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.
443 A symbol with ``internal`` or ``private`` linkage must have ``default``
451 All Global Variables, Functions and Aliases can have one of the following
455 "``dllimport``" causes the compiler to reference a function or variable via
456 a global pointer to a pointer that is set up by the DLL exporting the
457 symbol. On Microsoft Windows targets, the pointer name is formed by
458 combining ``__imp_`` and the function or variable name.
460 "``dllexport``" causes the compiler to provide a global pointer to a pointer
461 in a DLL, so that it can be referenced with the ``dllimport`` attribute. On
462 Microsoft Windows targets, the pointer name is formed by combining
463 ``__imp_`` and the function or variable name. Since this storage class
464 exists for defining a dll interface, the compiler, assembler and linker know
465 it is externally referenced and must refrain from deleting the symbol.
472 LLVM IR allows you to specify both "identified" and "literal" :ref:`structure
473 types <t_struct>`. Literal types are uniqued structurally, but identified types
474 are never uniqued. An :ref:`opaque structural type <t_opaque>` can also be used
475 to forward declare a type which is not yet available.
477 An example of a identified structure specification is:
481 %mytype = type { %mytype*, i32 }
483 Prior to the LLVM 3.0 release, identified types were structurally uniqued. Only
484 literal types are uniqued in recent versions of LLVM.
491 Global variables define regions of memory allocated at compilation time
494 Global variables definitions must be initialized, may have an explicit section
495 to be placed in, and may have an optional explicit alignment specified.
497 Global variables in other translation units can also be declared, in which
498 case they don't have an initializer.
500 A variable may be defined as ``thread_local``, which means that it will
501 not be shared by threads (each thread will have a separated copy of the
502 variable). Not all targets support thread-local variables. Optionally, a
503 TLS model may be specified:
506 For variables that are only used within the current shared library.
508 For variables in modules that will not be loaded dynamically.
510 For variables defined in the executable and only used within it.
512 The models correspond to the ELF TLS models; see `ELF Handling For
513 Thread-Local Storage <http://people.redhat.com/drepper/tls.pdf>`_ for
514 more information on under which circumstances the different models may
515 be used. The target may choose a different TLS model if the specified
516 model is not supported, or if a better choice of model can be made.
518 A variable may be defined as a global ``constant``, which indicates that
519 the contents of the variable will **never** be modified (enabling better
520 optimization, allowing the global data to be placed in the read-only
521 section of an executable, etc). Note that variables that need runtime
522 initialization cannot be marked ``constant`` as there is a store to the
525 LLVM explicitly allows *declarations* of global variables to be marked
526 constant, even if the final definition of the global is not. This
527 capability can be used to enable slightly better optimization of the
528 program, but requires the language definition to guarantee that
529 optimizations based on the 'constantness' are valid for the translation
530 units that do not include the definition.
532 As SSA values, global variables define pointer values that are in scope
533 (i.e. they dominate) all basic blocks in the program. Global variables
534 always define a pointer to their "content" type because they describe a
535 region of memory, and all memory objects in LLVM are accessed through
538 Global variables can be marked with ``unnamed_addr`` which indicates
539 that the address is not significant, only the content. Constants marked
540 like this can be merged with other constants if they have the same
541 initializer. Note that a constant with significant address *can* be
542 merged with a ``unnamed_addr`` constant, the result being a constant
543 whose address is significant.
545 A global variable may be declared to reside in a target-specific
546 numbered address space. For targets that support them, address spaces
547 may affect how optimizations are performed and/or what target
548 instructions are used to access the variable. The default address space
549 is zero. The address space qualifier must precede any other attributes.
551 LLVM allows an explicit section to be specified for globals. If the
552 target supports it, it will emit globals to the section specified.
554 By default, global initializers are optimized by assuming that global
555 variables defined within the module are not modified from their
556 initial values before the start of the global initializer. This is
557 true even for variables potentially accessible from outside the
558 module, including those with external linkage or appearing in
559 ``@llvm.used`` or dllexported variables. This assumption may be suppressed
560 by marking the variable with ``externally_initialized``.
562 An explicit alignment may be specified for a global, which must be a
563 power of 2. If not present, or if the alignment is set to zero, the
564 alignment of the global is set by the target to whatever it feels
565 convenient. If an explicit alignment is specified, the global is forced
566 to have exactly that alignment. Targets and optimizers are not allowed
567 to over-align the global if the global has an assigned section. In this
568 case, the extra alignment could be observable: for example, code could
569 assume that the globals are densely packed in their section and try to
570 iterate over them as an array, alignment padding would break this
573 Globals can also have a :ref:`DLL storage class <dllstorageclass>`.
577 [@<GlobalVarName> =] [Linkage] [Visibility] [DLLStorageClass] [ThreadLocal]
578 [AddrSpace] [unnamed_addr] [ExternallyInitialized]
579 <global | constant> <Type>
580 [, section "name"] [, align <Alignment>]
582 For example, the following defines a global in a numbered address space
583 with an initializer, section, and alignment:
587 @G = addrspace(5) constant float 1.0, section "foo", align 4
589 The following example just declares a global variable
593 @G = external global i32
595 The following example defines a thread-local global with the
596 ``initialexec`` TLS model:
600 @G = thread_local(initialexec) global i32 0, align 4
602 .. _functionstructure:
607 LLVM function definitions consist of the "``define``" keyword, an
608 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
609 style <visibility>`, an optional :ref:`DLL storage class <dllstorageclass>`,
610 an optional :ref:`calling convention <callingconv>`,
611 an optional ``unnamed_addr`` attribute, a return type, an optional
612 :ref:`parameter attribute <paramattrs>` for the return type, a function
613 name, a (possibly empty) argument list (each with optional :ref:`parameter
614 attributes <paramattrs>`), optional :ref:`function attributes <fnattrs>`,
615 an optional section, an optional alignment, an optional :ref:`garbage
616 collector name <gc>`, an optional :ref:`prefix <prefixdata>`, an opening
617 curly brace, a list of basic blocks, and a closing curly brace.
619 LLVM function declarations consist of the "``declare``" keyword, an
620 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
621 style <visibility>`, an optional :ref:`DLL storage class <dllstorageclass>`,
622 an optional :ref:`calling convention <callingconv>`,
623 an optional ``unnamed_addr`` attribute, a return type, an optional
624 :ref:`parameter attribute <paramattrs>` for the return type, a function
625 name, a possibly empty list of arguments, an optional alignment, an optional
626 :ref:`garbage collector name <gc>` and an optional :ref:`prefix <prefixdata>`.
628 A function definition contains a list of basic blocks, forming the CFG (Control
629 Flow Graph) for the function. Each basic block may optionally start with a label
630 (giving the basic block a symbol table entry), contains a list of instructions,
631 and ends with a :ref:`terminator <terminators>` instruction (such as a branch or
632 function return). If an explicit label is not provided, a block is assigned an
633 implicit numbered label, using the next value from the same counter as used for
634 unnamed temporaries (:ref:`see above<identifiers>`). For example, if a function
635 entry block does not have an explicit label, it will be assigned label "%0",
636 then the first unnamed temporary in that block will be "%1", etc.
638 The first basic block in a function is special in two ways: it is
639 immediately executed on entrance to the function, and it is not allowed
640 to have predecessor basic blocks (i.e. there can not be any branches to
641 the entry block of a function). Because the block can have no
642 predecessors, it also cannot have any :ref:`PHI nodes <i_phi>`.
644 LLVM allows an explicit section to be specified for functions. If the
645 target supports it, it will emit functions to the section specified.
647 An explicit alignment may be specified for a function. If not present,
648 or if the alignment is set to zero, the alignment of the function is set
649 by the target to whatever it feels convenient. If an explicit alignment
650 is specified, the function is forced to have at least that much
651 alignment. All alignments must be a power of 2.
653 If the ``unnamed_addr`` attribute is given, the address is know to not
654 be significant and two identical functions can be merged.
658 define [linkage] [visibility] [DLLStorageClass]
660 <ResultType> @<FunctionName> ([argument list])
661 [unnamed_addr] [fn Attrs] [section "name"] [align N]
662 [gc] [prefix Constant] { ... }
669 Aliases act as "second name" for the aliasee value (which can be either
670 function, global variable, another alias or bitcast of global value).
671 Aliases may have an optional :ref:`linkage type <linkage>`, an optional
672 :ref:`visibility style <visibility>`, and an optional :ref:`DLL storage class
677 @<Name> = [Visibility] [DLLStorageClass] alias [Linkage] <AliaseeTy> @<Aliasee>
679 The linkage must be one of ``private``, ``internal``, ``linkonce``, ``weak``,
680 ``linkonce_odr``, ``weak_odr``, ``external``. Note that some system linkers
681 might not correctly handle dropping a weak symbol that is aliased by a non-weak
684 Alias that are not ``unnamed_addr`` are guaranteed to have the same address as
687 The aliasee must be a definition.
689 Aliases are not allowed to point to aliases with linkages that can be
690 overridden. Since they are only a second name, the possibility of the
691 intermediate alias being overridden cannot be represented in an object file.
693 .. _namedmetadatastructure:
698 Named metadata is a collection of metadata. :ref:`Metadata
699 nodes <metadata>` (but not metadata strings) are the only valid
700 operands for a named metadata.
704 ; Some unnamed metadata nodes, which are referenced by the named metadata.
705 !0 = metadata !{metadata !"zero"}
706 !1 = metadata !{metadata !"one"}
707 !2 = metadata !{metadata !"two"}
709 !name = !{!0, !1, !2}
716 The return type and each parameter of a function type may have a set of
717 *parameter attributes* associated with them. Parameter attributes are
718 used to communicate additional information about the result or
719 parameters of a function. Parameter attributes are considered to be part
720 of the function, not of the function type, so functions with different
721 parameter attributes can have the same function type.
723 Parameter attributes are simple keywords that follow the type specified.
724 If multiple parameter attributes are needed, they are space separated.
729 declare i32 @printf(i8* noalias nocapture, ...)
730 declare i32 @atoi(i8 zeroext)
731 declare signext i8 @returns_signed_char()
733 Note that any attributes for the function result (``nounwind``,
734 ``readonly``) come immediately after the argument list.
736 Currently, only the following parameter attributes are defined:
739 This indicates to the code generator that the parameter or return
740 value should be zero-extended to the extent required by the target's
741 ABI (which is usually 32-bits, but is 8-bits for a i1 on x86-64) by
742 the caller (for a parameter) or the callee (for a return value).
744 This indicates to the code generator that the parameter or return
745 value should be sign-extended to the extent required by the target's
746 ABI (which is usually 32-bits) by the caller (for a parameter) or
747 the callee (for a return value).
749 This indicates that this parameter or return value should be treated
750 in a special target-dependent fashion during while emitting code for
751 a function call or return (usually, by putting it in a register as
752 opposed to memory, though some targets use it to distinguish between
753 two different kinds of registers). Use of this attribute is
756 This indicates that the pointer parameter should really be passed by
757 value to the function. The attribute implies that a hidden copy of
758 the pointee is made between the caller and the callee, so the callee
759 is unable to modify the value in the caller. This attribute is only
760 valid on LLVM pointer arguments. It is generally used to pass
761 structs and arrays by value, but is also valid on pointers to
762 scalars. The copy is considered to belong to the caller not the
763 callee (for example, ``readonly`` functions should not write to
764 ``byval`` parameters). This is not a valid attribute for return
767 The byval attribute also supports specifying an alignment with the
768 align attribute. It indicates the alignment of the stack slot to
769 form and the known alignment of the pointer specified to the call
770 site. If the alignment is not specified, then the code generator
771 makes a target-specific assumption.
777 The ``inalloca`` argument attribute allows the caller to take the
778 address of outgoing stack arguments. An ``inalloca`` argument must
779 be a pointer to stack memory produced by an ``alloca`` instruction.
780 The alloca, or argument allocation, must also be tagged with the
781 inalloca keyword. Only the past argument may have the ``inalloca``
782 attribute, and that argument is guaranteed to be passed in memory.
784 An argument allocation may be used by a call at most once because
785 the call may deallocate it. The ``inalloca`` attribute cannot be
786 used in conjunction with other attributes that affect argument
787 storage, like ``inreg``, ``nest``, ``sret``, or ``byval``. The
788 ``inalloca`` attribute also disables LLVM's implicit lowering of
789 large aggregate return values, which means that frontend authors
790 must lower them with ``sret`` pointers.
792 When the call site is reached, the argument allocation must have
793 been the most recent stack allocation that is still live, or the
794 results are undefined. It is possible to allocate additional stack
795 space after an argument allocation and before its call site, but it
796 must be cleared off with :ref:`llvm.stackrestore
799 See :doc:`InAlloca` for more information on how to use this
803 This indicates that the pointer parameter specifies the address of a
804 structure that is the return value of the function in the source
805 program. This pointer must be guaranteed by the caller to be valid:
806 loads and stores to the structure may be assumed by the callee
807 not to trap and to be properly aligned. This may only be applied to
808 the first parameter. This is not a valid attribute for return
814 This indicates that pointer values :ref:`based <pointeraliasing>` on
815 the argument or return value do not alias pointer values which are
816 not *based* on it, ignoring certain "irrelevant" dependencies. For a
817 call to the parent function, dependencies between memory references
818 from before or after the call and from those during the call are
819 "irrelevant" to the ``noalias`` keyword for the arguments and return
820 value used in that call. The caller shares the responsibility with
821 the callee for ensuring that these requirements are met. For further
822 details, please see the discussion of the NoAlias response in :ref:`alias
823 analysis <Must, May, or No>`.
825 Note that this definition of ``noalias`` is intentionally similar
826 to the definition of ``restrict`` in C99 for function arguments,
827 though it is slightly weaker.
829 For function return values, C99's ``restrict`` is not meaningful,
830 while LLVM's ``noalias`` is.
832 This indicates that the callee does not make any copies of the
833 pointer that outlive the callee itself. This is not a valid
834 attribute for return values.
839 This indicates that the pointer parameter can be excised using the
840 :ref:`trampoline intrinsics <int_trampoline>`. This is not a valid
841 attribute for return values and can only be applied to one parameter.
844 This indicates that the function always returns the argument as its return
845 value. This is an optimization hint to the code generator when generating
846 the caller, allowing tail call optimization and omission of register saves
847 and restores in some cases; it is not checked or enforced when generating
848 the callee. The parameter and the function return type must be valid
849 operands for the :ref:`bitcast instruction <i_bitcast>`. This is not a
850 valid attribute for return values and can only be applied to one parameter.
853 This indicates that the parameter or return pointer is not null. This
854 attribute may only be applied to pointer typed parameters. This is not
855 checked or enforced by LLVM, the caller must ensure that the pointer
856 passed in is non-null, or the callee must ensure that the returned pointer
861 Garbage Collector Names
862 -----------------------
864 Each function may specify a garbage collector name, which is simply a
869 define void @f() gc "name" { ... }
871 The compiler declares the supported values of *name*. Specifying a
872 collector which will cause the compiler to alter its output in order to
873 support the named garbage collection algorithm.
880 Prefix data is data associated with a function which the code generator
881 will emit immediately before the function body. The purpose of this feature
882 is to allow frontends to associate language-specific runtime metadata with
883 specific functions and make it available through the function pointer while
884 still allowing the function pointer to be called. To access the data for a
885 given function, a program may bitcast the function pointer to a pointer to
886 the constant's type. This implies that the IR symbol points to the start
889 To maintain the semantics of ordinary function calls, the prefix data must
890 have a particular format. Specifically, it must begin with a sequence of
891 bytes which decode to a sequence of machine instructions, valid for the
892 module's target, which transfer control to the point immediately succeeding
893 the prefix data, without performing any other visible action. This allows
894 the inliner and other passes to reason about the semantics of the function
895 definition without needing to reason about the prefix data. Obviously this
896 makes the format of the prefix data highly target dependent.
898 Prefix data is laid out as if it were an initializer for a global variable
899 of the prefix data's type. No padding is automatically placed between the
900 prefix data and the function body. If padding is required, it must be part
903 A trivial example of valid prefix data for the x86 architecture is ``i8 144``,
904 which encodes the ``nop`` instruction:
908 define void @f() prefix i8 144 { ... }
910 Generally prefix data can be formed by encoding a relative branch instruction
911 which skips the metadata, as in this example of valid prefix data for the
912 x86_64 architecture, where the first two bytes encode ``jmp .+10``:
916 %0 = type <{ i8, i8, i8* }>
918 define void @f() prefix %0 <{ i8 235, i8 8, i8* @md}> { ... }
920 A function may have prefix data but no body. This has similar semantics
921 to the ``available_externally`` linkage in that the data may be used by the
922 optimizers but will not be emitted in the object file.
929 Attribute groups are groups of attributes that are referenced by objects within
930 the IR. They are important for keeping ``.ll`` files readable, because a lot of
931 functions will use the same set of attributes. In the degenerative case of a
932 ``.ll`` file that corresponds to a single ``.c`` file, the single attribute
933 group will capture the important command line flags used to build that file.
935 An attribute group is a module-level object. To use an attribute group, an
936 object references the attribute group's ID (e.g. ``#37``). An object may refer
937 to more than one attribute group. In that situation, the attributes from the
938 different groups are merged.
940 Here is an example of attribute groups for a function that should always be
941 inlined, has a stack alignment of 4, and which shouldn't use SSE instructions:
945 ; Target-independent attributes:
946 attributes #0 = { alwaysinline alignstack=4 }
948 ; Target-dependent attributes:
949 attributes #1 = { "no-sse" }
951 ; Function @f has attributes: alwaysinline, alignstack=4, and "no-sse".
952 define void @f() #0 #1 { ... }
959 Function attributes are set to communicate additional information about
960 a function. Function attributes are considered to be part of the
961 function, not of the function type, so functions with different function
962 attributes can have the same function type.
964 Function attributes are simple keywords that follow the type specified.
965 If multiple attributes are needed, they are space separated. For
970 define void @f() noinline { ... }
971 define void @f() alwaysinline { ... }
972 define void @f() alwaysinline optsize { ... }
973 define void @f() optsize { ... }
976 This attribute indicates that, when emitting the prologue and
977 epilogue, the backend should forcibly align the stack pointer.
978 Specify the desired alignment, which must be a power of two, in
981 This attribute indicates that the inliner should attempt to inline
982 this function into callers whenever possible, ignoring any active
983 inlining size threshold for this caller.
985 This indicates that the callee function at a call site should be
986 recognized as a built-in function, even though the function's declaration
987 uses the ``nobuiltin`` attribute. This is only valid at call sites for
988 direct calls to functions which are declared with the ``nobuiltin``
991 This attribute indicates that this function is rarely called. When
992 computing edge weights, basic blocks post-dominated by a cold
993 function call are also considered to be cold; and, thus, given low
996 This attribute indicates that the source code contained a hint that
997 inlining this function is desirable (such as the "inline" keyword in
998 C/C++). It is just a hint; it imposes no requirements on the
1001 This attribute suggests that optimization passes and code generator
1002 passes make choices that keep the code size of this function as small
1003 as possible and perform optimizations that may sacrifice runtime
1004 performance in order to minimize the size of the generated code.
1006 This attribute disables prologue / epilogue emission for the
1007 function. This can have very system-specific consequences.
1009 This indicates that the callee function at a call site is not recognized as
1010 a built-in function. LLVM will retain the original call and not replace it
1011 with equivalent code based on the semantics of the built-in function, unless
1012 the call site uses the ``builtin`` attribute. This is valid at call sites
1013 and on function declarations and definitions.
1015 This attribute indicates that calls to the function cannot be
1016 duplicated. A call to a ``noduplicate`` function may be moved
1017 within its parent function, but may not be duplicated within
1018 its parent function.
1020 A function containing a ``noduplicate`` call may still
1021 be an inlining candidate, provided that the call is not
1022 duplicated by inlining. That implies that the function has
1023 internal linkage and only has one call site, so the original
1024 call is dead after inlining.
1026 This attributes disables implicit floating point instructions.
1028 This attribute indicates that the inliner should never inline this
1029 function in any situation. This attribute may not be used together
1030 with the ``alwaysinline`` attribute.
1032 This attribute suppresses lazy symbol binding for the function. This
1033 may make calls to the function faster, at the cost of extra program
1034 startup time if the function is not called during program startup.
1036 This attribute indicates that the code generator should not use a
1037 red zone, even if the target-specific ABI normally permits it.
1039 This function attribute indicates that the function never returns
1040 normally. This produces undefined behavior at runtime if the
1041 function ever does dynamically return.
1043 This function attribute indicates that the function never returns
1044 with an unwind or exceptional control flow. If the function does
1045 unwind, its runtime behavior is undefined.
1047 This function attribute indicates that the function is not optimized
1048 by any optimization or code generator passes with the
1049 exception of interprocedural optimization passes.
1050 This attribute cannot be used together with the ``alwaysinline``
1051 attribute; this attribute is also incompatible
1052 with the ``minsize`` attribute and the ``optsize`` attribute.
1054 This attribute requires the ``noinline`` attribute to be specified on
1055 the function as well, so the function is never inlined into any caller.
1056 Only functions with the ``alwaysinline`` attribute are valid
1057 candidates for inlining into the body of this function.
1059 This attribute suggests that optimization passes and code generator
1060 passes make choices that keep the code size of this function low,
1061 and otherwise do optimizations specifically to reduce code size as
1062 long as they do not significantly impact runtime performance.
1064 On a function, this attribute indicates that the function computes its
1065 result (or decides to unwind an exception) based strictly on its arguments,
1066 without dereferencing any pointer arguments or otherwise accessing
1067 any mutable state (e.g. memory, control registers, etc) visible to
1068 caller functions. It does not write through any pointer arguments
1069 (including ``byval`` arguments) and never changes any state visible
1070 to callers. This means that it cannot unwind exceptions by calling
1071 the ``C++`` exception throwing methods.
1073 On an argument, this attribute indicates that the function does not
1074 dereference that pointer argument, even though it may read or write the
1075 memory that the pointer points to if accessed through other pointers.
1077 On a function, this attribute indicates that the function does not write
1078 through any pointer arguments (including ``byval`` arguments) or otherwise
1079 modify any state (e.g. memory, control registers, etc) visible to
1080 caller functions. It may dereference pointer arguments and read
1081 state that may be set in the caller. A readonly function always
1082 returns the same value (or unwinds an exception identically) when
1083 called with the same set of arguments and global state. It cannot
1084 unwind an exception by calling the ``C++`` exception throwing
1087 On an argument, this attribute indicates that the function does not write
1088 through this pointer argument, even though it may write to the memory that
1089 the pointer points to.
1091 This attribute indicates that this function can return twice. The C
1092 ``setjmp`` is an example of such a function. The compiler disables
1093 some optimizations (like tail calls) in the caller of these
1095 ``sanitize_address``
1096 This attribute indicates that AddressSanitizer checks
1097 (dynamic address safety analysis) are enabled for this function.
1099 This attribute indicates that MemorySanitizer checks (dynamic detection
1100 of accesses to uninitialized memory) are enabled for this function.
1102 This attribute indicates that ThreadSanitizer checks
1103 (dynamic thread safety analysis) are enabled for this function.
1105 This attribute indicates that the function should emit a stack
1106 smashing protector. It is in the form of a "canary" --- a random value
1107 placed on the stack before the local variables that's checked upon
1108 return from the function to see if it has been overwritten. A
1109 heuristic is used to determine if a function needs stack protectors
1110 or not. The heuristic used will enable protectors for functions with:
1112 - Character arrays larger than ``ssp-buffer-size`` (default 8).
1113 - Aggregates containing character arrays larger than ``ssp-buffer-size``.
1114 - Calls to alloca() with variable sizes or constant sizes greater than
1115 ``ssp-buffer-size``.
1117 Variables that are identified as requiring a protector will be arranged
1118 on the stack such that they are adjacent to the stack protector guard.
1120 If a function that has an ``ssp`` attribute is inlined into a
1121 function that doesn't have an ``ssp`` attribute, then the resulting
1122 function will have an ``ssp`` attribute.
1124 This attribute indicates that the function should *always* emit a
1125 stack smashing protector. This overrides the ``ssp`` function
1128 Variables that are identified as requiring a protector will be arranged
1129 on the stack such that they are adjacent to the stack protector guard.
1130 The specific layout rules are:
1132 #. Large arrays and structures containing large arrays
1133 (``>= ssp-buffer-size``) are closest to the stack protector.
1134 #. Small arrays and structures containing small arrays
1135 (``< ssp-buffer-size``) are 2nd closest to the protector.
1136 #. Variables that have had their address taken are 3rd closest to the
1139 If a function that has an ``sspreq`` attribute is inlined into a
1140 function that doesn't have an ``sspreq`` attribute or which has an
1141 ``ssp`` or ``sspstrong`` attribute, then the resulting function will have
1142 an ``sspreq`` attribute.
1144 This attribute indicates that the function should emit a stack smashing
1145 protector. This attribute causes a strong heuristic to be used when
1146 determining if a function needs stack protectors. The strong heuristic
1147 will enable protectors for functions with:
1149 - Arrays of any size and type
1150 - Aggregates containing an array of any size and type.
1151 - Calls to alloca().
1152 - Local variables that have had their address taken.
1154 Variables that are identified as requiring a protector will be arranged
1155 on the stack such that they are adjacent to the stack protector guard.
1156 The specific layout rules are:
1158 #. Large arrays and structures containing large arrays
1159 (``>= ssp-buffer-size``) are closest to the stack protector.
1160 #. Small arrays and structures containing small arrays
1161 (``< ssp-buffer-size``) are 2nd closest to the protector.
1162 #. Variables that have had their address taken are 3rd closest to the
1165 This overrides the ``ssp`` function attribute.
1167 If a function that has an ``sspstrong`` attribute is inlined into a
1168 function that doesn't have an ``sspstrong`` attribute, then the
1169 resulting function will have an ``sspstrong`` attribute.
1171 This attribute indicates that the ABI being targeted requires that
1172 an unwind table entry be produce for this function even if we can
1173 show that no exceptions passes by it. This is normally the case for
1174 the ELF x86-64 abi, but it can be disabled for some compilation
1179 Module-Level Inline Assembly
1180 ----------------------------
1182 Modules may contain "module-level inline asm" blocks, which corresponds
1183 to the GCC "file scope inline asm" blocks. These blocks are internally
1184 concatenated by LLVM and treated as a single unit, but may be separated
1185 in the ``.ll`` file if desired. The syntax is very simple:
1187 .. code-block:: llvm
1189 module asm "inline asm code goes here"
1190 module asm "more can go here"
1192 The strings can contain any character by escaping non-printable
1193 characters. The escape sequence used is simply "\\xx" where "xx" is the
1194 two digit hex code for the number.
1196 The inline asm code is simply printed to the machine code .s file when
1197 assembly code is generated.
1199 .. _langref_datalayout:
1204 A module may specify a target specific data layout string that specifies
1205 how data is to be laid out in memory. The syntax for the data layout is
1208 .. code-block:: llvm
1210 target datalayout = "layout specification"
1212 The *layout specification* consists of a list of specifications
1213 separated by the minus sign character ('-'). Each specification starts
1214 with a letter and may include other information after the letter to
1215 define some aspect of the data layout. The specifications accepted are
1219 Specifies that the target lays out data in big-endian form. That is,
1220 the bits with the most significance have the lowest address
1223 Specifies that the target lays out data in little-endian form. That
1224 is, the bits with the least significance have the lowest address
1227 Specifies the natural alignment of the stack in bits. Alignment
1228 promotion of stack variables is limited to the natural stack
1229 alignment to avoid dynamic stack realignment. The stack alignment
1230 must be a multiple of 8-bits. If omitted, the natural stack
1231 alignment defaults to "unspecified", which does not prevent any
1232 alignment promotions.
1233 ``p[n]:<size>:<abi>:<pref>``
1234 This specifies the *size* of a pointer and its ``<abi>`` and
1235 ``<pref>``\erred alignments for address space ``n``. All sizes are in
1236 bits. The address space, ``n`` is optional, and if not specified,
1237 denotes the default address space 0. The value of ``n`` must be
1238 in the range [1,2^23).
1239 ``i<size>:<abi>:<pref>``
1240 This specifies the alignment for an integer type of a given bit
1241 ``<size>``. The value of ``<size>`` must be in the range [1,2^23).
1242 ``v<size>:<abi>:<pref>``
1243 This specifies the alignment for a vector type of a given bit
1245 ``f<size>:<abi>:<pref>``
1246 This specifies the alignment for a floating point type of a given bit
1247 ``<size>``. Only values of ``<size>`` that are supported by the target
1248 will work. 32 (float) and 64 (double) are supported on all targets; 80
1249 or 128 (different flavors of long double) are also supported on some
1252 This specifies the alignment for an object of aggregate type.
1254 If present, specifies that llvm names are mangled in the output. The
1257 * ``e``: ELF mangling: Private symbols get a ``.L`` prefix.
1258 * ``m``: Mips mangling: Private symbols get a ``$`` prefix.
1259 * ``o``: Mach-O mangling: Private symbols get ``L`` prefix. Other
1260 symbols get a ``_`` prefix.
1261 * ``w``: Windows COFF prefix: Similar to Mach-O, but stdcall and fastcall
1262 functions also get a suffix based on the frame size.
1263 ``n<size1>:<size2>:<size3>...``
1264 This specifies a set of native integer widths for the target CPU in
1265 bits. For example, it might contain ``n32`` for 32-bit PowerPC,
1266 ``n32:64`` for PowerPC 64, or ``n8:16:32:64`` for X86-64. Elements of
1267 this set are considered to support most general arithmetic operations
1270 On every specification that takes a ``<abi>:<pref>``, specifying the
1271 ``<pref>`` alignment is optional. If omitted, the preceding ``:``
1272 should be omitted too and ``<pref>`` will be equal to ``<abi>``.
1274 When constructing the data layout for a given target, LLVM starts with a
1275 default set of specifications which are then (possibly) overridden by
1276 the specifications in the ``datalayout`` keyword. The default
1277 specifications are given in this list:
1279 - ``E`` - big endian
1280 - ``p:64:64:64`` - 64-bit pointers with 64-bit alignment.
