i32 0, i32 2 ; yields %struct.RT*:%t2
5235 %t3 = getelementptr %struct.RT* %t2, i32 0, i32 1 ; yields [10 x [20 x i32]]*:%t3
5236 %t4 = getelementptr [10 x [20 x i32]]* %t3, i32 0, i32 5 ; yields [20 x i32]*:%t4
5237 %t5 = getelementptr [20 x i32]* %t4, i32 0, i32 13 ; yields i32*:%t5
5241 If the ``inbounds`` keyword is present, the result value of the
5242 ``getelementptr`` is a :ref:`poison value <poisonvalues>` if the base
5243 pointer is not an *in bounds* address of an allocated object, or if any
5244 of the addresses that would be formed by successive addition of the
5245 offsets implied by the indices to the base address with infinitely
5246 precise signed arithmetic are not an *in bounds* address of that
5247 allocated object. The *in bounds* addresses for an allocated object are
5248 all the addresses that point into the object, plus the address one byte
5249 past the end. In cases where the base is a vector of pointers the
5250 ``inbounds`` keyword applies to each of the computations element-wise.
5252 If the ``inbounds`` keyword is not present, the offsets are added to the
5253 base address with silently-wrapping two's complement arithmetic. If the
5254 offsets have a different width from the pointer, they are sign-extended
5255 or truncated to the width of the pointer. The result value of the
5256 ``getelementptr`` may be outside the object pointed to by the base
5257 pointer. The result value may not necessarily be used to access memory
5258 though, even if it happens to point into allocated storage. See the
5259 :ref:`Pointer Aliasing Rules <pointeraliasing>` section for more
5262 The getelementptr instruction is often confusing. For some more insight
5263 into how it works, see :doc:`the getelementptr FAQ <GetElementPtr>`.
5268 .. code-block:: llvm
5270 ; yields [12 x i8]*:aptr
5271 %aptr = getelementptr {i32, [12 x i8]}* %saptr, i64 0, i32 1
5273 %vptr = getelementptr {i32, <2 x i8>}* %svptr, i64 0, i32 1, i32 1
5275 %eptr = getelementptr [12 x i8]* %aptr, i64 0, i32 1
5277 %iptr = getelementptr [10 x i32]* @arr, i16 0, i16 0
5279 In cases where the pointer argument is a vector of pointers, each index
5280 must be a vector with the same number of elements. For example:
5282 .. code-block:: llvm
5284 %A = getelementptr <4 x i8*> %ptrs, <4 x i64> %offsets,
5286 Conversion Operations
5287 ---------------------
5289 The instructions in this category are the conversion instructions
5290 (casting) which all take a single operand and a type. They perform
5291 various bit conversions on the operand.
5293 '``trunc .. to``' Instruction
5294 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5301 <result> = trunc <ty> <value> to <ty2> ; yields ty2
5306 The '``trunc``' instruction truncates its operand to the type ``ty2``.
5311 The '``trunc``' instruction takes a value to trunc, and a type to trunc
5312 it to. Both types must be of :ref:`integer <t_integer>` types, or vectors
5313 of the same number of integers. The bit size of the ``value`` must be
5314 larger than the bit size of the destination type, ``ty2``. Equal sized
5315 types are not allowed.
5320 The '``trunc``' instruction truncates the high order bits in ``value``
5321 and converts the remaining bits to ``ty2``. Since the source size must
5322 be larger than the destination size, ``trunc`` cannot be a *no-op cast*.
5323 It will always truncate bits.
5328 .. code-block:: llvm
5330 %X = trunc i32 257 to i8 ; yields i8:1
5331 %Y = trunc i32 123 to i1 ; yields i1:true
5332 %Z = trunc i32 122 to i1 ; yields i1:false
5333 %W = trunc <2 x i16> <i16 8, i16 7> to <2 x i8> ; yields <i8 8, i8 7>
5335 '``zext .. to``' Instruction
5336 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5343 <result> = zext <ty> <value> to <ty2> ; yields ty2
5348 The '``zext``' instruction zero extends its operand to type ``ty2``.
5353 The '``zext``' instruction takes a value to cast, and a type to cast it
5354 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
5355 the same number of integers. The bit size of the ``value`` must be
5356 smaller than the bit size of the destination type, ``ty2``.
5361 The ``zext`` fills the high order bits of the ``value`` with zero bits
5362 until it reaches the size of the destination type, ``ty2``.
5364 When zero extending from i1, the result will always be either 0 or 1.
5369 .. code-block:: llvm
5371 %X = zext i32 257 to i64 ; yields i64:257
5372 %Y = zext i1 true to i32 ; yields i32:1
5373 %Z = zext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5375 '``sext .. to``' Instruction
5376 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5383 <result> = sext <ty> <value> to <ty2> ; yields ty2
5388 The '``sext``' sign extends ``value`` to the type ``ty2``.
5393 The '``sext``' instruction takes a value to cast, and a type to cast it
5394 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
5395 the same number of integers. The bit size of the ``value`` must be
5396 smaller than the bit size of the destination type, ``ty2``.
5401 The '``sext``' instruction performs a sign extension by copying the sign
5402 bit (highest order bit) of the ``value`` until it reaches the bit size
5403 of the type ``ty2``.
5405 When sign extending from i1, the extension always results in -1 or 0.
5410 .. code-block:: llvm
5412 %X = sext i8 -1 to i16 ; yields i16 :65535
5413 %Y = sext i1 true to i32 ; yields i32:-1
5414 %Z = sext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5416 '``fptrunc .. to``' Instruction
5417 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5424 <result> = fptrunc <ty> <value> to <ty2> ; yields ty2
5429 The '``fptrunc``' instruction truncates ``value`` to type ``ty2``.
5434 The '``fptrunc``' instruction takes a :ref:`floating point <t_floating>`
5435 value to cast and a :ref:`floating point <t_floating>` type to cast it to.
5436 The size of ``value`` must be larger than the size of ``ty2``. This
5437 implies that ``fptrunc`` cannot be used to make a *no-op cast*.
5442 The '``fptrunc``' instruction truncates a ``value`` from a larger
5443 :ref:`floating point <t_floating>` type to a smaller :ref:`floating
5444 point <t_floating>` type. If the value cannot fit within the
5445 destination type, ``ty2``, then the results are undefined.
5450 .. code-block:: llvm
5452 %X = fptrunc double 123.0 to float ; yields float:123.0
5453 %Y = fptrunc double 1.0E+300 to float ; yields undefined
5455 '``fpext .. to``' Instruction
5456 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5463 <result> = fpext <ty> <value> to <ty2> ; yields ty2
5468 The '``fpext``' extends a floating point ``value`` to a larger floating
5474 The '``fpext``' instruction takes a :ref:`floating point <t_floating>`
5475 ``value`` to cast, and a :ref:`floating point <t_floating>` type to cast it
5476 to. The source type must be smaller than the destination type.
5481 The '``fpext``' instruction extends the ``value`` from a smaller
5482 :ref:`floating point <t_floating>` type to a larger :ref:`floating
5483 point <t_floating>` type. The ``fpext`` cannot be used to make a
5484 *no-op cast* because it always changes bits. Use ``bitcast`` to make a
5485 *no-op cast* for a floating point cast.
5490 .. code-block:: llvm
5492 %X = fpext float 3.125 to double ; yields double:3.125000e+00
5493 %Y = fpext double %X to fp128 ; yields fp128:0xL00000000000000004000900000000000
5495 '``fptoui .. to``' Instruction
5496 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5503 <result> = fptoui <ty> <value> to <ty2> ; yields ty2
5508 The '``fptoui``' converts a floating point ``value`` to its unsigned
5509 integer equivalent of type ``ty2``.
5514 The '``fptoui``' instruction takes a value to cast, which must be a
5515 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5516 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5517 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5518 type with the same number of elements as ``ty``
5523 The '``fptoui``' instruction converts its :ref:`floating
5524 point <t_floating>` operand into the nearest (rounding towards zero)
5525 unsigned integer value. If the value cannot fit in ``ty2``, the results
5531 .. code-block:: llvm
5533 %X = fptoui double 123.0 to i32 ; yields i32:123
5534 %Y = fptoui float 1.0E+300 to i1 ; yields undefined:1
5535 %Z = fptoui float 1.04E+17 to i8 ; yields undefined:1
5537 '``fptosi .. to``' Instruction
5538 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5545 <result> = fptosi <ty> <value> to <ty2> ; yields ty2
5550 The '``fptosi``' instruction converts :ref:`floating point <t_floating>`
5551 ``value`` to type ``ty2``.
5556 The '``fptosi``' instruction takes a value to cast, which must be a
5557 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5558 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5559 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5560 type with the same number of elements as ``ty``
5565 The '``fptosi``' instruction converts its :ref:`floating
5566 point <t_floating>` operand into the nearest (rounding towards zero)
5567 signed integer value. If the value cannot fit in ``ty2``, the results
5573 .. code-block:: llvm
5575 %X = fptosi double -123.0 to i32 ; yields i32:-123
5576 %Y = fptosi float 1.0E-247 to i1 ; yields undefined:1
5577 %Z = fptosi float 1.04E+17 to i8 ; yields undefined:1
5579 '``uitofp .. to``' Instruction
5580 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5587 <result> = uitofp <ty> <value> to <ty2> ; yields ty2
5592 The '``uitofp``' instruction regards ``value`` as an unsigned integer
5593 and converts that value to the ``ty2`` type.
5598 The '``uitofp``' instruction takes a value to cast, which must be a
5599 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
5600 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
5601 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
5602 type with the same number of elements as ``ty``
5607 The '``uitofp``' instruction interprets its operand as an unsigned
5608 integer quantity and converts it to the corresponding floating point
5609 value. If the value cannot fit in the floating point value, the results
5615 .. code-block:: llvm
5617 %X = uitofp i32 257 to float ; yields float:257.0
5618 %Y = uitofp i8 -1 to double ; yields double:255.0
5620 '``sitofp .. to``' Instruction
5621 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5628 <result> = sitofp <ty> <value> to <ty2> ; yields ty2
5633 The '``sitofp``' instruction regards ``value`` as a signed integer and
5634 converts that value to the ``ty2`` type.
5639 The '``sitofp``' instruction takes a value to cast, which must be a
5640 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
5641 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
5642 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
5643 type with the same number of elements as ``ty``
5648 The '``sitofp``' instruction interprets its operand as a signed integer
5649 quantity and converts it to the corresponding floating point value. If
5650 the value cannot fit in the floating point value, the results are
5656 .. code-block:: llvm
5658 %X = sitofp i32 257 to float ; yields float:257.0
5659 %Y = sitofp i8 -1 to double ; yields double:-1.0
5663 '``ptrtoint .. to``' Instruction
5664 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5671 <result> = ptrtoint <ty> <value> to <ty2> ; yields ty2
5676 The '``ptrtoint``' instruction converts the pointer or a vector of
5677 pointers ``value`` to the integer (or vector of integers) type ``ty2``.
5682 The '``ptrtoint``' instruction takes a ``value`` to cast, which must be
5683 a a value of type :ref:`pointer <t_pointer>` or a vector of pointers, and a
5684 type to cast it to ``ty2``, which must be an :ref:`integer <t_integer>` or
5685 a vector of integers type.
5690 The '``ptrtoint``' instruction converts ``value`` to integer type
5691 ``ty2`` by interpreting the pointer value as an integer and either
5692 truncating or zero extending that value to the size of the integer type.
5693 If ``value`` is smaller than ``ty2`` then a zero extension is done. If
5694 ``value`` is larger than ``ty2`` then a truncation is done. If they are
5695 the same size, then nothing is done (*no-op cast*) other than a type
5701 .. code-block:: llvm
5703 %X = ptrtoint i32* %P to i8 ; yields truncation on 32-bit architecture
5704 %Y = ptrtoint i32* %P to i64 ; yields zero extension on 32-bit architecture
5705 %Z = ptrtoint <4 x i32*> %P to <4 x i64>; yields vector zero extension for a vector of addresses on 32-bit architecture
5709 '``inttoptr .. to``' Instruction
5710 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5717 <result> = inttoptr <ty> <value> to <ty2> ; yields ty2
5722 The '``inttoptr``' instruction converts an integer ``value`` to a
5723 pointer type, ``ty2``.
5728 The '``inttoptr``' instruction takes an :ref:`integer <t_integer>` value to
5729 cast, and a type to cast it to, which must be a :ref:`pointer <t_pointer>`
5735 The '``inttoptr``' instruction converts ``value`` to type ``ty2`` by
5736 applying either a zero extension or a truncation depending on the size
5737 of the integer ``value``. If ``value`` is larger than the size of a
5738 pointer then a truncation is done. If ``value`` is smaller than the size
5739 of a pointer then a zero extension is done. If they are the same size,
5740 nothing is done (*no-op cast*).