1281 - ``p[n]:64:64:64`` - Other address spaces are assumed to be the
1282 same as the default address space.
1283 - ``S0`` - natural stack alignment is unspecified
1284 - ``i1:8:8`` - i1 is 8-bit (byte) aligned
1285 - ``i8:8:8`` - i8 is 8-bit (byte) aligned
1286 - ``i16:16:16`` - i16 is 16-bit aligned
1287 - ``i32:32:32`` - i32 is 32-bit aligned
1288 - ``i64:32:64`` - i64 has ABI alignment of 32-bits but preferred
1289 alignment of 64-bits
1290 - ``f16:16:16`` - half is 16-bit aligned
1291 - ``f32:32:32`` - float is 32-bit aligned
1292 - ``f64:64:64`` - double is 64-bit aligned
1293 - ``f128:128:128`` - quad is 128-bit aligned
1294 - ``v64:64:64`` - 64-bit vector is 64-bit aligned
1295 - ``v128:128:128`` - 128-bit vector is 128-bit aligned
1296 - ``a:0:64`` - aggregates are 64-bit aligned
1298 When LLVM is determining the alignment for a given type, it uses the
1301 #. If the type sought is an exact match for one of the specifications,
1302 that specification is used.
1303 #. If no match is found, and the type sought is an integer type, then
1304 the smallest integer type that is larger than the bitwidth of the
1305 sought type is used. If none of the specifications are larger than
1306 the bitwidth then the largest integer type is used. For example,
1307 given the default specifications above, the i7 type will use the
1308 alignment of i8 (next largest) while both i65 and i256 will use the
1309 alignment of i64 (largest specified).
1310 #. If no match is found, and the type sought is a vector type, then the
1311 largest vector type that is smaller than the sought vector type will
1312 be used as a fall back. This happens because <128 x double> can be
1313 implemented in terms of 64 <2 x double>, for example.
1315 The function of the data layout string may not be what you expect.
1316 Notably, this is not a specification from the frontend of what alignment
1317 the code generator should use.
1319 Instead, if specified, the target data layout is required to match what
1320 the ultimate *code generator* expects. This string is used by the
1321 mid-level optimizers to improve code, and this only works if it matches
1322 what the ultimate code generator uses. If you would like to generate IR
1323 that does not embed this target-specific detail into the IR, then you
1324 don't have to specify the string. This will disable some optimizations
1325 that require precise layout information, but this also prevents those
1326 optimizations from introducing target specificity into the IR.
1333 A module may specify a target triple string that describes the target
1334 host. The syntax for the target triple is simply:
1336 .. code-block:: llvm
1338 target triple = "x86_64-apple-macosx10.7.0"
1340 The *target triple* string consists of a series of identifiers delimited
1341 by the minus sign character ('-'). The canonical forms are:
1345 ARCHITECTURE-VENDOR-OPERATING_SYSTEM
1346 ARCHITECTURE-VENDOR-OPERATING_SYSTEM-ENVIRONMENT
1348 This information is passed along to the backend so that it generates
1349 code for the proper architecture. It's possible to override this on the
1350 command line with the ``-mtriple`` command line option.
1352 .. _pointeraliasing:
1354 Pointer Aliasing Rules
1355 ----------------------
1357 Any memory access must be done through a pointer value associated with
1358 an address range of the memory access, otherwise the behavior is
1359 undefined. Pointer values are associated with address ranges according
1360 to the following rules:
1362 - A pointer value is associated with the addresses associated with any
1363 value it is *based* on.
1364 - An address of a global variable is associated with the address range
1365 of the variable's storage.
1366 - The result value of an allocation instruction is associated with the
1367 address range of the allocated storage.
1368 - A null pointer in the default address-space is associated with no
1370 - An integer constant other than zero or a pointer value returned from
1371 a function not defined within LLVM may be associated with address
1372 ranges allocated through mechanisms other than those provided by
1373 LLVM. Such ranges shall not overlap with any ranges of addresses
1374 allocated by mechanisms provided by LLVM.
1376 A pointer value is *based* on another pointer value according to the
1379 - A pointer value formed from a ``getelementptr`` operation is *based*
1380 on the first operand of the ``getelementptr``.
1381 - The result value of a ``bitcast`` is *based* on the operand of the
1383 - A pointer value formed by an ``inttoptr`` is *based* on all pointer
1384 values that contribute (directly or indirectly) to the computation of
1385 the pointer's value.
1386 - The "*based* on" relationship is transitive.
1388 Note that this definition of *"based"* is intentionally similar to the
1389 definition of *"based"* in C99, though it is slightly weaker.
1391 LLVM IR does not associate types with memory. The result type of a
1392 ``load`` merely indicates the size and alignment of the memory from
1393 which to load, as well as the interpretation of the value. The first
1394 operand type of a ``store`` similarly only indicates the size and
1395 alignment of the store.
1397 Consequently, type-based alias analysis, aka TBAA, aka
1398 ``-fstrict-aliasing``, is not applicable to general unadorned LLVM IR.
1399 :ref:`Metadata <metadata>` may be used to encode additional information
1400 which specialized optimization passes may use to implement type-based
1405 Volatile Memory Accesses
1406 ------------------------
1408 Certain memory accesses, such as :ref:`load <i_load>`'s,
1409 :ref:`store <i_store>`'s, and :ref:`llvm.memcpy <int_memcpy>`'s may be
1410 marked ``volatile``. The optimizers must not change the number of
1411 volatile operations or change their order of execution relative to other
1412 volatile operations. The optimizers *may* change the order of volatile
1413 operations relative to non-volatile operations. This is not Java's
1414 "volatile" and has no cross-thread synchronization behavior.
1416 IR-level volatile loads and stores cannot safely be optimized into
1417 llvm.memcpy or llvm.memmove intrinsics even when those intrinsics are
1418 flagged volatile. Likewise, the backend should never split or merge
1419 target-legal volatile load/store instructions.
1421 .. admonition:: Rationale
1423 Platforms may rely on volatile loads and stores of natively supported
1424 data width to be executed as single instruction. For example, in C
1425 this holds for an l-value of volatile primitive type with native
1426 hardware support, but not necessarily for aggregate types. The
1427 frontend upholds these expectations, which are intentionally
1428 unspecified in the IR. The rules above ensure that IR transformation
1429 do not violate the frontend's contract with the language.
1433 Memory Model for Concurrent Operations
1434 --------------------------------------
1436 The LLVM IR does not define any way to start parallel threads of
1437 execution or to register signal handlers. Nonetheless, there are
1438 platform-specific ways to create them, and we define LLVM IR's behavior
1439 in their presence. This model is inspired by the C++0x memory model.
1441 For a more informal introduction to this model, see the :doc:`Atomics`.
1443 We define a *happens-before* partial order as the least partial order
1446 - Is a superset of single-thread program order, and
1447 - When a *synchronizes-with* ``b``, includes an edge from ``a`` to
1448 ``b``. *Synchronizes-with* pairs are introduced by platform-specific
1449 techniques, like pthread locks, thread creation, thread joining,
1450 etc., and by atomic instructions. (See also :ref:`Atomic Memory Ordering
1451 Constraints <ordering>`).
1453 Note that program order does not introduce *happens-before* edges
1454 between a thread and signals executing inside that thread.
1456 Every (defined) read operation (load instructions, memcpy, atomic
1457 loads/read-modify-writes, etc.) R reads a series of bytes written by
1458 (defined) write operations (store instructions, atomic
1459 stores/read-modify-writes, memcpy, etc.). For the purposes of this
1460 section, initialized globals are considered to have a write of the
1461 initializer which is atomic and happens before any other read or write
1462 of the memory in question. For each byte of a read R, R\ :sub:`byte`
1463 may see any write to the same byte, except:
1465 - If write\ :sub:`1` happens before write\ :sub:`2`, and
1466 write\ :sub:`2` happens before R\ :sub:`byte`, then
1467 R\ :sub:`byte` does not see write\ :sub:`1`.
1468 - If R\ :sub:`byte` happens before write\ :sub:`3`, then
1469 R\ :sub:`byte` does not see write\ :sub:`3`.
1471 Given that definition, R\ :sub:`byte` is defined as follows:
1473 - If R is volatile, the result is target-dependent. (Volatile is
1474 supposed to give guarantees which can support ``sig_atomic_t`` in
1475 C/C++, and may be used for accesses to addresses which do not behave
1476 like normal memory. It does not generally provide cross-thread
1478 - Otherwise, if there is no write to the same byte that happens before
1479 R\ :sub:`byte`, R\ :sub:`byte` returns ``undef`` for that byte.
1480 - Otherwise, if R\ :sub:`byte` may see exactly one write,
1481 R\ :sub:`byte` returns the value written by that write.
1482 - Otherwise, if R is atomic, and all the writes R\ :sub:`byte` may
1483 see are atomic, it chooses one of the values written. See the :ref:`Atomic
1484 Memory Ordering Constraints <ordering>` section for additional
1485 constraints on how the choice is made.
1486 - Otherwise R\ :sub:`byte` returns ``undef``.
1488 R returns the value composed of the series of bytes it read. This
1489 implies that some bytes within the value may be ``undef`` **without**
1490 the entire value being ``undef``. Note that this only defines the
1491 semantics of the operation; it doesn't mean that targets will emit more
1492 than one instruction to read the series of bytes.
1494 Note that in cases where none of the atomic intrinsics are used, this
1495 model places only one restriction on IR transformations on top of what
1496 is required for single-threaded execution: introducing a store to a byte
1497 which might not otherwise be stored is not allowed in general.
1498 (Specifically, in the case where another thread might write to and read
1499 from an address, introducing a store can change a load that may see
1500 exactly one write into a load that may see multiple writes.)
1504 Atomic Memory Ordering Constraints
1505 ----------------------------------
1507 Atomic instructions (:ref:`cmpxchg <i_cmpxchg>`,
1508 :ref:`atomicrmw <i_atomicrmw>`, :ref:`fence <i_fence>`,
1509 :ref:`atomic load <i_load>`, and :ref:`atomic store <i_store>`) take
1510 ordering parameters that determine which other atomic instructions on
1511 the same address they *synchronize with*. These semantics are borrowed
1512 from Java and C++0x, but are somewhat more colloquial. If these
1513 descriptions aren't precise enough, check those specs (see spec
1514 references in the :doc:`atomics guide <Atomics>`).
1515 :ref:`fence <i_fence>` instructions treat these orderings somewhat
1516 differently since they don't take an address. See that instruction's
1517 documentation for details.
1519 For a simpler introduction to the ordering constraints, see the
1523 The set of values that can be read is governed by the happens-before
1524 partial order. A value cannot be read unless some operation wrote
1525 it. This is intended to provide a guarantee strong enough to model
1526 Java's non-volatile shared variables. This ordering cannot be
1527 specified for read-modify-write operations; it is not strong enough
1528 to make them atomic in any interesting way.
1530 In addition to the guarantees of ``unordered``, there is a single
1531 total order for modifications by ``monotonic`` operations on each
1532 address. All modification orders must be compatible with the
1533 happens-before order. There is no guarantee that the modification
1534 orders can be combined to a global total order for the whole program
1535 (and this often will not be possible). The read in an atomic
1536 read-modify-write operation (:ref:`cmpxchg <i_cmpxchg>` and
1537 :ref:`atomicrmw <i_atomicrmw>`) reads the value in the modification
1538 order immediately before the value it writes. If one atomic read
1539 happens before another atomic read of the same address, the later
1540 read must see the same value or a later value in the address's
1541 modification order. This disallows reordering of ``monotonic`` (or
1542 stronger) operations on the same address. If an address is written
1543 ``monotonic``-ally by one thread, and other threads ``monotonic``-ally
1544 read that address repeatedly, the other threads must eventually see
1545 the write. This corresponds to the C++0x/C1x
1546 ``memory_order_relaxed``.
1548 In addition to the guarantees of ``monotonic``, a
1549 *synchronizes-with* edge may be formed with a ``release`` operation.
1550 This is intended to model C++'s ``memory_order_acquire``.
1552 In addition to the guarantees of ``monotonic``, if this operation
1553 writes a value which is subsequently read by an ``acquire``
1554 operation, it *synchronizes-with* that operation. (This isn't a
1555 complete description; see the C++0x definition of a release
1556 sequence.) This corresponds to the C++0x/C1x
1557 ``memory_order_release``.
1558 ``acq_rel`` (acquire+release)
1559 Acts as both an ``acquire`` and ``release`` operation on its
1560 address. This corresponds to the C++0x/C1x ``memory_order_acq_rel``.
1561 ``seq_cst`` (sequentially consistent)
1562 In addition to the guarantees of ``acq_rel`` (``acquire`` for an
1563 operation which only reads, ``release`` for an operation which only
1564 writes), there is a global total order on all
1565 sequentially-consistent operations on all addresses, which is
1566 consistent with the *happens-before* partial order and with the
1567 modification orders of all the affected addresses. Each
1568 sequentially-consistent read sees the last preceding write to the
1569 same address in this global order. This corresponds to the C++0x/C1x
1570 ``memory_order_seq_cst`` and Java volatile.
1574 If an atomic operation is marked ``singlethread``, it only *synchronizes
1575 with* or participates in modification and seq\_cst total orderings with
1576 other operations running in the same thread (for example, in signal
1584 LLVM IR floating-point binary ops (:ref:`fadd <i_fadd>`,
1585 :ref:`fsub <i_fsub>`, :ref:`fmul <i_fmul>`, :ref:`fdiv <i_fdiv>`,
1586 :ref:`frem <i_frem>`) have the following flags that can set to enable
1587 otherwise unsafe floating point operations
1590 No NaNs - Allow optimizations to assume the arguments and result are not
1591 NaN. Such optimizations are required to retain defined behavior over
1592 NaNs, but the value of the result is undefined.
1595 No Infs - Allow optimizations to assume the arguments and result are not
1596 +/-Inf. Such optimizations are required to retain defined behavior over
1597 +/-Inf, but the value of the result is undefined.
1600 No Signed Zeros - Allow optimizations to treat the sign of a zero
1601 argument or result as insignificant.
1604 Allow Reciprocal - Allow optimizations to use the reciprocal of an
1605 argument rather than perform division.
1608 Fast - Allow algebraically equivalent transformations that may
1609 dramatically change results in floating point (e.g. reassociate). This
1610 flag implies all the others.
1617 The LLVM type system is one of the most important features of the
1618 intermediate representation. Being typed enables a number of
1619 optimizations to be performed on the intermediate representation
1620 directly, without having to do extra analyses on the side before the
1621 transformation. A strong type system makes it easier to read the
1622 generated code and enables novel analyses and transformations that are
1623 not feasible to perform on normal three address code representations.
1633 The void type does not represent any value and has no size.
1651 The function type can be thought of as a function signature. It consists of a
1652 return type and a list of formal parameter types. The return type of a function
1653 type is a void type or first class type --- except for :ref:`label <t_label>`
1654 and :ref:`metadata <t_metadata>` types.
1660 <returntype> (<parameter list>)
1662 ...where '``<parameter list>``' is a comma-separated list of type
1663 specifiers. Optionally, the parameter list may include a type ``...``, which
1664 indicates that the function takes a variable number of arguments. Variable
1665 argument functions can access their arguments with the :ref:`variable argument
1666 handling intrinsic <int_varargs>` functions. '``<returntype>``' is any type
1667 except :ref:`label <t_label>` and :ref:`metadata <t_metadata>`.
1671 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1672 | ``i32 (i32)`` | function taking an ``i32``, returning an ``i32`` |
1673 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1674 | ``float (i16, i32 *) *`` | :ref:`Pointer <t_pointer>` to a function that takes an ``i16`` and a :ref:`pointer <t_pointer>` to ``i32``, returning ``float``. |
1675 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1676 | ``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. |
1677 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1678 | ``{i32, i32} (i32)`` | A function taking an ``i32``, returning a :ref:`structure <t_struct>` containing two ``i32`` values |
1679 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1686 The :ref:`first class <t_firstclass>` types are perhaps the most important.
1687 Values of these types are the only ones which can be produced by
1695 These are the types that are valid in registers from CodeGen's perspective.
1704 The integer type is a very simple type that simply specifies an
1705 arbitrary bit width for the integer type desired. Any bit width from 1
1706 bit to 2\ :sup:`23`\ -1 (about 8 million) can be specified.
1714 The number of bits the integer will occupy is specified by the ``N``
1720 +----------------+------------------------------------------------+
1721 | ``i1`` | a single-bit integer. |
1722 +----------------+------------------------------------------------+
1723 | ``i32`` | a 32-bit integer. |
1724 +----------------+------------------------------------------------+
1725 | ``i1942652`` | a really big integer of over 1 million bits. |
1726 +----------------+------------------------------------------------+
1730 Floating Point Types
1731 """"""""""""""""""""
1740 - 16-bit floating point value
1743 - 32-bit floating point value
1746 - 64-bit floating point value
1749 - 128-bit floating point value (112-bit mantissa)
1752 - 80-bit floating point value (X87)
1755 - 128-bit floating point value (two 64-bits)
1762 The x86_mmx type represents a value held in an MMX register on an x86
1763 machine. The operations allowed on it are quite limited: parameters and
1764 return values, load and store, and bitcast. User-specified MMX
1765 instructions are represented as intrinsic or asm calls with arguments
1766 and/or results of this type. There are no arrays, vectors or constants
1783 The pointer type is used to specify memory locations. Pointers are
1784 commonly used to reference objects in memory.
1786 Pointer types may have an optional address space attribute defining the
1787 numbered address space where the pointed-to object resides. The default
1788 address space is number zero. The semantics of non-zero address spaces
1789 are target-specific.
1791 Note that LLVM does not permit pointers to void (``void*``) nor does it
1792 permit pointers to labels (``label*``). Use ``i8*`` instead.
1802 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1803 | ``[4 x i32]*`` | A :ref:`pointer <t_pointer>` to :ref:`array <t_array>` of four ``i32`` values. |
1804 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1805 | ``i32 (i32*) *`` | A :ref:`pointer <t_pointer>` to a :ref:`function <t_function>` that takes an ``i32*``, returning an ``i32``. |
1806 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1807 | ``i32 addrspace(5)*`` | A :ref:`pointer <t_pointer>` to an ``i32`` value that resides in address space #5. |
1808 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1817 A vector type is a simple derived type that represents a vector of
1818 elements. Vector types are used when multiple primitive data are
1819 operated in parallel using a single instruction (SIMD). A vector type
1820 requires a size (number of elements) and an underlying primitive data
1821 type. Vector types are considered :ref:`first class <t_firstclass>`.
1827 < <# elements> x <elementtype> >
1829 The number of elements is a constant integer value larger than 0;
1830 elementtype may be any integer or floating point type, or a pointer to
1831 these types. Vectors of size zero are not allowed.
1835 +-------------------+--------------------------------------------------+
1836 | ``<4 x i32>`` | Vector of 4 32-bit integer values. |
1837 +-------------------+--------------------------------------------------+
1838 | ``<8 x float>`` | Vector of 8 32-bit floating-point values. |
1839 +-------------------+--------------------------------------------------+
1840 | ``<2 x i64>`` | Vector of 2 64-bit integer values. |
1841 +-------------------+--------------------------------------------------+
1842 | ``<4 x i64*>`` | Vector of 4 pointers to 64-bit integer values. |
1843 +-------------------+--------------------------------------------------+
1852 The label type represents code labels.
1867 The metadata type represents embedded metadata. No derived types may be
1868 created from metadata except for :ref:`function <t_function>` arguments.
1881 Aggregate Types are a subset of derived types that can contain multiple
1882 member types. :ref:`Arrays <t_array>` and :ref:`structs <t_struct>` are
1883 aggregate types. :ref:`Vectors <t_vector>` are not considered to be
1893 The array type is a very simple derived type that arranges elements
1894 sequentially in memory. The array type requires a size (number of
1895 elements) and an underlying data type.
1901 [<# elements> x <elementtype>]
1903 The number of elements is a constant integer value; ``elementtype`` may
1904 be any type with a size.
1908 +------------------+--------------------------------------+
1909 | ``[40 x i32]`` | Array of 40 32-bit integer values. |
1910 +------------------+--------------------------------------+
1911 | ``[41 x i32]`` | Array of 41 32-bit integer values. |
1912 +------------------+--------------------------------------+
1913 | ``[4 x i8]`` | Array of 4 8-bit integer values. |
1914 +------------------+--------------------------------------+
1916 Here are some examples of multidimensional arrays:
1918 +-----------------------------+----------------------------------------------------------+
1919 | ``[3 x [4 x i32]]`` | 3x4 array of 32-bit integer values. |
1920 +-----------------------------+----------------------------------------------------------+
1921 | ``[12 x [10 x float]]`` | 12x10 array of single precision floating point values. |
1922 +-----------------------------+----------------------------------------------------------+
1923 | ``[2 x [3 x [4 x i16]]]`` | 2x3x4 array of 16-bit integer values. |
1924 +-----------------------------+----------------------------------------------------------+
1926 There is no restriction on indexing beyond the end of the array implied
1927 by a static type (though there are restrictions on indexing beyond the
1928 bounds of an allocated object in some cases). This means that
1929 single-dimension 'variable sized array' addressing can be implemented in
1930 LLVM with a zero length array type. An implementation of 'pascal style
1931 arrays' in LLVM could use the type "``{ i32, [0 x float]}``", for
1941 The structure type is used to represent a collection of data members
1942 together in memory. The elements of a structure may be any type that has
1945 Structures in memory are accessed using '``load``' and '``store``' by
1946 getting a pointer to a field with the '``getelementptr``' instruction.
1947 Structures in registers are accessed using the '``extractvalue``' and
1948 '``insertvalue``' instructions.
1950 Structures may optionally be "packed" structures, which indicate that
1951 the alignment of the struct is one byte, and that there is no padding
1952 between the elements. In non-packed structs, padding between field types
1953 is inserted as defined by the DataLayout string in the module, which is
1954 required to match what the underlying code generator expects.
1956 Structures can either be "literal" or "identified". A literal structure
1957 is defined inline with other types (e.g. ``{i32, i32}*``) whereas
1958 identified types are always defined at the top level with a name.
1959 Literal types are uniqued by their contents and can never be recursive
1960 or opaque since there is no way to write one. Identified types can be
1961 recursive, can be opaqued, and are never uniqued.
1967 %T1 = type { <type list> } ; Identified normal struct type
1968 %T2 = type <{ <type list> }> ; Identified packed struct type
1972 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1973 | ``{ i32, i32, i32 }`` | A triple of three ``i32`` values |
1974 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1975 | ``{ 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``. |
1976 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1977 | ``<{ i8, i32 }>`` | A packed struct known to be 5 bytes in size. |
1978 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1982 Opaque Structure Types
1983 """"""""""""""""""""""
1987 Opaque structure types are used to represent named structure types that
1988 do not have a body specified. This corresponds (for example) to the C
1989 notion of a forward declared structure.
2000 +--------------+-------------------+
2001 | ``opaque`` | An opaque type. |
2002 +--------------+-------------------+
2009 LLVM has several different basic types of constants. This section
2010 describes them all and their syntax.
2015 **Boolean constants**
2016 The two strings '``true``' and '``false``' are both valid constants
2018 **Integer constants**
2019 Standard integers (such as '4') are constants of the
2020 :ref:`integer <t_integer>` type. Negative numbers may be used with
2022 **Floating point constants**
2023 Floating point constants use standard decimal notation (e.g.
2024 123.421), exponential notation (e.g. 1.23421e+2), or a more precise
2025 hexadecimal notation (see below). The assembler requires the exact
2026 decimal value of a floating-point constant. For example, the
2027 assembler accepts 1.25 but rejects 1.3 because 1.3 is a repeating
2028 decimal in binary. Floating point constants must have a :ref:`floating
2029 point <t_floating>` type.
2030 **Null pointer constants**
2031 The identifier '``null``' is recognized as a null pointer constant
2032 and must be of :ref:`pointer type <t_pointer>`.
2034 The one non-intuitive notation for constants is the hexadecimal form of
2035 floating point constants. For example, the form
2036 '``double 0x432ff973cafa8000``' is equivalent to (but harder to read
2037 than) '``double 4.5e+15``'. The only time hexadecimal floating point
2038 constants are required (and the only time that they are generated by the
2039 disassembler) is when a floating point constant must be emitted but it
2040 cannot be represented as a decimal floating point number in a reasonable
2041 number of digits. For example, NaN's, infinities, and other special
2042 values are represented in their IEEE hexadecimal format so that assembly
2043 and disassembly do not cause any bits to change in the constants.
2045 When using the hexadecimal form, constants of types half, float, and
2046 double are represented using the 16-digit form shown above (which
2047 matches the IEEE754 representation for double); half and float values
2048 must, however, be exactly representable as IEEE 754 half and single
2049 precision, respectively. Hexadecimal format is always used for long
2050 double, and there are three forms of long double. The 80-bit format used
2051 by x86 is represented as ``0xK`` followed by 20 hexadecimal digits. The
2052 128-bit format used by PowerPC (two adjacent doubles) is represented by
2053 ``0xM`` followed by 32 hexadecimal digits. The IEEE 128-bit format is
2054 represented by ``0xL`` followed by 32 hexadecimal digits. Long doubles
2055 will only work if they match the long double format on your target.
2056 The IEEE 16-bit format (half precision) is represented by ``0xH``
2057 followed by 4 hexadecimal digits. All hexadecimal formats are big-endian
2058 (sign bit at the left).
2060 There are no constants of type x86_mmx.
2062 .. _complexconstants:
2067 Complex constants are a (potentially recursive) combination of simple
2068 constants and smaller complex constants.
2070 **Structure constants**
2071 Structure constants are represented with notation similar to
2072 structure type definitions (a comma separated list of elements,
2073 surrounded by braces (``{}``)). For example:
2074 "``{ i32 4, float 17.0, i32* @G }``", where "``@G``" is declared as
2075 "``@G = external global i32``". Structure constants must have
2076 :ref:`structure type <t_struct>`, and the number and types of elements
2077 must match those specified by the type.
2079 Array constants are represented with notation similar to array type
2080 definitions (a comma separated list of elements, surrounded by
2081 square brackets (``[]``)). For example:
2082 "``[ i32 42, i32 11, i32 74 ]``". Array constants must have
2083 :ref:`array type <t_array>`, and the number and types of elements must
2084 match those specified by the type.
2085 **Vector constants**
2086 Vector constants are represented with notation similar to vector
2087 type definitions (a comma separated list of elements, surrounded by
2088 less-than/greater-than's (``<>``)). For example:
2089 "``< i32 42, i32 11, i32 74, i32 100 >``". Vector constants
2090 must have :ref:`vector type <t_vector>`, and the number and types of
2091 elements must match those specified by the type.
2092 **Zero initialization**
2093 The string '``zeroinitializer``' can be used to zero initialize a
2094 value to zero of *any* type, including scalar and
2095 :ref:`aggregate <t_aggregate>` types. This is often used to avoid
2096 having to print large zero initializers (e.g. for large arrays) and
2097 is always exactly equivalent to using explicit zero initializers.
2099 A metadata node is a structure-like constant with :ref:`metadata
2100 type <t_metadata>`. For example:
2101 "``metadata !{ i32 0, metadata !"test" }``". Unlike other
2102 constants that are meant to be interpreted as part of the
2103 instruction stream, metadata is a place to attach additional
2104 information such as debug info.
2106 Global Variable and Function Addresses
2107 --------------------------------------
2109 The addresses of :ref:`global variables <globalvars>` and
2110 :ref:`functions <functionstructure>` are always implicitly valid
2111 (link-time) constants. These constants are explicitly referenced when
2112 the :ref:`identifier for the global <identifiers>` is used and always have
2113 :ref:`pointer <t_pointer>` type. For example, the following is a legal LLVM
2116 .. code-block:: llvm
2120 @Z = global [2 x i32*] [ i32* @X, i32* @Y ]
2127 The string '``undef``' can be used anywhere a constant is expected, and
2128 indicates that the user of the value may receive an unspecified
2129 bit-pattern. Undefined values may be of any type (other than '``label``'
2130 or '``void``') and be used anywhere a constant is permitted.
2132 Undefined values are useful because they indicate to the compiler that
2133 the program is well defined no matter what value is used. This gives the
2134 compiler more freedom to optimize. Here are some examples of
2135 (potentially surprising) transformations that are valid (in pseudo IR):
2137 .. code-block:: llvm
2147 This is safe because all of the output bits are affected by the undef
2148 bits. Any output bit can have a zero or one depending on the input bits.
2150 .. code-block:: llvm
2161 These logical operations have bits that are not always affected by the
2162 input. For example, if ``%X`` has a zero bit, then the output of the
2163 '``and``' operation will always be a zero for that bit, no matter what
2164 the corresponding bit from the '``undef``' is. As such, it is unsafe to
2165 optimize or assume that the result of the '``and``' is '``undef``'.
2166 However, it is safe to assume that all bits of the '``undef``' could be
2167 0, and optimize the '``and``' to 0. Likewise, it is safe to assume that
2168 all the bits of the '``undef``' operand to the '``or``' could be set,
2169 allowing the '``or``' to be folded to -1.
2171 .. code-block:: llvm
2173 %A = select undef, %X, %Y
2174 %B = select undef, 42, %Y
2175 %C = select %X, %Y, undef
2185 This set of examples shows that undefined '``select``' (and conditional
2186 branch) conditions can go *either way*, but they have to come from one
2187 of the two operands. In the ``%A`` example, if ``%X`` and ``%Y`` were
2188 both known to have a clear low bit, then ``%A`` would have to have a
2189 cleared low bit. However, in the ``%C`` example, the optimizer is
2190 allowed to assume that the '``undef``' operand could be the same as
2191 ``%Y``, allowing the whole '``select``' to be eliminated.