5745 .. code-block:: llvm
5747 %X = inttoptr i32 255 to i32* ; yields zero extension on 64-bit architecture
5748 %Y = inttoptr i32 255 to i32* ; yields no-op on 32-bit architecture
5749 %Z = inttoptr i64 0 to i32* ; yields truncation on 32-bit architecture
5750 %Z = inttoptr <4 x i32> %G to <4 x i8*>; yields truncation of vector G to four pointers
5754 '``bitcast .. to``' Instruction
5755 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5762 <result> = bitcast <ty> <value> to <ty2> ; yields ty2
5767 The '``bitcast``' instruction converts ``value`` to type ``ty2`` without
5773 The '``bitcast``' instruction takes a value to cast, which must be a
5774 non-aggregate first class value, and a type to cast it to, which must
5775 also be a non-aggregate :ref:`first class <t_firstclass>` type. The
5776 bit sizes of ``value`` and the destination type, ``ty2``, must be
5777 identical. If the source type is a pointer, the destination type must
5778 also be a pointer of the same size. This instruction supports bitwise
5779 conversion of vectors to integers and to vectors of other types (as
5780 long as they have the same size).
5785 The '``bitcast``' instruction converts ``value`` to type ``ty2``. It
5786 is always a *no-op cast* because no bits change with this
5787 conversion. The conversion is done as if the ``value`` had been stored
5788 to memory and read back as type ``ty2``. Pointer (or vector of
5789 pointers) types may only be converted to other pointer (or vector of
5790 pointers) types with the same address space through this instruction.
5791 To convert pointers to other types, use the :ref:`inttoptr <i_inttoptr>`
5792 or :ref:`ptrtoint <i_ptrtoint>` instructions first.
5797 .. code-block:: llvm
5799 %X = bitcast i8 255 to i8 ; yields i8 :-1
5800 %Y = bitcast i32* %x to sint* ; yields sint*:%x
5801 %Z = bitcast <2 x int> %V to i64; ; yields i64: %V
5802 %Z = bitcast <2 x i32*> %V to <2 x i64*> ; yields <2 x i64*>
5804 .. _i_addrspacecast:
5806 '``addrspacecast .. to``' Instruction
5807 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5814 <result> = addrspacecast <pty> <ptrval> to <pty2> ; yields pty2
5819 The '``addrspacecast``' instruction converts ``ptrval`` from ``pty`` in
5820 address space ``n`` to type ``pty2`` in address space ``m``.
5825 The '``addrspacecast``' instruction takes a pointer or vector of pointer value
5826 to cast and a pointer type to cast it to, which must have a different
5832 The '``addrspacecast``' instruction converts the pointer value
5833 ``ptrval`` to type ``pty2``. It can be a *no-op cast* or a complex
5834 value modification, depending on the target and the address space
5835 pair. Pointer conversions within the same address space must be
5836 performed with the ``bitcast`` instruction. Note that if the address space
5837 conversion is legal then both result and operand refer to the same memory
5843 .. code-block:: llvm
5845 %X = addrspacecast i32* %x to i32 addrspace(1)* ; yields i32 addrspace(1)*:%x
5846 %Y = addrspacecast i32 addrspace(1)* %y to i64 addrspace(2)* ; yields i64 addrspace(2)*:%y
5847 %Z = addrspacecast <4 x i32*> %z to <4 x float addrspace(3)*> ; yields <4 x float addrspace(3)*>:%z
5854 The instructions in this category are the "miscellaneous" instructions,
5855 which defy better classification.
5859 '``icmp``' Instruction
5860 ^^^^^^^^^^^^^^^^^^^^^^
5867 <result> = icmp <cond> <ty> <op1>, <op2> ; yields {i1} or {<N x i1>}:result
5872 The '``icmp``' instruction returns a boolean value or a vector of
5873 boolean values based on comparison of its two integer, integer vector,
5874 pointer, or pointer vector operands.
5879 The '``icmp``' instruction takes three operands. The first operand is
5880 the condition code indicating the kind of comparison to perform. It is
5881 not a value, just a keyword. The possible condition code are:
5884 #. ``ne``: not equal
5885 #. ``ugt``: unsigned greater than
5886 #. ``uge``: unsigned greater or equal
5887 #. ``ult``: unsigned less than
5888 #. ``ule``: unsigned less or equal
5889 #. ``sgt``: signed greater than
5890 #. ``sge``: signed greater or equal
5891 #. ``slt``: signed less than
5892 #. ``sle``: signed less or equal
5894 The remaining two arguments must be :ref:`integer <t_integer>` or
5895 :ref:`pointer <t_pointer>` or integer :ref:`vector <t_vector>` typed. They
5896 must also be identical types.
5901 The '``icmp``' compares ``op1`` and ``op2`` according to the condition
5902 code given as ``cond``. The comparison performed always yields either an
5903 :ref:`i1 <t_integer>` or vector of ``i1`` result, as follows:
5905 #. ``eq``: yields ``true`` if the operands are equal, ``false``
5906 otherwise. No sign interpretation is necessary or performed.
5907 #. ``ne``: yields ``true`` if the operands are unequal, ``false``
5908 otherwise. No sign interpretation is necessary or performed.
5909 #. ``ugt``: interprets the operands as unsigned values and yields
5910 ``true`` if ``op1`` is greater than ``op2``.
5911 #. ``uge``: interprets the operands as unsigned values and yields
5912 ``true`` if ``op1`` is greater than or equal to ``op2``.
5913 #. ``ult``: interprets the operands as unsigned values and yields
5914 ``true`` if ``op1`` is less than ``op2``.
5915 #. ``ule``: interprets the operands as unsigned values and yields
5916 ``true`` if ``op1`` is less than or equal to ``op2``.
5917 #. ``sgt``: interprets the operands as signed values and yields ``true``
5918 if ``op1`` is greater than ``op2``.
5919 #. ``sge``: interprets the operands as signed values and yields ``true``
5920 if ``op1`` is greater than or equal to ``op2``.
5921 #. ``slt``: interprets the operands as signed values and yields ``true``
5922 if ``op1`` is less than ``op2``.
5923 #. ``sle``: interprets the operands as signed values and yields ``true``
5924 if ``op1`` is less than or equal to ``op2``.
5926 If the operands are :ref:`pointer <t_pointer>` typed, the pointer values
5927 are compared as if they were integers.
5929 If the operands are integer vectors, then they are compared element by
5930 element. The result is an ``i1`` vector with the same number of elements
5931 as the values being compared. Otherwise, the result is an ``i1``.
5936 .. code-block:: llvm
5938 <result> = icmp eq i32 4, 5 ; yields: result=false
5939 <result> = icmp ne float* %X, %X ; yields: result=false
5940 <result> = icmp ult i16 4, 5 ; yields: result=true
5941 <result> = icmp sgt i16 4, 5 ; yields: result=false
5942 <result> = icmp ule i16 -4, 5 ; yields: result=false
5943 <result> = icmp sge i16 4, 5 ; yields: result=false
5945 Note that the code generator does not yet support vector types with the
5946 ``icmp`` instruction.
5950 '``fcmp``' Instruction
5951 ^^^^^^^^^^^^^^^^^^^^^^
5958 <result> = fcmp <cond> <ty> <op1>, <op2> ; yields {i1} or {<N x i1>}:result
5963 The '``fcmp``' instruction returns a boolean value or vector of boolean
5964 values based on comparison of its operands.
5966 If the operands are floating point scalars, then the result type is a
5967 boolean (:ref:`i1 <t_integer>`).
5969 If the operands are floating point vectors, then the result type is a
5970 vector of boolean with the same number of elements as the operands being
5976 The '``fcmp``' instruction takes three operands. The first operand is
5977 the condition code indicating the kind of comparison to perform. It is
5978 not a value, just a keyword. The possible condition code are:
5980 #. ``false``: no comparison, always returns false
5981 #. ``oeq``: ordered and equal
5982 #. ``ogt``: ordered and greater than
5983 #. ``oge``: ordered and greater than or equal
5984 #. ``olt``: ordered and less than
5985 #. ``ole``: ordered and less than or equal
5986 #. ``one``: ordered and not equal
5987 #. ``ord``: ordered (no nans)
5988 #. ``ueq``: unordered or equal
5989 #. ``ugt``: unordered or greater than
5990 #. ``uge``: unordered or greater than or equal
5991 #. ``ult``: unordered or less than
5992 #. ``ule``: unordered or less than or equal
5993 #. ``une``: unordered or not equal
5994 #. ``uno``: unordered (either nans)
5995 #. ``true``: no comparison, always returns true
5997 *Ordered* means that neither operand is a QNAN while *unordered* means
5998 that either operand may be a QNAN.
6000 Each of ``val1`` and ``val2`` arguments must be either a :ref:`floating
6001 point <t_floating>` type or a :ref:`vector <t_vector>` of floating point
6002 type. They must have identical types.
6007 The '``fcmp``' instruction compares ``op1`` and ``op2`` according to the
6008 condition code given as ``cond``. If the operands are vectors, then the
6009 vectors are compared element by element. Each comparison performed
6010 always yields an :ref:`i1 <t_integer>` result, as follows:
6012 #. ``false``: always yields ``false``, regardless of operands.
6013 #. ``oeq``: yields ``true`` if both operands are not a QNAN and ``op1``
6014 is equal to ``op2``.
6015 #. ``ogt``: yields ``true`` if both operands are not a QNAN and ``op1``
6016 is greater than ``op2``.
6017 #. ``oge``: yields ``true`` if both operands are not a QNAN and ``op1``
6018 is greater than or equal to ``op2``.
6019 #. ``olt``: yields ``true`` if both operands are not a QNAN and ``op1``
6020 is less than ``op2``.
6021 #. ``ole``: yields ``true`` if both operands are not a QNAN and ``op1``
6022 is less than or equal to ``op2``.
6023 #. ``one``: yields ``true`` if both operands are not a QNAN and ``op1``
6024 is not equal to ``op2``.
6025 #. ``ord``: yields ``true`` if both operands are not a QNAN.
6026 #. ``ueq``: yields ``true`` if either operand is a QNAN or ``op1`` is
6028 #. ``ugt``: yields ``true`` if either operand is a QNAN or ``op1`` is
6029 greater than ``op2``.
6030 #. ``uge``: yields ``true`` if either operand is a QNAN or ``op1`` is
6031 greater than or equal to ``op2``.
6032 #. ``ult``: yields ``true`` if either operand is a QNAN or ``op1`` is
6034 #. ``ule``: yields ``true`` if either operand is a QNAN or ``op1`` is
6035 less than or equal to ``op2``.
6036 #. ``une``: yields ``true`` if either operand is a QNAN or ``op1`` is
6037 not equal to ``op2``.
6038 #. ``uno``: yields ``true`` if either operand is a QNAN.
6039 #. ``true``: always yields ``true``, regardless of operands.
6044 .. code-block:: llvm
6046 <result> = fcmp oeq float 4.0, 5.0 ; yields: result=false
6047 <result> = fcmp one float 4.0, 5.0 ; yields: result=true
6048 <result> = fcmp olt float 4.0, 5.0 ; yields: result=true
6049 <result> = fcmp ueq double 1.0, 2.0 ; yields: result=false
6051 Note that the code generator does not yet support vector types with the
6052 ``fcmp`` instruction.
6056 '``phi``' Instruction
6057 ^^^^^^^^^^^^^^^^^^^^^
6064 <result> = phi <ty> [ <val0>, <label0>], ...
6069 The '``phi``' instruction is used to implement the φ node in the SSA
6070 graph representing the function.
6075 The type of the incoming values is specified with the first type field.
6076 After this, the '``phi``' instruction takes a list of pairs as
6077 arguments, with one pair for each predecessor basic block of the current
6078 block. Only values of :ref:`first class <t_firstclass>` type may be used as
6079 the value arguments to the PHI node. Only labels may be used as the
6082 There must be no non-phi instructions between the start of a basic block
6083 and the PHI instructions: i.e. PHI instructions must be first in a basic
6086 For the purposes of the SSA form, the use of each incoming value is
6087 deemed to occur on the edge from the corresponding predecessor block to
6088 the current block (but after any definition of an '``invoke``'
6089 instruction's return value on the same edge).
6094 At runtime, the '``phi``' instruction logically takes on the value
6095 specified by the pair corresponding to the predecessor basic block that
6096 executed just prior to the current block.
6101 .. code-block:: llvm
6103 Loop: ; Infinite loop that counts from 0 on up...
6104 %indvar = phi i32 [ 0, %LoopHeader ], [ %nextindvar, %Loop ]
6105 %nextindvar = add i32 %indvar, 1
6110 '``select``' Instruction
6111 ^^^^^^^^^^^^^^^^^^^^^^^^
6118 <result> = select selty <cond>, <ty> <val1>, <ty> <val2> ; yields ty
6120 selty is either i1 or {<N x i1>}
6125 The '``select``' instruction is used to choose one value based on a
6126 condition, without IR-level branching.
6131 The '``select``' instruction requires an 'i1' value or a vector of 'i1'
6132 values indicating the condition, and two values of the same :ref:`first
6133 class <t_firstclass>` type. If the val1/val2 are vectors and the
6134 condition is a scalar, then entire vectors are selected, not individual
6140 If the condition is an i1 and it evaluates to 1, the instruction returns
6141 the first value argument; otherwise, it returns the second value
6144 If the condition is a vector of i1, then the value arguments must be
6145 vectors of the same size, and the selection is done element by element.
6150 .. code-block:: llvm
6152 %X = select i1 true, i8 17, i8 42 ; yields i8:17
6156 '``call``' Instruction
6157 ^^^^^^^^^^^^^^^^^^^^^^
6164 <result> = [tail | musttail] call [cconv] [ret attrs] <ty> [<fnty>*] <fnptrval>(<function args>) [fn attrs]
6169 The '``call``' instruction represents a simple function call.
6174 This instruction requires several arguments:
6176 #. The optional ``tail`` and ``musttail`` markers indicate that the optimizers
6177 should perform tail call optimization. The ``tail`` marker is a hint that
6178 `can be ignored <CodeGenerator.html#sibcallopt>`_. The ``musttail`` marker
6179 means that the call must be tail call optimized in order for the program to
6180 be correct. The ``musttail`` marker provides these guarantees:
6182 #. The call will not cause unbounded stack growth if it is part of a
6183 recursive cycle in the call graph.
6184 #. Arguments with the :ref:`inalloca <attr_inalloca>` attribute are
6187 Both markers imply that the callee does not access allocas or varargs from
6188 the caller. Calls marked ``musttail`` must obey the following additional
6191 - The call must immediately precede a :ref:`ret <i_ret>` instruction,
6192 or a pointer bitcast followed by a ret instruction.
6193 - The ret instruction must return the (possibly bitcasted) value
6194 produced by the call or void.
6195 - The caller and callee prototypes must match. Pointer types of
6196 parameters or return types may differ in pointee type, but not
6198 - The calling conventions of the caller and callee must match.
6199 - All ABI-impacting function attributes, such as sret, byval, inreg,
6200 returned, and inalloca, must match.
6202 Tail call optimization for calls marked ``tail`` is guaranteed to occur if
6203 the following conditions are met:
6205 - Caller and callee both have the calling convention ``fastcc``.
6206 - The call is in tail position (ret immediately follows call and ret
6207 uses value of call or is void).
6208 - Option ``-tailcallopt`` is enabled, or
6209 ``llvm::GuaranteedTailCallOpt`` is ``true``.
6210 - `Platform specific constraints are
6211 met. <CodeGenerator.html#tailcallopt>`_
6213 #. The optional "cconv" marker indicates which :ref:`calling
6214 convention <callingconv>` the call should use. If none is
6215 specified, the call defaults to using C calling conventions. The
6216 calling convention of the call must match the calling convention of
6217 the target function, or else the behavior is undefined.
6218 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
6219 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
6221 #. '``ty``': the type of the call instruction itself which is also the
6222 type of the return value. Functions that return no value are marked
6224 #. '``fnty``': shall be the signature of the pointer to function value
6225 being invoked. The argument types must match the types implied by
6226 this signature. This type can be omitted if the function is not
6227 varargs and if the function type does not return a pointer to a
6229 #. '``fnptrval``': An LLVM value containing a pointer to a function to
6230 be invoked. In most cases, this is a direct function invocation, but
6231 indirect ``call``'s are just as possible, calling an arbitrary pointer
6233 #. '``function args``': argument list whose types match the function
6234 signature argument types and parameter attributes. All arguments must
6235 be of :ref:`first class <t_firstclass>` type. If the function signature
6236 indicates the function accepts a variable number of arguments, the
6237 extra arguments can be specified.
6238 #. The optional :ref:`function attributes <fnattrs>` list. Only
6239 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
6240 attributes are valid here.
6245 The '``call``' instruction is used to cause control flow to transfer to
6246 a specified function, with its incoming arguments bound to the specified
6247 values. Upon a '``ret``' instruction in the called function, control
6248 flow continues with the instruction after the function call, and the
6249 return value of the function is bound to the result argument.
6254 .. code-block:: llvm
6256 %retval = call i32 @test(i32 %argc)
6257 call i32 (i8*, ...)* @printf(i8* %msg, i32 12, i8 42) ; yields i32
6258 %X = tail call i32 @foo() ; yields i32
6259 %Y = tail call fastcc i32 @foo() ; yields i32
6260 call void %foo(i8 97 signext)
6262 %struct.A = type { i32, i8 }
6263 %r = call %struct.A @foo() ; yields { 32, i8 }
6264 %gr = extractvalue %struct.A %r, 0 ; yields i32
6265 %gr1 = extractvalue %struct.A %r, 1 ; yields i8
6266 %Z = call void @foo() noreturn ; indicates that %foo never returns normally
6267 %ZZ = call zeroext i32 @bar() ; Return value is %zero extended
6269 llvm treats calls to some functions with names and arguments that match
6270 the standard C99 library as being the C99 library functions, and may
6271 perform optimizations or generate code for them under that assumption.
6272 This is something we'd like to change in the future to provide better
6273 support for freestanding environments and non-C-based languages.
6277 '``va_arg``' Instruction
6278 ^^^^^^^^^^^^^^^^^^^^^^^^
6285 <resultval> = va_arg <va_list*> <arglist>, <argty>
6290 The '``va_arg``' instruction is used to access arguments passed through
6291 the "variable argument" area of a function call. It is used to implement
6292 the ``va_arg`` macro in C.
6297 This instruction takes a ``va_list*`` value and the type of the
6298 argument. It returns a value of the specified argument type and
6299 increments the ``va_list`` to point to the next argument. The actual
6300 type of ``va_list`` is target specific.
6305 The '``va_arg``' instruction loads an argument of the specified type
6306 from the specified ``va_list`` and causes the ``va_list`` to point to
6307 the next argument. For more information, see the variable argument
6308 handling :ref:`Intrinsic Functions <int_varargs>`.
6310 It is legal for this instruction to be called in a function which does
6311 not take a variable number of arguments, for example, the ``vfprintf``
6314 ``va_arg`` is an LLVM instruction instead of an :ref:`intrinsic
6315 function <intrinsics>` because it takes a type as an argument.
6320 See the :ref:`variable argument processing <int_varargs>` section.
6322 Note that the code generator does not yet fully support va\_arg on many
6323 targets. Also, it does not currently support va\_arg with aggregate
6324 types on any target.
6328 '``landingpad``' Instruction
6329 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6336 <resultval> = landingpad <resultty> personality <type> <pers_fn> <clause>+
6337 <resultval> = landingpad <resultty> personality <type> <pers_fn> cleanup <clause>*
6339 <clause> := catch <type> <value>
6340 <clause> := filter <array constant type> <array constant>
6345 The '``landingpad``' instruction is used by `LLVM's exception handling
6346 system <ExceptionHandling.html#overview>`_ to specify that a basic block
6347 is a landing pad --- one where the exception lands, and corresponds to the
6348 code found in the ``catch`` portion of a ``try``/``catch`` sequence. It
6349 defines values supplied by the personality function (``pers_fn``) upon
6350 re-entry to the function. The ``resultval`` has the type ``resultty``.
6355 This instruction takes a ``pers_fn`` value. This is the personality
6356 function associated with the unwinding mechanism. The optional
6357 ``cleanup`` flag indicates that the landing pad block is a cleanup.
6359 A ``clause`` begins with the clause type --- ``catch`` or ``filter`` --- and
6360 contains the global variable representing the "type" that may be caught
6361 or filtered respectively. Unlike the ``catch`` clause, the ``filter``
6362 clause takes an array constant as its argument. Use
6363 "``[0 x i8**] undef``" for a filter which cannot throw. The
6364 '``landingpad``' instruction must contain *at least* one ``clause`` or
6365 the ``cleanup`` flag.
6370 The '``landingpad``' instruction defines the values which are set by the
6371 personality function (``pers_fn``) upon re-entry to the function, and
6372 therefore the "result type" of the ``landingpad`` instruction. As with
6373 calling conventions, how the personality function results are
6374 represented in LLVM IR is target specific.
6376 The clauses are applied in order from top to bottom. If two
6377 ``landingpad`` instructions are merged together through inlining, the
6378 clauses from the calling function are appended to the list of clauses.
6379 When the call stack is being unwound due to an exception being thrown,
6380 the exception is compared against each ``clause`` in turn. If it doesn't
6381 match any of the clauses, and the ``cleanup`` flag is not set, then
6382 unwinding continues further up the call stack.
6384 The ``landingpad`` instruction has several restrictions:
6386 - A landing pad block is a basic block which is the unwind destination
6387 of an '``invoke``' instruction.
6388 - A landing pad block must have a '``landingpad``' instruction as its
6389 first non-PHI instruction.
6390 - There can be only one '``landingpad``' instruction within the landing
6392 - A basic block that is not a landing pad block may not include a
6393 '``landingpad``' instruction.
6394 - All '``landingpad``' instructions in a function must have the same
6395 personality function.
6400 .. code-block:: llvm
6402 ;; A landing pad which can catch an integer.
6403 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6405 ;; A landing pad that is a cleanup.
6406 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6408 ;; A landing pad which can catch an integer and can only throw a double.
6409 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6411 filter [1 x i8**] [@_ZTId]
6418 LLVM supports the notion of an "intrinsic function". These functions
6419 have well known names and semantics and are required to follow certain
6420 restrictions. Overall, these intrinsics represent an extension mechanism
6421 for the LLVM language that does not require changing all of the
6422 transformations in LLVM when adding to the language (or the bitcode
6423 reader/writer, the parser, etc...).
6425 Intrinsic function names must all start with an "``llvm.``" prefix. This
6426 prefix is reserved in LLVM for intrinsic names; thus, function names may
6427 not begin with this prefix. Intrinsic functions must always be external
6428 functions: you cannot define the body of intrinsic functions. Intrinsic
6429 functions may only be used in call or invoke instructions: it is illegal
6430 to take the address of an intrinsic function. Additionally, because
6431 intrinsic functions are part of the LLVM language, it is required if any
6432 are added that they be documented here.
6434 Some intrinsic functions can be overloaded, i.e., the intrinsic
6435 represents a family of functions that perform the same operation but on
6436 different data types. Because LLVM can represent over 8 million
6437 different integer types, overloading is used commonly to allow an
6438 intrinsic function to operate on any integer type. One or more of the
6439 argument types or the result type can be overloaded to accept any
6440 integer type. Argument types may also be defined as exactly matching a
6441 previous argument's type or the result type. This allows an intrinsic
6442 function which accepts multiple arguments, but needs all of them to be
6443 of the same type, to only be overloaded with respect to a single
6444 argument or the result.
6446 Overloaded intrinsics will have the names of its overloaded argument
6447 types encoded into its function name, each preceded by a period. Only
6448 those types which are overloaded result in a name suffix. Arguments
6449 whose type is matched against another type do not. For example, the
6450 ``llvm.ctpop`` function can take an integer of any width and returns an
6451 integer of exactly the same integer width. This leads to a family of
6452 functions such as ``i8 @llvm.ctpop.i8(i8 %val)`` and
6453 ``i29 @llvm.ctpop.i29(i29 %val)``. Only one type, the return type, is
6454 overloaded, and only one type suffix is required. Because the argument's
6455 type is matched against the return type, it does not require its own
6458 To learn how to add an intrinsic function, please see the `Extending
6459 LLVM Guide <ExtendingLLVM.html>`_.
6463 Variable Argument Handling Intrinsics
6464 -------------------------------------
6466 Variable argument support is defined in LLVM with the
6467 :ref:`va_arg <i_va_arg>` instruction and these three intrinsic
6468 functions. These functions are related to the similarly named macros
6469 defined in the ``<stdarg.h>`` header file.
6471 All of these functions operate on arguments that use a target-specific
6472 value type "``va_list``". The LLVM assembly language reference manual
6473 does not define what this type is, so all transformations should be
6474 prepared to handle these functions regardless of the type used.
6476 This example shows how the :ref:`va_arg <i_va_arg>` instruction and the
6477 variable argument handling intrinsic functions are used.
6479 .. code-block:: llvm
6481 define i32 @test(i32 %X, ...) {
6482 ; Initialize variable argument processing
6484 %ap2 = bitcast i8** %ap to i8*
6485 call void @llvm.va_start(i8* %ap2)
6487 ; Read a single integer argument
6488 %tmp = va_arg i8** %ap, i32
6490 ; Demonstrate usage of llvm.va_copy and llvm.va_end
6492 %aq2 = bitcast i8** %aq to i8*
6493 call void @llvm.va_copy(i8* %aq2, i8* %ap2)
6494 call void @llvm.va_end(i8* %aq2)
6496 ; Stop processing of arguments.
6497 call void @llvm.va_end(i8* %ap2)
6501 declare void @llvm.va_start(i8*)
6502 declare void @llvm.va_copy(i8*, i8*)
6503 declare void @llvm.va_end(i8*)
6507 '``llvm.va_start``' Intrinsic
6508 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6515 declare void @llvm.va_start(i8* <arglist>)
6520 The '``llvm.va_start``' intrinsic initializes ``*<arglist>`` for
6521 subsequent use by ``va_arg``.
6526 The argument is a pointer to a ``va_list`` element to initialize.
6531 The '``llvm.va_start``' intrinsic works just like the ``va_start`` macro
6532 available in C. In a target-dependent way, it initializes the
6533 ``va_list`` element to which the argument points, so that the next call
6534 to ``va_arg`` will produce the first variable argument passed to the
6535 function. Unlike the C ``va_start`` macro, this intrinsic does not need
6536 to know the last argument of the function as the compiler can figure
6539 '``llvm.va_end``' Intrinsic
6540 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6547 declare void @llvm.va_end(i8* <arglist>)
6552 The '``llvm.va_end``' intrinsic destroys ``*<arglist>``, which has been
6553 initialized previously with ``llvm.va_start`` or ``llvm.va_copy``.