2193 .. code-block:: llvm
2195 %A = xor undef, undef
2212 This example points out that two '``undef``' operands are not
2213 necessarily the same. This can be surprising to people (and also matches
2214 C semantics) where they assume that "``X^X``" is always zero, even if
2215 ``X`` is undefined. This isn't true for a number of reasons, but the
2216 short answer is that an '``undef``' "variable" can arbitrarily change
2217 its value over its "live range". This is true because the variable
2218 doesn't actually *have a live range*. Instead, the value is logically
2219 read from arbitrary registers that happen to be around when needed, so
2220 the value is not necessarily consistent over time. In fact, ``%A`` and
2221 ``%C`` need to have the same semantics or the core LLVM "replace all
2222 uses with" concept would not hold.
2224 .. code-block:: llvm
2232 These examples show the crucial difference between an *undefined value*
2233 and *undefined behavior*. An undefined value (like '``undef``') is
2234 allowed to have an arbitrary bit-pattern. This means that the ``%A``
2235 operation can be constant folded to '``undef``', because the '``undef``'
2236 could be an SNaN, and ``fdiv`` is not (currently) defined on SNaN's.
2237 However, in the second example, we can make a more aggressive
2238 assumption: because the ``undef`` is allowed to be an arbitrary value,
2239 we are allowed to assume that it could be zero. Since a divide by zero
2240 has *undefined behavior*, we are allowed to assume that the operation
2241 does not execute at all. This allows us to delete the divide and all
2242 code after it. Because the undefined operation "can't happen", the
2243 optimizer can assume that it occurs in dead code.
2245 .. code-block:: llvm
2247 a: store undef -> %X
2248 b: store %X -> undef
2253 These examples reiterate the ``fdiv`` example: a store *of* an undefined
2254 value can be assumed to not have any effect; we can assume that the
2255 value is overwritten with bits that happen to match what was already
2256 there. However, a store *to* an undefined location could clobber
2257 arbitrary memory, therefore, it has undefined behavior.
2264 Poison values are similar to :ref:`undef values <undefvalues>`, however
2265 they also represent the fact that an instruction or constant expression
2266 which cannot evoke side effects has nevertheless detected a condition
2267 which results in undefined behavior.
2269 There is currently no way of representing a poison value in the IR; they
2270 only exist when produced by operations such as :ref:`add <i_add>` with
2273 Poison value behavior is defined in terms of value *dependence*:
2275 - Values other than :ref:`phi <i_phi>` nodes depend on their operands.
2276 - :ref:`Phi <i_phi>` nodes depend on the operand corresponding to
2277 their dynamic predecessor basic block.
2278 - Function arguments depend on the corresponding actual argument values
2279 in the dynamic callers of their functions.
2280 - :ref:`Call <i_call>` instructions depend on the :ref:`ret <i_ret>`
2281 instructions that dynamically transfer control back to them.
2282 - :ref:`Invoke <i_invoke>` instructions depend on the
2283 :ref:`ret <i_ret>`, :ref:`resume <i_resume>`, or exception-throwing
2284 call instructions that dynamically transfer control back to them.
2285 - Non-volatile loads and stores depend on the most recent stores to all
2286 of the referenced memory addresses, following the order in the IR
2287 (including loads and stores implied by intrinsics such as
2288 :ref:`@llvm.memcpy <int_memcpy>`.)
2289 - An instruction with externally visible side effects depends on the
2290 most recent preceding instruction with externally visible side
2291 effects, following the order in the IR. (This includes :ref:`volatile
2292 operations <volatile>`.)
2293 - An instruction *control-depends* on a :ref:`terminator
2294 instruction <terminators>` if the terminator instruction has
2295 multiple successors and the instruction is always executed when
2296 control transfers to one of the successors, and may not be executed
2297 when control is transferred to another.
2298 - Additionally, an instruction also *control-depends* on a terminator
2299 instruction if the set of instructions it otherwise depends on would
2300 be different if the terminator had transferred control to a different
2302 - Dependence is transitive.
2304 Poison Values have the same behavior as :ref:`undef values <undefvalues>`,
2305 with the additional affect that any instruction which has a *dependence*
2306 on a poison value has undefined behavior.
2308 Here are some examples:
2310 .. code-block:: llvm
2313 %poison = sub nuw i32 0, 1 ; Results in a poison value.
2314 %still_poison = and i32 %poison, 0 ; 0, but also poison.
2315 %poison_yet_again = getelementptr i32* @h, i32 %still_poison
2316 store i32 0, i32* %poison_yet_again ; memory at @h[0] is poisoned
2318 store i32 %poison, i32* @g ; Poison value stored to memory.
2319 %poison2 = load i32* @g ; Poison value loaded back from memory.
2321 store volatile i32 %poison, i32* @g ; External observation; undefined behavior.
2323 %narrowaddr = bitcast i32* @g to i16*
2324 %wideaddr = bitcast i32* @g to i64*
2325 %poison3 = load i16* %narrowaddr ; Returns a poison value.
2326 %poison4 = load i64* %wideaddr ; Returns a poison value.
2328 %cmp = icmp slt i32 %poison, 0 ; Returns a poison value.
2329 br i1 %cmp, label %true, label %end ; Branch to either destination.
2332 store volatile i32 0, i32* @g ; This is control-dependent on %cmp, so
2333 ; it has undefined behavior.
2337 %p = phi i32 [ 0, %entry ], [ 1, %true ]
2338 ; Both edges into this PHI are
2339 ; control-dependent on %cmp, so this
2340 ; always results in a poison value.
2342 store volatile i32 0, i32* @g ; This would depend on the store in %true
2343 ; if %cmp is true, or the store in %entry
2344 ; otherwise, so this is undefined behavior.
2346 br i1 %cmp, label %second_true, label %second_end
2347 ; The same branch again, but this time the
2348 ; true block doesn't have side effects.
2355 store volatile i32 0, i32* @g ; This time, the instruction always depends
2356 ; on the store in %end. Also, it is
2357 ; control-equivalent to %end, so this is
2358 ; well-defined (ignoring earlier undefined
2359 ; behavior in this example).
2363 Addresses of Basic Blocks
2364 -------------------------
2366 ``blockaddress(@function, %block)``
2368 The '``blockaddress``' constant computes the address of the specified
2369 basic block in the specified function, and always has an ``i8*`` type.
2370 Taking the address of the entry block is illegal.
2372 This value only has defined behavior when used as an operand to the
2373 ':ref:`indirectbr <i_indirectbr>`' instruction, or for comparisons
2374 against null. Pointer equality tests between labels addresses results in
2375 undefined behavior --- though, again, comparison against null is ok, and
2376 no label is equal to the null pointer. This may be passed around as an
2377 opaque pointer sized value as long as the bits are not inspected. This
2378 allows ``ptrtoint`` and arithmetic to be performed on these values so
2379 long as the original value is reconstituted before the ``indirectbr``
2382 Finally, some targets may provide defined semantics when using the value
2383 as the operand to an inline assembly, but that is target specific.
2387 Constant Expressions
2388 --------------------
2390 Constant expressions are used to allow expressions involving other
2391 constants to be used as constants. Constant expressions may be of any
2392 :ref:`first class <t_firstclass>` type and may involve any LLVM operation
2393 that does not have side effects (e.g. load and call are not supported).
2394 The following is the syntax for constant expressions:
2396 ``trunc (CST to TYPE)``
2397 Truncate a constant to another type. The bit size of CST must be
2398 larger than the bit size of TYPE. Both types must be integers.
2399 ``zext (CST to TYPE)``
2400 Zero extend a constant to another type. The bit size of CST must be
2401 smaller than the bit size of TYPE. Both types must be integers.
2402 ``sext (CST to TYPE)``
2403 Sign extend a constant to another type. The bit size of CST must be
2404 smaller than the bit size of TYPE. Both types must be integers.
2405 ``fptrunc (CST to TYPE)``
2406 Truncate a floating point constant to another floating point type.
2407 The size of CST must be larger than the size of TYPE. Both types
2408 must be floating point.
2409 ``fpext (CST to TYPE)``
2410 Floating point extend a constant to another type. The size of CST
2411 must be smaller or equal to the size of TYPE. Both types must be
2413 ``fptoui (CST to TYPE)``
2414 Convert a floating point constant to the corresponding unsigned
2415 integer constant. TYPE must be a scalar or vector integer type. CST
2416 must be of scalar or vector floating point type. Both CST and TYPE
2417 must be scalars, or vectors of the same number of elements. If the
2418 value won't fit in the integer type, the results are undefined.
2419 ``fptosi (CST to TYPE)``
2420 Convert a floating point constant to the corresponding signed
2421 integer constant. TYPE must be a scalar or vector integer type. CST
2422 must be of scalar or vector floating point type. Both CST and TYPE
2423 must be scalars, or vectors of the same number of elements. If the
2424 value won't fit in the integer type, the results are undefined.
2425 ``uitofp (CST to TYPE)``
2426 Convert an unsigned integer constant to the corresponding floating
2427 point constant. TYPE must be a scalar or vector floating point type.
2428 CST must be of scalar or vector integer type. Both CST and TYPE must
2429 be scalars, or vectors of the same number of elements. If the value
2430 won't fit in the floating point type, the results are undefined.
2431 ``sitofp (CST to TYPE)``
2432 Convert a signed integer constant to the corresponding floating
2433 point constant. TYPE must be a scalar or vector floating point type.
2434 CST must be of scalar or vector integer type. Both CST and TYPE must
2435 be scalars, or vectors of the same number of elements. If the value
2436 won't fit in the floating point type, the results are undefined.
2437 ``ptrtoint (CST to TYPE)``
2438 Convert a pointer typed constant to the corresponding integer
2439 constant. ``TYPE`` must be an integer type. ``CST`` must be of
2440 pointer type. The ``CST`` value is zero extended, truncated, or
2441 unchanged to make it fit in ``TYPE``.
2442 ``inttoptr (CST to TYPE)``
2443 Convert an integer constant to a pointer constant. TYPE must be a
2444 pointer type. CST must be of integer type. The CST value is zero
2445 extended, truncated, or unchanged to make it fit in a pointer size.
2446 This one is *really* dangerous!
2447 ``bitcast (CST to TYPE)``
2448 Convert a constant, CST, to another TYPE. The constraints of the
2449 operands are the same as those for the :ref:`bitcast
2450 instruction <i_bitcast>`.
2451 ``addrspacecast (CST to TYPE)``
2452 Convert a constant pointer or constant vector of pointer, CST, to another
2453 TYPE in a different address space. The constraints of the operands are the
2454 same as those for the :ref:`addrspacecast instruction <i_addrspacecast>`.
2455 ``getelementptr (CSTPTR, IDX0, IDX1, ...)``, ``getelementptr inbounds (CSTPTR, IDX0, IDX1, ...)``
2456 Perform the :ref:`getelementptr operation <i_getelementptr>` on
2457 constants. As with the :ref:`getelementptr <i_getelementptr>`
2458 instruction, the index list may have zero or more indexes, which are
2459 required to make sense for the type of "CSTPTR".
2460 ``select (COND, VAL1, VAL2)``
2461 Perform the :ref:`select operation <i_select>` on constants.
2462 ``icmp COND (VAL1, VAL2)``
2463 Performs the :ref:`icmp operation <i_icmp>` on constants.
2464 ``fcmp COND (VAL1, VAL2)``
2465 Performs the :ref:`fcmp operation <i_fcmp>` on constants.
2466 ``extractelement (VAL, IDX)``
2467 Perform the :ref:`extractelement operation <i_extractelement>` on
2469 ``insertelement (VAL, ELT, IDX)``
2470 Perform the :ref:`insertelement operation <i_insertelement>` on
2472 ``shufflevector (VEC1, VEC2, IDXMASK)``
2473 Perform the :ref:`shufflevector operation <i_shufflevector>` on
2475 ``extractvalue (VAL, IDX0, IDX1, ...)``
2476 Perform the :ref:`extractvalue operation <i_extractvalue>` on
2477 constants. The index list is interpreted in a similar manner as
2478 indices in a ':ref:`getelementptr <i_getelementptr>`' operation. At
2479 least one index value must be specified.
2480 ``insertvalue (VAL, ELT, IDX0, IDX1, ...)``
2481 Perform the :ref:`insertvalue operation <i_insertvalue>` on constants.
2482 The index list is interpreted in a similar manner as indices in a
2483 ':ref:`getelementptr <i_getelementptr>`' operation. At least one index
2484 value must be specified.
2485 ``OPCODE (LHS, RHS)``
2486 Perform the specified operation of the LHS and RHS constants. OPCODE
2487 may be any of the :ref:`binary <binaryops>` or :ref:`bitwise
2488 binary <bitwiseops>` operations. The constraints on operands are
2489 the same as those for the corresponding instruction (e.g. no bitwise
2490 operations on floating point values are allowed).
2497 Inline Assembler Expressions
2498 ----------------------------
2500 LLVM supports inline assembler expressions (as opposed to :ref:`Module-Level
2501 Inline Assembly <moduleasm>`) through the use of a special value. This
2502 value represents the inline assembler as a string (containing the
2503 instructions to emit), a list of operand constraints (stored as a
2504 string), a flag that indicates whether or not the inline asm expression
2505 has side effects, and a flag indicating whether the function containing
2506 the asm needs to align its stack conservatively. An example inline
2507 assembler expression is:
2509 .. code-block:: llvm
2511 i32 (i32) asm "bswap $0", "=r,r"
2513 Inline assembler expressions may **only** be used as the callee operand
2514 of a :ref:`call <i_call>` or an :ref:`invoke <i_invoke>` instruction.
2515 Thus, typically we have:
2517 .. code-block:: llvm
2519 %X = call i32 asm "bswap $0", "=r,r"(i32 %Y)
2521 Inline asms with side effects not visible in the constraint list must be
2522 marked as having side effects. This is done through the use of the
2523 '``sideeffect``' keyword, like so:
2525 .. code-block:: llvm
2527 call void asm sideeffect "eieio", ""()
2529 In some cases inline asms will contain code that will not work unless
2530 the stack is aligned in some way, such as calls or SSE instructions on
2531 x86, yet will not contain code that does that alignment within the asm.
2532 The compiler should make conservative assumptions about what the asm
2533 might contain and should generate its usual stack alignment code in the
2534 prologue if the '``alignstack``' keyword is present:
2536 .. code-block:: llvm
2538 call void asm alignstack "eieio", ""()
2540 Inline asms also support using non-standard assembly dialects. The
2541 assumed dialect is ATT. When the '``inteldialect``' keyword is present,
2542 the inline asm is using the Intel dialect. Currently, ATT and Intel are
2543 the only supported dialects. An example is:
2545 .. code-block:: llvm
2547 call void asm inteldialect "eieio", ""()
2549 If multiple keywords appear the '``sideeffect``' keyword must come
2550 first, the '``alignstack``' keyword second and the '``inteldialect``'
2556 The call instructions that wrap inline asm nodes may have a
2557 "``!srcloc``" MDNode attached to it that contains a list of constant
2558 integers. If present, the code generator will use the integer as the
2559 location cookie value when report errors through the ``LLVMContext``
2560 error reporting mechanisms. This allows a front-end to correlate backend
2561 errors that occur with inline asm back to the source code that produced
2564 .. code-block:: llvm
2566 call void asm sideeffect "something bad", ""(), !srcloc !42
2568 !42 = !{ i32 1234567 }
2570 It is up to the front-end to make sense of the magic numbers it places
2571 in the IR. If the MDNode contains multiple constants, the code generator
2572 will use the one that corresponds to the line of the asm that the error
2577 Metadata Nodes and Metadata Strings
2578 -----------------------------------
2580 LLVM IR allows metadata to be attached to instructions in the program
2581 that can convey extra information about the code to the optimizers and
2582 code generator. One example application of metadata is source-level
2583 debug information. There are two metadata primitives: strings and nodes.
2584 All metadata has the ``metadata`` type and is identified in syntax by a
2585 preceding exclamation point ('``!``').
2587 A metadata string is a string surrounded by double quotes. It can
2588 contain any character by escaping non-printable characters with
2589 "``\xx``" where "``xx``" is the two digit hex code. For example:
2592 Metadata nodes are represented with notation similar to structure
2593 constants (a comma separated list of elements, surrounded by braces and
2594 preceded by an exclamation point). Metadata nodes can have any values as
2595 their operand. For example:
2597 .. code-block:: llvm
2599 !{ metadata !"test\00", i32 10}
2601 A :ref:`named metadata <namedmetadatastructure>` is a collection of
2602 metadata nodes, which can be looked up in the module symbol table. For
2605 .. code-block:: llvm
2607 !foo = metadata !{!4, !3}
2609 Metadata can be used as function arguments. Here ``llvm.dbg.value``
2610 function is using two metadata arguments:
2612 .. code-block:: llvm
2614 call void @llvm.dbg.value(metadata !24, i64 0, metadata !25)
2616 Metadata can be attached with an instruction. Here metadata ``!21`` is
2617 attached to the ``add`` instruction using the ``!dbg`` identifier:
2619 .. code-block:: llvm
2621 %indvar.next = add i64 %indvar, 1, !dbg !21
2623 More information about specific metadata nodes recognized by the
2624 optimizers and code generator is found below.
2629 In LLVM IR, memory does not have types, so LLVM's own type system is not
2630 suitable for doing TBAA. Instead, metadata is added to the IR to
2631 describe a type system of a higher level language. This can be used to
2632 implement typical C/C++ TBAA, but it can also be used to implement
2633 custom alias analysis behavior for other languages.
2635 The current metadata format is very simple. TBAA metadata nodes have up
2636 to three fields, e.g.:
2638 .. code-block:: llvm
2640 !0 = metadata !{ metadata !"an example type tree" }
2641 !1 = metadata !{ metadata !"int", metadata !0 }
2642 !2 = metadata !{ metadata !"float", metadata !0 }
2643 !3 = metadata !{ metadata !"const float", metadata !2, i64 1 }
2645 The first field is an identity field. It can be any value, usually a
2646 metadata string, which uniquely identifies the type. The most important
2647 name in the tree is the name of the root node. Two trees with different
2648 root node names are entirely disjoint, even if they have leaves with
2651 The second field identifies the type's parent node in the tree, or is
2652 null or omitted for a root node. A type is considered to alias all of
2653 its descendants and all of its ancestors in the tree. Also, a type is
2654 considered to alias all types in other trees, so that bitcode produced
2655 from multiple front-ends is handled conservatively.
2657 If the third field is present, it's an integer which if equal to 1
2658 indicates that the type is "constant" (meaning
2659 ``pointsToConstantMemory`` should return true; see `other useful
2660 AliasAnalysis methods <AliasAnalysis.html#OtherItfs>`_).
2662 '``tbaa.struct``' Metadata
2663 ^^^^^^^^^^^^^^^^^^^^^^^^^^
2665 The :ref:`llvm.memcpy <int_memcpy>` is often used to implement
2666 aggregate assignment operations in C and similar languages, however it
2667 is defined to copy a contiguous region of memory, which is more than
2668 strictly necessary for aggregate types which contain holes due to
2669 padding. Also, it doesn't contain any TBAA information about the fields
2672 ``!tbaa.struct`` metadata can describe which memory subregions in a
2673 memcpy are padding and what the TBAA tags of the struct are.
2675 The current metadata format is very simple. ``!tbaa.struct`` metadata
2676 nodes are a list of operands which are in conceptual groups of three.
2677 For each group of three, the first operand gives the byte offset of a
2678 field in bytes, the second gives its size in bytes, and the third gives
2681 .. code-block:: llvm
2683 !4 = metadata !{ i64 0, i64 4, metadata !1, i64 8, i64 4, metadata !2 }
2685 This describes a struct with two fields. The first is at offset 0 bytes
2686 with size 4 bytes, and has tbaa tag !1. The second is at offset 8 bytes
2687 and has size 4 bytes and has tbaa tag !2.
2689 Note that the fields need not be contiguous. In this example, there is a
2690 4 byte gap between the two fields. This gap represents padding which
2691 does not carry useful data and need not be preserved.
2693 '``fpmath``' Metadata
2694 ^^^^^^^^^^^^^^^^^^^^^
2696 ``fpmath`` metadata may be attached to any instruction of floating point
2697 type. It can be used to express the maximum acceptable error in the
2698 result of that instruction, in ULPs, thus potentially allowing the
2699 compiler to use a more efficient but less accurate method of computing
2700 it. ULP is defined as follows:
2702 If ``x`` is a real number that lies between two finite consecutive
2703 floating-point numbers ``a`` and ``b``, without being equal to one
2704 of them, then ``ulp(x) = |b - a|``, otherwise ``ulp(x)`` is the
2705 distance between the two non-equal finite floating-point numbers
2706 nearest ``x``. Moreover, ``ulp(NaN)`` is ``NaN``.
2708 The metadata node shall consist of a single positive floating point
2709 number representing the maximum relative error, for example:
2711 .. code-block:: llvm
2713 !0 = metadata !{ float 2.5 } ; maximum acceptable inaccuracy is 2.5 ULPs
2715 '``range``' Metadata
2716 ^^^^^^^^^^^^^^^^^^^^
2718 ``range`` metadata may be attached only to loads of integer types. It
2719 expresses the possible ranges the loaded value is in. The ranges are
2720 represented with a flattened list of integers. The loaded value is known
2721 to be in the union of the ranges defined by each consecutive pair. Each
2722 pair has the following properties:
2724 - The type must match the type loaded by the instruction.
2725 - The pair ``a,b`` represents the range ``[a,b)``.
2726 - Both ``a`` and ``b`` are constants.
2727 - The range is allowed to wrap.
2728 - The range should not represent the full or empty set. That is,
2731 In addition, the pairs must be in signed order of the lower bound and
2732 they must be non-contiguous.
2736 .. code-block:: llvm
2738 %a = load i8* %x, align 1, !range !0 ; Can only be 0 or 1
2739 %b = load i8* %y, align 1, !range !1 ; Can only be 255 (-1), 0 or 1
2740 %c = load i8* %z, align 1, !range !2 ; Can only be 0, 1, 3, 4 or 5
2741 %d = load i8* %z, align 1, !range !3 ; Can only be -2, -1, 3, 4 or 5
2743 !0 = metadata !{ i8 0, i8 2 }
2744 !1 = metadata !{ i8 255, i8 2 }
2745 !2 = metadata !{ i8 0, i8 2, i8 3, i8 6 }
2746 !3 = metadata !{ i8 -2, i8 0, i8 3, i8 6 }
2751 It is sometimes useful to attach information to loop constructs. Currently,
2752 loop metadata is implemented as metadata attached to the branch instruction
2753 in the loop latch block. This type of metadata refer to a metadata node that is
2754 guaranteed to be separate for each loop. The loop identifier metadata is
2755 specified with the name ``llvm.loop``.
2757 The loop identifier metadata is implemented using a metadata that refers to
2758 itself to avoid merging it with any other identifier metadata, e.g.,
2759 during module linkage or function inlining. That is, each loop should refer
2760 to their own identification metadata even if they reside in separate functions.
2761 The following example contains loop identifier metadata for two separate loop
2764 .. code-block:: llvm
2766 !0 = metadata !{ metadata !0 }
2767 !1 = metadata !{ metadata !1 }
2769 The loop identifier metadata can be used to specify additional per-loop
2770 metadata. Any operands after the first operand can be treated as user-defined
2771 metadata. For example the ``llvm.vectorizer.unroll`` metadata is understood
2772 by the loop vectorizer to indicate how many times to unroll the loop:
2774 .. code-block:: llvm
2776 br i1 %exitcond, label %._crit_edge, label %.lr.ph, !llvm.loop !0
2778 !0 = metadata !{ metadata !0, metadata !1 }
2779 !1 = metadata !{ metadata !"llvm.vectorizer.unroll", i32 2 }
2784 Metadata types used to annotate memory accesses with information helpful
2785 for optimizations are prefixed with ``llvm.mem``.
2787 '``llvm.mem.parallel_loop_access``' Metadata
2788 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2790 The ``llvm.mem.parallel_loop_access`` metadata refers to a loop identifier,
2791 or metadata containing a list of loop identifiers for nested loops.
2792 The metadata is attached to memory accessing instructions and denotes that
2793 no loop carried memory dependence exist between it and other instructions denoted
2794 with the same loop identifier.
2796 Precisely, given two instructions ``m1`` and ``m2`` that both have the
2797 ``llvm.mem.parallel_loop_access`` metadata, with ``L1`` and ``L2`` being the
2798 set of loops associated with that metadata, respectively, then there is no loop
2799 carried dependence between ``m1`` and ``m2`` for loops ``L1`` or
2802 As a special case, if all memory accessing instructions in a loop have
2803 ``llvm.mem.parallel_loop_access`` metadata that refers to that loop, then the
2804 loop has no loop carried memory dependences and is considered to be a parallel
2807 Note that if not all memory access instructions have such metadata referring to
2808 the loop, then the loop is considered not being trivially parallel. Additional
2809 memory dependence analysis is required to make that determination. As a fail
2810 safe mechanism, this causes loops that were originally parallel to be considered
2811 sequential (if optimization passes that are unaware of the parallel semantics
2812 insert new memory instructions into the loop body).
2814 Example of a loop that is considered parallel due to its correct use of
2815 both ``llvm.loop`` and ``llvm.mem.parallel_loop_access``
2816 metadata types that refer to the same loop identifier metadata.
2818 .. code-block:: llvm
2822 %val0 = load i32* %arrayidx, !llvm.mem.parallel_loop_access !0
2824 store i32 %val0, i32* %arrayidx1, !llvm.mem.parallel_loop_access !0
2826 br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !0
2830 !0 = metadata !{ metadata !0 }
2832 It is also possible to have nested parallel loops. In that case the
2833 memory accesses refer to a list of loop identifier metadata nodes instead of
2834 the loop identifier metadata node directly:
2836 .. code-block:: llvm
2840 %val1 = load i32* %arrayidx3, !llvm.mem.parallel_loop_access !2
2842 br label %inner.for.body
2846 %val0 = load i32* %arrayidx1, !llvm.mem.parallel_loop_access !0
2848 store i32 %val0, i32* %arrayidx2, !llvm.mem.parallel_loop_access !0
2850 br i1 %exitcond, label %inner.for.end, label %inner.for.body, !llvm.loop !1
2854 store i32 %val1, i32* %arrayidx4, !llvm.mem.parallel_loop_access !2
2856 br i1 %exitcond, label %outer.for.end, label %outer.for.body, !llvm.loop !2
2858 outer.for.end: ; preds = %for.body
2860 !0 = metadata !{ metadata !1, metadata !2 } ; a list of loop identifiers
2861 !1 = metadata !{ metadata !1 } ; an identifier for the inner loop
2862 !2 = metadata !{ metadata !2 } ; an identifier for the outer loop
2864 '``llvm.vectorizer``'
2865 ^^^^^^^^^^^^^^^^^^^^^
2867 Metadata prefixed with ``llvm.vectorizer`` is used to control per-loop
2868 vectorization parameters such as vectorization factor and unroll factor.
2870 ``llvm.vectorizer`` metadata should be used in conjunction with ``llvm.loop``
2871 loop identification metadata.
2873 '``llvm.vectorizer.unroll``' Metadata
2874 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2876 This metadata instructs the loop vectorizer to unroll the specified
2877 loop exactly ``N`` times.
2879 The first operand is the string ``llvm.vectorizer.unroll`` and the second
2880 operand is an integer specifying the unroll factor. For example:
2882 .. code-block:: llvm
2884 !0 = metadata !{ metadata !"llvm.vectorizer.unroll", i32 4 }
2886 Note that setting ``llvm.vectorizer.unroll`` to 1 disables unrolling of the
2889 If ``llvm.vectorizer.unroll`` is set to 0 then the amount of unrolling will be
2890 determined automatically.
2892 '``llvm.vectorizer.width``' Metadata
2893 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2895 This metadata sets the target width of the vectorizer to ``N``. Without
2896 this metadata, the vectorizer will choose a width automatically.
2897 Regardless of this metadata, the vectorizer will only vectorize loops if
2898 it believes it is valid to do so.
2900 The first operand is the string ``llvm.vectorizer.width`` and the second
2901 operand is an integer specifying the width. For example:
2903 .. code-block:: llvm
2905 !0 = metadata !{ metadata !"llvm.vectorizer.width", i32 4 }
2907 Note that setting ``llvm.vectorizer.width`` to 1 disables vectorization of the
2910 If ``llvm.vectorizer.width`` is set to 0 then the width will be determined
2913 Module Flags Metadata
2914 =====================
2916 Information about the module as a whole is difficult to convey to LLVM's
2917 subsystems. The LLVM IR isn't sufficient to transmit this information.
2918 The ``llvm.module.flags`` named metadata exists in order to facilitate
2919 this. These flags are in the form of key / value pairs --- much like a
2920 dictionary --- making it easy for any subsystem who cares about a flag to
2923 The ``llvm.module.flags`` metadata contains a list of metadata triplets.
2924 Each triplet has the following form:
2926 - The first element is a *behavior* flag, which specifies the behavior
2927 when two (or more) modules are merged together, and it encounters two
2928 (or more) metadata with the same ID. The supported behaviors are
2930 - The second element is a metadata string that is a unique ID for the
2931 metadata. Each module may only have one flag entry for each unique ID (not
2932 including entries with the **Require** behavior).
2933 - The third element is the value of the flag.
2935 When two (or more) modules are merged together, the resulting
2936 ``llvm.module.flags`` metadata is the union of the modules' flags. That is, for
2937 each unique metadata ID string, there will be exactly one entry in the merged
2938 modules ``llvm.module.flags`` metadata table, and the value for that entry will
2939 be determined by the merge behavior flag, as described below. The only exception
2940 is that entries with the *Require* behavior are always preserved.
2942 The following behaviors are supported:
2953 Emits an error if two values disagree, otherwise the resulting value
2954 is that of the operands.
2958 Emits a warning if two values disagree. The result value will be the
2959 operand for the flag from the first module being linked.
2963 Adds a requirement that another module flag be present and have a
2964 specified value after linking is performed. The value must be a
2965 metadata pair, where the first element of the pair is the ID of the
2966 module flag to be restricted, and the second element of the pair is
2967 the value the module flag should be restricted to. This behavior can
2968 be used to restrict the allowable results (via triggering of an
2969 error) of linking IDs with the **Override** behavior.