6558 The argument is a pointer to a ``va_list`` to destroy.
6563 The '``llvm.va_end``' intrinsic works just like the ``va_end`` macro
6564 available in C. In a target-dependent way, it destroys the ``va_list``
6565 element to which the argument points. Calls to
6566 :ref:`llvm.va_start <int_va_start>` and
6567 :ref:`llvm.va_copy <int_va_copy>` must be matched exactly with calls to
6572 '``llvm.va_copy``' Intrinsic
6573 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6580 declare void @llvm.va_copy(i8* <destarglist>, i8* <srcarglist>)
6585 The '``llvm.va_copy``' intrinsic copies the current argument position
6586 from the source argument list to the destination argument list.
6591 The first argument is a pointer to a ``va_list`` element to initialize.
6592 The second argument is a pointer to a ``va_list`` element to copy from.
6597 The '``llvm.va_copy``' intrinsic works just like the ``va_copy`` macro
6598 available in C. In a target-dependent way, it copies the source
6599 ``va_list`` element into the destination ``va_list`` element. This
6600 intrinsic is necessary because the `` llvm.va_start`` intrinsic may be
6601 arbitrarily complex and require, for example, memory allocation.
6603 Accurate Garbage Collection Intrinsics
6604 --------------------------------------
6606 LLVM support for `Accurate Garbage Collection <GarbageCollection.html>`_
6607 (GC) requires the implementation and generation of these intrinsics.
6608 These intrinsics allow identification of :ref:`GC roots on the
6609 stack <int_gcroot>`, as well as garbage collector implementations that
6610 require :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers.
6611 Front-ends for type-safe garbage collected languages should generate
6612 these intrinsics to make use of the LLVM garbage collectors. For more
6613 details, see `Accurate Garbage Collection with
6614 LLVM <GarbageCollection.html>`_.
6616 The garbage collection intrinsics only operate on objects in the generic
6617 address space (address space zero).
6621 '``llvm.gcroot``' Intrinsic
6622 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6629 declare void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
6634 The '``llvm.gcroot``' intrinsic declares the existence of a GC root to
6635 the code generator, and allows some metadata to be associated with it.
6640 The first argument specifies the address of a stack object that contains
6641 the root pointer. The second pointer (which must be either a constant or
6642 a global value address) contains the meta-data to be associated with the
6648 At runtime, a call to this intrinsic stores a null pointer into the
6649 "ptrloc" location. At compile-time, the code generator generates
6650 information to allow the runtime to find the pointer at GC safe points.
6651 The '``llvm.gcroot``' intrinsic may only be used in a function which
6652 :ref:`specifies a GC algorithm <gc>`.
6656 '``llvm.gcread``' Intrinsic
6657 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6664 declare i8* @llvm.gcread(i8* %ObjPtr, i8** %Ptr)
6669 The '``llvm.gcread``' intrinsic identifies reads of references from heap
6670 locations, allowing garbage collector implementations that require read
6676 The second argument is the address to read from, which should be an
6677 address allocated from the garbage collector. The first object is a
6678 pointer to the start of the referenced object, if needed by the language
6679 runtime (otherwise null).
6684 The '``llvm.gcread``' intrinsic has the same semantics as a load
6685 instruction, but may be replaced with substantially more complex code by
6686 the garbage collector runtime, as needed. The '``llvm.gcread``'
6687 intrinsic may only be used in a function which :ref:`specifies a GC
6692 '``llvm.gcwrite``' Intrinsic
6693 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6700 declare void @llvm.gcwrite(i8* %P1, i8* %Obj, i8** %P2)
6705 The '``llvm.gcwrite``' intrinsic identifies writes of references to heap
6706 locations, allowing garbage collector implementations that require write
6707 barriers (such as generational or reference counting collectors).
6712 The first argument is the reference to store, the second is the start of
6713 the object to store it to, and the third is the address of the field of
6714 Obj to store to. If the runtime does not require a pointer to the
6715 object, Obj may be null.
6720 The '``llvm.gcwrite``' intrinsic has the same semantics as a store
6721 instruction, but may be replaced with substantially more complex code by
6722 the garbage collector runtime, as needed. The '``llvm.gcwrite``'
6723 intrinsic may only be used in a function which :ref:`specifies a GC
6726 Code Generator Intrinsics
6727 -------------------------
6729 These intrinsics are provided by LLVM to expose special features that
6730 may only be implemented with code generator support.
6732 '``llvm.returnaddress``' Intrinsic
6733 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6740 declare i8 *@llvm.returnaddress(i32 <level>)
6745 The '``llvm.returnaddress``' intrinsic attempts to compute a
6746 target-specific value indicating the return address of the current
6747 function or one of its callers.
6752 The argument to this intrinsic indicates which function to return the
6753 address for. Zero indicates the calling function, one indicates its
6754 caller, etc. The argument is **required** to be a constant integer
6760 The '``llvm.returnaddress``' intrinsic either returns a pointer
6761 indicating the return address of the specified call frame, or zero if it
6762 cannot be identified. The value returned by this intrinsic is likely to
6763 be incorrect or 0 for arguments other than zero, so it should only be
6764 used for debugging purposes.
6766 Note that calling this intrinsic does not prevent function inlining or
6767 other aggressive transformations, so the value returned may not be that
6768 of the obvious source-language caller.
6770 '``llvm.frameaddress``' Intrinsic
6771 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6778 declare i8* @llvm.frameaddress(i32 <level>)
6783 The '``llvm.frameaddress``' intrinsic attempts to return the
6784 target-specific frame pointer value for the specified stack frame.
6789 The argument to this intrinsic indicates which function to return the
6790 frame pointer for. Zero indicates the calling function, one indicates
6791 its caller, etc. The argument is **required** to be a constant integer
6797 The '``llvm.frameaddress``' intrinsic either returns a pointer
6798 indicating the frame address of the specified call frame, or zero if it
6799 cannot be identified. The value returned by this intrinsic is likely to
6800 be incorrect or 0 for arguments other than zero, so it should only be
6801 used for debugging purposes.
6803 Note that calling this intrinsic does not prevent function inlining or
6804 other aggressive transformations, so the value returned may not be that
6805 of the obvious source-language caller.
6807 .. _int_read_register:
6808 .. _int_write_register:
6810 '``llvm.read_register``' and '``llvm.write_register``' Intrinsics
6811 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6818 declare i32 @llvm.read_register.i32(metadata)
6819 declare i64 @llvm.read_register.i64(metadata)
6820 declare void @llvm.write_register.i32(metadata, i32 @value)
6821 declare void @llvm.write_register.i64(metadata, i64 @value)
6822 !0 = metadata !{metadata !"sp\00"}
6827 The '``llvm.read_register``' and '``llvm.write_register``' intrinsics
6828 provides access to the named register. The register must be valid on
6829 the architecture being compiled to. The type needs to be compatible
6830 with the register being read.
6835 The '``llvm.read_register``' intrinsic returns the current value of the
6836 register, where possible. The '``llvm.write_register``' intrinsic sets
6837 the current value of the register, where possible.
6839 This is useful to implement named register global variables that need
6840 to always be mapped to a specific register, as is common practice on
6841 bare-metal programs including OS kernels.
6843 The compiler doesn't check for register availability or use of the used
6844 register in surrounding code, including inline assembly. Because of that,
6845 allocatable registers are not supported.
6847 Warning: So far it only works with the stack pointer on selected
6848 architectures (ARM, ARM64, x86_64 and AArch64). Significant amount of
6849 work is needed to support other registers and even more so, allocatable
6854 '``llvm.stacksave``' Intrinsic
6855 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6862 declare i8* @llvm.stacksave()
6867 The '``llvm.stacksave``' intrinsic is used to remember the current state
6868 of the function stack, for use with
6869 :ref:`llvm.stackrestore <int_stackrestore>`. This is useful for
6870 implementing language features like scoped automatic variable sized
6876 This intrinsic returns a opaque pointer value that can be passed to
6877 :ref:`llvm.stackrestore <int_stackrestore>`. When an
6878 ``llvm.stackrestore`` intrinsic is executed with a value saved from
6879 ``llvm.stacksave``, it effectively restores the state of the stack to
6880 the state it was in when the ``llvm.stacksave`` intrinsic executed. In
6881 practice, this pops any :ref:`alloca <i_alloca>` blocks from the stack that
6882 were allocated after the ``llvm.stacksave`` was executed.
6884 .. _int_stackrestore:
6886 '``llvm.stackrestore``' Intrinsic
6887 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6894 declare void @llvm.stackrestore(i8* %ptr)
6899 The '``llvm.stackrestore``' intrinsic is used to restore the state of
6900 the function stack to the state it was in when the corresponding
6901 :ref:`llvm.stacksave <int_stacksave>` intrinsic executed. This is
6902 useful for implementing language features like scoped automatic variable
6903 sized arrays in C99.
6908 See the description for :ref:`llvm.stacksave <int_stacksave>`.
6910 '``llvm.prefetch``' Intrinsic
6911 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6918 declare void @llvm.prefetch(i8* <address>, i32 <rw>, i32 <locality>, i32 <cache type>)
6923 The '``llvm.prefetch``' intrinsic is a hint to the code generator to
6924 insert a prefetch instruction if supported; otherwise, it is a noop.
6925 Prefetches have no effect on the behavior of the program but can change
6926 its performance characteristics.
6931 ``address`` is the address to be prefetched, ``rw`` is the specifier
6932 determining if the fetch should be for a read (0) or write (1), and
6933 ``locality`` is a temporal locality specifier ranging from (0) - no
6934 locality, to (3) - extremely local keep in cache. The ``cache type``
6935 specifies whether the prefetch is performed on the data (1) or
6936 instruction (0) cache. The ``rw``, ``locality`` and ``cache type``
6937 arguments must be constant integers.
6942 This intrinsic does not modify the behavior of the program. In
6943 particular, prefetches cannot trap and do not produce a value. On
6944 targets that support this intrinsic, the prefetch can provide hints to
6945 the processor cache for better performance.
6947 '``llvm.pcmarker``' Intrinsic
6948 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6955 declare void @llvm.pcmarker(i32 <id>)
6960 The '``llvm.pcmarker``' intrinsic is a method to export a Program
6961 Counter (PC) in a region of code to simulators and other tools. The
6962 method is target specific, but it is expected that the marker will use
6963 exported symbols to transmit the PC of the marker. The marker makes no
6964 guarantees that it will remain with any specific instruction after
6965 optimizations. It is possible that the presence of a marker will inhibit
6966 optimizations. The intended use is to be inserted after optimizations to
6967 allow correlations of simulation runs.
6972 ``id`` is a numerical id identifying the marker.
6977 This intrinsic does not modify the behavior of the program. Backends
6978 that do not support this intrinsic may ignore it.
6980 '``llvm.readcyclecounter``' Intrinsic
6981 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6988 declare i64 @llvm.readcyclecounter()
6993 The '``llvm.readcyclecounter``' intrinsic provides access to the cycle
6994 counter register (or similar low latency, high accuracy clocks) on those
6995 targets that support it. On X86, it should map to RDTSC. On Alpha, it
6996 should map to RPCC. As the backing counters overflow quickly (on the
6997 order of 9 seconds on alpha), this should only be used for small
7003 When directly supported, reading the cycle counter should not modify any
7004 memory. Implementations are allowed to either return a application
7005 specific value or a system wide value. On backends without support, this
7006 is lowered to a constant 0.
7008 Note that runtime support may be conditional on the privilege-level code is
7009 running at and the host platform.
7011 '``llvm.clear_cache``' Intrinsic
7012 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7019 declare void @llvm.clear_cache(i8*, i8*)
7024 The '``llvm.clear_cache``' intrinsic ensures visibility of modifications
7025 in the specified range to the execution unit of the processor. On
7026 targets with non-unified instruction and data cache, the implementation
7027 flushes the instruction cache.
7032 On platforms with coherent instruction and data caches (e.g. x86), this
7033 intrinsic is a nop. On platforms with non-coherent instruction and data
7034 cache (e.g. ARM, MIPS), the intrinsic is lowered either to appropriate
7035 instructions or a system call, if cache flushing requires special
7038 The default behavior is to emit a call to ``__clear_cache`` from the run
7041 This instrinsic does *not* empty the instruction pipeline. Modifications
7042 of the current function are outside the scope of the intrinsic.
7044 Standard C Library Intrinsics
7045 -----------------------------
7047 LLVM provides intrinsics for a few important standard C library
7048 functions. These intrinsics allow source-language front-ends to pass
7049 information about the alignment of the pointer arguments to the code
7050 generator, providing opportunity for more efficient code generation.
7054 '``llvm.memcpy``' Intrinsic
7055 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7060 This is an overloaded intrinsic. You can use ``llvm.memcpy`` on any
7061 integer bit width and for different address spaces. Not all targets
7062 support all bit widths however.