2973 Uses the specified value, regardless of the behavior or value of the
2974 other module. If both modules specify **Override**, but the values
2975 differ, an error will be emitted.
2979 Appends the two values, which are required to be metadata nodes.
2983 Appends the two values, which are required to be metadata
2984 nodes. However, duplicate entries in the second list are dropped
2985 during the append operation.
2987 It is an error for a particular unique flag ID to have multiple behaviors,
2988 except in the case of **Require** (which adds restrictions on another metadata
2989 value) or **Override**.
2991 An example of module flags:
2993 .. code-block:: llvm
2995 !0 = metadata !{ i32 1, metadata !"foo", i32 1 }
2996 !1 = metadata !{ i32 4, metadata !"bar", i32 37 }
2997 !2 = metadata !{ i32 2, metadata !"qux", i32 42 }
2998 !3 = metadata !{ i32 3, metadata !"qux",
3000 metadata !"foo", i32 1
3003 !llvm.module.flags = !{ !0, !1, !2, !3 }
3005 - Metadata ``!0`` has the ID ``!"foo"`` and the value '1'. The behavior
3006 if two or more ``!"foo"`` flags are seen is to emit an error if their
3007 values are not equal.
3009 - Metadata ``!1`` has the ID ``!"bar"`` and the value '37'. The
3010 behavior if two or more ``!"bar"`` flags are seen is to use the value
3013 - Metadata ``!2`` has the ID ``!"qux"`` and the value '42'. The
3014 behavior if two or more ``!"qux"`` flags are seen is to emit a
3015 warning if their values are not equal.
3017 - Metadata ``!3`` has the ID ``!"qux"`` and the value:
3021 metadata !{ metadata !"foo", i32 1 }
3023 The behavior is to emit an error if the ``llvm.module.flags`` does not
3024 contain a flag with the ID ``!"foo"`` that has the value '1' after linking is
3027 Objective-C Garbage Collection Module Flags Metadata
3028 ----------------------------------------------------
3030 On the Mach-O platform, Objective-C stores metadata about garbage
3031 collection in a special section called "image info". The metadata
3032 consists of a version number and a bitmask specifying what types of
3033 garbage collection are supported (if any) by the file. If two or more
3034 modules are linked together their garbage collection metadata needs to
3035 be merged rather than appended together.
3037 The Objective-C garbage collection module flags metadata consists of the
3038 following key-value pairs:
3047 * - ``Objective-C Version``
3048 - **[Required]** --- The Objective-C ABI version. Valid values are 1 and 2.
3050 * - ``Objective-C Image Info Version``
3051 - **[Required]** --- The version of the image info section. Currently
3054 * - ``Objective-C Image Info Section``
3055 - **[Required]** --- The section to place the metadata. Valid values are
3056 ``"__OBJC, __image_info, regular"`` for Objective-C ABI version 1, and
3057 ``"__DATA,__objc_imageinfo, regular, no_dead_strip"`` for
3058 Objective-C ABI version 2.
3060 * - ``Objective-C Garbage Collection``
3061 - **[Required]** --- Specifies whether garbage collection is supported or
3062 not. Valid values are 0, for no garbage collection, and 2, for garbage
3063 collection supported.
3065 * - ``Objective-C GC Only``
3066 - **[Optional]** --- Specifies that only garbage collection is supported.
3067 If present, its value must be 6. This flag requires that the
3068 ``Objective-C Garbage Collection`` flag have the value 2.
3070 Some important flag interactions:
3072 - If a module with ``Objective-C Garbage Collection`` set to 0 is
3073 merged with a module with ``Objective-C Garbage Collection`` set to
3074 2, then the resulting module has the
3075 ``Objective-C Garbage Collection`` flag set to 0.
3076 - A module with ``Objective-C Garbage Collection`` set to 0 cannot be
3077 merged with a module with ``Objective-C GC Only`` set to 6.
3079 Automatic Linker Flags Module Flags Metadata
3080 --------------------------------------------
3082 Some targets support embedding flags to the linker inside individual object
3083 files. Typically this is used in conjunction with language extensions which
3084 allow source files to explicitly declare the libraries they depend on, and have
3085 these automatically be transmitted to the linker via object files.
3087 These flags are encoded in the IR using metadata in the module flags section,
3088 using the ``Linker Options`` key. The merge behavior for this flag is required
3089 to be ``AppendUnique``, and the value for the key is expected to be a metadata
3090 node which should be a list of other metadata nodes, each of which should be a
3091 list of metadata strings defining linker options.
3093 For example, the following metadata section specifies two separate sets of
3094 linker options, presumably to link against ``libz`` and the ``Cocoa``
3097 !0 = metadata !{ i32 6, metadata !"Linker Options",
3099 metadata !{ metadata !"-lz" },
3100 metadata !{ metadata !"-framework", metadata !"Cocoa" } } }
3101 !llvm.module.flags = !{ !0 }
3103 The metadata encoding as lists of lists of options, as opposed to a collapsed
3104 list of options, is chosen so that the IR encoding can use multiple option
3105 strings to specify e.g., a single library, while still having that specifier be
3106 preserved as an atomic element that can be recognized by a target specific
3107 assembly writer or object file emitter.
3109 Each individual option is required to be either a valid option for the target's
3110 linker, or an option that is reserved by the target specific assembly writer or
3111 object file emitter. No other aspect of these options is defined by the IR.
3113 .. _intrinsicglobalvariables:
3115 Intrinsic Global Variables
3116 ==========================
3118 LLVM has a number of "magic" global variables that contain data that
3119 affect code generation or other IR semantics. These are documented here.
3120 All globals of this sort should have a section specified as
3121 "``llvm.metadata``". This section and all globals that start with
3122 "``llvm.``" are reserved for use by LLVM.
3126 The '``llvm.used``' Global Variable
3127 -----------------------------------
3129 The ``@llvm.used`` global is an array which has
3130 :ref:`appending linkage <linkage_appending>`. This array contains a list of
3131 pointers to named global variables, functions and aliases which may optionally
3132 have a pointer cast formed of bitcast or getelementptr. For example, a legal
3135 .. code-block:: llvm
3140 @llvm.used = appending global [2 x i8*] [
3142 i8* bitcast (i32* @Y to i8*)
3143 ], section "llvm.metadata"
3145 If a symbol appears in the ``@llvm.used`` list, then the compiler, assembler,
3146 and linker are required to treat the symbol as if there is a reference to the
3147 symbol that it cannot see (which is why they have to be named). For example, if
3148 a variable has internal linkage and no references other than that from the
3149 ``@llvm.used`` list, it cannot be deleted. This is commonly used to represent
3150 references from inline asms and other things the compiler cannot "see", and
3151 corresponds to "``attribute((used))``" in GNU C.
3153 On some targets, the code generator must emit a directive to the
3154 assembler or object file to prevent the assembler and linker from
3155 molesting the symbol.
3157 .. _gv_llvmcompilerused:
3159 The '``llvm.compiler.used``' Global Variable
3160 --------------------------------------------
3162 The ``@llvm.compiler.used`` directive is the same as the ``@llvm.used``
3163 directive, except that it only prevents the compiler from touching the
3164 symbol. On targets that support it, this allows an intelligent linker to
3165 optimize references to the symbol without being impeded as it would be
3168 This is a rare construct that should only be used in rare circumstances,
3169 and should not be exposed to source languages.
3171 .. _gv_llvmglobalctors:
3173 The '``llvm.global_ctors``' Global Variable
3174 -------------------------------------------
3176 .. code-block:: llvm
3178 %0 = type { i32, void ()*, i8* }
3179 @llvm.global_ctors = appending global [1 x %0] [%0 { i32 65535, void ()* @ctor, i8* @data }]
3181 The ``@llvm.global_ctors`` array contains a list of constructor
3182 functions, priorities, and an optional associated global or function.
3183 The functions referenced by this array will be called in ascending order
3184 of priority (i.e. lowest first) when the module is loaded. The order of
3185 functions with the same priority is not defined.
3187 If the third field is present, non-null, and points to a global variable
3188 or function, the initializer function will only run if the associated
3189 data from the current module is not discarded.
3191 .. _llvmglobaldtors:
3193 The '``llvm.global_dtors``' Global Variable
3194 -------------------------------------------
3196 .. code-block:: llvm
3198 %0 = type { i32, void ()*, i8* }
3199 @llvm.global_dtors = appending global [1 x %0] [%0 { i32 65535, void ()* @dtor, i8* @data }]
3201 The ``@llvm.global_dtors`` array contains a list of destructor
3202 functions, priorities, and an optional associated global or function.
3203 The functions referenced by this array will be called in descending
3204 order of priority (i.e. highest first) when the module is loaded. The
3205 order of functions with the same priority is not defined.
3207 If the third field is present, non-null, and points to a global variable
3208 or function, the destructor function will only run if the associated
3209 data from the current module is not discarded.
3211 Instruction Reference
3212 =====================
3214 The LLVM instruction set consists of several different classifications
3215 of instructions: :ref:`terminator instructions <terminators>`, :ref:`binary
3216 instructions <binaryops>`, :ref:`bitwise binary
3217 instructions <bitwiseops>`, :ref:`memory instructions <memoryops>`, and
3218 :ref:`other instructions <otherops>`.
3222 Terminator Instructions
3223 -----------------------
3225 As mentioned :ref:`previously <functionstructure>`, every basic block in a
3226 program ends with a "Terminator" instruction, which indicates which
3227 block should be executed after the current block is finished. These
3228 terminator instructions typically yield a '``void``' value: they produce
3229 control flow, not values (the one exception being the
3230 ':ref:`invoke <i_invoke>`' instruction).
3232 The terminator instructions are: ':ref:`ret <i_ret>`',
3233 ':ref:`br <i_br>`', ':ref:`switch <i_switch>`',
3234 ':ref:`indirectbr <i_indirectbr>`', ':ref:`invoke <i_invoke>`',
3235 ':ref:`resume <i_resume>`', and ':ref:`unreachable <i_unreachable>`'.
3239 '``ret``' Instruction
3240 ^^^^^^^^^^^^^^^^^^^^^
3247 ret <type> <value> ; Return a value from a non-void function
3248 ret void ; Return from void function
3253 The '``ret``' instruction is used to return control flow (and optionally
3254 a value) from a function back to the caller.
3256 There are two forms of the '``ret``' instruction: one that returns a
3257 value and then causes control flow, and one that just causes control
3263 The '``ret``' instruction optionally accepts a single argument, the
3264 return value. The type of the return value must be a ':ref:`first
3265 class <t_firstclass>`' type.
3267 A function is not :ref:`well formed <wellformed>` if it it has a non-void
3268 return type and contains a '``ret``' instruction with no return value or
3269 a return value with a type that does not match its type, or if it has a
3270 void return type and contains a '``ret``' instruction with a return
3276 When the '``ret``' instruction is executed, control flow returns back to
3277 the calling function's context. If the caller is a
3278 ":ref:`call <i_call>`" instruction, execution continues at the
3279 instruction after the call. If the caller was an
3280 ":ref:`invoke <i_invoke>`" instruction, execution continues at the
3281 beginning of the "normal" destination block. If the instruction returns
3282 a value, that value shall set the call or invoke instruction's return
3288 .. code-block:: llvm
3290 ret i32 5 ; Return an integer value of 5
3291 ret void ; Return from a void function
3292 ret { i32, i8 } { i32 4, i8 2 } ; Return a struct of values 4 and 2
3296 '``br``' Instruction
3297 ^^^^^^^^^^^^^^^^^^^^
3304 br i1 <cond>, label <iftrue>, label <iffalse>
3305 br label <dest> ; Unconditional branch
3310 The '``br``' instruction is used to cause control flow to transfer to a
3311 different basic block in the current function. There are two forms of
3312 this instruction, corresponding to a conditional branch and an
3313 unconditional branch.
3318 The conditional branch form of the '``br``' instruction takes a single
3319 '``i1``' value and two '``label``' values. The unconditional form of the
3320 '``br``' instruction takes a single '``label``' value as a target.
3325 Upon execution of a conditional '``br``' instruction, the '``i1``'
3326 argument is evaluated. If the value is ``true``, control flows to the
3327 '``iftrue``' ``label`` argument. If "cond" is ``false``, control flows
3328 to the '``iffalse``' ``label`` argument.
3333 .. code-block:: llvm
3336 %cond = icmp eq i32 %a, %b
3337 br i1 %cond, label %IfEqual, label %IfUnequal
3345 '``switch``' Instruction
3346 ^^^^^^^^^^^^^^^^^^^^^^^^
3353 switch <intty> <value>, label <defaultdest> [ <intty> <val>, label <dest> ... ]
3358 The '``switch``' instruction is used to transfer control flow to one of
3359 several different places. It is a generalization of the '``br``'
3360 instruction, allowing a branch to occur to one of many possible
3366 The '``switch``' instruction uses three parameters: an integer
3367 comparison value '``value``', a default '``label``' destination, and an
3368 array of pairs of comparison value constants and '``label``'s. The table
3369 is not allowed to contain duplicate constant entries.
3374 The ``switch`` instruction specifies a table of values and destinations.
3375 When the '``switch``' instruction is executed, this table is searched
3376 for the given value. If the value is found, control flow is transferred
3377 to the corresponding destination; otherwise, control flow is transferred
3378 to the default destination.
3383 Depending on properties of the target machine and the particular
3384 ``switch`` instruction, this instruction may be code generated in
3385 different ways. For example, it could be generated as a series of
3386 chained conditional branches or with a lookup table.
3391 .. code-block:: llvm
3393 ; Emulate a conditional br instruction
3394 %Val = zext i1 %value to i32
3395 switch i32 %Val, label %truedest [ i32 0, label %falsedest ]
3397 ; Emulate an unconditional br instruction
3398 switch i32 0, label %dest [ ]
3400 ; Implement a jump table:
3401 switch i32 %val, label %otherwise [ i32 0, label %onzero
3403 i32 2, label %ontwo ]
3407 '``indirectbr``' Instruction
3408 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3415 indirectbr <somety>* <address>, [ label <dest1>, label <dest2>, ... ]
3420 The '``indirectbr``' instruction implements an indirect branch to a
3421 label within the current function, whose address is specified by
3422 "``address``". Address must be derived from a
3423 :ref:`blockaddress <blockaddress>` constant.
3428 The '``address``' argument is the address of the label to jump to. The
3429 rest of the arguments indicate the full set of possible destinations
3430 that the address may point to. Blocks are allowed to occur multiple
3431 times in the destination list, though this isn't particularly useful.
3433 This destination list is required so that dataflow analysis has an
3434 accurate understanding of the CFG.
3439 Control transfers to the block specified in the address argument. All
3440 possible destination blocks must be listed in the label list, otherwise
3441 this instruction has undefined behavior. This implies that jumps to
3442 labels defined in other functions have undefined behavior as well.
3447 This is typically implemented with a jump through a register.
3452 .. code-block:: llvm
3454 indirectbr i8* %Addr, [ label %bb1, label %bb2, label %bb3 ]
3458 '``invoke``' Instruction
3459 ^^^^^^^^^^^^^^^^^^^^^^^^
3466 <result> = invoke [cconv] [ret attrs] <ptr to function ty> <function ptr val>(<function args>) [fn attrs]
3467 to label <normal label> unwind label <exception label>
3472 The '``invoke``' instruction causes control to transfer to a specified
3473 function, with the possibility of control flow transfer to either the
3474 '``normal``' label or the '``exception``' label. If the callee function
3475 returns with the "``ret``" instruction, control flow will return to the
3476 "normal" label. If the callee (or any indirect callees) returns via the
3477 ":ref:`resume <i_resume>`" instruction or other exception handling
3478 mechanism, control is interrupted and continued at the dynamically
3479 nearest "exception" label.
3481 The '``exception``' label is a `landing
3482 pad <ExceptionHandling.html#overview>`_ for the exception. As such,
3483 '``exception``' label is required to have the
3484 ":ref:`landingpad <i_landingpad>`" instruction, which contains the
3485 information about the behavior of the program after unwinding happens,
3486 as its first non-PHI instruction. The restrictions on the
3487 "``landingpad``" instruction's tightly couples it to the "``invoke``"
3488 instruction, so that the important information contained within the
3489 "``landingpad``" instruction can't be lost through normal code motion.
3494 This instruction requires several arguments:
3496 #. The optional "cconv" marker indicates which :ref:`calling
3497 convention <callingconv>` the call should use. If none is
3498 specified, the call defaults to using C calling conventions.
3499 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
3500 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
3502 #. '``ptr to function ty``': shall be the signature of the pointer to
3503 function value being invoked. In most cases, this is a direct
3504 function invocation, but indirect ``invoke``'s are just as possible,
3505 branching off an arbitrary pointer to function value.
3506 #. '``function ptr val``': An LLVM value containing a pointer to a
3507 function to be invoked.
3508 #. '``function args``': argument list whose types match the function
3509 signature argument types and parameter attributes. All arguments must
3510 be of :ref:`first class <t_firstclass>` type. If the function signature
3511 indicates the function accepts a variable number of arguments, the
3512 extra arguments can be specified.
3513 #. '``normal label``': the label reached when the called function
3514 executes a '``ret``' instruction.
3515 #. '``exception label``': the label reached when a callee returns via
3516 the :ref:`resume <i_resume>` instruction or other exception handling
3518 #. The optional :ref:`function attributes <fnattrs>` list. Only
3519 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
3520 attributes are valid here.
3525 This instruction is designed to operate as a standard '``call``'
3526 instruction in most regards. The primary difference is that it
3527 establishes an association with a label, which is used by the runtime
3528 library to unwind the stack.
3530 This instruction is used in languages with destructors to ensure that
3531 proper cleanup is performed in the case of either a ``longjmp`` or a
3532 thrown exception. Additionally, this is important for implementation of
3533 '``catch``' clauses in high-level languages that support them.
3535 For the purposes of the SSA form, the definition of the value returned
3536 by the '``invoke``' instruction is deemed to occur on the edge from the
3537 current block to the "normal" label. If the callee unwinds then no
3538 return value is available.
3543 .. code-block:: llvm
3545 %retval = invoke i32 @Test(i32 15) to label %Continue
3546 unwind label %TestCleanup ; {i32}:retval set
3547 %retval = invoke coldcc i32 %Testfnptr(i32 15) to label %Continue
3548 unwind label %TestCleanup ; {i32}:retval set
3552 '``resume``' Instruction
3553 ^^^^^^^^^^^^^^^^^^^^^^^^
3560 resume <type> <value>
3565 The '``resume``' instruction is a terminator instruction that has no
3571 The '``resume``' instruction requires one argument, which must have the
3572 same type as the result of any '``landingpad``' instruction in the same
3578 The '``resume``' instruction resumes propagation of an existing
3579 (in-flight) exception whose unwinding was interrupted with a
3580 :ref:`landingpad <i_landingpad>` instruction.
3585 .. code-block:: llvm
3587 resume { i8*, i32 } %exn
3591 '``unreachable``' Instruction
3592 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3604 The '``unreachable``' instruction has no defined semantics. This
3605 instruction is used to inform the optimizer that a particular portion of
3606 the code is not reachable. This can be used to indicate that the code
3607 after a no-return function cannot be reached, and other facts.
3612 The '``unreachable``' instruction has no defined semantics.
3619 Binary operators are used to do most of the computation in a program.
3620 They require two operands of the same type, execute an operation on
3621 them, and produce a single value. The operands might represent multiple
3622 data, as is the case with the :ref:`vector <t_vector>` data type. The
3623 result value has the same type as its operands.
3625 There are several different binary operators:
3629 '``add``' Instruction
3630 ^^^^^^^^^^^^^^^^^^^^^
3637 <result> = add <ty> <op1>, <op2> ; yields {ty}:result
3638 <result> = add nuw <ty> <op1>, <op2> ; yields {ty}:result
3639 <result> = add nsw <ty> <op1>, <op2> ; yields {ty}:result
3640 <result> = add nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3645 The '``add``' instruction returns the sum of its two operands.
3650 The two arguments to the '``add``' instruction must be
3651 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3652 arguments must have identical types.
3657 The value produced is the integer sum of the two operands.
3659 If the sum has unsigned overflow, the result returned is the
3660 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
3663 Because LLVM integers use a two's complement representation, this
3664 instruction is appropriate for both signed and unsigned integers.
3666 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3667 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3668 result value of the ``add`` is a :ref:`poison value <poisonvalues>` if
3669 unsigned and/or signed overflow, respectively, occurs.
3674 .. code-block:: llvm
3676 <result> = add i32 4, %var ; yields {i32}:result = 4 + %var
3680 '``fadd``' Instruction
3681 ^^^^^^^^^^^^^^^^^^^^^^
3688 <result> = fadd [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3693 The '``fadd``' instruction returns the sum of its two operands.
3698 The two arguments to the '``fadd``' instruction must be :ref:`floating
3699 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3700 Both arguments must have identical types.
3705 The value produced is the floating point sum of the two operands. This
3706 instruction can also take any number of :ref:`fast-math flags <fastmath>`,
3707 which are optimization hints to enable otherwise unsafe floating point
3713 .. code-block:: llvm
3715 <result> = fadd float 4.0, %var ; yields {float}:result = 4.0 + %var
3717 '``sub``' Instruction
3718 ^^^^^^^^^^^^^^^^^^^^^
3725 <result> = sub <ty> <op1>, <op2> ; yields {ty}:result
3726 <result> = sub nuw <ty> <op1>, <op2> ; yields {ty}:result
3727 <result> = sub nsw <ty> <op1>, <op2> ; yields {ty}:result
3728 <result> = sub nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3733 The '``sub``' instruction returns the difference of its two operands.
3735 Note that the '``sub``' instruction is used to represent the '``neg``'
3736 instruction present in most other intermediate representations.
3741 The two arguments to the '``sub``' instruction must be
3742 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3743 arguments must have identical types.
3748 The value produced is the integer difference of the two operands.
3750 If the difference has unsigned overflow, the result returned is the
3751 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
3754 Because LLVM integers use a two's complement representation, this
3755 instruction is appropriate for both signed and unsigned integers.
3757 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3758 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3759 result value of the ``sub`` is a :ref:`poison value <poisonvalues>` if
3760 unsigned and/or signed overflow, respectively, occurs.
3765 .. code-block:: llvm
3767 <result> = sub i32 4, %var ; yields {i32}:result = 4 - %var
3768 <result> = sub i32 0, %val ; yields {i32}:result = -%var
3772 '``fsub``' Instruction
3773 ^^^^^^^^^^^^^^^^^^^^^^
3780 <result> = fsub [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3785 The '``fsub``' instruction returns the difference of its two operands.
3787 Note that the '``fsub``' instruction is used to represent the '``fneg``'
3788 instruction present in most other intermediate representations.
3793 The two arguments to the '``fsub``' instruction must be :ref:`floating
3794 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3795 Both arguments must have identical types.
3800 The value produced is the floating point difference of the two operands.
3801 This instruction can also take any number of :ref:`fast-math
3802 flags <fastmath>`, which are optimization hints to enable otherwise
3803 unsafe floating point optimizations:
3808 .. code-block:: llvm
3810 <result> = fsub float 4.0, %var ; yields {float}:result = 4.0 - %var
3811 <result> = fsub float -0.0, %val ; yields {float}:result = -%var
3813 '``mul``' Instruction
3814 ^^^^^^^^^^^^^^^^^^^^^
3821 <result> = mul <ty> <op1>, <op2> ; yields {ty}:result
3822 <result> = mul nuw <ty> <op1>, <op2> ; yields {ty}:result
3823 <result> = mul nsw <ty> <op1>, <op2> ; yields {ty}:result
3824 <result> = mul nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3829 The '``mul``' instruction returns the product of its two operands.
3834 The two arguments to the '``mul``' instruction must be
3835 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3836 arguments must have identical types.
3841 The value produced is the integer product of the two operands.
3843 If the result of the multiplication has unsigned overflow, the result
3844 returned is the mathematical result modulo 2\ :sup:`n`\ , where n is the
3845 bit width of the result.
3847 Because LLVM integers use a two's complement representation, and the
3848 result is the same width as the operands, this instruction returns the
3849 correct result for both signed and unsigned integers. If a full product
3850 (e.g. ``i32`` * ``i32`` -> ``i64``) is needed, the operands should be
3851 sign-extended or zero-extended as appropriate to the width of the full
3854 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3855 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3856 result value of the ``mul`` is a :ref:`poison value <poisonvalues>` if
3857 unsigned and/or signed overflow, respectively, occurs.
3862 .. code-block:: llvm
3864 <result> = mul i32 4, %var ; yields {i32}:result = 4 * %var
3868 '``fmul``' Instruction
3869 ^^^^^^^^^^^^^^^^^^^^^^
3876 <result> = fmul [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3881 The '``fmul``' instruction returns the product of its two operands.
3886 The two arguments to the '``fmul``' instruction must be :ref:`floating
3887 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3888 Both arguments must have identical types.
3893 The value produced is the floating point product of the two operands.
3894 This instruction can also take any number of :ref:`fast-math
3895 flags <fastmath>`, which are optimization hints to enable otherwise
3896 unsafe floating point optimizations:
3901 .. code-block:: llvm
3903 <result> = fmul float 4.0, %var ; yields {float}:result = 4.0 * %var
3905 '``udiv``' Instruction
3906 ^^^^^^^^^^^^^^^^^^^^^^
3913 <result> = udiv <ty> <op1>, <op2> ; yields {ty}:result
3914 <result> = udiv exact <ty> <op1>, <op2> ; yields {ty}:result
3919 The '``udiv``' instruction returns the quotient of its two operands.
3924 The two arguments to the '``udiv``' instruction must be
3925 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3926 arguments must have identical types.
3931 The value produced is the unsigned integer quotient of the two operands.
3933 Note that unsigned integer division and signed integer division are
3934 distinct operations; for signed integer division, use '``sdiv``'.
3936 Division by zero leads to undefined behavior.
3938 If the ``exact`` keyword is present, the result value of the ``udiv`` is
3939 a :ref:`poison value <poisonvalues>` if %op1 is not a multiple of %op2 (as
3940 such, "((a udiv exact b) mul b) == a").
3945 .. code-block:: llvm
3947 <result> = udiv i32 4, %var ; yields {i32}:result = 4 / %var
3949 '``sdiv``' Instruction
3950 ^^^^^^^^^^^^^^^^^^^^^^
3957 <result> = sdiv <ty> <op1>, <op2> ; yields {ty}:result
3958 <result> = sdiv exact <ty> <op1>, <op2> ; yields {ty}:result
3963 The '``sdiv``' instruction returns the quotient of its two operands.
3968 The two arguments to the '``sdiv``' instruction must be
3969 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3970 arguments must have identical types.
3975 The value produced is the signed integer quotient of the two operands
3976 rounded towards zero.
3978 Note that signed integer division and unsigned integer division are
3979 distinct operations; for unsigned integer division, use '``udiv``'.
3981 Division by zero leads to undefined behavior. Overflow also leads to
3982 undefined behavior; this is a rare case, but can occur, for example, by
3983 doing a 32-bit division of -2147483648 by -1.
3985 If the ``exact`` keyword is present, the result value of the ``sdiv`` is
3986 a :ref:`poison value <poisonvalues>` if the result would be rounded.
3991 .. code-block:: llvm
3993 <result> = sdiv i32 4, %var ; yields {i32}:result = 4 / %var
3997 '``fdiv``' Instruction
3998 ^^^^^^^^^^^^^^^^^^^^^^
4005 <result> = fdiv [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
4010 The '``fdiv``' instruction returns the quotient of its two operands.
4015 The two arguments to the '``fdiv``' instruction must be :ref:`floating
4016 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4017 Both arguments must have identical types.
4022 The value produced is the floating point quotient of the two operands.
4023 This instruction can also take any number of :ref:`fast-math
4024 flags <fastmath>`, which are optimization hints to enable otherwise
4025 unsafe floating point optimizations:
4030 .. code-block:: llvm
4032 <result> = fdiv float 4.0, %var ; yields {float}:result = 4.0 / %var
4034 '``urem``' Instruction
4035 ^^^^^^^^^^^^^^^^^^^^^^
4042 <result> = urem <ty> <op1>, <op2> ; yields {ty}:result
4047 The '``urem``' instruction returns the remainder from the unsigned
4048 division of its two arguments.
4053 The two arguments to the '``urem``' instruction must be
4054 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4055 arguments must have identical types.
4060 This instruction returns the unsigned integer *remainder* of a division.
4061 This instruction always performs an unsigned division to get the
4064 Note that unsigned integer remainder and signed integer remainder are
4065 distinct operations; for signed integer remainder, use '``srem``'.
4067 Taking the remainder of a division by zero leads to undefined behavior.
4072 .. code-block:: llvm
4074 <result> = urem i32 4, %var ; yields {i32}:result = 4 % %var
4076 '``srem``' Instruction
4077 ^^^^^^^^^^^^^^^^^^^^^^
4084 <result> = srem <ty> <op1>, <op2> ; yields {ty}:result
4089 The '``srem``' instruction returns the remainder from the signed
4090 division of its two operands. This instruction can also take
4091 :ref:`vector <t_vector>` versions of the values in which case the elements
4097 The two arguments to the '``srem``' instruction must be
4098 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4099 arguments must have identical types.
4104 This instruction returns the *remainder* of a division (where the result
4105 is either zero or has the same sign as the dividend, ``op1``), not the
4106 *modulo* operator (where the result is either zero or has the same sign
4107 as the divisor, ``op2``) of a value. For more information about the
4108 difference, see `The Math
4109 Forum <http://mathforum.org/dr.math/problems/anne.4.28.99.html>`_. For a
4110 table of how this is implemented in various languages, please see
4112 operation <http://en.wikipedia.org/wiki/Modulo_operation>`_.
4114 Note that signed integer remainder and unsigned integer remainder are
4115 distinct operations; for unsigned integer remainder, use '``urem``'.
4117 Taking the remainder of a division by zero leads to undefined behavior.
4118 Overflow also leads to undefined behavior; this is a rare case, but can
4119 occur, for example, by taking the remainder of a 32-bit division of
4120 -2147483648 by -1. (The remainder doesn't actually overflow, but this
4121 rule lets srem be implemented using instructions that return both the
4122 result of the division and the remainder.)
4127 .. code-block:: llvm
4129 <result> = srem i32 4, %var ; yields {i32}:result = 4 % %var
4133 '``frem``' Instruction
4134 ^^^^^^^^^^^^^^^^^^^^^^
4141 <result> = frem [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
4146 The '``frem``' instruction returns the remainder from the division of
4152 The two arguments to the '``frem``' instruction must be :ref:`floating
4153 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4154 Both arguments must have identical types.
4159 This instruction returns the *remainder* of a division. The remainder
4160 has the same sign as the dividend. This instruction can also take any
4161 number of :ref:`fast-math flags <fastmath>`, which are optimization hints
4162 to enable otherwise unsafe floating point optimizations:
4167 .. code-block:: llvm
4169 <result> = frem float 4.0, %var ; yields {float}:result = 4.0 % %var
4173 Bitwise Binary Operations
4174 -------------------------
4176 Bitwise binary operators are used to do various forms of bit-twiddling
4177 in a program. They are generally very efficient instructions and can
4178 commonly be strength reduced from other instructions. They require two
4179 operands of the same type, execute an operation on them, and produce a
4180 single value. The resulting value is the same type as its operands.
4182 '``shl``' Instruction
4183 ^^^^^^^^^^^^^^^^^^^^^
4190 <result> = shl <ty> <op1>, <op2> ; yields {ty}:result
4191 <result> = shl nuw <ty> <op1>, <op2> ; yields {ty}:result
4192 <result> = shl nsw <ty> <op1>, <op2> ; yields {ty}:result
4193 <result> = shl nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
4198 The '``shl``' instruction returns the first operand shifted to the left
4199 a specified number of bits.
4204 Both arguments to the '``shl``' instruction must be the same
4205 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4206 '``op2``' is treated as an unsigned value.
4211 The value produced is ``op1`` \* 2\ :sup:`op2` mod 2\ :sup:`n`,
4212 where ``n`` is the width of the result. If ``op2`` is (statically or
4213 dynamically) negative or equal to or larger than the number of bits in
4214 ``op1``, the result is undefined. If the arguments are vectors, each
4215 vector element of ``op1`` is shifted by the corresponding shift amount
4218 If the ``nuw`` keyword is present, then the shift produces a :ref:`poison
4219 value <poisonvalues>` if it shifts out any non-zero bits. If the
4220 ``nsw`` keyword is present, then the shift produces a :ref:`poison
4221 value <poisonvalues>` if it shifts out any bits that disagree with the
4222 resultant sign bit. As such, NUW/NSW have the same semantics as they
4223 would if the shift were expressed as a mul instruction with the same
4224 nsw/nuw bits in (mul %op1, (shl 1, %op2)).
4229 .. code-block:: llvm
4231 <result> = shl i32 4, %var ; yields {i32}: 4 << %var
4232 <result> = shl i32 4, 2 ; yields {i32}: 16
4233 <result> = shl i32 1, 10 ; yields {i32}: 1024
4234 <result> = shl i32 1, 32 ; undefined
4235 <result> = shl <2 x i32> < i32 1, i32 1>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 2, i32 4>
4237 '``lshr``' Instruction
4238 ^^^^^^^^^^^^^^^^^^^^^^
4245 <result> = lshr <ty> <op1>, <op2> ; yields {ty}:result
4246 <result> = lshr exact <ty> <op1>, <op2> ; yields {ty}:result
4251 The '``lshr``' instruction (logical shift right) returns the first
4252 operand shifted to the right a specified number of bits with zero fill.
4257 Both arguments to the '``lshr``' instruction must be the same
4258 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4259 '``op2``' is treated as an unsigned value.
4264 This instruction always performs a logical shift right operation. The
4265 most significant bits of the result will be filled with zero bits after
4266 the shift. If ``op2`` is (statically or dynamically) equal to or larger
4267 than the number of bits in ``op1``, the result is undefined. If the
4268 arguments are vectors, each vector element of ``op1`` is shifted by the
4269 corresponding shift amount in ``op2``.
4271 If the ``exact`` keyword is present, the result value of the ``lshr`` is
4272 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
4278 .. code-block:: llvm
4280 <result> = lshr i32 4, 1 ; yields {i32}:result = 2
4281 <result> = lshr i32 4, 2 ; yields {i32}:result = 1
4282 <result> = lshr i8 4, 3 ; yields {i8}:result = 0
4283 <result> = lshr i8 -2, 1 ; yields {i8}:result = 0x7F
4284 <result> = lshr i32 1, 32 ; undefined
4285 <result> = lshr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 0x7FFFFFFF, i32 1>
4287 '``ashr``' Instruction
4288 ^^^^^^^^^^^^^^^^^^^^^^
4295 <result> = ashr <ty> <op1>, <op2> ; yields {ty}:result
4296 <result> = ashr exact <ty> <op1>, <op2> ; yields {ty}:result
4301 The '``ashr``' instruction (arithmetic shift right) returns the first
4302 operand shifted to the right a specified number of bits with sign
4308 Both arguments to the '``ashr``' instruction must be the same
4309 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4310 '``op2``' is treated as an unsigned value.
4315 This instruction always performs an arithmetic shift right operation,
4316 The most significant bits of the result will be filled with the sign bit
4317 of ``op1``. If ``op2`` is (statically or dynamically) equal to or larger
4318 than the number of bits in ``op1``, the result is undefined. If the
4319 arguments are vectors, each vector element of ``op1`` is shifted by the
4320 corresponding shift amount in ``op2``.
4322 If the ``exact`` keyword is present, the result value of the ``ashr`` is
4323 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
4329 .. code-block:: llvm
4331 <result> = ashr i32 4, 1 ; yields {i32}:result = 2
4332 <result> = ashr i32 4, 2 ; yields {i32}:result = 1
4333 <result> = ashr i8 4, 3 ; yields {i8}:result = 0
4334 <result> = ashr i8 -2, 1 ; yields {i8}:result = -1
4335 <result> = ashr i32 1, 32 ; undefined
4336 <result> = ashr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 3> ; yields: result=<2 x i32> < i32 -1, i32 0>
4338 '``and``' Instruction
4339 ^^^^^^^^^^^^^^^^^^^^^
4346 <result> = and <ty> <op1>, <op2> ; yields {ty}:result
4351 The '``and``' instruction returns the bitwise logical and of its two
4357 The two arguments to the '``and``' instruction must be
4358 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4359 arguments must have identical types.
4364 The truth table used for the '``and``' instruction is:
4381 .. code-block:: llvm
4383 <result> = and i32 4, %var ; yields {i32}:result = 4 & %var
4384 <result> = and i32 15, 40 ; yields {i32}:result = 8
4385 <result> = and i32 4, 8 ; yields {i32}:result = 0
4387 '``or``' Instruction
4388 ^^^^^^^^^^^^^^^^^^^^
4395 <result> = or <ty> <op1>, <op2> ; yields {ty}:result
4400 The '``or``' instruction returns the bitwise logical inclusive or of its
4406 The two arguments to the '``or``' instruction must be
4407 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4408 arguments must have identical types.
4413 The truth table used for the '``or``' instruction is:
4432 <result> = or i32 4, %var ; yields {i32}:result = 4 | %var
4433 <result> = or i32 15, 40 ; yields {i32}:result = 47
4434 <result> = or i32 4, 8 ; yields {i32}:result = 12
4436 '``xor``' Instruction
4437 ^^^^^^^^^^^^^^^^^^^^^
4444 <result> = xor <ty> <op1>, <op2> ; yields {ty}:result
4449 The '``xor``' instruction returns the bitwise logical exclusive or of
4450 its two operands. The ``xor`` is used to implement the "one's
4451 complement" operation, which is the "~" operator in C.
4456 The two arguments to the '``xor``' instruction must be
4457 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4458 arguments must have identical types.
4463 The truth table used for the '``xor``' instruction is:
4480 .. code-block:: llvm
4482 <result> = xor i32 4, %var ; yields {i32}:result = 4 ^ %var
4483 <result> = xor i32 15, 40 ; yields {i32}:result = 39
4484 <result> = xor i32 4, 8 ; yields {i32}:result = 12
4485 <result> = xor i32 %V, -1 ; yields {i32}:result = ~%V
4490 LLVM supports several instructions to represent vector operations in a
4491 target-independent manner. These instructions cover the element-access
4492 and vector-specific operations needed to process vectors effectively.
4493 While LLVM does directly support these vector operations, many
4494 sophisticated algorithms will want to use target-specific intrinsics to
4495 take full advantage of a specific target.
4497 .. _i_extractelement:
4499 '``extractelement``' Instruction
4500 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4507 <result> = extractelement <n x <ty>> <val>, <ty2> <idx> ; yields <ty>
4512 The '``extractelement``' instruction extracts a single scalar element
4513 from a vector at a specified index.
4518 The first operand of an '``extractelement``' instruction is a value of
4519 :ref:`vector <t_vector>` type. The second operand is an index indicating
4520 the position from which to extract the element. The index may be a
4521 variable of any integer type.
4526 The result is a scalar of the same type as the element type of ``val``.
4527 Its value is the value at position ``idx`` of ``val``. If ``idx``
4528 exceeds the length of ``val``, the results are undefined.
4533 .. code-block:: llvm
4535 <result> = extractelement <4 x i32> %vec, i32 0 ; yields i32
4537 .. _i_insertelement:
4539 '``insertelement``' Instruction
4540 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4547 <result> = insertelement <n x <ty>> <val>, <ty> <elt>, <ty2> <idx> ; yields <n x <ty>>
4552 The '``insertelement``' instruction inserts a scalar element into a
4553 vector at a specified index.
4558 The first operand of an '``insertelement``' instruction is a value of
4559 :ref:`vector <t_vector>` type. The second operand is a scalar value whose
4560 type must equal the element type of the first operand. The third operand
4561 is an index indicating the position at which to insert the value. The
4562 index may be a variable of any integer type.
4567 The result is a vector of the same type as ``val``. Its element values
4568 are those of ``val`` except at position ``idx``, where it gets the value
4569 ``elt``. If ``idx`` exceeds the length of ``val``, the results are
4575 .. code-block:: llvm
4577 <result> = insertelement <4 x i32> %vec, i32 1, i32 0 ; yields <4 x i32>
4579 .. _i_shufflevector:
4581 '``shufflevector``' Instruction
4582 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4589 <result> = shufflevector <n x <ty>> <v1>, <n x <ty>> <v2>, <m x i32> <mask> ; yields <m x <ty>>
4594 The '``shufflevector``' instruction constructs a permutation of elements
4595 from two input vectors, returning a vector with the same element type as
4596 the input and length that is the same as the shuffle mask.
4601 The first two operands of a '``shufflevector``' instruction are vectors
4602 with the same type. The third argument is a shuffle mask whose element
4603 type is always 'i32'. The result of the instruction is a vector whose
4604 length is the same as the shuffle mask and whose element type is the
4605 same as the element type of the first two operands.
4607 The shuffle mask operand is required to be a constant vector with either
4608 constant integer or undef values.
4613 The elements of the two input vectors are numbered from left to right
4614 across both of the vectors. The shuffle mask operand specifies, for each
4615 element of the result vector, which element of the two input vectors the
4616 result element gets. The element selector may be undef (meaning "don't
4617 care") and the second operand may be undef if performing a shuffle from
4623 .. code-block:: llvm
4625 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4626 <4 x i32> <i32 0, i32 4, i32 1, i32 5> ; yields <4 x i32>
4627 <result> = shufflevector <4 x i32> %v1, <4 x i32> undef,
4628 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32> - Identity shuffle.
4629 <result> = shufflevector <8 x i32> %v1, <8 x i32> undef,
4630 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32>
4631 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4632 <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7 > ; yields <8 x i32>
4634 Aggregate Operations
4635 --------------------
4637 LLVM supports several instructions for working with
4638 :ref:`aggregate <t_aggregate>` values.
4642 '``extractvalue``' Instruction
4643 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4650 <result> = extractvalue <aggregate type> <val>, <idx>{, <idx>}*
4655 The '``extractvalue``' instruction extracts the value of a member field
4656 from an :ref:`aggregate <t_aggregate>` value.
4661 The first operand of an '``extractvalue``' instruction is a value of
4662 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The operands are
4663 constant indices to specify which value to extract in a similar manner
4664 as indices in a '``getelementptr``' instruction.
4666 The major differences to ``getelementptr`` indexing are:
4668 - Since the value being indexed is not a pointer, the first index is
4669 omitted and assumed to be zero.
4670 - At least one index must be specified.
4671 - Not only struct indices but also array indices must be in bounds.
4676 The result is the value at the position in the aggregate specified by
4682 .. code-block:: llvm
4684 <result> = extractvalue {i32, float} %agg, 0 ; yields i32
4688 '``insertvalue``' Instruction
4689 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4696 <result> = insertvalue <aggregate type> <val>, <ty> <elt>, <idx>{, <idx>}* ; yields <aggregate type>
4701 The '``insertvalue``' instruction inserts a value into a member field in
4702 an :ref:`aggregate <t_aggregate>` value.
4707 The first operand of an '``insertvalue``' instruction is a value of
4708 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The second operand is
4709 a first-class value to insert. The following operands are constant
4710 indices indicating the position at which to insert the value in a
4711 similar manner as indices in a '``extractvalue``' instruction. The value
4712 to insert must have the same type as the value identified by the
4718 The result is an aggregate of the same type as ``val``. Its value is
4719 that of ``val`` except that the value at the position specified by the
4720 indices is that of ``elt``.
4725 .. code-block:: llvm
4727 %agg1 = insertvalue {i32, float} undef, i32 1, 0 ; yields {i32 1, float undef}
4728 %agg2 = insertvalue {i32, float} %agg1, float %val, 1 ; yields {i32 1, float %val}
4729 %agg3 = insertvalue {i32, {float}} %agg1, float %val, 1, 0 ; yields {i32 1, float %val}
4733 Memory Access and Addressing Operations
4734 ---------------------------------------
4736 A key design point of an SSA-based representation is how it represents
4737 memory. In LLVM, no memory locations are in SSA form, which makes things
4738 very simple. This section describes how to read, write, and allocate
4743 '``alloca``' Instruction
4744 ^^^^^^^^^^^^^^^^^^^^^^^^
4751 <result> = alloca [inalloca] <type> [, <ty> <NumElements>] [, align <alignment>] ; yields {type*}:result
4756 The '``alloca``' instruction allocates memory on the stack frame of the
4757 currently executing function, to be automatically released when this
4758 function returns to its caller. The object is always allocated in the
4759 generic address space (address space zero).
4764 The '``alloca``' instruction allocates ``sizeof(<type>)*NumElements``
4765 bytes of memory on the runtime stack, returning a pointer of the
4766 appropriate type to the program. If "NumElements" is specified, it is
4767 the number of elements allocated, otherwise "NumElements" is defaulted
4768 to be one. If a constant alignment is specified, the value result of the
4769 allocation is guaranteed to be aligned to at least that boundary. If not
4770 specified, or if zero, the target can choose to align the allocation on
4771 any convenient boundary compatible with the type.
4773 '``type``' may be any sized type.
4778 Memory is allocated; a pointer is returned. The operation is undefined
4779 if there is insufficient stack space for the allocation. '``alloca``'d
4780 memory is automatically released when the function returns. The
4781 '``alloca``' instruction is commonly used to represent automatic
4782 variables that must have an address available. When the function returns
4783 (either with the ``ret`` or ``resume`` instructions), the memory is
4784 reclaimed. Allocating zero bytes is legal, but the result is undefined.
4785 The order in which memory is allocated (ie., which way the stack grows)
4791 .. code-block:: llvm
4793 %ptr = alloca i32 ; yields {i32*}:ptr
4794 %ptr = alloca i32, i32 4 ; yields {i32*}:ptr
4795 %ptr = alloca i32, i32 4, align 1024 ; yields {i32*}:ptr
4796 %ptr = alloca i32, align 1024 ; yields {i32*}:ptr
4800 '``load``' Instruction
4801 ^^^^^^^^^^^^^^^^^^^^^^
4808 <result> = load [volatile] <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>][, !invariant.load !<index>]
4809 <result> = load atomic [volatile] <ty>* <pointer> [singlethread] <ordering>, align <alignment>
4810 !<index> = !{ i32 1 }
4815 The '``load``' instruction is used to read from memory.
4820 The argument to the ``load`` instruction specifies the memory address
4821 from which to load. The pointer must point to a :ref:`first
4822 class <t_firstclass>` type. If the ``load`` is marked as ``volatile``,
4823 then the optimizer is not allowed to modify the number or order of
4824 execution of this ``load`` with other :ref:`volatile
4825 operations <volatile>`.
4827 If the ``load`` is marked as ``atomic``, it takes an extra
4828 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
4829 ``release`` and ``acq_rel`` orderings are not valid on ``load``
4830 instructions. Atomic loads produce :ref:`defined <memmodel>` results
4831 when they may see multiple atomic stores. The type of the pointee must
4832 be an integer type whose bit width is a power of two greater than or
4833 equal to eight and less than or equal to a target-specific size limit.
4834 ``align`` must be explicitly specified on atomic loads, and the load has
4835 undefined behavior if the alignment is not set to a value which is at
4836 least the size in bytes of the pointee. ``!nontemporal`` does not have
4837 any defined semantics for atomic loads.
4839 The optional constant ``align`` argument specifies the alignment of the
4840 operation (that is, the alignment of the memory address). A value of 0
4841 or an omitted ``align`` argument means that the operation has the ABI
4842 alignment for the target. It is the responsibility of the code emitter
4843 to ensure that the alignment information is correct. Overestimating the
4844 alignment results in undefined behavior. Underestimating the alignment
4845 may produce less efficient code. An alignment of 1 is always safe.
4847 The optional ``!nontemporal`` metadata must reference a single
4848 metadata name ``<index>`` corresponding to a metadata node with one
4849 ``i32`` entry of value 1. The existence of the ``!nontemporal``
4850 metadata on the instruction tells the optimizer and code generator
4851 that this load is not expected to be reused in the cache. The code
4852 generator may select special instructions to save cache bandwidth, such
4853 as the ``MOVNT`` instruction on x86.
4855 The optional ``!invariant.load`` metadata must reference a single
4856 metadata name ``<index>`` corresponding to a metadata node with no
4857 entries. The existence of the ``!invariant.load`` metadata on the
4858 instruction tells the optimizer and code generator that this load
4859 address points to memory which does not change value during program
4860 execution. The optimizer may then move this load around, for example, by
4861 hoisting it out of loops using loop invariant code motion.
4866 The location of memory pointed to is loaded. If the value being loaded
4867 is of scalar type then the number of bytes read does not exceed the
4868 minimum number of bytes needed to hold all bits of the type. For
4869 example, loading an ``i24`` reads at most three bytes. When loading a
4870 value of a type like ``i20`` with a size that is not an integral number
4871 of bytes, the result is undefined if the value was not originally
4872 written using a store of the same type.
4877 .. code-block:: llvm
4879 %ptr = alloca i32 ; yields {i32*}:ptr
4880 store i32 3, i32* %ptr ; yields {void}
4881 %val = load i32* %ptr ; yields {i32}:val = i32 3
4885 '``store``' Instruction
4886 ^^^^^^^^^^^^^^^^^^^^^^^
4893 store [volatile] <ty> <value>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>] ; yields {void}
4894 store atomic [volatile] <ty> <value>, <ty>* <pointer> [singlethread] <ordering>, align <alignment> ; yields {void}
4899 The '``store``' instruction is used to write to memory.
4904 There are two arguments to the ``store`` instruction: a value to store
4905 and an address at which to store it. The type of the ``<pointer>``
4906 operand must be a pointer to the :ref:`first class <t_firstclass>` type of
4907 the ``<value>`` operand. If the ``store`` is marked as ``volatile``,
4908 then the optimizer is not allowed to modify the number or order of
4909 execution of this ``store`` with other :ref:`volatile
4910 operations <volatile>`.
4912 If the ``store`` is marked as ``atomic``, it takes an extra
4913 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
4914 ``acquire`` and ``acq_rel`` orderings aren't valid on ``store``
4915 instructions. Atomic loads produce :ref:`defined <memmodel>` results
4916 when they may see multiple atomic stores. The type of the pointee must
4917 be an integer type whose bit width is a power of two greater than or
4918 equal to eight and less than or equal to a target-specific size limit.
4919 ``align`` must be explicitly specified on atomic stores, and the store
4920 has undefined behavior if the alignment is not set to a value which is
4921 at least the size in bytes of the pointee. ``!nontemporal`` does not
4922 have any defined semantics for atomic stores.
4924 The optional constant ``align`` argument specifies the alignment of the
4925 operation (that is, the alignment of the memory address). A value of 0
4926 or an omitted ``align`` argument means that the operation has the ABI
4927 alignment for the target. It is the responsibility of the code emitter
4928 to ensure that the alignment information is correct. Overestimating the
4929 alignment results in undefined behavior. Underestimating the
4930 alignment may produce less efficient code. An alignment of 1 is always
4933 The optional ``!nontemporal`` metadata must reference a single metadata
4934 name ``<index>`` corresponding to a metadata node with one ``i32`` entry of
4935 value 1. The existence of the ``!nontemporal`` metadata on the instruction
4936 tells the optimizer and code generator that this load is not expected to
4937 be reused in the cache. The code generator may select special
4938 instructions to save cache bandwidth, such as the MOVNT instruction on
4944 The contents of memory are updated to contain ``<value>`` at the
4945 location specified by the ``<pointer>`` operand. If ``<value>`` is
4946 of scalar type then the number of bytes written does not exceed the
4947 minimum number of bytes needed to hold all bits of the type. For
4948 example, storing an ``i24`` writes at most three bytes. When writing a
4949 value of a type like ``i20`` with a size that is not an integral number
4950 of bytes, it is unspecified what happens to the extra bits that do not
4951 belong to the type, but they will typically be overwritten.
4956 .. code-block:: llvm
4958 %ptr = alloca i32 ; yields {i32*}:ptr
4959 store i32 3, i32* %ptr ; yields {void}
4960 %val = load i32* %ptr ; yields {i32}:val = i32 3
4964 '``fence``' Instruction
4965 ^^^^^^^^^^^^^^^^^^^^^^^
4972 fence [singlethread] <ordering> ; yields {void}
4977 The '``fence``' instruction is used to introduce happens-before edges
4983 '``fence``' instructions take an :ref:`ordering <ordering>` argument which
4984 defines what *synchronizes-with* edges they add. They can only be given
4985 ``acquire``, ``release``, ``acq_rel``, and ``seq_cst`` orderings.
4990 A fence A which has (at least) ``release`` ordering semantics
4991 *synchronizes with* a fence B with (at least) ``acquire`` ordering
4992 semantics if and only if there exist atomic operations X and Y, both
4993 operating on some atomic object M, such that A is sequenced before X, X
4994 modifies M (either directly or through some side effect of a sequence
4995 headed by X), Y is sequenced before B, and Y observes M. This provides a
4996 *happens-before* dependency between A and B. Rather than an explicit
4997 ``fence``, one (but not both) of the atomic operations X or Y might
4998 provide a ``release`` or ``acquire`` (resp.) ordering constraint and
4999 still *synchronize-with* the explicit ``fence`` and establish the
5000 *happens-before* edge.
5002 A ``fence`` which has ``seq_cst`` ordering, in addition to having both
5003 ``acquire`` and ``release`` semantics specified above, participates in
5004 the global program order of other ``seq_cst`` operations and/or fences.
5006 The optional ":ref:`singlethread <singlethread>`" argument specifies
5007 that the fence only synchronizes with other fences in the same thread.
5008 (This is useful for interacting with signal handlers.)
5013 .. code-block:: llvm
5015 fence acquire ; yields {void}
5016 fence singlethread seq_cst ; yields {void}
5020 '``cmpxchg``' Instruction
5021 ^^^^^^^^^^^^^^^^^^^^^^^^^
5028 cmpxchg [volatile] <ty>* <pointer>, <ty> <cmp>, <ty> <new> [singlethread] <success ordering> <failure ordering> ; yields {ty}
5033 The '``cmpxchg``' instruction is used to atomically modify memory. It
5034 loads a value in memory and compares it to a given value. If they are
5035 equal, it stores a new value into the memory.
5040 There are three arguments to the '``cmpxchg``' instruction: an address
5041 to operate on, a value to compare to the value currently be at that
5042 address, and a new value to place at that address if the compared values
5043 are equal. The type of '<cmp>' must be an integer type whose bit width
5044 is a power of two greater than or equal to eight and less than or equal
5045 to a target-specific size limit. '<cmp>' and '<new>' must have the same
5046 type, and the type of '<pointer>' must be a pointer to that type. If the
5047 ``cmpxchg`` is marked as ``volatile``, then the optimizer is not allowed
5048 to modify the number or order of execution of this ``cmpxchg`` with
5049 other :ref:`volatile operations <volatile>`.
5051 The success and failure :ref:`ordering <ordering>` arguments specify how this
5052 ``cmpxchg`` synchronizes with other atomic operations. The both ordering
5053 parameters must be at least ``monotonic``, the ordering constraint on failure
5054 must be no stronger than that on success, and the failure ordering cannot be
5055 either ``release`` or ``acq_rel``.
5057 The optional "``singlethread``" argument declares that the ``cmpxchg``
5058 is only atomic with respect to code (usually signal handlers) running in
5059 the same thread as the ``cmpxchg``. Otherwise the cmpxchg is atomic with
5060 respect to all other code in the system.
5062 The pointer passed into cmpxchg must have alignment greater than or
5063 equal to the size in memory of the operand.
5068 The contents of memory at the location specified by the '``<pointer>``'
5069 operand is read and compared to '``<cmp>``'; if the read value is the
5070 equal, '``<new>``' is written. The original value at the location is
5073 A successful ``cmpxchg`` is a read-modify-write instruction for the purpose of
5074 identifying release sequences. A failed ``cmpxchg`` is equivalent to an atomic
5075 load with an ordering parameter determined the second ordering parameter.
5080 .. code-block:: llvm
5083 %orig = atomic load i32* %ptr unordered ; yields {i32}
5087 %cmp = phi i32 [ %orig, %entry ], [%old, %loop]
5088 %squared = mul i32 %cmp, %cmp
5089 %old = cmpxchg i32* %ptr, i32 %cmp, i32 %squared acq_rel monotonic ; yields {i32}
5090 %success = icmp eq i32 %cmp, %old
5091 br i1 %success, label %done, label %loop
5098 '``atomicrmw``' Instruction
5099 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
5106 atomicrmw [volatile] <operation> <ty>* <pointer>, <ty> <value> [singlethread] <ordering> ; yields {ty}
5111 The '``atomicrmw``' instruction is used to atomically modify memory.
5116 There are three arguments to the '``atomicrmw``' instruction: an
5117 operation to apply, an address whose value to modify, an argument to the
5118 operation. The operation must be one of the following keywords:
5132 The type of '<value>' must be an integer type whose bit width is a power
5133 of two greater than or equal to eight and less than or equal to a
5134 target-specific size limit. The type of the '``<pointer>``' operand must
5135 be a pointer to that type. If the ``atomicrmw`` is marked as
5136 ``volatile``, then the optimizer is not allowed to modify the number or
5137 order of execution of this ``atomicrmw`` with other :ref:`volatile
5138 operations <volatile>`.
5143 The contents of memory at the location specified by the '``<pointer>``'
5144 operand are atomically read, modified, and written back. The original
5145 value at the location is returned. The modification is specified by the
5148 - xchg: ``*ptr = val``
5149 - add: ``*ptr = *ptr + val``
5150 - sub: ``*ptr = *ptr - val``
5151 - and: ``*ptr = *ptr & val``
5152 - nand: ``*ptr = ~(*ptr & val)``
5153 - or: ``*ptr = *ptr | val``
5154 - xor: ``*ptr = *ptr ^ val``
5155 - max: ``*ptr = *ptr > val ? *ptr : val`` (using a signed comparison)
5156 - min: ``*ptr = *ptr < val ? *ptr : val`` (using a signed comparison)
5157 - umax: ``*ptr = *ptr > val ? *ptr : val`` (using an unsigned
5159 - umin: ``*ptr = *ptr < val ? *ptr : val`` (using an unsigned
5165 .. code-block:: llvm
5167 %old = atomicrmw add i32* %ptr, i32 1 acquire ; yields {i32}
5169 .. _i_getelementptr:
5171 '``getelementptr``' Instruction
5172 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5179 <result> = getelementptr <pty>* <ptrval>{, <ty> <idx>}*
5180 <result> = getelementptr inbounds <pty>* <ptrval>{, <ty> <idx>}*
5181 <result> = getelementptr <ptr vector> ptrval, <vector index type> idx
5186 The '``getelementptr``' instruction is used to get the address of a
5187 subelement of an :ref:`aggregate <t_aggregate>` data structure. It performs
5188 address calculation only and does not access memory.
5193 The first argument is always a pointer or a vector of pointers, and
5194 forms the basis of the calculation. The remaining arguments are indices
5195 that indicate which of the elements of the aggregate object are indexed.
5196 The interpretation of each index is dependent on the type being indexed
5197 into. The first index always indexes the pointer value given as the
5198 first argument, the second index indexes a value of the type pointed to
5199 (not necessarily the value directly pointed to, since the first index
5200 can be non-zero), etc. The first type indexed into must be a pointer
5201 value, subsequent types can be arrays, vectors, and structs. Note that
5202 subsequent types being indexed into can never be pointers, since that
5203 would require loading the pointer before continuing calculation.