7066 declare void @llvm.memcpy.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
7067 i32 <len>, i32 <align>, i1 <isvolatile>)
7068 declare void @llvm.memcpy.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
7069 i64 <len>, i32 <align>, i1 <isvolatile>)
7074 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
7075 source location to the destination location.
7077 Note that, unlike the standard libc function, the ``llvm.memcpy.*``
7078 intrinsics do not return a value, takes extra alignment/isvolatile
7079 arguments and the pointers can be in specified address spaces.
7084 The first argument is a pointer to the destination, the second is a
7085 pointer to the source. The third argument is an integer argument
7086 specifying the number of bytes to copy, the fourth argument is the
7087 alignment of the source and destination locations, and the fifth is a
7088 boolean indicating a volatile access.
7090 If the call to this intrinsic has an alignment value that is not 0 or 1,
7091 then the caller guarantees that both the source and destination pointers
7092 are aligned to that boundary.
7094 If the ``isvolatile`` parameter is ``true``, the ``llvm.memcpy`` call is
7095 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
7096 very cleanly specified and it is unwise to depend on it.
7101 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
7102 source location to the destination location, which are not allowed to
7103 overlap. It copies "len" bytes of memory over. If the argument is known
7104 to be aligned to some boundary, this can be specified as the fourth
7105 argument, otherwise it should be set to 0 or 1 (both meaning no alignment).
7107 '``llvm.memmove``' Intrinsic
7108 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7113 This is an overloaded intrinsic. You can use llvm.memmove on any integer
7114 bit width and for different address space. Not all targets support all
7119 declare void @llvm.memmove.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
7120 i32 <len>, i32 <align>, i1 <isvolatile>)
7121 declare void @llvm.memmove.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
7122 i64 <len>, i32 <align>, i1 <isvolatile>)
7127 The '``llvm.memmove.*``' intrinsics move a block of memory from the
7128 source location to the destination location. It is similar to the
7129 '``llvm.memcpy``' intrinsic but allows the two memory locations to
7132 Note that, unlike the standard libc function, the ``llvm.memmove.*``
7133 intrinsics do not return a value, takes extra alignment/isvolatile
7134 arguments and the pointers can be in specified address spaces.
7139 The first argument is a pointer to the destination, the second is a
7140 pointer to the source. The third argument is an integer argument
7141 specifying the number of bytes to copy, the fourth argument is the
7142 alignment of the source and destination locations, and the fifth is a
7143 boolean indicating a volatile access.
7145 If the call to this intrinsic has an alignment value that is not 0 or 1,
7146 then the caller guarantees that the source and destination pointers are
7147 aligned to that boundary.
7149 If the ``isvolatile`` parameter is ``true``, the ``llvm.memmove`` call
7150 is a :ref:`volatile operation <volatile>`. The detailed access behavior is
7151 not very cleanly specified and it is unwise to depend on it.
7156 The '``llvm.memmove.*``' intrinsics copy a block of memory from the
7157 source location to the destination location, which may overlap. It
7158 copies "len" bytes of memory over. If the argument is known to be
7159 aligned to some boundary, this can be specified as the fourth argument,
7160 otherwise it should be set to 0 or 1 (both meaning no alignment).
7162 '``llvm.memset.*``' Intrinsics
7163 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7168 This is an overloaded intrinsic. You can use llvm.memset on any integer
7169 bit width and for different address spaces. However, not all targets
7170 support all bit widths.
7174 declare void @llvm.memset.p0i8.i32(i8* <dest>, i8 <val>,
7175 i32 <len>, i32 <align>, i1 <isvolatile>)
7176 declare void @llvm.memset.p0i8.i64(i8* <dest>, i8 <val>,
7177 i64 <len>, i32 <align>, i1 <isvolatile>)
7182 The '``llvm.memset.*``' intrinsics fill a block of memory with a
7183 particular byte value.
7185 Note that, unlike the standard libc function, the ``llvm.memset``
7186 intrinsic does not return a value and takes extra alignment/volatile
7187 arguments. Also, the destination can be in an arbitrary address space.
7192 The first argument is a pointer to the destination to fill, the second
7193 is the byte value with which to fill it, the third argument is an
7194 integer argument specifying the number of bytes to fill, and the fourth
7195 argument is the known alignment of the destination location.
7197 If the call to this intrinsic has an alignment value that is not 0 or 1,
7198 then the caller guarantees that the destination pointer is aligned to
7201 If the ``isvolatile`` parameter is ``true``, the ``llvm.memset`` call is
7202 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
7203 very cleanly specified and it is unwise to depend on it.
7208 The '``llvm.memset.*``' intrinsics fill "len" bytes of memory starting
7209 at the destination location. If the argument is known to be aligned to
7210 some boundary, this can be specified as the fourth argument, otherwise
7211 it should be set to 0 or 1 (both meaning no alignment).
7213 '``llvm.sqrt.*``' Intrinsic
7214 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7219 This is an overloaded intrinsic. You can use ``llvm.sqrt`` on any
7220 floating point or vector of floating point type. Not all targets support
7225 declare float @llvm.sqrt.f32(float %Val)
7226 declare double @llvm.sqrt.f64(double %Val)
7227 declare x86_fp80 @llvm.sqrt.f80(x86_fp80 %Val)
7228 declare fp128 @llvm.sqrt.f128(fp128 %Val)
7229 declare ppc_fp128 @llvm.sqrt.ppcf128(ppc_fp128 %Val)
7234 The '``llvm.sqrt``' intrinsics return the sqrt of the specified operand,
7235 returning the same value as the libm '``sqrt``' functions would. Unlike
7236 ``sqrt`` in libm, however, ``llvm.sqrt`` has undefined behavior for
7237 negative numbers other than -0.0 (which allows for better optimization,
7238 because there is no need to worry about errno being set).
7239 ``llvm.sqrt(-0.0)`` is defined to return -0.0 like IEEE sqrt.
7244 The argument and return value are floating point numbers of the same
7250 This function returns the sqrt of the specified operand if it is a
7251 nonnegative floating point number.
7253 '``llvm.powi.*``' Intrinsic
7254 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7259 This is an overloaded intrinsic. You can use ``llvm.powi`` on any
7260 floating point or vector of floating point type. Not all targets support
7265 declare float @llvm.powi.f32(float %Val, i32 %power)
7266 declare double @llvm.powi.f64(double %Val, i32 %power)
7267 declare x86_fp80 @llvm.powi.f80(x86_fp80 %Val, i32 %power)
7268 declare fp128 @llvm.powi.f128(fp128 %Val, i32 %power)
7269 declare ppc_fp128 @llvm.powi.ppcf128(ppc_fp128 %Val, i32 %power)
7274 The '``llvm.powi.*``' intrinsics return the first operand raised to the
7275 specified (positive or negative) power. The order of evaluation of
7276 multiplications is not defined. When a vector of floating point type is
7277 used, the second argument remains a scalar integer value.
7282 The second argument is an integer power, and the first is a value to
7283 raise to that power.
7288 This function returns the first value raised to the second power with an
7289 unspecified sequence of rounding operations.
7291 '``llvm.sin.*``' Intrinsic
7292 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7297 This is an overloaded intrinsic. You can use ``llvm.sin`` on any
7298 floating point or vector of floating point type. Not all targets support
7303 declare float @llvm.sin.f32(float %Val)
7304 declare double @llvm.sin.f64(double %Val)
7305 declare x86_fp80 @llvm.sin.f80(x86_fp80 %Val)
7306 declare fp128 @llvm.sin.f128(fp128 %Val)
7307 declare ppc_fp128 @llvm.sin.ppcf128(ppc_fp128 %Val)
7312 The '``llvm.sin.*``' intrinsics return the sine of the operand.
7317 The argument and return value are floating point numbers of the same
7323 This function returns the sine of the specified operand, returning the
7324 same values as the libm ``sin`` functions would, and handles error
7325 conditions in the same way.
7327 '``llvm.cos.*``' Intrinsic
7328 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7333 This is an overloaded intrinsic. You can use ``llvm.cos`` on any
7334 floating point or vector of floating point type. Not all targets support
7339 declare float @llvm.cos.f32(float %Val)
7340 declare double @llvm.cos.f64(double %Val)
7341 declare x86_fp80 @llvm.cos.f80(x86_fp80 %Val)
7342 declare fp128 @llvm.cos.f128(fp128 %Val)
7343 declare ppc_fp128 @llvm.cos.ppcf128(ppc_fp128 %Val)
7348 The '``llvm.cos.*``' intrinsics return the cosine of the operand.
7353 The argument and return value are floating point numbers of the same
7359 This function returns the cosine of the specified operand, returning the
7360 same values as the libm ``cos`` functions would, and handles error
7361 conditions in the same way.
7363 '``llvm.pow.*``' Intrinsic
7364 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7369 This is an overloaded intrinsic. You can use ``llvm.pow`` on any
7370 floating point or vector of floating point type. Not all targets support
7375 declare float @llvm.pow.f32(float %Val, float %Power)
7376 declare double @llvm.pow.f64(double %Val, double %Power)
7377 declare x86_fp80 @llvm.pow.f80(x86_fp80 %Val, x86_fp80 %Power)
7378 declare fp128 @llvm.pow.f128(fp128 %Val, fp128 %Power)
7379 declare ppc_fp128 @llvm.pow.ppcf128(ppc_fp128 %Val, ppc_fp128 Power)
7384 The '``llvm.pow.*``' intrinsics return the first operand raised to the
7385 specified (positive or negative) power.
7390 The second argument is a floating point power, and the first is a value
7391 to raise to that power.
7396 This function returns the first value raised to the second power,
7397 returning the same values as the libm ``pow`` functions would, and
7398 handles error conditions in the same way.
7400 '``llvm.exp.*``' Intrinsic
7401 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7406 This is an overloaded intrinsic. You can use ``llvm.exp`` on any
7407 floating point or vector of floating point type. Not all targets support
7412 declare float @llvm.exp.f32(float %Val)
7413 declare double @llvm.exp.f64(double %Val)
7414 declare x86_fp80 @llvm.exp.f80(x86_fp80 %Val)
7415 declare fp128 @llvm.exp.f128(fp128 %Val)
7416 declare ppc_fp128 @llvm.exp.ppcf128(ppc_fp128 %Val)
7421 The '``llvm.exp.*``' intrinsics perform the exp function.
7426 The argument and return value are floating point numbers of the same
7432 This function returns the same values as the libm ``exp`` functions
7433 would, and handles error conditions in the same way.
7435 '``llvm.exp2.*``' Intrinsic
7436 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7441 This is an overloaded intrinsic. You can use ``llvm.exp2`` on any
7442 floating point or vector of floating point type. Not all targets support
7447 declare float @llvm.exp2.f32(float %Val)
7448 declare double @llvm.exp2.f64(double %Val)
7449 declare x86_fp80 @llvm.exp2.f80(x86_fp80 %Val)
7450 declare fp128 @llvm.exp2.f128(fp128 %Val)
7451 declare ppc_fp128 @llvm.exp2.ppcf128(ppc_fp128 %Val)
7456 The '``llvm.exp2.*``' intrinsics perform the exp2 function.
7461 The argument and return value are floating point numbers of the same
7467 This function returns the same values as the libm ``exp2`` functions
7468 would, and handles error conditions in the same way.
7470 '``llvm.log.*``' Intrinsic
7471 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7476 This is an overloaded intrinsic. You can use ``llvm.log`` on any
7477 floating point or vector of floating point type. Not all targets support
7482 declare float @llvm.log.f32(float %Val)
7483 declare double @llvm.log.f64(double %Val)
7484 declare x86_fp80 @llvm.log.f80(x86_fp80 %Val)
7485 declare fp128 @llvm.log.f128(fp128 %Val)
7486 declare ppc_fp128 @llvm.log.ppcf128(ppc_fp128 %Val)
7491 The '``llvm.log.*``' intrinsics perform the log function.
7496 The argument and return value are floating point numbers of the same
7502 This function returns the same values as the libm ``log`` functions
7503 would, and handles error conditions in the same way.
7505 '``llvm.log10.*``' Intrinsic
7506 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7511 This is an overloaded intrinsic. You can use ``llvm.log10`` on any
7512 floating point or vector of floating point type. Not all targets support
7517 declare float @llvm.log10.f32(float %Val)
7518 declare double @llvm.log10.f64(double %Val)
7519 declare x86_fp80 @llvm.log10.f80(x86_fp80 %Val)
7520 declare fp128 @llvm.log10.f128(fp128 %Val)
7521 declare ppc_fp128 @llvm.log10.ppcf128(ppc_fp128 %Val)
7526 The '``llvm.log10.*``' intrinsics perform the log10 function.
7531 The argument and return value are floating point numbers of the same
7537 This function returns the same values as the libm ``log10`` functions
7538 would, and handles error conditions in the same way.