5205 The type of each index argument depends on the type it is indexing into.
5206 When indexing into a (optionally packed) structure, only ``i32`` integer
5207 **constants** are allowed (when using a vector of indices they must all
5208 be the **same** ``i32`` integer constant). When indexing into an array,
5209 pointer or vector, integers of any width are allowed, and they are not
5210 required to be constant. These integers are treated as signed values
5213 For example, let's consider a C code fragment and how it gets compiled
5229 int *foo(struct ST *s) {
5230 return &s[1].Z.B[5][13];
5233 The LLVM code generated by Clang is:
5235 .. code-block:: llvm
5237 %struct.RT = type { i8, [10 x [20 x i32]], i8 }
5238 %struct.ST = type { i32, double, %struct.RT }
5240 define i32* @foo(%struct.ST* %s) nounwind uwtable readnone optsize ssp {
5242 %arrayidx = getelementptr inbounds %struct.ST* %s, i64 1, i32 2, i32 1, i64 5, i64 13
5249 In the example above, the first index is indexing into the
5250 '``%struct.ST*``' type, which is a pointer, yielding a '``%struct.ST``'
5251 = '``{ i32, double, %struct.RT }``' type, a structure. The second index
5252 indexes into the third element of the structure, yielding a
5253 '``%struct.RT``' = '``{ i8 , [10 x [20 x i32]], i8 }``' type, another
5254 structure. The third index indexes into the second element of the
5255 structure, yielding a '``[10 x [20 x i32]]``' type, an array. The two
5256 dimensions of the array are subscripted into, yielding an '``i32``'
5257 type. The '``getelementptr``' instruction returns a pointer to this
5258 element, thus computing a value of '``i32*``' type.
5260 Note that it is perfectly legal to index partially through a structure,
5261 returning a pointer to an inner element. Because of this, the LLVM code
5262 for the given testcase is equivalent to:
5264 .. code-block:: llvm
5266 define i32* @foo(%struct.ST* %s) {
5267 %t1 = getelementptr %struct.ST* %s, i32 1 ; yields %struct.ST*:%t1
5268 %t2 = getelementptr %struct.ST* %t1, i32 0, i32 2 ; yields %struct.RT*:%t2
5269 %t3 = getelementptr %struct.RT* %t2, i32 0, i32 1 ; yields [10 x [20 x i32]]*:%t3
5270 %t4 = getelementptr [10 x [20 x i32]]* %t3, i32 0, i32 5 ; yields [20 x i32]*:%t4
5271 %t5 = getelementptr [20 x i32]* %t4, i32 0, i32 13 ; yields i32*:%t5
5275 If the ``inbounds`` keyword is present, the result value of the
5276 ``getelementptr`` is a :ref:`poison value <poisonvalues>` if the base
5277 pointer is not an *in bounds* address of an allocated object, or if any
5278 of the addresses that would be formed by successive addition of the
5279 offsets implied by the indices to the base address with infinitely
5280 precise signed arithmetic are not an *in bounds* address of that
5281 allocated object. The *in bounds* addresses for an allocated object are
5282 all the addresses that point into the object, plus the address one byte
5283 past the end. In cases where the base is a vector of pointers the
5284 ``inbounds`` keyword applies to each of the computations element-wise.
5286 If the ``inbounds`` keyword is not present, the offsets are added to the
5287 base address with silently-wrapping two's complement arithmetic. If the
5288 offsets have a different width from the pointer, they are sign-extended
5289 or truncated to the width of the pointer. The result value of the
5290 ``getelementptr`` may be outside the object pointed to by the base
5291 pointer. The result value may not necessarily be used to access memory
5292 though, even if it happens to point into allocated storage. See the
5293 :ref:`Pointer Aliasing Rules <pointeraliasing>` section for more
5296 The getelementptr instruction is often confusing. For some more insight
5297 into how it works, see :doc:`the getelementptr FAQ <GetElementPtr>`.
5302 .. code-block:: llvm
5304 ; yields [12 x i8]*:aptr
5305 %aptr = getelementptr {i32, [12 x i8]}* %saptr, i64 0, i32 1
5307 %vptr = getelementptr {i32, <2 x i8>}* %svptr, i64 0, i32 1, i32 1
5309 %eptr = getelementptr [12 x i8]* %aptr, i64 0, i32 1
5311 %iptr = getelementptr [10 x i32]* @arr, i16 0, i16 0
5313 In cases where the pointer argument is a vector of pointers, each index
5314 must be a vector with the same number of elements. For example:
5316 .. code-block:: llvm
5318 %A = getelementptr <4 x i8*> %ptrs, <4 x i64> %offsets,
5320 Conversion Operations
5321 ---------------------
5323 The instructions in this category are the conversion instructions
5324 (casting) which all take a single operand and a type. They perform
5325 various bit conversions on the operand.
5327 '``trunc .. to``' Instruction
5328 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5335 <result> = trunc <ty> <value> to <ty2> ; yields ty2
5340 The '``trunc``' instruction truncates its operand to the type ``ty2``.
5345 The '``trunc``' instruction takes a value to trunc, and a type to trunc
5346 it to. Both types must be of :ref:`integer <t_integer>` types, or vectors
5347 of the same number of integers. The bit size of the ``value`` must be
5348 larger than the bit size of the destination type, ``ty2``. Equal sized
5349 types are not allowed.
5354 The '``trunc``' instruction truncates the high order bits in ``value``
5355 and converts the remaining bits to ``ty2``. Since the source size must
5356 be larger than the destination size, ``trunc`` cannot be a *no-op cast*.
5357 It will always truncate bits.
5362 .. code-block:: llvm
5364 %X = trunc i32 257 to i8 ; yields i8:1
5365 %Y = trunc i32 123 to i1 ; yields i1:true
5366 %Z = trunc i32 122 to i1 ; yields i1:false
5367 %W = trunc <2 x i16> <i16 8, i16 7> to <2 x i8> ; yields <i8 8, i8 7>
5369 '``zext .. to``' Instruction
5370 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5377 <result> = zext <ty> <value> to <ty2> ; yields ty2
5382 The '``zext``' instruction zero extends its operand to type ``ty2``.
5387 The '``zext``' instruction takes a value to cast, and a type to cast it
5388 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
5389 the same number of integers. The bit size of the ``value`` must be
5390 smaller than the bit size of the destination type, ``ty2``.
5395 The ``zext`` fills the high order bits of the ``value`` with zero bits
5396 until it reaches the size of the destination type, ``ty2``.
5398 When zero extending from i1, the result will always be either 0 or 1.
5403 .. code-block:: llvm
5405 %X = zext i32 257 to i64 ; yields i64:257
5406 %Y = zext i1 true to i32 ; yields i32:1
5407 %Z = zext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5409 '``sext .. to``' Instruction
5410 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5417 <result> = sext <ty> <value> to <ty2> ; yields ty2
5422 The '``sext``' sign extends ``value`` to the type ``ty2``.
5427 The '``sext``' instruction takes a value to cast, and a type to cast it
5428 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
5429 the same number of integers. The bit size of the ``value`` must be
5430 smaller than the bit size of the destination type, ``ty2``.
5435 The '``sext``' instruction performs a sign extension by copying the sign
5436 bit (highest order bit) of the ``value`` until it reaches the bit size
5437 of the type ``ty2``.
5439 When sign extending from i1, the extension always results in -1 or 0.
5444 .. code-block:: llvm
5446 %X = sext i8 -1 to i16 ; yields i16 :65535
5447 %Y = sext i1 true to i32 ; yields i32:-1
5448 %Z = sext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5450 '``fptrunc .. to``' Instruction
5451 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5458 <result> = fptrunc <ty> <value> to <ty2> ; yields ty2
5463 The '``fptrunc``' instruction truncates ``value`` to type ``ty2``.
5468 The '``fptrunc``' instruction takes a :ref:`floating point <t_floating>`
5469 value to cast and a :ref:`floating point <t_floating>` type to cast it to.
5470 The size of ``value`` must be larger than the size of ``ty2``. This
5471 implies that ``fptrunc`` cannot be used to make a *no-op cast*.
5476 The '``fptrunc``' instruction truncates a ``value`` from a larger
5477 :ref:`floating point <t_floating>` type to a smaller :ref:`floating
5478 point <t_floating>` type. If the value cannot fit within the
5479 destination type, ``ty2``, then the results are undefined.
5484 .. code-block:: llvm
5486 %X = fptrunc double 123.0 to float ; yields float:123.0
5487 %Y = fptrunc double 1.0E+300 to float ; yields undefined
5489 '``fpext .. to``' Instruction
5490 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5497 <result> = fpext <ty> <value> to <ty2> ; yields ty2
5502 The '``fpext``' extends a floating point ``value`` to a larger floating
5508 The '``fpext``' instruction takes a :ref:`floating point <t_floating>`
5509 ``value`` to cast, and a :ref:`floating point <t_floating>` type to cast it
5510 to. The source type must be smaller than the destination type.
5515 The '``fpext``' instruction extends the ``value`` from a smaller
5516 :ref:`floating point <t_floating>` type to a larger :ref:`floating
5517 point <t_floating>` type. The ``fpext`` cannot be used to make a
5518 *no-op cast* because it always changes bits. Use ``bitcast`` to make a
5519 *no-op cast* for a floating point cast.
5524 .. code-block:: llvm
5526 %X = fpext float 3.125 to double ; yields double:3.125000e+00
5527 %Y = fpext double %X to fp128 ; yields fp128:0xL00000000000000004000900000000000
5529 '``fptoui .. to``' Instruction
5530 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5537 <result> = fptoui <ty> <value> to <ty2> ; yields ty2
5542 The '``fptoui``' converts a floating point ``value`` to its unsigned
5543 integer equivalent of type ``ty2``.
5548 The '``fptoui``' instruction takes a value to cast, which must be a
5549 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5550 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5551 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5552 type with the same number of elements as ``ty``
5557 The '``fptoui``' instruction converts its :ref:`floating
5558 point <t_floating>` operand into the nearest (rounding towards zero)
5559 unsigned integer value. If the value cannot fit in ``ty2``, the results
5565 .. code-block:: llvm
5567 %X = fptoui double 123.0 to i32 ; yields i32:123
5568 %Y = fptoui float 1.0E+300 to i1 ; yields undefined:1
5569 %Z = fptoui float 1.04E+17 to i8 ; yields undefined:1
5571 '``fptosi .. to``' Instruction
5572 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5579 <result> = fptosi <ty> <value> to <ty2> ; yields ty2
5584 The '``fptosi``' instruction converts :ref:`floating point <t_floating>`
5585 ``value`` to type ``ty2``.
5590 The '``fptosi``' instruction takes a value to cast, which must be a
5591 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5592 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5593 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5594 type with the same number of elements as ``ty``
5599 The '``fptosi``' instruction converts its :ref:`floating
5600 point <t_floating>` operand into the nearest (rounding towards zero)
5601 signed integer value. If the value cannot fit in ``ty2``, the results
5607 .. code-block:: llvm
5609 %X = fptosi double -123.0 to i32 ; yields i32:-123
5610 %Y = fptosi float 1.0E-247 to i1 ; yields undefined:1
5611 %Z = fptosi float 1.04E+17 to i8 ; yields undefined:1
5613 '``uitofp .. to``' Instruction
5614 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5621 <result> = uitofp <ty> <value> to <ty2> ; yields ty2
5626 The '``uitofp``' instruction regards ``value`` as an unsigned integer
5627 and converts that value to the ``ty2`` type.
5632 The '``uitofp``' instruction takes a value to cast, which must be a
5633 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
5634 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
5635 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
5636 type with the same number of elements as ``ty``
5641 The '``uitofp``' instruction interprets its operand as an unsigned
5642 integer quantity and converts it to the corresponding floating point
5643 value. If the value cannot fit in the floating point value, the results
5649 .. code-block:: llvm
5651 %X = uitofp i32 257 to float ; yields float:257.0
5652 %Y = uitofp i8 -1 to double ; yields double:255.0
5654 '``sitofp .. to``' Instruction
5655 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5662 <result> = sitofp <ty> <value> to <ty2> ; yields ty2
5667 The '``sitofp``' instruction regards ``value`` as a signed integer and
5668 converts that value to the ``ty2`` type.
5673 The '``sitofp``' instruction takes a value to cast, which must be a
5674 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
5675 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
5676 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
5677 type with the same number of elements as ``ty``
5682 The '``sitofp``' instruction interprets its operand as a signed integer
5683 quantity and converts it to the corresponding floating point value. If
5684 the value cannot fit in the floating point value, the results are
5690 .. code-block:: llvm
5692 %X = sitofp i32 257 to float ; yields float:257.0
5693 %Y = sitofp i8 -1 to double ; yields double:-1.0
5697 '``ptrtoint .. to``' Instruction
5698 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5705 <result> = ptrtoint <ty> <value> to <ty2> ; yields ty2
5710 The '``ptrtoint``' instruction converts the pointer or a vector of
5711 pointers ``value`` to the integer (or vector of integers) type ``ty2``.
5716 The '``ptrtoint``' instruction takes a ``value`` to cast, which must be
5717 a a value of type :ref:`pointer <t_pointer>` or a vector of pointers, and a
5718 type to cast it to ``ty2``, which must be an :ref:`integer <t_integer>` or
5719 a vector of integers type.
5724 The '``ptrtoint``' instruction converts ``value`` to integer type
5725 ``ty2`` by interpreting the pointer value as an integer and either
5726 truncating or zero extending that value to the size of the integer type.
5727 If ``value`` is smaller than ``ty2`` then a zero extension is done. If
5728 ``value`` is larger than ``ty2`` then a truncation is done. If they are
5729 the same size, then nothing is done (*no-op cast*) other than a type
5735 .. code-block:: llvm
5737 %X = ptrtoint i32* %P to i8 ; yields truncation on 32-bit architecture
5738 %Y = ptrtoint i32* %P to i64 ; yields zero extension on 32-bit architecture
5739 %Z = ptrtoint <4 x i32*> %P to <4 x i64>; yields vector zero extension for a vector of addresses on 32-bit architecture
5743 '``inttoptr .. to``' Instruction
5744 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5751 <result> = inttoptr <ty> <value> to <ty2> ; yields ty2
5756 The '``inttoptr``' instruction converts an integer ``value`` to a
5757 pointer type, ``ty2``.
5762 The '``inttoptr``' instruction takes an :ref:`integer <t_integer>` value to
5763 cast, and a type to cast it to, which must be a :ref:`pointer <t_pointer>`
5769 The '``inttoptr``' instruction converts ``value`` to type ``ty2`` by
5770 applying either a zero extension or a truncation depending on the size
5771 of the integer ``value``. If ``value`` is larger than the size of a
5772 pointer then a truncation is done. If ``value`` is smaller than the size
5773 of a pointer then a zero extension is done. If they are the same size,
5774 nothing is done (*no-op cast*).
5779 .. code-block:: llvm
5781 %X = inttoptr i32 255 to i32* ; yields zero extension on 64-bit architecture
5782 %Y = inttoptr i32 255 to i32* ; yields no-op on 32-bit architecture
5783 %Z = inttoptr i64 0 to i32* ; yields truncation on 32-bit architecture
5784 %Z = inttoptr <4 x i32> %G to <4 x i8*>; yields truncation of vector G to four pointers
5788 '``bitcast .. to``' Instruction
5789 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5796 <result> = bitcast <ty> <value> to <ty2> ; yields ty2
5801 The '``bitcast``' instruction converts ``value`` to type ``ty2`` without
5807 The '``bitcast``' instruction takes a value to cast, which must be a
5808 non-aggregate first class value, and a type to cast it to, which must
5809 also be a non-aggregate :ref:`first class <t_firstclass>` type. The
5810 bit sizes of ``value`` and the destination type, ``ty2``, must be
5811 identical. If the source type is a pointer, the destination type must
5812 also be a pointer of the same size. This instruction supports bitwise
5813 conversion of vectors to integers and to vectors of other types (as
5814 long as they have the same size).
5819 The '``bitcast``' instruction converts ``value`` to type ``ty2``. It
5820 is always a *no-op cast* because no bits change with this
5821 conversion. The conversion is done as if the ``value`` had been stored
5822 to memory and read back as type ``ty2``. Pointer (or vector of
5823 pointers) types may only be converted to other pointer (or vector of
5824 pointers) types with the same address space through this instruction.
5825 To convert pointers to other types, use the :ref:`inttoptr <i_inttoptr>`
5826 or :ref:`ptrtoint <i_ptrtoint>` instructions first.
5831 .. code-block:: llvm
5833 %X = bitcast i8 255 to i8 ; yields i8 :-1
5834 %Y = bitcast i32* %x to sint* ; yields sint*:%x
5835 %Z = bitcast <2 x int> %V to i64; ; yields i64: %V
5836 %Z = bitcast <2 x i32*> %V to <2 x i64*> ; yields <2 x i64*>
5838 .. _i_addrspacecast:
5840 '``addrspacecast .. to``' Instruction
5841 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5848 <result> = addrspacecast <pty> <ptrval> to <pty2> ; yields pty2
5853 The '``addrspacecast``' instruction converts ``ptrval`` from ``pty`` in
5854 address space ``n`` to type ``pty2`` in address space ``m``.
5859 The '``addrspacecast``' instruction takes a pointer or vector of pointer value
5860 to cast and a pointer type to cast it to, which must have a different
5866 The '``addrspacecast``' instruction converts the pointer value
5867 ``ptrval`` to type ``pty2``. It can be a *no-op cast* or a complex
5868 value modification, depending on the target and the address space
5869 pair. Pointer conversions within the same address space must be
5870 performed with the ``bitcast`` instruction. Note that if the address space
5871 conversion is legal then both result and operand refer to the same memory
5877 .. code-block:: llvm
5879 %X = addrspacecast i32* %x to i32 addrspace(1)* ; yields i32 addrspace(1)*:%x
5880 %Y = addrspacecast i32 addrspace(1)* %y to i64 addrspace(2)* ; yields i64 addrspace(2)*:%y
5881 %Z = addrspacecast <4 x i32*> %z to <4 x float addrspace(3)*> ; yields <4 x float addrspace(3)*>:%z
5888 The instructions in this category are the "miscellaneous" instructions,
5889 which defy better classification.
5893 '``icmp``' Instruction
5894 ^^^^^^^^^^^^^^^^^^^^^^
5901 <result> = icmp <cond> <ty> <op1>, <op2> ; yields {i1} or {<N x i1>}:result
5906 The '``icmp``' instruction returns a boolean value or a vector of
5907 boolean values based on comparison of its two integer, integer vector,
5908 pointer, or pointer vector operands.
5913 The '``icmp``' instruction takes three operands. The first operand is
5914 the condition code indicating the kind of comparison to perform. It is
5915 not a value, just a keyword. The possible condition code are:
5918 #. ``ne``: not equal
5919 #. ``ugt``: unsigned greater than
5920 #. ``uge``: unsigned greater or equal
5921 #. ``ult``: unsigned less than
5922 #. ``ule``: unsigned less or equal
5923 #. ``sgt``: signed greater than
5924 #. ``sge``: signed greater or equal
5925 #. ``slt``: signed less than
5926 #. ``sle``: signed less or equal
5928 The remaining two arguments must be :ref:`integer <t_integer>` or
5929 :ref:`pointer <t_pointer>` or integer :ref:`vector <t_vector>` typed. They
5930 must also be identical types.
5935 The '``icmp``' compares ``op1`` and ``op2`` according to the condition
5936 code given as ``cond``. The comparison performed always yields either an
5937 :ref:`i1 <t_integer>` or vector of ``i1`` result, as follows:
5939 #. ``eq``: yields ``true`` if the operands are equal, ``false``
5940 otherwise. No sign interpretation is necessary or performed.
5941 #. ``ne``: yields ``true`` if the operands are unequal, ``false``
5942 otherwise. No sign interpretation is necessary or performed.
5943 #. ``ugt``: interprets the operands as unsigned values and yields
5944 ``true`` if ``op1`` is greater than ``op2``.
5945 #. ``uge``: interprets the operands as unsigned values and yields
5946 ``true`` if ``op1`` is greater than or equal to ``op2``.
5947 #. ``ult``: interprets the operands as unsigned values and yields
5948 ``true`` if ``op1`` is less than ``op2``.
5949 #. ``ule``: interprets the operands as unsigned values and yields
5950 ``true`` if ``op1`` is less than or equal to ``op2``.
5951 #. ``sgt``: interprets the operands as signed values and yields ``true``
5952 if ``op1`` is greater than ``op2``.
5953 #. ``sge``: interprets the operands as signed values and yields ``true``
5954 if ``op1`` is greater than or equal to ``op2``.
5955 #. ``slt``: interprets the operands as signed values and yields ``true``
5956 if ``op1`` is less than ``op2``.
5957 #. ``sle``: interprets the operands as signed values and yields ``true``
5958 if ``op1`` is less than or equal to ``op2``.
5960 If the operands are :ref:`pointer <t_pointer>` typed, the pointer values
5961 are compared as if they were integers.
5963 If the operands are integer vectors, then they are compared element by
5964 element. The result is an ``i1`` vector with the same number of elements
5965 as the values being compared. Otherwise, the result is an ``i1``.
5970 .. code-block:: llvm
5972 <result> = icmp eq i32 4, 5 ; yields: result=false
5973 <result> = icmp ne float* %X, %X ; yields: result=false
5974 <result> = icmp ult i16 4, 5 ; yields: result=true
5975 <result> = icmp sgt i16 4, 5 ; yields: result=false
5976 <result> = icmp ule i16 -4, 5 ; yields: result=false
5977 <result> = icmp sge i16 4, 5 ; yields: result=false
5979 Note that the code generator does not yet support vector types with the
5980 ``icmp`` instruction.
5984 '``fcmp``' Instruction
5985 ^^^^^^^^^^^^^^^^^^^^^^
5992 <result> = fcmp <cond> <ty> <op1>, <op2> ; yields {i1} or {<N x i1>}:result
5997 The '``fcmp``' instruction returns a boolean value or vector of boolean
5998 values based on comparison of its operands.
6000 If the operands are floating point scalars, then the result type is a
6001 boolean (:ref:`i1 <t_integer>`).
6003 If the operands are floating point vectors, then the result type is a
6004 vector of boolean with the same number of elements as the operands being
6010 The '``fcmp``' instruction takes three operands. The first operand is
6011 the condition code indicating the kind of comparison to perform. It is
6012 not a value, just a keyword. The possible condition code are:
6014 #. ``false``: no comparison, always returns false
6015 #. ``oeq``: ordered and equal
6016 #. ``ogt``: ordered and greater than
6017 #. ``oge``: ordered and greater than or equal
6018 #. ``olt``: ordered and less than
6019 #. ``ole``: ordered and less than or equal
6020 #. ``one``: ordered and not equal
6021 #. ``ord``: ordered (no nans)
6022 #. ``ueq``: unordered or equal
6023 #. ``ugt``: unordered or greater than
6024 #. ``uge``: unordered or greater than or equal
6025 #. ``ult``: unordered or less than
6026 #. ``ule``: unordered or less than or equal
6027 #. ``une``: unordered or not equal
6028 #. ``uno``: unordered (either nans)
6029 #. ``true``: no comparison, always returns true
6031 *Ordered* means that neither operand is a QNAN while *unordered* means
6032 that either operand may be a QNAN.
6034 Each of ``val1`` and ``val2`` arguments must be either a :ref:`floating
6035 point <t_floating>` type or a :ref:`vector <t_vector>` of floating point
6036 type. They must have identical types.
6041 The '``fcmp``' instruction compares ``op1`` and ``op2`` according to the
6042 condition code given as ``cond``. If the operands are vectors, then the
6043 vectors are compared element by element. Each comparison performed
6044 always yields an :ref:`i1 <t_integer>` result, as follows:
6046 #. ``false``: always yields ``false``, regardless of operands.
6047 #. ``oeq``: yields ``true`` if both operands are not a QNAN and ``op1``
6048 is equal to ``op2``.
6049 #. ``ogt``: yields ``true`` if both operands are not a QNAN and ``op1``
6050 is greater than ``op2``.
6051 #. ``oge``: yields ``true`` if both operands are not a QNAN and ``op1``
6052 is greater than or equal to ``op2``.
6053 #. ``olt``: yields ``true`` if both operands are not a QNAN and ``op1``
6054 is less than ``op2``.
6055 #. ``ole``: yields ``true`` if both operands are not a QNAN and ``op1``
6056 is less than or equal to ``op2``.
6057 #. ``one``: yields ``true`` if both operands are not a QNAN and ``op1``
6058 is not equal to ``op2``.
6059 #. ``ord``: yields ``true`` if both operands are not a QNAN.
6060 #. ``ueq``: yields ``true`` if either operand is a QNAN or ``op1`` is
6062 #. ``ugt``: yields ``true`` if either operand is a QNAN or ``op1`` is
6063 greater than ``op2``.
6064 #. ``uge``: yields ``true`` if either operand is a QNAN or ``op1`` is
6065 greater than or equal to ``op2``.
6066 #. ``ult``: yields ``true`` if either operand is a QNAN or ``op1`` is
6068 #. ``ule``: yields ``true`` if either operand is a QNAN or ``op1`` is
6069 less than or equal to ``op2``.
6070 #. ``une``: yields ``true`` if either operand is a QNAN or ``op1`` is
6071 not equal to ``op2``.
6072 #. ``uno``: yields ``true`` if either operand is a QNAN.
6073 #. ``true``: always yields ``true``, regardless of operands.
6078 .. code-block:: llvm
6080 <result> = fcmp oeq float 4.0, 5.0 ; yields: result=false
6081 <result> = fcmp one float 4.0, 5.0 ; yields: result=true
6082 <result> = fcmp olt float 4.0, 5.0 ; yields: result=true
6083 <result> = fcmp ueq double 1.0, 2.0 ; yields: result=false
6085 Note that the code generator does not yet support vector types with the
6086 ``fcmp`` instruction.
6090 '``phi``' Instruction
6091 ^^^^^^^^^^^^^^^^^^^^^
6098 <result> = phi <ty> [ <val0>, <label0>], ...
6103 The '``phi``' instruction is used to implement the φ node in the SSA
6104 graph representing the function.
6109 The type of the incoming values is specified with the first type field.
6110 After this, the '``phi``' instruction takes a list of pairs as
6111 arguments, with one pair for each predecessor basic block of the current
6112 block. Only values of :ref:`first class <t_firstclass>` type may be used as
6113 the value arguments to the PHI node. Only labels may be used as the
6116 There must be no non-phi instructions between the start of a basic block
6117 and the PHI instructions: i.e. PHI instructions must be first in a basic
6120 For the purposes of the SSA form, the use of each incoming value is
6121 deemed to occur on the edge from the corresponding predecessor block to
6122 the current block (but after any definition of an '``invoke``'
6123 instruction's return value on the same edge).
6128 At runtime, the '``phi``' instruction logically takes on the value
6129 specified by the pair corresponding to the predecessor basic block that
6130 executed just prior to the current block.
6135 .. code-block:: llvm
6137 Loop: ; Infinite loop that counts from 0 on up...
6138 %indvar = phi i32 [ 0, %LoopHeader ], [ %nextindvar, %Loop ]
6139 %nextindvar = add i32 %indvar, 1
6144 '``select``' Instruction
6145 ^^^^^^^^^^^^^^^^^^^^^^^^
6152 <result> = select selty <cond>, <ty> <val1>, <ty> <val2> ; yields ty
6154 selty is either i1 or {<N x i1>}
6159 The '``select``' instruction is used to choose one value based on a
6160 condition, without IR-level branching.
6165 The '``select``' instruction requires an 'i1' value or a vector of 'i1'
6166 values indicating the condition, and two values of the same :ref:`first
6167 class <t_firstclass>` type. If the val1/val2 are vectors and the
6168 condition is a scalar, then entire vectors are selected, not individual
6174 If the condition is an i1 and it evaluates to 1, the instruction returns
6175 the first value argument; otherwise, it returns the second value
6178 If the condition is a vector of i1, then the value arguments must be
6179 vectors of the same size, and the selection is done element by element.
6184 .. code-block:: llvm
6186 %X = select i1 true, i8 17, i8 42 ; yields i8:17
6190 '``call``' Instruction
6191 ^^^^^^^^^^^^^^^^^^^^^^
6198 <result> = [tail | musttail] call [cconv] [ret attrs] <ty> [<fnty>*] <fnptrval>(<function args>) [fn attrs]
6203 The '``call``' instruction represents a simple function call.
6208 This instruction requires several arguments:
6210 #. The optional ``tail`` and ``musttail`` markers indicate that the optimizers
6211 should perform tail call optimization. The ``tail`` marker is a hint that
6212 `can be ignored <CodeGenerator.html#sibcallopt>`_. The ``musttail`` marker
6213 means that the call must be tail call optimized in order for the program to
6214 be correct. The ``musttail`` marker provides these guarantees:
6216 #. The call will not cause unbounded stack growth if it is part of a
6217 recursive cycle in the call graph.
6218 #. Arguments with the :ref:`inalloca <attr_inalloca>` attribute are
6221 Both markers imply that the callee does not access allocas or varargs from
6222 the caller. Calls marked ``musttail`` must obey the following additional
6225 - The call must immediately precede a :ref:`ret <i_ret>` instruction,
6226 or a pointer bitcast followed by a ret instruction.
6227 - The ret instruction must return the (possibly bitcasted) value
6228 produced by the call or void.
6229 - The caller and callee prototypes must match. Pointer types of
6230 parameters or return types may differ in pointee type, but not
6232 - The calling conventions of the caller and callee must match.
6233 - All ABI-impacting function attributes, such as sret, byval, inreg,
6234 returned, and inalloca, must match.
6236 Tail call optimization for calls marked ``tail`` is guaranteed to occur if
6237 the following conditions are met:
6239 - Caller and callee both have the calling convention ``fastcc``.
6240 - The call is in tail position (ret immediately follows call and ret
6241 uses value of call or is void).
6242 - Option ``-tailcallopt`` is enabled, or
6243 ``llvm::GuaranteedTailCallOpt`` is ``true``.
6244 - `Platform specific constraints are
6245 met. <CodeGenerator.html#tailcallopt>`_
6247 #. The optional "cconv" marker indicates which :ref:`calling
6248 convention <callingconv>` the call should use. If none is
6249 specified, the call defaults to using C calling conventions. The
6250 calling convention of the call must match the calling convention of
6251 the target function, or else the behavior is undefined.