7540 '``llvm.log2.*``' Intrinsic
7541 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7546 This is an overloaded intrinsic. You can use ``llvm.log2`` on any
7547 floating point or vector of floating point type. Not all targets support
7552 declare float @llvm.log2.f32(float %Val)
7553 declare double @llvm.log2.f64(double %Val)
7554 declare x86_fp80 @llvm.log2.f80(x86_fp80 %Val)
7555 declare fp128 @llvm.log2.f128(fp128 %Val)
7556 declare ppc_fp128 @llvm.log2.ppcf128(ppc_fp128 %Val)
7561 The '``llvm.log2.*``' intrinsics perform the log2 function.
7566 The argument and return value are floating point numbers of the same
7572 This function returns the same values as the libm ``log2`` functions
7573 would, and handles error conditions in the same way.
7575 '``llvm.fma.*``' Intrinsic
7576 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7581 This is an overloaded intrinsic. You can use ``llvm.fma`` on any
7582 floating point or vector of floating point type. Not all targets support
7587 declare float @llvm.fma.f32(float %a, float %b, float %c)
7588 declare double @llvm.fma.f64(double %a, double %b, double %c)
7589 declare x86_fp80 @llvm.fma.f80(x86_fp80 %a, x86_fp80 %b, x86_fp80 %c)
7590 declare fp128 @llvm.fma.f128(fp128 %a, fp128 %b, fp128 %c)
7591 declare ppc_fp128 @llvm.fma.ppcf128(ppc_fp128 %a, ppc_fp128 %b, ppc_fp128 %c)
7596 The '``llvm.fma.*``' intrinsics perform the fused multiply-add
7602 The argument and return value are floating point numbers of the same
7608 This function returns the same values as the libm ``fma`` functions
7609 would, and does not set errno.
7611 '``llvm.fabs.*``' Intrinsic
7612 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7617 This is an overloaded intrinsic. You can use ``llvm.fabs`` on any
7618 floating point or vector of floating point type. Not all targets support
7623 declare float @llvm.fabs.f32(float %Val)
7624 declare double @llvm.fabs.f64(double %Val)
7625 declare x86_fp80 @llvm.fabs.f80(x86_fp80 %Val)
7626 declare fp128 @llvm.fabs.f128(fp128 %Val)
7627 declare ppc_fp128 @llvm.fabs.ppcf128(ppc_fp128 %Val)
7632 The '``llvm.fabs.*``' intrinsics return the absolute value of the
7638 The argument and return value are floating point numbers of the same
7644 This function returns the same values as the libm ``fabs`` functions
7645 would, and handles error conditions in the same way.
7647 '``llvm.copysign.*``' Intrinsic
7648 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7653 This is an overloaded intrinsic. You can use ``llvm.copysign`` on any
7654 floating point or vector of floating point type. Not all targets support
7659 declare float @llvm.copysign.f32(float %Mag, float %Sgn)
7660 declare double @llvm.copysign.f64(double %Mag, double %Sgn)
7661 declare x86_fp80 @llvm.copysign.f80(x86_fp80 %Mag, x86_fp80 %Sgn)
7662 declare fp128 @llvm.copysign.f128(fp128 %Mag, fp128 %Sgn)
7663 declare ppc_fp128 @llvm.copysign.ppcf128(ppc_fp128 %Mag, ppc_fp128 %Sgn)
7668 The '``llvm.copysign.*``' intrinsics return a value with the magnitude of the
7669 first operand and the sign of the second operand.
7674 The arguments and return value are floating point numbers of the same
7680 This function returns the same values as the libm ``copysign``
7681 functions would, and handles error conditions in the same way.
7683 '``llvm.floor.*``' Intrinsic
7684 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7689 This is an overloaded intrinsic. You can use ``llvm.floor`` on any
7690 floating point or vector of floating point type. Not all targets support
7695 declare float @llvm.floor.f32(float %Val)
7696 declare double @llvm.floor.f64(double %Val)
7697 declare x86_fp80 @llvm.floor.f80(x86_fp80 %Val)
7698 declare fp128 @llvm.floor.f128(fp128 %Val)
7699 declare ppc_fp128 @llvm.floor.ppcf128(ppc_fp128 %Val)
7704 The '``llvm.floor.*``' intrinsics return the floor of the operand.
7709 The argument and return value are floating point numbers of the same
7715 This function returns the same values as the libm ``floor`` functions
7716 would, and handles error conditions in the same way.
7718 '``llvm.ceil.*``' Intrinsic
7719 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7724 This is an overloaded intrinsic. You can use ``llvm.ceil`` on any
7725 floating point or vector of floating point type. Not all targets support
7730 declare float @llvm.ceil.f32(float %Val)
7731 declare double @llvm.ceil.f64(double %Val)
7732 declare x86_fp80 @llvm.ceil.f80(x86_fp80 %Val)
7733 declare fp128 @llvm.ceil.f128(fp128 %Val)
7734 declare ppc_fp128 @llvm.ceil.ppcf128(ppc_fp128 %Val)
7739 The '``llvm.ceil.*``' intrinsics return the ceiling of the operand.
7744 The argument and return value are floating point numbers of the same
7750 This function returns the same values as the libm ``ceil`` functions
7751 would, and handles error conditions in the same way.
7753 '``llvm.trunc.*``' Intrinsic
7754 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7759 This is an overloaded intrinsic. You can use ``llvm.trunc`` on any
7760 floating point or vector of floating point type. Not all targets support
7765 declare float @llvm.trunc.f32(float %Val)
7766 declare double @llvm.trunc.f64(double %Val)
7767 declare x86_fp80 @llvm.trunc.f80(x86_fp80 %Val)
7768 declare fp128 @llvm.trunc.f128(fp128 %Val)
7769 declare ppc_fp128 @llvm.trunc.ppcf128(ppc_fp128 %Val)
7774 The '``llvm.trunc.*``' intrinsics returns the operand rounded to the
7775 nearest integer not larger in magnitude than the operand.
7780 The argument and return value are floating point numbers of the same
7786 This function returns the same values as the libm ``trunc`` functions
7787 would, and handles error conditions in the same way.
7789 '``llvm.rint.*``' Intrinsic
7790 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7795 This is an overloaded intrinsic. You can use ``llvm.rint`` on any
7796 floating point or vector of floating point type. Not all targets support
7801 declare float @llvm.rint.f32(float %Val)
7802 declare double @llvm.rint.f64(double %Val)
7803 declare x86_fp80 @llvm.rint.f80(x86_fp80 %Val)
7804 declare fp128 @llvm.rint.f128(fp128 %Val)
7805 declare ppc_fp128 @llvm.rint.ppcf128(ppc_fp128 %Val)
7810 The '``llvm.rint.*``' intrinsics returns the operand rounded to the
7811 nearest integer. It may raise an inexact floating-point exception if the
7812 operand isn't an integer.
7817 The argument and return value are floating point numbers of the same
7823 This function returns the same values as the libm ``rint`` functions
7824 would, and handles error conditions in the same way.
7826 '``llvm.nearbyint.*``' Intrinsic
7827 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7832 This is an overloaded intrinsic. You can use ``llvm.nearbyint`` on any
7833 floating point or vector of floating point type. Not all targets support
7838 declare float @llvm.nearbyint.f32(float %Val)
7839 declare double @llvm.nearbyint.f64(double %Val)
7840 declare x86_fp80 @llvm.nearbyint.f80(x86_fp80 %Val)
7841 declare fp128 @llvm.nearbyint.f128(fp128 %Val)
7842 declare ppc_fp128 @llvm.nearbyint.ppcf128(ppc_fp128 %Val)
7847 The '``llvm.nearbyint.*``' intrinsics returns the operand rounded to the
7853 The argument and return value are floating point numbers of the same
7859 This function returns the same values as the libm ``nearbyint``
7860 functions would, and handles error conditions in the same way.
7862 '``llvm.round.*``' Intrinsic
7863 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7868 This is an overloaded intrinsic. You can use ``llvm.round`` on any
7869 floating point or vector of floating point type. Not all targets support
7874 declare float @llvm.round.f32(float %Val)
7875 declare double @llvm.round.f64(double %Val)
7876 declare x86_fp80 @llvm.round.f80(x86_fp80 %Val)
7877 declare fp128 @llvm.round.f128(fp128 %Val)
7878 declare ppc_fp128 @llvm.round.ppcf128(ppc_fp128 %Val)
7883 The '``llvm.round.*``' intrinsics returns the operand rounded to the
7889 The argument and return value are floating point numbers of the same
7895 This function returns the same values as the libm ``round``
7896 functions would, and handles error conditions in the same way.
7898 Bit Manipulation Intrinsics
7899 ---------------------------
7901 LLVM provides intrinsics for a few important bit manipulation
7902 operations. These allow efficient code generation for some algorithms.
7904 '``llvm.bswap.*``' Intrinsics
7905 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7910 This is an overloaded intrinsic function. You can use bswap on any
7911 integer type that is an even number of bytes (i.e. BitWidth % 16 == 0).
7915 declare i16 @llvm.bswap.i16(i16 <id>)
7916 declare i32 @llvm.bswap.i32(i32 <id>)
7917 declare i64 @llvm.bswap.i64(i64 <id>)
7922 The '``llvm.bswap``' family of intrinsics is used to byte swap integer
7923 values with an even number of bytes (positive multiple of 16 bits).
7924 These are useful for performing operations on data that is not in the
7925 target's native byte order.
7930 The ``llvm.bswap.i16`` intrinsic returns an i16 value that has the high
7931 and low byte of the input i16 swapped. Similarly, the ``llvm.bswap.i32``
7932 intrinsic returns an i32 value that has the four bytes of the input i32
7933 swapped, so that if the input bytes are numbered 0, 1, 2, 3 then the
7934 returned i32 will have its bytes in 3, 2, 1, 0 order. The
7935 ``llvm.bswap.i48``, ``llvm.bswap.i64`` and other intrinsics extend this
7936 concept to additional even-byte lengths (6 bytes, 8 bytes and more,
7939 '``llvm.ctpop.*``' Intrinsic
7940 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7945 This is an overloaded intrinsic. You can use llvm.ctpop on any integer
7946 bit width, or on any vector with integer elements. Not all targets
7947 support all bit widths or vector types, however.
7951 declare i8 @llvm.ctpop.i8(i8 <src>)
7952 declare i16 @llvm.ctpop.i16(i16 <src>)
7953 declare i32 @llvm.ctpop.i32(i32 <src>)
7954 declare i64 @llvm.ctpop.i64(i64 <src>)
7955 declare i256 @llvm.ctpop.i256(i256 <src>)
7956 declare <2 x i32> @llvm.ctpop.v2i32(<2 x i32> <src>)
7961 The '``llvm.ctpop``' family of intrinsics counts the number of bits set
7967 The only argument is the value to be counted. The argument may be of any
7968 integer type, or a vector with integer elements. The return type must
7969 match the argument type.
7974 The '``llvm.ctpop``' intrinsic counts the 1's in a variable, or within
7975 each element of a vector.
7977 '``llvm.ctlz.*``' Intrinsic
7978 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7983 This is an overloaded intrinsic. You can use ``llvm.ctlz`` on any
7984 integer bit width, or any vector whose elements are integers. Not all
7985 targets support all bit widths or vector types, however.
7989 declare i8 @llvm.ctlz.i8 (i8 <src>, i1 <is_zero_undef>)
7990 declare i16 @llvm.ctlz.i16 (i16 <src>, i1 <is_zero_undef>)
7991 declare i32 @llvm.ctlz.i32 (i32 <src>, i1 <is_zero_undef>)
7992 declare i64 @llvm.ctlz.i64 (i64 <src>, i1 <is_zero_undef>)
7993 declare i256 @llvm.ctlz.i256(i256 <src>, i1 <is_zero_undef>)
7994 declase <2 x i32> @llvm.ctlz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
7999 The '``llvm.ctlz``' family of intrinsic functions counts the number of
8000 leading zeros in a variable.
8005 The first argument is the value to be counted. This argument may be of
8006 any integer type, or a vectory with integer element type. The return
8007 type must match the first argument type.
8009 The second argument must be a constant and is a flag to indicate whether
8010 the intrinsic should ensure that a zero as the first argument produces a
8011 defined result. Historically some architectures did not provide a
8012 defined result for zero values as efficiently, and many algorithms are
8013 now predicated on avoiding zero-value inputs.
8018 The '``llvm.ctlz``' intrinsic counts the leading (most significant)
8019 zeros in a variable, or within each element of the vector. If
8020 ``src == 0`` then the result is the size in bits of the type of ``src``
8021 if ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
8022 ``llvm.ctlz(i32 2) = 30``.
8024 '``llvm.cttz.*``' Intrinsic
8025 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8030 This is an overloaded intrinsic. You can use ``llvm.cttz`` on any
8031 integer bit width, or any vector of integer elements. Not all targets
8032 support all bit widths or vector types, however.
8036 declare i8 @llvm.cttz.i8 (i8 <src>, i1 <is_zero_undef>)
8037 declare i16 @llvm.cttz.i16 (i16 <src>, i1 <is_zero_undef>)
8038 declare i32 @llvm.cttz.i32 (i32 <src>, i1 <is_zero_undef>)
8039 declare i64 @llvm.cttz.i64 (i64 <src>, i1 <is_zero_undef>)
8040 declare i256 @llvm.cttz.i256(i256 <src>, i1 <is_zero_undef>)
8041 declase <2 x i32> @llvm.cttz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
8046 The '``llvm.cttz``' family of intrinsic functions counts the number of
8052 The first argument is the value to be counted. This argument may be of
8053 any integer type, or a vectory with integer element type. The return
8054 type must match the first argument type.