6252 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
6253 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
6255 #. '``ty``': the type of the call instruction itself which is also the
6256 type of the return value. Functions that return no value are marked
6258 #. '``fnty``': shall be the signature of the pointer to function value
6259 being invoked. The argument types must match the types implied by
6260 this signature. This type can be omitted if the function is not
6261 varargs and if the function type does not return a pointer to a
6263 #. '``fnptrval``': An LLVM value containing a pointer to a function to
6264 be invoked. In most cases, this is a direct function invocation, but
6265 indirect ``call``'s are just as possible, calling an arbitrary pointer
6267 #. '``function args``': argument list whose types match the function
6268 signature argument types and parameter attributes. All arguments must
6269 be of :ref:`first class <t_firstclass>` type. If the function signature
6270 indicates the function accepts a variable number of arguments, the
6271 extra arguments can be specified.
6272 #. The optional :ref:`function attributes <fnattrs>` list. Only
6273 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
6274 attributes are valid here.
6279 The '``call``' instruction is used to cause control flow to transfer to
6280 a specified function, with its incoming arguments bound to the specified
6281 values. Upon a '``ret``' instruction in the called function, control
6282 flow continues with the instruction after the function call, and the
6283 return value of the function is bound to the result argument.
6288 .. code-block:: llvm
6290 %retval = call i32 @test(i32 %argc)
6291 call i32 (i8*, ...)* @printf(i8* %msg, i32 12, i8 42) ; yields i32
6292 %X = tail call i32 @foo() ; yields i32
6293 %Y = tail call fastcc i32 @foo() ; yields i32
6294 call void %foo(i8 97 signext)
6296 %struct.A = type { i32, i8 }
6297 %r = call %struct.A @foo() ; yields { 32, i8 }
6298 %gr = extractvalue %struct.A %r, 0 ; yields i32
6299 %gr1 = extractvalue %struct.A %r, 1 ; yields i8
6300 %Z = call void @foo() noreturn ; indicates that %foo never returns normally
6301 %ZZ = call zeroext i32 @bar() ; Return value is %zero extended
6303 llvm treats calls to some functions with names and arguments that match
6304 the standard C99 library as being the C99 library functions, and may
6305 perform optimizations or generate code for them under that assumption.
6306 This is something we'd like to change in the future to provide better
6307 support for freestanding environments and non-C-based languages.
6311 '``va_arg``' Instruction
6312 ^^^^^^^^^^^^^^^^^^^^^^^^
6319 <resultval> = va_arg <va_list*> <arglist>, <argty>
6324 The '``va_arg``' instruction is used to access arguments passed through
6325 the "variable argument" area of a function call. It is used to implement
6326 the ``va_arg`` macro in C.
6331 This instruction takes a ``va_list*`` value and the type of the
6332 argument. It returns a value of the specified argument type and
6333 increments the ``va_list`` to point to the next argument. The actual
6334 type of ``va_list`` is target specific.
6339 The '``va_arg``' instruction loads an argument of the specified type
6340 from the specified ``va_list`` and causes the ``va_list`` to point to
6341 the next argument. For more information, see the variable argument
6342 handling :ref:`Intrinsic Functions <int_varargs>`.
6344 It is legal for this instruction to be called in a function which does
6345 not take a variable number of arguments, for example, the ``vfprintf``
6348 ``va_arg`` is an LLVM instruction instead of an :ref:`intrinsic
6349 function <intrinsics>` because it takes a type as an argument.
6354 See the :ref:`variable argument processing <int_varargs>` section.
6356 Note that the code generator does not yet fully support va\_arg on many
6357 targets. Also, it does not currently support va\_arg with aggregate
6358 types on any target.
6362 '``landingpad``' Instruction
6363 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6370 <resultval> = landingpad <resultty> personality <type> <pers_fn> <clause>+
6371 <resultval> = landingpad <resultty> personality <type> <pers_fn> cleanup <clause>*
6373 <clause> := catch <type> <value>
6374 <clause> := filter <array constant type> <array constant>
6379 The '``landingpad``' instruction is used by `LLVM's exception handling
6380 system <ExceptionHandling.html#overview>`_ to specify that a basic block
6381 is a landing pad --- one where the exception lands, and corresponds to the
6382 code found in the ``catch`` portion of a ``try``/``catch`` sequence. It
6383 defines values supplied by the personality function (``pers_fn``) upon
6384 re-entry to the function. The ``resultval`` has the type ``resultty``.
6389 This instruction takes a ``pers_fn`` value. This is the personality
6390 function associated with the unwinding mechanism. The optional
6391 ``cleanup`` flag indicates that the landing pad block is a cleanup.
6393 A ``clause`` begins with the clause type --- ``catch`` or ``filter`` --- and
6394 contains the global variable representing the "type" that may be caught
6395 or filtered respectively. Unlike the ``catch`` clause, the ``filter``
6396 clause takes an array constant as its argument. Use
6397 "``[0 x i8**] undef``" for a filter which cannot throw. The
6398 '``landingpad``' instruction must contain *at least* one ``clause`` or
6399 the ``cleanup`` flag.
6404 The '``landingpad``' instruction defines the values which are set by the
6405 personality function (``pers_fn``) upon re-entry to the function, and
6406 therefore the "result type" of the ``landingpad`` instruction. As with
6407 calling conventions, how the personality function results are
6408 represented in LLVM IR is target specific.
6410 The clauses are applied in order from top to bottom. If two
6411 ``landingpad`` instructions are merged together through inlining, the
6412 clauses from the calling function are appended to the list of clauses.
6413 When the call stack is being unwound due to an exception being thrown,
6414 the exception is compared against each ``clause`` in turn. If it doesn't
6415 match any of the clauses, and the ``cleanup`` flag is not set, then
6416 unwinding continues further up the call stack.
6418 The ``landingpad`` instruction has several restrictions:
6420 - A landing pad block is a basic block which is the unwind destination
6421 of an '``invoke``' instruction.
6422 - A landing pad block must have a '``landingpad``' instruction as its
6423 first non-PHI instruction.
6424 - There can be only one '``landingpad``' instruction within the landing
6426 - A basic block that is not a landing pad block may not include a
6427 '``landingpad``' instruction.
6428 - All '``landingpad``' instructions in a function must have the same
6429 personality function.
6434 .. code-block:: llvm
6436 ;; A landing pad which can catch an integer.
6437 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6439 ;; A landing pad that is a cleanup.
6440 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6442 ;; A landing pad which can catch an integer and can only throw a double.
6443 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6445 filter [1 x i8**] [@_ZTId]
6452 LLVM supports the notion of an "intrinsic function". These functions
6453 have well known names and semantics and are required to follow certain
6454 restrictions. Overall, these intrinsics represent an extension mechanism
6455 for the LLVM language that does not require changing all of the
6456 transformations in LLVM when adding to the language (or the bitcode
6457 reader/writer, the parser, etc...).
6459 Intrinsic function names must all start with an "``llvm.``" prefix. This
6460 prefix is reserved in LLVM for intrinsic names; thus, function names may
6461 not begin with this prefix. Intrinsic functions must always be external
6462 functions: you cannot define the body of intrinsic functions. Intrinsic
6463 functions may only be used in call or invoke instructions: it is illegal
6464 to take the address of an intrinsic function. Additionally, because
6465 intrinsic functions are part of the LLVM language, it is required if any
6466 are added that they be documented here.
6468 Some intrinsic functions can be overloaded, i.e., the intrinsic
6469 represents a family of functions that perform the same operation but on
6470 different data types. Because LLVM can represent over 8 million
6471 different integer types, overloading is used commonly to allow an
6472 intrinsic function to operate on any integer type. One or more of the
6473 argument types or the result type can be overloaded to accept any
6474 integer type. Argument types may also be defined as exactly matching a
6475 previous argument's type or the result type. This allows an intrinsic
6476 function which accepts multiple arguments, but needs all of them to be
6477 of the same type, to only be overloaded with respect to a single
6478 argument or the result.
6480 Overloaded intrinsics will have the names of its overloaded argument
6481 types encoded into its function name, each preceded by a period. Only
6482 those types which are overloaded result in a name suffix. Arguments
6483 whose type is matched against another type do not. For example, the
6484 ``llvm.ctpop`` function can take an integer of any width and returns an
6485 integer of exactly the same integer width. This leads to a family of
6486 functions such as ``i8 @llvm.ctpop.i8(i8 %val)`` and
6487 ``i29 @llvm.ctpop.i29(i29 %val)``. Only one type, the return type, is
6488 overloaded, and only one type suffix is required. Because the argument's
6489 type is matched against the return type, it does not require its own
6492 To learn how to add an intrinsic function, please see the `Extending
6493 LLVM Guide <ExtendingLLVM.html>`_.
6497 Variable Argument Handling Intrinsics
6498 -------------------------------------
6500 Variable argument support is defined in LLVM with the
6501 :ref:`va_arg <i_va_arg>` instruction and these three intrinsic
6502 functions. These functions are related to the similarly named macros
6503 defined in the ``<stdarg.h>`` header file.
6505 All of these functions operate on arguments that use a target-specific
6506 value type "``va_list``". The LLVM assembly language reference manual
6507 does not define what this type is, so all transformations should be
6508 prepared to handle these functions regardless of the type used.
6510 This example shows how the :ref:`va_arg <i_va_arg>` instruction and the
6511 variable argument handling intrinsic functions are used.
6513 .. code-block:: llvm
6515 define i32 @test(i32 %X, ...) {
6516 ; Initialize variable argument processing
6518 %ap2 = bitcast i8** %ap to i8*
6519 call void @llvm.va_start(i8* %ap2)
6521 ; Read a single integer argument
6522 %tmp = va_arg i8** %ap, i32
6524 ; Demonstrate usage of llvm.va_copy and llvm.va_end
6526 %aq2 = bitcast i8** %aq to i8*
6527 call void @llvm.va_copy(i8* %aq2, i8* %ap2)
6528 call void @llvm.va_end(i8* %aq2)
6530 ; Stop processing of arguments.
6531 call void @llvm.va_end(i8* %ap2)
6535 declare void @llvm.va_start(i8*)
6536 declare void @llvm.va_copy(i8*, i8*)
6537 declare void @llvm.va_end(i8*)
6541 '``llvm.va_start``' Intrinsic
6542 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6549 declare void @llvm.va_start(i8* <arglist>)
6554 The '``llvm.va_start``' intrinsic initializes ``*<arglist>`` for
6555 subsequent use by ``va_arg``.
6560 The argument is a pointer to a ``va_list`` element to initialize.
6565 The '``llvm.va_start``' intrinsic works just like the ``va_start`` macro
6566 available in C. In a target-dependent way, it initializes the
6567 ``va_list`` element to which the argument points, so that the next call
6568 to ``va_arg`` will produce the first variable argument passed to the
6569 function. Unlike the C ``va_start`` macro, this intrinsic does not need
6570 to know the last argument of the function as the compiler can figure
6573 '``llvm.va_end``' Intrinsic
6574 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6581 declare void @llvm.va_end(i8* <arglist>)
6586 The '``llvm.va_end``' intrinsic destroys ``*<arglist>``, which has been
6587 initialized previously with ``llvm.va_start`` or ``llvm.va_copy``.
6592 The argument is a pointer to a ``va_list`` to destroy.
6597 The '``llvm.va_end``' intrinsic works just like the ``va_end`` macro
6598 available in C. In a target-dependent way, it destroys the ``va_list``
6599 element to which the argument points. Calls to
6600 :ref:`llvm.va_start <int_va_start>` and
6601 :ref:`llvm.va_copy <int_va_copy>` must be matched exactly with calls to
6606 '``llvm.va_copy``' Intrinsic
6607 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6614 declare void @llvm.va_copy(i8* <destarglist>, i8* <srcarglist>)
6619 The '``llvm.va_copy``' intrinsic copies the current argument position
6620 from the source argument list to the destination argument list.
6625 The first argument is a pointer to a ``va_list`` element to initialize.
6626 The second argument is a pointer to a ``va_list`` element to copy from.
6631 The '``llvm.va_copy``' intrinsic works just like the ``va_copy`` macro
6632 available in C. In a target-dependent way, it copies the source
6633 ``va_list`` element into the destination ``va_list`` element. This
6634 intrinsic is necessary because the `` llvm.va_start`` intrinsic may be
6635 arbitrarily complex and require, for example, memory allocation.
6637 Accurate Garbage Collection Intrinsics
6638 --------------------------------------
6640 LLVM support for `Accurate Garbage Collection <GarbageCollection.html>`_
6641 (GC) requires the implementation and generation of these intrinsics.
6642 These intrinsics allow identification of :ref:`GC roots on the
6643 stack <int_gcroot>`, as well as garbage collector implementations that
6644 require :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers.
6645 Front-ends for type-safe garbage collected languages should generate
6646 these intrinsics to make use of the LLVM garbage collectors. For more
6647 details, see `Accurate Garbage Collection with
6648 LLVM <GarbageCollection.html>`_.
6650 The garbage collection intrinsics only operate on objects in the generic
6651 address space (address space zero).
6655 '``llvm.gcroot``' Intrinsic
6656 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6663 declare void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
6668 The '``llvm.gcroot``' intrinsic declares the existence of a GC root to
6669 the code generator, and allows some metadata to be associated with it.
6674 The first argument specifies the address of a stack object that contains
6675 the root pointer. The second pointer (which must be either a constant or
6676 a global value address) contains the meta-data to be associated with the
6682 At runtime, a call to this intrinsic stores a null pointer into the
6683 "ptrloc" location. At compile-time, the code generator generates
6684 information to allow the runtime to find the pointer at GC safe points.
6685 The '``llvm.gcroot``' intrinsic may only be used in a function which
6686 :ref:`specifies a GC algorithm <gc>`.
6690 '``llvm.gcread``' Intrinsic
6691 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6698 declare i8* @llvm.gcread(i8* %ObjPtr, i8** %Ptr)
6703 The '``llvm.gcread``' intrinsic identifies reads of references from heap
6704 locations, allowing garbage collector implementations that require read
6710 The second argument is the address to read from, which should be an
6711 address allocated from the garbage collector. The first object is a
6712 pointer to the start of the referenced object, if needed by the language
6713 runtime (otherwise null).
6718 The '``llvm.gcread``' intrinsic has the same semantics as a load
6719 instruction, but may be replaced with substantially more complex code by
6720 the garbage collector runtime, as needed. The '``llvm.gcread``'
6721 intrinsic may only be used in a function which :ref:`specifies a GC
6726 '``llvm.gcwrite``' Intrinsic
6727 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6734 declare void @llvm.gcwrite(i8* %P1, i8* %Obj, i8** %P2)
6739 The '``llvm.gcwrite``' intrinsic identifies writes of references to heap
6740 locations, allowing garbage collector implementations that require write
6741 barriers (such as generational or reference counting collectors).
6746 The first argument is the reference to store, the second is the start of
6747 the object to store it to, and the third is the address of the field of
6748 Obj to store to. If the runtime does not require a pointer to the
6749 object, Obj may be null.
6754 The '``llvm.gcwrite``' intrinsic has the same semantics as a store
6755 instruction, but may be replaced with substantially more complex code by
6756 the garbage collector runtime, as needed. The '``llvm.gcwrite``'
6757 intrinsic may only be used in a function which :ref:`specifies a GC
6760 Code Generator Intrinsics
6761 -------------------------
6763 These intrinsics are provided by LLVM to expose special features that
6764 may only be implemented with code generator support.
6766 '``llvm.returnaddress``' Intrinsic
6767 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6774 declare i8 *@llvm.returnaddress(i32 <level>)
6779 The '``llvm.returnaddress``' intrinsic attempts to compute a
6780 target-specific value indicating the return address of the current
6781 function or one of its callers.
6786 The argument to this intrinsic indicates which function to return the
6787 address for. Zero indicates the calling function, one indicates its
6788 caller, etc. The argument is **required** to be a constant integer
6794 The '``llvm.returnaddress``' intrinsic either returns a pointer
6795 indicating the return address of the specified call frame, or zero if it
6796 cannot be identified. The value returned by this intrinsic is likely to
6797 be incorrect or 0 for arguments other than zero, so it should only be
6798 used for debugging purposes.
6800 Note that calling this intrinsic does not prevent function inlining or
6801 other aggressive transformations, so the value returned may not be that
6802 of the obvious source-language caller.
6804 '``llvm.frameaddress``' Intrinsic
6805 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6812 declare i8* @llvm.frameaddress(i32 <level>)
6817 The '``llvm.frameaddress``' intrinsic attempts to return the
6818 target-specific frame pointer value for the specified stack frame.
6823 The argument to this intrinsic indicates which function to return the
6824 frame pointer for. Zero indicates the calling function, one indicates
6825 its caller, etc. The argument is **required** to be a constant integer
6831 The '``llvm.frameaddress``' intrinsic either returns a pointer
6832 indicating the frame address of the specified call frame, or zero if it
6833 cannot be identified. The value returned by this intrinsic is likely to
6834 be incorrect or 0 for arguments other than zero, so it should only be
6835 used for debugging purposes.
6837 Note that calling this intrinsic does not prevent function inlining or
6838 other aggressive transformations, so the value returned may not be that
6839 of the obvious source-language caller.
6841 .. _int_read_register:
6842 .. _int_write_register:
6844 '``llvm.read_register``' and '``llvm.write_register``' Intrinsics
6845 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6852 declare i32 @llvm.read_register.i32(metadata)
6853 declare i64 @llvm.read_register.i64(metadata)
6854 declare void @llvm.write_register.i32(metadata, i32 @value)
6855 declare void @llvm.write_register.i64(metadata, i64 @value)
6856 !0 = metadata !{metadata !"sp\00"}
6861 The '``llvm.read_register``' and '``llvm.write_register``' intrinsics
6862 provides access to the named register. The register must be valid on
6863 the architecture being compiled to. The type needs to be compatible
6864 with the register being read.
6869 The '``llvm.read_register``' intrinsic returns the current value of the
6870 register, where possible. The '``llvm.write_register``' intrinsic sets
6871 the current value of the register, where possible.
6873 This is useful to implement named register global variables that need
6874 to always be mapped to a specific register, as is common practice on
6875 bare-metal programs including OS kernels.
6877 The compiler doesn't check for register availability or use of the used
6878 register in surrounding code, including inline assembly. Because of that,
6879 allocatable registers are not supported.
6881 Warning: So far it only works with the stack pointer on selected
6882 architectures (ARM, AArch64, PowerPC and x86_64). Significant amount of
6883 work is needed to support other registers and even more so, allocatable
6888 '``llvm.stacksave``' Intrinsic
6889 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6896 declare i8* @llvm.stacksave()
6901 The '``llvm.stacksave``' intrinsic is used to remember the current state
6902 of the function stack, for use with
6903 :ref:`llvm.stackrestore <int_stackrestore>`. This is useful for
6904 implementing language features like scoped automatic variable sized
6910 This intrinsic returns a opaque pointer value that can be passed to
6911 :ref:`llvm.stackrestore <int_stackrestore>`. When an
6912 ``llvm.stackrestore`` intrinsic is executed with a value saved from
6913 ``llvm.stacksave``, it effectively restores the state of the stack to
6914 the state it was in when the ``llvm.stacksave`` intrinsic executed. In
6915 practice, this pops any :ref:`alloca <i_alloca>` blocks from the stack that
6916 were allocated after the ``llvm.stacksave`` was executed.
6918 .. _int_stackrestore:
6920 '``llvm.stackrestore``' Intrinsic
6921 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6928 declare void @llvm.stackrestore(i8* %ptr)
6933 The '``llvm.stackrestore``' intrinsic is used to restore the state of
6934 the function stack to the state it was in when the corresponding
6935 :ref:`llvm.stacksave <int_stacksave>` intrinsic executed. This is
6936 useful for implementing language features like scoped automatic variable
6937 sized arrays in C99.
6942 See the description for :ref:`llvm.stacksave <int_stacksave>`.
6944 '``llvm.prefetch``' Intrinsic
6945 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6952 declare void @llvm.prefetch(i8* <address>, i32 <rw>, i32 <locality>, i32 <cache type>)
6957 The '``llvm.prefetch``' intrinsic is a hint to the code generator to
6958 insert a prefetch instruction if supported; otherwise, it is a noop.
6959 Prefetches have no effect on the behavior of the program but can change
6960 its performance characteristics.
6965 ``address`` is the address to be prefetched, ``rw`` is the specifier
6966 determining if the fetch should be for a read (0) or write (1), and
6967 ``locality`` is a temporal locality specifier ranging from (0) - no
6968 locality, to (3) - extremely local keep in cache. The ``cache type``
6969 specifies whether the prefetch is performed on the data (1) or
6970 instruction (0) cache. The ``rw``, ``locality`` and ``cache type``
6971 arguments must be constant integers.
6976 This intrinsic does not modify the behavior of the program. In
6977 particular, prefetches cannot trap and do not produce a value. On
6978 targets that support this intrinsic, the prefetch can provide hints to
6979 the processor cache for better performance.
6981 '``llvm.pcmarker``' Intrinsic
6982 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6989 declare void @llvm.pcmarker(i32 <id>)
6994 The '``llvm.pcmarker``' intrinsic is a method to export a Program
6995 Counter (PC) in a region of code to simulators and other tools. The
6996 method is target specific, but it is expected that the marker will use
6997 exported symbols to transmit the PC of the marker. The marker makes no
6998 guarantees that it will remain with any specific instruction after
6999 optimizations. It is possible that the presence of a marker will inhibit
7000 optimizations. The intended use is to be inserted after optimizations to
7001 allow correlations of simulation runs.
7006 ``id`` is a numerical id identifying the marker.
7011 This intrinsic does not modify the behavior of the program. Backends
7012 that do not support this intrinsic may ignore it.
7014 '``llvm.readcyclecounter``' Intrinsic
7015 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7022 declare i64 @llvm.readcyclecounter()
7027 The '``llvm.readcyclecounter``' intrinsic provides access to the cycle
7028 counter register (or similar low latency, high accuracy clocks) on those
7029 targets that support it. On X86, it should map to RDTSC. On Alpha, it
7030 should map to RPCC. As the backing counters overflow quickly (on the
7031 order of 9 seconds on alpha), this should only be used for small
7037 When directly supported, reading the cycle counter should not modify any
7038 memory. Implementations are allowed to either return a application
7039 specific value or a system wide value. On backends without support, this
7040 is lowered to a constant 0.
7042 Note that runtime support may be conditional on the privilege-level code is
7043 running at and the host platform.
7045 '``llvm.clear_cache``' Intrinsic
7046 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7053 declare void @llvm.clear_cache(i8*, i8*)
7058 The '``llvm.clear_cache``' intrinsic ensures visibility of modifications
7059 in the specified range to the execution unit of the processor. On
7060 targets with non-unified instruction and data cache, the implementation
7061 flushes the instruction cache.
7066 On platforms with coherent instruction and data caches (e.g. x86), this
7067 intrinsic is a nop. On platforms with non-coherent instruction and data
7068 cache (e.g. ARM, MIPS), the intrinsic is lowered either to appropriate
7069 instructions or a system call, if cache flushing requires special
7072 The default behavior is to emit a call to ``__clear_cache`` from the run
7075 This instrinsic does *not* empty the instruction pipeline. Modifications
7076 of the current function are outside the scope of the intrinsic.
7078 Standard C Library Intrinsics
7079 -----------------------------
7081 LLVM provides intrinsics for a few important standard C library
7082 functions. These intrinsics allow source-language front-ends to pass
7083 information about the alignment of the pointer arguments to the code
7084 generator, providing opportunity for more efficient code generation.
7088 '``llvm.memcpy``' Intrinsic
7089 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7094 This is an overloaded intrinsic. You can use ``llvm.memcpy`` on any
7095 integer bit width and for different address spaces. Not all targets
7096 support all bit widths however.
7100 declare void @llvm.memcpy.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
7101 i32 <len>, i32 <align>, i1 <isvolatile>)
7102 declare void @llvm.memcpy.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
7103 i64 <len>, i32 <align>, i1 <isvolatile>)
7108 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
7109 source location to the destination location.
7111 Note that, unlike the standard libc function, the ``llvm.memcpy.*``
7112 intrinsics do not return a value, takes extra alignment/isvolatile
7113 arguments and the pointers can be in specified address spaces.
7118 The first argument is a pointer to the destination, the second is a
7119 pointer to the source. The third argument is an integer argument
7120 specifying the number of bytes to copy, the fourth argument is the
7121 alignment of the source and destination locations, and the fifth is a
7122 boolean indicating a volatile access.
7124 If the call to this intrinsic has an alignment value that is not 0 or 1,
7125 then the caller guarantees that both the source and destination pointers
7126 are aligned to that boundary.
7128 If the ``isvolatile`` parameter is ``true``, the ``llvm.memcpy`` call is
7129 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
7130 very cleanly specified and it is unwise to depend on it.
7135 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
7136 source location to the destination location, which are not allowed to
7137 overlap. It copies "len" bytes of memory over. If the argument is known
7138 to be aligned to some boundary, this can be specified as the fourth
7139 argument, otherwise it should be set to 0 or 1 (both meaning no alignment).
7141 '``llvm.memmove``' Intrinsic
7142 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7147 This is an overloaded intrinsic. You can use llvm.memmove on any integer
7148 bit width and for different address space. Not all targets support all
7153 declare void @llvm.memmove.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
7154 i32 <len>, i32 <align>, i1 <isvolatile>)
7155 declare void @llvm.memmove.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
7156 i64 <len>, i32 <align>, i1 <isvolatile>)
7161 The '``llvm.memmove.*``' intrinsics move a block of memory from the
7162 source location to the destination location. It is similar to the
7163 '``llvm.memcpy``' intrinsic but allows the two memory locations to
7166 Note that, unlike the standard libc function, the ``llvm.memmove.*``
7167 intrinsics do not return a value, takes extra alignment/isvolatile
7168 arguments and the pointers can be in specified address spaces.
7173 The first argument is a pointer to the destination, the second is a
7174 pointer to the source. The third argument is an integer argument
7175 specifying the number of bytes to copy, the fourth argument is the
7176 alignment of the source and destination locations, and the fifth is a
7177 boolean indicating a volatile access.
7179 If the call to this intrinsic has an alignment value that is not 0 or 1,
7180 then the caller guarantees that the source and destination pointers are
7181 aligned to that boundary.
7183 If the ``isvolatile`` parameter is ``true``, the ``llvm.memmove`` call
7184 is a :ref:`volatile operation <volatile>`. The detailed access behavior is
7185 not very cleanly specified and it is unwise to depend on it.
7190 The '``llvm.memmove.*``' intrinsics copy a block of memory from the
7191 source location to the destination location, which may overlap. It
7192 copies "len" bytes of memory over. If the argument is known to be
7193 aligned to some boundary, this can be specified as the fourth argument,
7194 otherwise it should be set to 0 or 1 (both meaning no alignment).
7196 '``llvm.memset.*``' Intrinsics
7197 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7202 This is an overloaded intrinsic. You can use llvm.memset on any integer
7203 bit width and for different address spaces. However, not all targets
7204 support all bit widths.
7208 declare void @llvm.memset.p0i8.i32(i8* <dest>, i8 <val>,
7209 i32 <len>, i32 <align>, i1 <isvolatile>)
7210 declare void @llvm.memset.p0i8.i64(i8* <dest>, i8 <val>,
7211 i64 <len>, i32 <align>, i1 <isvolatile>)
7216 The '``llvm.memset.*``' intrinsics fill a block of memory with a
7217 particular byte value.
7219 Note that, unlike the standard libc function, the ``llvm.memset``
7220 intrinsic does not return a value and takes extra alignment/volatile
7221 arguments. Also, the destination can be in an arbitrary address space.
7226 The first argument is a pointer to the destination to fill, the second
7227 is the byte value with which to fill it, the third argument is an
7228 integer argument specifying the number of bytes to fill, and the fourth
7229 argument is the known alignment of the destination location.
7231 If the call to this intrinsic has an alignment value that is not 0 or 1,
7232 then the caller guarantees that the destination pointer is aligned to
7235 If the ``isvolatile`` parameter is ``true``, the ``llvm.memset`` call is
7236 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
7237 very cleanly specified and it is unwise to depend on it.
7242 The '``llvm.memset.*``' intrinsics fill "len" bytes of memory starting
7243 at the destination location. If the argument is known to be aligned to
7244 some boundary, this can be specified as the fourth argument, otherwise
7245 it should be set to 0 or 1 (both meaning no alignment).
7247 '``llvm.sqrt.*``' Intrinsic
7248 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7253 This is an overloaded intrinsic. You can use ``llvm.sqrt`` on any
7254 floating point or vector of floating point type. Not all targets support
7259 declare float @llvm.sqrt.f32(float %Val)
7260 declare double @llvm.sqrt.f64(double %Val)
7261 declare x86_fp80 @llvm.sqrt.f80(x86_fp80 %Val)
7262 declare fp128 @llvm.sqrt.f128(fp128 %Val)
7263 declare ppc_fp128 @llvm.sqrt.ppcf128(ppc_fp128 %Val)
7268 The '``llvm.sqrt``' intrinsics return the sqrt of the specified operand,
7269 returning the same value as the libm '``sqrt``' functions would. Unlike
7270 ``sqrt`` in libm, however, ``llvm.sqrt`` has undefined behavior for
7271 negative numbers other than -0.0 (which allows for better optimization,
7272 because there is no need to worry about errno being set).
7273 ``llvm.sqrt(-0.0)`` is defined to return -0.0 like IEEE sqrt.
7278 The argument and return value are floating point numbers of the same
7284 This function returns the sqrt of the specified operand if it is a
7285 nonnegative floating point number.
7287 '``llvm.powi.*``' Intrinsic
7288 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7293 This is an overloaded intrinsic. You can use ``llvm.powi`` on any
7294 floating point or vector of floating point type. Not all targets support
7299 declare float @llvm.powi.f32(float %Val, i32 %power)
7300 declare double @llvm.powi.f64(double %Val, i32 %power)
7301 declare x86_fp80 @llvm.powi.f80(x86_fp80 %Val, i32 %power)
7302 declare fp128 @llvm.powi.f128(fp128 %Val, i32 %power)
7303 declare ppc_fp128 @llvm.powi.ppcf128(ppc_fp128 %Val, i32 %power)
7308 The '``llvm.powi.*``' intrinsics return the first operand raised to the
7309 specified (positive or negative) power. The order of evaluation of
7310 multiplications is not defined. When a vector of floating point type is
7311 used, the second argument remains a scalar integer value.