8056 The second argument must be a constant and is a flag to indicate whether
8057 the intrinsic should ensure that a zero as the first argument produces a
8058 defined result. Historically some architectures did not provide a
8059 defined result for zero values as efficiently, and many algorithms are
8060 now predicated on avoiding zero-value inputs.
8065 The '``llvm.cttz``' intrinsic counts the trailing (least significant)
8066 zeros in a variable, or within each element of a vector. If ``src == 0``
8067 then the result is the size in bits of the type of ``src`` if
8068 ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
8069 ``llvm.cttz(2) = 1``.
8071 Arithmetic with Overflow Intrinsics
8072 -----------------------------------
8074 LLVM provides intrinsics for some arithmetic with overflow operations.
8076 '``llvm.sadd.with.overflow.*``' Intrinsics
8077 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8082 This is an overloaded intrinsic. You can use ``llvm.sadd.with.overflow``
8083 on any integer bit width.
8087 declare {i16, i1} @llvm.sadd.with.overflow.i16(i16 %a, i16 %b)
8088 declare {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
8089 declare {i64, i1} @llvm.sadd.with.overflow.i64(i64 %a, i64 %b)
8094 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
8095 a signed addition of the two arguments, and indicate whether an overflow
8096 occurred during the signed summation.
8101 The arguments (%a and %b) and the first element of the result structure
8102 may be of integer types of any bit width, but they must have the same
8103 bit width. The second element of the result structure must be of type
8104 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
8110 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
8111 a signed addition of the two variables. They return a structure --- the
8112 first element of which is the signed summation, and the second element
8113 of which is a bit specifying if the signed summation resulted in an
8119 .. code-block:: llvm
8121 %res = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
8122 %sum = extractvalue {i32, i1} %res, 0
8123 %obit = extractvalue {i32, i1} %res, 1
8124 br i1 %obit, label %overflow, label %normal
8126 '``llvm.uadd.with.overflow.*``' Intrinsics
8127 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8132 This is an overloaded intrinsic. You can use ``llvm.uadd.with.overflow``
8133 on any integer bit width.
8137 declare {i16, i1} @llvm.uadd.with.overflow.i16(i16 %a, i16 %b)
8138 declare {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
8139 declare {i64, i1} @llvm.uadd.with.overflow.i64(i64 %a, i64 %b)
8144 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
8145 an unsigned addition of the two arguments, and indicate whether a carry
8146 occurred during the unsigned summation.
8151 The arguments (%a and %b) and the first element of the result structure
8152 may be of integer types of any bit width, but they must have the same
8153 bit width. The second element of the result structure must be of type
8154 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8160 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
8161 an unsigned addition of the two arguments. They return a structure --- the
8162 first element of which is the sum, and the second element of which is a
8163 bit specifying if the unsigned summation resulted in a carry.
8168 .. code-block:: llvm
8170 %res = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
8171 %sum = extractvalue {i32, i1} %res, 0
8172 %obit = extractvalue {i32, i1} %res, 1
8173 br i1 %obit, label %carry, label %normal
8175 '``llvm.ssub.with.overflow.*``' Intrinsics
8176 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8181 This is an overloaded intrinsic. You can use ``llvm.ssub.with.overflow``
8182 on any integer bit width.
8186 declare {i16, i1} @llvm.ssub.with.overflow.i16(i16 %a, i16 %b)
8187 declare {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
8188 declare {i64, i1} @llvm.ssub.with.overflow.i64(i64 %a, i64 %b)
8193 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
8194 a signed subtraction of the two arguments, and indicate whether an
8195 overflow occurred during the signed subtraction.
8200 The arguments (%a and %b) and the first element of the result structure
8201 may be of integer types of any bit width, but they must have the same
8202 bit width. The second element of the result structure must be of type
8203 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
8209 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
8210 a signed subtraction of the two arguments. They return a structure --- the
8211 first element of which is the subtraction, and the second element of
8212 which is a bit specifying if the signed subtraction resulted in an
8218 .. code-block:: llvm
8220 %res = call {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
8221 %sum = extractvalue {i32, i1} %res, 0
8222 %obit = extractvalue {i32, i1} %res, 1
8223 br i1 %obit, label %overflow, label %normal
8225 '``llvm.usub.with.overflow.*``' Intrinsics
8226 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8231 This is an overloaded intrinsic. You can use ``llvm.usub.with.overflow``
8232 on any integer bit width.
8236 declare {i16, i1} @llvm.usub.with.overflow.i16(i16 %a, i16 %b)
8237 declare {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
8238 declare {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
8243 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
8244 an unsigned subtraction of the two arguments, and indicate whether an
8245 overflow occurred during the unsigned subtraction.
8250 The arguments (%a and %b) and the first element of the result structure
8251 may be of integer types of any bit width, but they must have the same
8252 bit width. The second element of the result structure must be of type
8253 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8259 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
8260 an unsigned subtraction of the two arguments. They return a structure ---
8261 the first element of which is the subtraction, and the second element of
8262 which is a bit specifying if the unsigned subtraction resulted in an
8268 .. code-block:: llvm
8270 %res = call {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
8271 %sum = extractvalue {i32, i1} %res, 0
8272 %obit = extractvalue {i32, i1} %res, 1
8273 br i1 %obit, label %overflow, label %normal
8275 '``llvm.smul.with.overflow.*``' Intrinsics
8276 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8281 This is an overloaded intrinsic. You can use ``llvm.smul.with.overflow``
8282 on any integer bit width.
8286 declare {i16, i1} @llvm.smul.with.overflow.i16(i16 %a, i16 %b)
8287 declare {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
8288 declare {i64, i1} @llvm.smul.with.overflow.i64(i64 %a, i64 %b)
8293 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
8294 a signed multiplication of the two arguments, and indicate whether an
8295 overflow occurred during the signed multiplication.
8300 The arguments (%a and %b) and the first element of the result structure
8301 may be of integer types of any bit width, but they must have the same
8302 bit width. The second element of the result structure must be of type
8303 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
8309 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
8310 a signed multiplication of the two arguments. They return a structure ---
8311 the first element of which is the multiplication, and the second element
8312 of which is a bit specifying if the signed multiplication resulted in an
8318 .. code-block:: llvm
8320 %res = call {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
8321 %sum = extractvalue {i32, i1} %res, 0
8322 %obit = extractvalue {i32, i1} %res, 1
8323 br i1 %obit, label %overflow, label %normal
8325 '``llvm.umul.with.overflow.*``' Intrinsics
8326 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8331 This is an overloaded intrinsic. You can use ``llvm.umul.with.overflow``
8332 on any integer bit width.
8336 declare {i16, i1} @llvm.umul.with.overflow.i16(i16 %a, i16 %b)
8337 declare {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
8338 declare {i64, i1} @llvm.umul.with.overflow.i64(i64 %a, i64 %b)
8343 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
8344 a unsigned multiplication of the two arguments, and indicate whether an
8345 overflow occurred during the unsigned multiplication.
8350 The arguments (%a and %b) and the first element of the result structure
8351 may be of integer types of any bit width, but they must have the same
8352 bit width. The second element of the result structure must be of type
8353 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8359 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
8360 an unsigned multiplication of the two arguments. They return a structure ---
8361 the first element of which is the multiplication, and the second
8362 element of which is a bit specifying if the unsigned multiplication
8363 resulted in an overflow.
8368 .. code-block:: llvm
8370 %res = call {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
8371 %sum = extractvalue {i32, i1} %res, 0
8372 %obit = extractvalue {i32, i1} %res, 1
8373 br i1 %obit, label %overflow, label %normal
8375 Specialised Arithmetic Intrinsics
8376 ---------------------------------
8378 '``llvm.fmuladd.*``' Intrinsic
8379 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8386 declare float @llvm.fmuladd.f32(float %a, float %b, float %c)
8387 declare double @llvm.fmuladd.f64(double %a, double %b, double %c)
8392 The '``llvm.fmuladd.*``' intrinsic functions represent multiply-add
8393 expressions that can be fused if the code generator determines that (a) the
8394 target instruction set has support for a fused operation, and (b) that the
8395 fused operation is more efficient than the equivalent, separate pair of mul
8396 and add instructions.
8401 The '``llvm.fmuladd.*``' intrinsics each take three arguments: two
8402 multiplicands, a and b, and an addend c.
8411 %0 = call float @llvm.fmuladd.f32(%a, %b, %c)
8413 is equivalent to the expression a \* b + c, except that rounding will
8414 not be performed between the multiplication and addition steps if the
8415 code generator fuses the operations. Fusion is not guaranteed, even if
8416 the target platform supports it. If a fused multiply-add is required the
8417 corresponding llvm.fma.\* intrinsic function should be used
8418 instead. This never sets errno, just as '``llvm.fma.*``'.
8423 .. code-block:: llvm
8425 %r2 = call float @llvm.fmuladd.f32(float %a, float %b, float %c) ; yields {float}:r2 = (a * b) + c
8427 Half Precision Floating Point Intrinsics
8428 ----------------------------------------
8430 For most target platforms, half precision floating point is a
8431 storage-only format. This means that it is a dense encoding (in memory)
8432 but does not support computation in the format.
8434 This means that code must first load the half-precision floating point
8435 value as an i16, then convert it to float with
8436 :ref:`llvm.convert.from.fp16 <int_convert_from_fp16>`. Computation can
8437 then be performed on the float value (including extending to double
8438 etc). To store the value back to memory, it is first converted to float
8439 if needed, then converted to i16 with
8440 :ref:`llvm.convert.to.fp16 <int_convert_to_fp16>`, then storing as an
8443 .. _int_convert_to_fp16:
8445 '``llvm.convert.to.fp16``' Intrinsic
8446 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8453 declare i16 @llvm.convert.to.fp16(f32 %a)
8458 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion
8459 from single precision floating point format to half precision floating
8465 The intrinsic function contains single argument - the value to be
8471 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion
8472 from single precision floating point format to half precision floating
8473 point format. The return value is an ``i16`` which contains the
8479 .. code-block:: llvm
8481 %res = call i16 @llvm.convert.to.fp16(f32 %a)
8482 store i16 %res, i16* @x, align 2
8484 .. _int_convert_from_fp16:
8486 '``llvm.convert.from.fp16``' Intrinsic
8487 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8494 declare f32 @llvm.convert.from.fp16(i16 %a)
8499 The '``llvm.convert.from.fp16``' intrinsic function performs a
8500 conversion from half precision floating point format to single precision
8501 floating point format.
8506 The intrinsic function contains single argument - the value to be
8512 The '``llvm.convert.from.fp16``' intrinsic function performs a
8513 conversion from half single precision floating point format to single
8514 precision floating point format. The input half-float value is
8515 represented by an ``i16`` value.
8520 .. code-block:: llvm
8522 %a = load i16* @x, align 2
8523 %res = call f32 @llvm.convert.from.fp16(i16 %a)
8528 The LLVM debugger intrinsics (which all start with ``llvm.dbg.``
8529 prefix), are described in the `LLVM Source Level
8530 Debugging <SourceLevelDebugging.html#format_common_intrinsics>`_
8533 Exception Handling Intrinsics
8534 -----------------------------
8536 The LLVM exception handling intrinsics (which all start with
8537 ``llvm.eh.`` prefix), are described in the `LLVM Exception
8538 Handling <ExceptionHandling.html#format_common_intrinsics>`_ document.
8542 Trampoline Intrinsics
8543 ---------------------
8545 These intrinsics make it possible to excise one parameter, marked with
8546 the :ref:`nest <nest>` attribute, from a function. The result is a
8547 callable function pointer lacking the nest parameter - the caller does
8548 not need to provide a value for it. Instead, the value to use is stored
8549 in advance in a "trampoline", a block of memory usually allocated on the
8550 stack, which also contains code to splice the nest value into the
8551 argument list. This is used to implement the GCC nested function address
8554 For example, if the function is ``i32 f(i8* nest %c, i32 %x, i32 %y)``
8555 then the resulting function pointer has signature ``i32 (i32, i32)*``.
8556 It can be created as follows:
8558 .. code-block:: llvm
8560 %tramp = alloca [10 x i8], align 4 ; size and alignment only correct for X86
8561 %tramp1 = getelementptr [10 x i8]* %tramp, i32 0, i32 0
8562 call i8* @llvm.init.trampoline(i8* %tramp1, i8* bitcast (i32 (i8*, i32, i32)* @f to i8*), i8* %nval)
8563 %p = call i8* @llvm.adjust.trampoline(i8* %tramp1)
8564 %fp = bitcast i8* %p to i32 (i32, i32)*
8566 The call ``%val = call i32 %fp(i32 %x, i32 %y)`` is then equivalent to
8567 ``%val = call i32 %f(i8* %nval, i32 %x, i32 %y)``.
8571 '``llvm.init.trampoline``' Intrinsic
8572 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8579 declare void @llvm.init.trampoline(i8* <tramp>, i8* <func>, i8* <nval>)
8584 This fills the memory pointed to by ``tramp`` with executable code,
8585 turning it into a trampoline.
8590 The ``llvm.init.trampoline`` intrinsic takes three arguments, all
8591 pointers. The ``tramp`` argument must point to a sufficiently large and
8592 sufficiently aligned block of memory; this memory is written to by the
8593 intrinsic. Note that the size and the alignment are target-specific -
8594 LLVM currently provides no portable way of determining them, so a
8595 front-end that generates this intrinsic needs to have some
8596 target-specific knowledge. The ``func`` argument must hold a function
8597 bitcast to an ``i8*``.
8602 The block of memory pointed to by ``tramp`` is filled with target
8603 dependent code, turning it into a function. Then ``tramp`` needs to be
8604 passed to :ref:`llvm.adjust.trampoline <int_at>` to get a pointer which can
8605 be :ref:`bitcast (to a new function) and called <int_trampoline>`. The new
8606 function's signature is the same as that of ``func`` with any arguments
8607 marked with the ``nest`` attribute removed. At most one such ``nest``
8608 argument is allowed, and it must be of pointer type. Calling the new
8609 function is equivalent to calling ``func`` with the same argument list,
8610 but with ``nval`` used for the missing ``nest`` argument. If, after
8611 calling ``llvm.init.trampoline``, the memory pointed to by ``tramp`` is
8612 modified, then the effect of any later call to the returned function
8613 pointer is undefined.
8617 '``llvm.adjust.trampoline``' Intrinsic
8618 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8625 declare i8* @llvm.adjust.trampoline(i8* <tramp>)
8630 This performs any required machine-specific adjustment to the address of
8631 a trampoline (passed as ``tramp``).
8636 ``tramp`` must point to a block of memory which already has trampoline
8637 code filled in by a previous call to
8638 :ref:`llvm.init.trampoline <int_it>`.
8643 On some architectures the address of the code to be executed needs to be
8644 different to the address where the trampoline is actually stored. This
8645 intrinsic returns the executable address corresponding to ``tramp``
8646 after performing the required machine specific adjustments. The pointer
8647 returned can then be :ref:`bitcast and executed <int_trampoline>`.
8652 This class of intrinsics exists to information about the lifetime of
8653 memory objects and ranges where variables are immutable.
8657 '``llvm.lifetime.start``' Intrinsic
8658 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8665 declare void @llvm.lifetime.start(i64 <size>, i8* nocapture <ptr>)
8670 The '``llvm.lifetime.start``' intrinsic specifies the start of a memory
8676 The first argument is a constant integer representing the size of the
8677 object, or -1 if it is variable sized. The second argument is a pointer
8683 This intrinsic indicates that before this point in the code, the value
8684 of the memory pointed to by ``ptr`` is dead. This means that it is known
8685 to never be used and has an undefined value. A load from the pointer
8686 that precedes this intrinsic can be replaced with ``'undef'``.
8690 '``llvm.lifetime.end``' Intrinsic
8691 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8698 declare void @llvm.lifetime.end(i64 <size>, i8* nocapture <ptr>)
8703 The '``llvm.lifetime.end``' intrinsic specifies the end of a memory
8709 The first argument is a constant integer representing the size of the
8710 object, or -1 if it is variable sized. The second argument is a pointer
8716 This intrinsic indicates that after this point in the code, the value of
8717 the memory pointed to by ``ptr`` is dead. This means that it is known to
8718 never be used and has an undefined value. Any stores into the memory
8719 object following this intrinsic may be removed as dead.
8721 '``llvm.invariant.start``' Intrinsic
8722 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8729 declare {}* @llvm.invariant.start(i64 <size>, i8* nocapture <ptr>)
8734 The '``llvm.invariant.start``' intrinsic specifies that the contents of
8735 a memory object will not change.
8740 The first argument is a constant integer representing the size of the
8741 object, or -1 if it is variable sized. The second argument is a pointer
8747 This intrinsic indicates that until an ``llvm.invariant.end`` that uses
8748 the return value, the referenced memory location is constant and
8751 '``llvm.invariant.end``' Intrinsic
8752 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8759 declare void @llvm.invariant.end({}* <start>, i64 <size>, i8* nocapture <ptr>)
8764 The '``llvm.invariant.end``' intrinsic specifies that the contents of a
8765 memory object are mutable.
8770 The first argument is the matching ``llvm.invariant.start`` intrinsic.
8771 The second argument is a constant integer representing the size of the
8772 object, or -1 if it is variable sized and the third argument is a
8773 pointer to the object.
8778 This intrinsic indicates that the memory is mutable again.
8783 This class of intrinsics is designed to be generic and has no specific
8786 '``llvm.var.annotation``' Intrinsic
8787 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8794 declare void @llvm.var.annotation(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
8799 The '``llvm.var.annotation``' intrinsic.
8804 The first argument is a pointer to a value, the second is a pointer to a
8805 global string, the third is a pointer to a global string which is the
8806 source file name, and the last argument is the line number.
8811 This intrinsic allows annotation of local variables with arbitrary
8812 strings. This can be useful for special purpose optimizations that want
8813 to look for these annotations. These have no other defined use; they are
8814 ignored by code generation and optimization.
8816 '``llvm.ptr.annotation.*``' Intrinsic
8817 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8822 This is an overloaded intrinsic. You can use '``llvm.ptr.annotation``' on a
8823 pointer to an integer of any width. *NOTE* you must specify an address space for
8824 the pointer. The identifier for the default address space is the integer
8829 declare i8* @llvm.ptr.annotation.p<address space>i8(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
8830 declare i16* @llvm.ptr.annotation.p<address space>i16(i16* <val>, i8* <str>, i8* <str>, i32 <int>)
8831 declare i32* @llvm.ptr.annotation.p<address space>i32(i32* <val>, i8* <str>, i8* <str>, i32 <int>)
8832 declare i64* @llvm.ptr.annotation.p<address space>i64(i64* <val>, i8* <str>, i8* <str>, i32 <int>)
8833 declare i256* @llvm.ptr.annotation.p<address space>i256(i256* <val>, i8* <str>, i8* <str>, i32 <int>)
8838 The '``llvm.ptr.annotation``' intrinsic.
8843 The first argument is a pointer to an integer value of arbitrary bitwidth
8844 (result of some expression), the second is a pointer to a global string, the
8845 third is a pointer to a global string which is the source file name, and the
8846 last argument is the line number. It returns the value of the first argument.
8851 This intrinsic allows annotation of a pointer to an integer with arbitrary
8852 strings. This can be useful for special purpose optimizations that want to look
8853 for these annotations. These have no other defined use; they are ignored by code
8854 generation and optimization.
8856 '``llvm.annotation.*``' Intrinsic
8857 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8862 This is an overloaded intrinsic. You can use '``llvm.annotation``' on
8863 any integer bit width.
8867 declare i8 @llvm.annotation.i8(i8 <val>, i8* <str>, i8* <str>, i32 <int>)
8868 declare i16 @llvm.annotation.i16(i16 <val>, i8* <str>, i8* <str>, i32 <int>)
8869 declare i32 @llvm.annotation.i32(i32 <val>, i8* <str>, i8* <str>, i32 <int>)
8870 declare i64 @llvm.annotation.i64(i64 <val>, i8* <str>, i8* <str>, i32 <int>)
8871 declare i256 @llvm.annotation.i256(i256 <val>, i8* <str>, i8* <str>, i32 <int>)
8876 The '``llvm.annotation``' intrinsic.
8881 The first argument is an integer value (result of some expression), the
8882 second is a pointer to a global string, the third is a pointer to a
8883 global string which is the source file name, and the last argument is
8884 the line number. It returns the value of the first argument.
8889 This intrinsic allows annotations to be put on arbitrary expressions
8890 with arbitrary strings. This can be useful for special purpose
8891 optimizations that want to look for these annotations. These have no
8892 other defined use; they are ignored by code generation and optimization.
8894 '``llvm.trap``' Intrinsic
8895 ^^^^^^^^^^^^^^^^^^^^^^^^^
8902 declare void @llvm.trap() noreturn nounwind
8907 The '``llvm.trap``' intrinsic.
8917 This intrinsic is lowered to the target dependent trap instruction. If
8918 the target does not have a trap instruction, this intrinsic will be
8919 lowered to a call of the ``abort()`` function.
8921 '``llvm.debugtrap``' Intrinsic
8922 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8929 declare void @llvm.debugtrap() nounwind
8934 The '``llvm.debugtrap``' intrinsic.
8944 This intrinsic is lowered to code which is intended to cause an
8945 execution trap with the intention of requesting the attention of a
8948 '``llvm.stackprotector``' Intrinsic
8949 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8956 declare void @llvm.stackprotector(i8* <guard>, i8** <slot>)
8961 The ``llvm.stackprotector`` intrinsic takes the ``guard`` and stores it
8962 onto the stack at ``slot``. The stack slot is adjusted to ensure that it
8963 is placed on the stack before local variables.
8968 The ``llvm.stackprotector`` intrinsic requires two pointer arguments.
8969 The first argument is the value loaded from the stack guard
8970 ``@__stack_chk_guard``. The second variable is an ``alloca`` that has
8971 enough space to hold the value of the guard.
8976 This intrinsic causes the prologue/epilogue inserter to force the position of
8977 the ``AllocaInst`` stack slot to be before local variables on the stack. This is
8978 to ensure that if a local variable on the stack is overwritten, it will destroy
8979 the value of the guard. When the function exits, the guard on the stack is
8980 checked against the original guard by ``llvm.stackprotectorcheck``. If they are
8981 different, then ``llvm.stackprotectorcheck`` causes the program to abort by
8982 calling the ``__stack_chk_fail()`` function.
8984 '``llvm.stackprotectorcheck``' Intrinsic
8985 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8992 declare void @llvm.stackprotectorcheck(i8** <guard>)
8997 The ``llvm.stackprotectorcheck`` intrinsic compares ``guard`` against an already
8998 created stack protector and if they are not equal calls the
8999 ``__stack_chk_fail()`` function.
9004 The ``llvm.stackprotectorcheck`` intrinsic requires one pointer argument, the
9005 the variable ``@__stack_chk_guard``.
9010 This intrinsic is provided to perform the stack protector check by comparing
9011 ``guard`` with the stack slot created by ``llvm.stackprotector`` and if the
9012 values do not match call the ``__stack_chk_fail()`` function.
9014 The reason to provide this as an IR level intrinsic instead of implementing it
9015 via other IR operations is that in order to perform this operation at the IR
9016 level without an intrinsic, one would need to create additional basic blocks to
9017 handle the success/failure cases. This makes it difficult to stop the stack
9018 protector check from disrupting sibling tail calls in Codegen. With this
9019 intrinsic, we are able to generate the stack protector basic blocks late in
9020 codegen after the tail call decision has occurred.
9022 '``llvm.objectsize``' Intrinsic
9023 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9030 declare i32 @llvm.objectsize.i32(i8* <object>, i1 <min>)
9031 declare i64 @llvm.objectsize.i64(i8* <object>, i1 <min>)
9036 The ``llvm.objectsize`` intrinsic is designed to provide information to
9037 the optimizers to determine at compile time whether a) an operation
9038 (like memcpy) will overflow a buffer that corresponds to an object, or
9039 b) that a runtime check for overflow isn't necessary. An object in this
9040 context means an allocation of a specific class, structure, array, or
9046 The ``llvm.objectsize`` intrinsic takes two arguments. The first
9047 argument is a pointer to or into the ``object``. The second argument is
9048 a boolean and determines whether ``llvm.objectsize`` returns 0 (if true)
9049 or -1 (if false) when the object size is unknown. The second argument
9050 only accepts constants.
9055 The ``llvm.objectsize`` intrinsic is lowered to a constant representing
9056 the size of the object concerned. If the size cannot be determined at
9057 compile time, ``llvm.objectsize`` returns ``i32/i64 -1 or 0`` (depending
9058 on the ``min`` argument).
9060 '``llvm.expect``' Intrinsic
9061 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9066 This is an overloaded intrinsic. You can use ``llvm.expect`` on any
9071 declare i1 @llvm.expect.i1(i1 <val>, i1 <expected_val>)
9072 declare i32 @llvm.expect.i32(i32 <val>, i32 <expected_val>)
9073 declare i64 @llvm.expect.i64(i64 <val>, i64 <expected_val>)
9078 The ``llvm.expect`` intrinsic provides information about expected (the
9079 most probable) value of ``val``, which can be used by optimizers.
9084 The ``llvm.expect`` intrinsic takes two arguments. The first argument is
9085 a value. The second argument is an expected value, this needs to be a
9086 constant value, variables are not allowed.
9091 This intrinsic is lowered to the ``val``.
9093 '``llvm.donothing``' Intrinsic
9094 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9101 declare void @llvm.donothing() nounwind readnone
9106 The ``llvm.donothing`` intrinsic doesn't perform any operation. It's the
9107 only intrinsic that can be called with an invoke instruction.
9117 This intrinsic does nothing, and it's removed by optimizers and ignored
9120 Stack Map Intrinsics
9121 --------------------
9123 LLVM provides experimental intrinsics to support runtime patching
9124 mechanisms commonly desired in dynamic language JITs. These intrinsics
9125 are described in :doc:`StackMaps`.