7316 The second argument is an integer power, and the first is a value to
7317 raise to that power.
7322 This function returns the first value raised to the second power with an
7323 unspecified sequence of rounding operations.
7325 '``llvm.sin.*``' Intrinsic
7326 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7331 This is an overloaded intrinsic. You can use ``llvm.sin`` on any
7332 floating point or vector of floating point type. Not all targets support
7337 declare float @llvm.sin.f32(float %Val)
7338 declare double @llvm.sin.f64(double %Val)
7339 declare x86_fp80 @llvm.sin.f80(x86_fp80 %Val)
7340 declare fp128 @llvm.sin.f128(fp128 %Val)
7341 declare ppc_fp128 @llvm.sin.ppcf128(ppc_fp128 %Val)
7346 The '``llvm.sin.*``' intrinsics return the sine of the operand.
7351 The argument and return value are floating point numbers of the same
7357 This function returns the sine of the specified operand, returning the
7358 same values as the libm ``sin`` functions would, and handles error
7359 conditions in the same way.
7361 '``llvm.cos.*``' Intrinsic
7362 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7367 This is an overloaded intrinsic. You can use ``llvm.cos`` on any
7368 floating point or vector of floating point type. Not all targets support
7373 declare float @llvm.cos.f32(float %Val)
7374 declare double @llvm.cos.f64(double %Val)
7375 declare x86_fp80 @llvm.cos.f80(x86_fp80 %Val)
7376 declare fp128 @llvm.cos.f128(fp128 %Val)
7377 declare ppc_fp128 @llvm.cos.ppcf128(ppc_fp128 %Val)
7382 The '``llvm.cos.*``' intrinsics return the cosine of the operand.
7387 The argument and return value are floating point numbers of the same
7393 This function returns the cosine of the specified operand, returning the
7394 same values as the libm ``cos`` functions would, and handles error
7395 conditions in the same way.
7397 '``llvm.pow.*``' Intrinsic
7398 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7403 This is an overloaded intrinsic. You can use ``llvm.pow`` on any
7404 floating point or vector of floating point type. Not all targets support
7409 declare float @llvm.pow.f32(float %Val, float %Power)
7410 declare double @llvm.pow.f64(double %Val, double %Power)
7411 declare x86_fp80 @llvm.pow.f80(x86_fp80 %Val, x86_fp80 %Power)
7412 declare fp128 @llvm.pow.f128(fp128 %Val, fp128 %Power)
7413 declare ppc_fp128 @llvm.pow.ppcf128(ppc_fp128 %Val, ppc_fp128 Power)
7418 The '``llvm.pow.*``' intrinsics return the first operand raised to the
7419 specified (positive or negative) power.
7424 The second argument is a floating point power, and the first is a value
7425 to raise to that power.
7430 This function returns the first value raised to the second power,
7431 returning the same values as the libm ``pow`` functions would, and
7432 handles error conditions in the same way.
7434 '``llvm.exp.*``' Intrinsic
7435 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7440 This is an overloaded intrinsic. You can use ``llvm.exp`` on any
7441 floating point or vector of floating point type. Not all targets support
7446 declare float @llvm.exp.f32(float %Val)
7447 declare double @llvm.exp.f64(double %Val)
7448 declare x86_fp80 @llvm.exp.f80(x86_fp80 %Val)
7449 declare fp128 @llvm.exp.f128(fp128 %Val)
7450 declare ppc_fp128 @llvm.exp.ppcf128(ppc_fp128 %Val)
7455 The '``llvm.exp.*``' intrinsics perform the exp function.
7460 The argument and return value are floating point numbers of the same
7466 This function returns the same values as the libm ``exp`` functions
7467 would, and handles error conditions in the same way.
7469 '``llvm.exp2.*``' Intrinsic
7470 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7475 This is an overloaded intrinsic. You can use ``llvm.exp2`` on any
7476 floating point or vector of floating point type. Not all targets support
7481 declare float @llvm.exp2.f32(float %Val)
7482 declare double @llvm.exp2.f64(double %Val)
7483 declare x86_fp80 @llvm.exp2.f80(x86_fp80 %Val)
7484 declare fp128 @llvm.exp2.f128(fp128 %Val)
7485 declare ppc_fp128 @llvm.exp2.ppcf128(ppc_fp128 %Val)
7490 The '``llvm.exp2.*``' intrinsics perform the exp2 function.
7495 The argument and return value are floating point numbers of the same
7501 This function returns the same values as the libm ``exp2`` functions
7502 would, and handles error conditions in the same way.
7504 '``llvm.log.*``' Intrinsic
7505 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7510 This is an overloaded intrinsic. You can use ``llvm.log`` on any
7511 floating point or vector of floating point type. Not all targets support
7516 declare float @llvm.log.f32(float %Val)
7517 declare double @llvm.log.f64(double %Val)
7518 declare x86_fp80 @llvm.log.f80(x86_fp80 %Val)
7519 declare fp128 @llvm.log.f128(fp128 %Val)
7520 declare ppc_fp128 @llvm.log.ppcf128(ppc_fp128 %Val)
7525 The '``llvm.log.*``' intrinsics perform the log function.
7530 The argument and return value are floating point numbers of the same
7536 This function returns the same values as the libm ``log`` functions
7537 would, and handles error conditions in the same way.
7539 '``llvm.log10.*``' Intrinsic
7540 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7545 This is an overloaded intrinsic. You can use ``llvm.log10`` on any
7546 floating point or vector of floating point type. Not all targets support
7551 declare float @llvm.log10.f32(float %Val)
7552 declare double @llvm.log10.f64(double %Val)
7553 declare x86_fp80 @llvm.log10.f80(x86_fp80 %Val)
7554 declare fp128 @llvm.log10.f128(fp128 %Val)
7555 declare ppc_fp128 @llvm.log10.ppcf128(ppc_fp128 %Val)
7560 The '``llvm.log10.*``' intrinsics perform the log10 function.
7565 The argument and return value are floating point numbers of the same
7571 This function returns the same values as the libm ``log10`` functions
7572 would, and handles error conditions in the same way.
7574 '``llvm.log2.*``' Intrinsic
7575 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7580 This is an overloaded intrinsic. You can use ``llvm.log2`` on any
7581 floating point or vector of floating point type. Not all targets support
7586 declare float @llvm.log2.f32(float %Val)
7587 declare double @llvm.log2.f64(double %Val)
7588 declare x86_fp80 @llvm.log2.f80(x86_fp80 %Val)
7589 declare fp128 @llvm.log2.f128(fp128 %Val)
7590 declare ppc_fp128 @llvm.log2.ppcf128(ppc_fp128 %Val)
7595 The '``llvm.log2.*``' intrinsics perform the log2 function.
7600 The argument and return value are floating point numbers of the same
7606 This function returns the same values as the libm ``log2`` functions
7607 would, and handles error conditions in the same way.
7609 '``llvm.fma.*``' Intrinsic
7610 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7615 This is an overloaded intrinsic. You can use ``llvm.fma`` on any
7616 floating point or vector of floating point type. Not all targets support
7621 declare float @llvm.fma.f32(float %a, float %b, float %c)
7622 declare double @llvm.fma.f64(double %a, double %b, double %c)
7623 declare x86_fp80 @llvm.fma.f80(x86_fp80 %a, x86_fp80 %b, x86_fp80 %c)
7624 declare fp128 @llvm.fma.f128(fp128 %a, fp128 %b, fp128 %c)
7625 declare ppc_fp128 @llvm.fma.ppcf128(ppc_fp128 %a, ppc_fp128 %b, ppc_fp128 %c)
7630 The '``llvm.fma.*``' intrinsics perform the fused multiply-add
7636 The argument and return value are floating point numbers of the same
7642 This function returns the same values as the libm ``fma`` functions
7643 would, and does not set errno.
7645 '``llvm.fabs.*``' Intrinsic
7646 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7651 This is an overloaded intrinsic. You can use ``llvm.fabs`` on any
7652 floating point or vector of floating point type. Not all targets support
7657 declare float @llvm.fabs.f32(float %Val)
7658 declare double @llvm.fabs.f64(double %Val)
7659 declare x86_fp80 @llvm.fabs.f80(x86_fp80 %Val)
7660 declare fp128 @llvm.fabs.f128(fp128 %Val)
7661 declare ppc_fp128 @llvm.fabs.ppcf128(ppc_fp128 %Val)
7666 The '``llvm.fabs.*``' intrinsics return the absolute value of the
7672 The argument and return value are floating point numbers of the same
7678 This function returns the same values as the libm ``fabs`` functions
7679 would, and handles error conditions in the same way.
7681 '``llvm.copysign.*``' Intrinsic
7682 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7687 This is an overloaded intrinsic. You can use ``llvm.copysign`` on any
7688 floating point or vector of floating point type. Not all targets support
7693 declare float @llvm.copysign.f32(float %Mag, float %Sgn)
7694 declare double @llvm.copysign.f64(double %Mag, double %Sgn)
7695 declare x86_fp80 @llvm.copysign.f80(x86_fp80 %Mag, x86_fp80 %Sgn)
7696 declare fp128 @llvm.copysign.f128(fp128 %Mag, fp128 %Sgn)
7697 declare ppc_fp128 @llvm.copysign.ppcf128(ppc_fp128 %Mag, ppc_fp128 %Sgn)
7702 The '``llvm.copysign.*``' intrinsics return a value with the magnitude of the
7703 first operand and the sign of the second operand.
7708 The arguments and return value are floating point numbers of the same
7714 This function returns the same values as the libm ``copysign``
7715 functions would, and handles error conditions in the same way.
7717 '``llvm.floor.*``' Intrinsic
7718 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7723 This is an overloaded intrinsic. You can use ``llvm.floor`` on any
7724 floating point or vector of floating point type. Not all targets support
7729 declare float @llvm.floor.f32(float %Val)
7730 declare double @llvm.floor.f64(double %Val)
7731 declare x86_fp80 @llvm.floor.f80(x86_fp80 %Val)
7732 declare fp128 @llvm.floor.f128(fp128 %Val)
7733 declare ppc_fp128 @llvm.floor.ppcf128(ppc_fp128 %Val)
7738 The '``llvm.floor.*``' intrinsics return the floor of the operand.
7743 The argument and return value are floating point numbers of the same
7749 This function returns the same values as the libm ``floor`` functions
7750 would, and handles error conditions in the same way.
7752 '``llvm.ceil.*``' Intrinsic
7753 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7758 This is an overloaded intrinsic. You can use ``llvm.ceil`` on any
7759 floating point or vector of floating point type. Not all targets support
7764 declare float @llvm.ceil.f32(float %Val)
7765 declare double @llvm.ceil.f64(double %Val)
7766 declare x86_fp80 @llvm.ceil.f80(x86_fp80 %Val)
7767 declare fp128 @llvm.ceil.f128(fp128 %Val)
7768 declare ppc_fp128 @llvm.ceil.ppcf128(ppc_fp128 %Val)
7773 The '``llvm.ceil.*``' intrinsics return the ceiling of the operand.
7778 The argument and return value are floating point numbers of the same
7784 This function returns the same values as the libm ``ceil`` functions
7785 would, and handles error conditions in the same way.
7787 '``llvm.trunc.*``' Intrinsic
7788 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7793 This is an overloaded intrinsic. You can use ``llvm.trunc`` on any
7794 floating point or vector of floating point type. Not all targets support
7799 declare float @llvm.trunc.f32(float %Val)
7800 declare double @llvm.trunc.f64(double %Val)
7801 declare x86_fp80 @llvm.trunc.f80(x86_fp80 %Val)
7802 declare fp128 @llvm.trunc.f128(fp128 %Val)
7803 declare ppc_fp128 @llvm.trunc.ppcf128(ppc_fp128 %Val)
7808 The '``llvm.trunc.*``' intrinsics returns the operand rounded to the
7809 nearest integer not larger in magnitude than the operand.
7814 The argument and return value are floating point numbers of the same
7820 This function returns the same values as the libm ``trunc`` functions
7821 would, and handles error conditions in the same way.
7823 '``llvm.rint.*``' Intrinsic
7824 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7829 This is an overloaded intrinsic. You can use ``llvm.rint`` on any
7830 floating point or vector of floating point type. Not all targets support
7835 declare float @llvm.rint.f32(float %Val)
7836 declare double @llvm.rint.f64(double %Val)
7837 declare x86_fp80 @llvm.rint.f80(x86_fp80 %Val)
7838 declare fp128 @llvm.rint.f128(fp128 %Val)
7839 declare ppc_fp128 @llvm.rint.ppcf128(ppc_fp128 %Val)
7844 The '``llvm.rint.*``' intrinsics returns the operand rounded to the
7845 nearest integer. It may raise an inexact floating-point exception if the
7846 operand isn't an integer.
7851 The argument and return value are floating point numbers of the same
7857 This function returns the same values as the libm ``rint`` functions
7858 would, and handles error conditions in the same way.
7860 '``llvm.nearbyint.*``' Intrinsic
7861 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7866 This is an overloaded intrinsic. You can use ``llvm.nearbyint`` on any
7867 floating point or vector of floating point type. Not all targets support
7872 declare float @llvm.nearbyint.f32(float %Val)
7873 declare double @llvm.nearbyint.f64(double %Val)
7874 declare x86_fp80 @llvm.nearbyint.f80(x86_fp80 %Val)
7875 declare fp128 @llvm.nearbyint.f128(fp128 %Val)
7876 declare ppc_fp128 @llvm.nearbyint.ppcf128(ppc_fp128 %Val)
7881 The '``llvm.nearbyint.*``' intrinsics returns the operand rounded to the
7887 The argument and return value are floating point numbers of the same
7893 This function returns the same values as the libm ``nearbyint``
7894 functions would, and handles error conditions in the same way.
7896 '``llvm.round.*``' Intrinsic
7897 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7902 This is an overloaded intrinsic. You can use ``llvm.round`` on any
7903 floating point or vector of floating point type. Not all targets support
7908 declare float @llvm.round.f32(float %Val)
7909 declare double @llvm.round.f64(double %Val)
7910 declare x86_fp80 @llvm.round.f80(x86_fp80 %Val)
7911 declare fp128 @llvm.round.f128(fp128 %Val)
7912 declare ppc_fp128 @llvm.round.ppcf128(ppc_fp128 %Val)
7917 The '``llvm.round.*``' intrinsics returns the operand rounded to the
7923 The argument and return value are floating point numbers of the same
7929 This function returns the same values as the libm ``round``
7930 functions would, and handles error conditions in the same way.
7932 Bit Manipulation Intrinsics
7933 ---------------------------
7935 LLVM provides intrinsics for a few important bit manipulation
7936 operations. These allow efficient code generation for some algorithms.
7938 '``llvm.bswap.*``' Intrinsics
7939 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7944 This is an overloaded intrinsic function. You can use bswap on any
7945 integer type that is an even number of bytes (i.e. BitWidth % 16 == 0).
7949 declare i16 @llvm.bswap.i16(i16 <id>)
7950 declare i32 @llvm.bswap.i32(i32 <id>)
7951 declare i64 @llvm.bswap.i64(i64 <id>)
7956 The '``llvm.bswap``' family of intrinsics is used to byte swap integer
7957 values with an even number of bytes (positive multiple of 16 bits).
7958 These are useful for performing operations on data that is not in the
7959 target's native byte order.
7964 The ``llvm.bswap.i16`` intrinsic returns an i16 value that has the high
7965 and low byte of the input i16 swapped. Similarly, the ``llvm.bswap.i32``
7966 intrinsic returns an i32 value that has the four bytes of the input i32
7967 swapped, so that if the input bytes are numbered 0, 1, 2, 3 then the
7968 returned i32 will have its bytes in 3, 2, 1, 0 order. The
7969 ``llvm.bswap.i48``, ``llvm.bswap.i64`` and other intrinsics extend this
7970 concept to additional even-byte lengths (6 bytes, 8 bytes and more,
7973 '``llvm.ctpop.*``' Intrinsic
7974 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7979 This is an overloaded intrinsic. You can use llvm.ctpop on any integer
7980 bit width, or on any vector with integer elements. Not all targets
7981 support all bit widths or vector types, however.
7985 declare i8 @llvm.ctpop.i8(i8 <src>)
7986 declare i16 @llvm.ctpop.i16(i16 <src>)
7987 declare i32 @llvm.ctpop.i32(i32 <src>)
7988 declare i64 @llvm.ctpop.i64(i64 <src>)
7989 declare i256 @llvm.ctpop.i256(i256 <src>)
7990 declare <2 x i32> @llvm.ctpop.v2i32(<2 x i32> <src>)
7995 The '``llvm.ctpop``' family of intrinsics counts the number of bits set
8001 The only argument is the value to be counted. The argument may be of any
8002 integer type, or a vector with integer elements. The return type must
8003 match the argument type.
8008 The '``llvm.ctpop``' intrinsic counts the 1's in a variable, or within
8009 each element of a vector.
8011 '``llvm.ctlz.*``' Intrinsic
8012 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8017 This is an overloaded intrinsic. You can use ``llvm.ctlz`` on any
8018 integer bit width, or any vector whose elements are integers. Not all
8019 targets support all bit widths or vector types, however.
8023 declare i8 @llvm.ctlz.i8 (i8 <src>, i1 <is_zero_undef>)
8024 declare i16 @llvm.ctlz.i16 (i16 <src>, i1 <is_zero_undef>)
8025 declare i32 @llvm.ctlz.i32 (i32 <src>, i1 <is_zero_undef>)
8026 declare i64 @llvm.ctlz.i64 (i64 <src>, i1 <is_zero_undef>)
8027 declare i256 @llvm.ctlz.i256(i256 <src>, i1 <is_zero_undef>)
8028 declase <2 x i32> @llvm.ctlz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
8033 The '``llvm.ctlz``' family of intrinsic functions counts the number of
8034 leading zeros in a variable.
8039 The first argument is the value to be counted. This argument may be of
8040 any integer type, or a vectory with integer element type. The return
8041 type must match the first argument type.
8043 The second argument must be a constant and is a flag to indicate whether
8044 the intrinsic should ensure that a zero as the first argument produces a
8045 defined result. Historically some architectures did not provide a
8046 defined result for zero values as efficiently, and many algorithms are
8047 now predicated on avoiding zero-value inputs.
8052 The '``llvm.ctlz``' intrinsic counts the leading (most significant)
8053 zeros in a variable, or within each element of the vector. If
8054 ``src == 0`` then the result is the size in bits of the type of ``src``
8055 if ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
8056 ``llvm.ctlz(i32 2) = 30``.
8058 '``llvm.cttz.*``' Intrinsic
8059 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8064 This is an overloaded intrinsic. You can use ``llvm.cttz`` on any
8065 integer bit width, or any vector of integer elements. Not all targets
8066 support all bit widths or vector types, however.
8070 declare i8 @llvm.cttz.i8 (i8 <src>, i1 <is_zero_undef>)
8071 declare i16 @llvm.cttz.i16 (i16 <src>, i1 <is_zero_undef>)
8072 declare i32 @llvm.cttz.i32 (i32 <src>, i1 <is_zero_undef>)
8073 declare i64 @llvm.cttz.i64 (i64 <src>, i1 <is_zero_undef>)
8074 declare i256 @llvm.cttz.i256(i256 <src>, i1 <is_zero_undef>)
8075 declase <2 x i32> @llvm.cttz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
8080 The '``llvm.cttz``' family of intrinsic functions counts the number of
8086 The first argument is the value to be counted. This argument may be of
8087 any integer type, or a vectory with integer element type. The return
8088 type must match the first argument type.
8090 The second argument must be a constant and is a flag to indicate whether
8091 the intrinsic should ensure that a zero as the first argument produces a
8092 defined result. Historically some architectures did not provide a
8093 defined result for zero values as efficiently, and many algorithms are
8094 now predicated on avoiding zero-value inputs.
8099 The '``llvm.cttz``' intrinsic counts the trailing (least significant)
8100 zeros in a variable, or within each element of a vector. If ``src == 0``
8101 then the result is the size in bits of the type of ``src`` if
8102 ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
8103 ``llvm.cttz(2) = 1``.
8105 Arithmetic with Overflow Intrinsics
8106 -----------------------------------
8108 LLVM provides intrinsics for some arithmetic with overflow operations.
8110 '``llvm.sadd.with.overflow.*``' Intrinsics
8111 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8116 This is an overloaded intrinsic. You can use ``llvm.sadd.with.overflow``
8117 on any integer bit width.
8121 declare {i16, i1} @llvm.sadd.with.overflow.i16(i16 %a, i16 %b)
8122 declare {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
8123 declare {i64, i1} @llvm.sadd.with.overflow.i64(i64 %a, i64 %b)
8128 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
8129 a signed addition of the two arguments, and indicate whether an overflow
8130 occurred during the signed summation.
8135 The arguments (%a and %b) and the first element of the result structure
8136 may be of integer types of any bit width, but they must have the same
8137 bit width. The second element of the result structure must be of type
8138 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
8144 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
8145 a signed addition of the two variables. They return a structure --- the
8146 first element of which is the signed summation, and the second element
8147 of which is a bit specifying if the signed summation resulted in an
8153 .. code-block:: llvm
8155 %res = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
8156 %sum = extractvalue {i32, i1} %res, 0
8157 %obit = extractvalue {i32, i1} %res, 1
8158 br i1 %obit, label %overflow, label %normal
8160 '``llvm.uadd.with.overflow.*``' Intrinsics
8161 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8166 This is an overloaded intrinsic. You can use ``llvm.uadd.with.overflow``
8167 on any integer bit width.
8171 declare {i16, i1} @llvm.uadd.with.overflow.i16(i16 %a, i16 %b)
8172 declare {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
8173 declare {i64, i1} @llvm.uadd.with.overflow.i64(i64 %a, i64 %b)
8178 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
8179 an unsigned addition of the two arguments, and indicate whether a carry
8180 occurred during the unsigned summation.
8185 The arguments (%a and %b) and the first element of the result structure
8186 may be of integer types of any bit width, but they must have the same
8187 bit width. The second element of the result structure must be of type
8188 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8194 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
8195 an unsigned addition of the two arguments. They return a structure --- the
8196 first element of which is the sum, and the second element of which is a
8197 bit specifying if the unsigned summation resulted in a carry.
8202 .. code-block:: llvm
8204 %res = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
8205 %sum = extractvalue {i32, i1} %res, 0
8206 %obit = extractvalue {i32, i1} %res, 1
8207 br i1 %obit, label %carry, label %normal
8209 '``llvm.ssub.with.overflow.*``' Intrinsics
8210 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8215 This is an overloaded intrinsic. You can use ``llvm.ssub.with.overflow``
8216 on any integer bit width.
8220 declare {i16, i1} @llvm.ssub.with.overflow.i16(i16 %a, i16 %b)
8221 declare {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
8222 declare {i64, i1} @llvm.ssub.with.overflow.i64(i64 %a, i64 %b)
8227 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
8228 a signed subtraction of the two arguments, and indicate whether an
8229 overflow occurred during the signed subtraction.
8234 The arguments (%a and %b) and the first element of the result structure
8235 may be of integer types of any bit width, but they must have the same
8236 bit width. The second element of the result structure must be of type
8237 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
8243 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
8244 a signed subtraction of the two arguments. They return a structure --- the
8245 first element of which is the subtraction, and the second element of
8246 which is a bit specifying if the signed subtraction resulted in an
8252 .. code-block:: llvm
8254 %res = call {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
8255 %sum = extractvalue {i32, i1} %res, 0
8256 %obit = extractvalue {i32, i1} %res, 1
8257 br i1 %obit, label %overflow, label %normal
8259 '``llvm.usub.with.overflow.*``' Intrinsics
8260 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8265 This is an overloaded intrinsic. You can use ``llvm.usub.with.overflow``
8266 on any integer bit width.
8270 declare {i16, i1} @llvm.usub.with.overflow.i16(i16 %a, i16 %b)
8271 declare {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
8272 declare {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
8277 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
8278 an unsigned subtraction of the two arguments, and indicate whether an
8279 overflow occurred during the unsigned subtraction.
8284 The arguments (%a and %b) and the first element of the result structure
8285 may be of integer types of any bit width, but they must have the same
8286 bit width. The second element of the result structure must be of type
8287 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8293 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
8294 an unsigned subtraction of the two arguments. They return a structure ---
8295 the first element of which is the subtraction, and the second element of
8296 which is a bit specifying if the unsigned subtraction resulted in an
8302 .. code-block:: llvm
8304 %res = call {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
8305 %sum = extractvalue {i32, i1} %res, 0
8306 %obit = extractvalue {i32, i1} %res, 1
8307 br i1 %obit, label %overflow, label %normal
8309 '``llvm.smul.with.overflow.*``' Intrinsics
8310 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8315 This is an overloaded intrinsic. You can use ``llvm.smul.with.overflow``
8316 on any integer bit width.
8320 declare {i16, i1} @llvm.smul.with.overflow.i16(i16 %a, i16 %b)
8321 declare {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
8322 declare {i64, i1} @llvm.smul.with.overflow.i64(i64 %a, i64 %b)
8327 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
8328 a signed multiplication of the two arguments, and indicate whether an
8329 overflow occurred during the signed multiplication.
8334 The arguments (%a and %b) and the first element of the result structure
8335 may be of integer types of any bit width, but they must have the same
8336 bit width. The second element of the result structure must be of type
8337 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
8343 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
8344 a signed multiplication of the two arguments. They return a structure ---
8345 the first element of which is the multiplication, and the second element
8346 of which is a bit specifying if the signed multiplication resulted in an
8352 .. code-block:: llvm
8354 %res = call {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
8355 %sum = extractvalue {i32, i1} %res, 0
8356 %obit = extractvalue {i32, i1} %res, 1
8357 br i1 %obit, label %overflow, label %normal
8359 '``llvm.umul.with.overflow.*``' Intrinsics
8360 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8365 This is an overloaded intrinsic. You can use ``llvm.umul.with.overflow``
8366 on any integer bit width.
8370 declare {i16, i1} @llvm.umul.with.overflow.i16(i16 %a, i16 %b)
8371 declare {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
8372 declare {i64, i1} @llvm.umul.with.overflow.i64(i64 %a, i64 %b)
8377 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
8378 a unsigned multiplication of the two arguments, and indicate whether an
8379 overflow occurred during the unsigned multiplication.
8384 The arguments (%a and %b) and the first element of the result structure
8385 may be of integer types of any bit width, but they must have the same
8386 bit width. The second element of the result structure must be of type
8387 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8393 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
8394 an unsigned multiplication of the two arguments. They return a structure ---
8395 the first element of which is the multiplication, and the second
8396 element of which is a bit specifying if the unsigned multiplication
8397 resulted in an overflow.
8402 .. code-block:: llvm
8404 %res = call {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
8405 %sum = extractvalue {i32, i1} %res, 0
8406 %obit = extractvalue {i32, i1} %res, 1
8407 br i1 %obit, label %overflow, label %normal
8409 Specialised Arithmetic Intrinsics
8410 ---------------------------------
8412 '``llvm.fmuladd.*``' Intrinsic
8413 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8420 declare float @llvm.fmuladd.f32(float %a, float %b, float %c)
8421 declare double @llvm.fmuladd.f64(double %a, double %b, double %c)
8426 The '``llvm.fmuladd.*``' intrinsic functions represent multiply-add
8427 expressions that can be fused if the code generator determines that (a) the
8428 target instruction set has support for a fused operation, and (b) that the
8429 fused operation is more efficient than the equivalent, separate pair of mul
8430 and add instructions.
8435 The '``llvm.fmuladd.*``' intrinsics each take three arguments: two
8436 multiplicands, a and b, and an addend c.
8445 %0 = call float @llvm.fmuladd.f32(%a, %b, %c)
8447 is equivalent to the expression a \* b + c, except that rounding will
8448 not be performed between the multiplication and addition steps if the
8449 code generator fuses the operations. Fusion is not guaranteed, even if
8450 the target platform supports it. If a fused multiply-add is required the
8451 corresponding llvm.fma.\* intrinsic function should be used
8452 instead. This never sets errno, just as '``llvm.fma.*``'.
8457 .. code-block:: llvm
8459 %r2 = call float @llvm.fmuladd.f32(float %a, float %b, float %c) ; yields {float}:r2 = (a * b) + c
8461 Half Precision Floating Point Intrinsics
8462 ----------------------------------------
8464 For most target platforms, half precision floating point is a
8465 storage-only format. This means that it is a dense encoding (in memory)
8466 but does not support computation in the format.
8468 This means that code must first load the half-precision floating point
8469 value as an i16, then convert it to float with
8470 :ref:`llvm.convert.from.fp16 <int_convert_from_fp16>`. Computation can
8471 then be performed on the float value (including extending to double
8472 etc). To store the value back to memory, it is first converted to float
8473 if needed, then converted to i16 with
8474 :ref:`llvm.convert.to.fp16 <int_convert_to_fp16>`, then storing as an
8477 .. _int_convert_to_fp16:
8479 '``llvm.convert.to.fp16``' Intrinsic
8480 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8487 declare i16 @llvm.convert.to.fp16(f32 %a)
8492 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion
8493 from single precision floating point format to half precision floating
8499 The intrinsic function contains single argument - the value to be
8505 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion
8506 from single precision floating point format to half precision floating
8507 point format. The return value is an ``i16`` which contains the
8513 .. code-block:: llvm
8515 %res = call i16 @llvm.convert.to.fp16(f32 %a)
8516 store i16 %res, i16* @x, align 2
8518 .. _int_convert_from_fp16:
8520 '``llvm.convert.from.fp16``' Intrinsic
8521 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8528 declare f32 @llvm.convert.from.fp16(i16 %a)
8533 The '``llvm.convert.from.fp16``' intrinsic function performs a
8534 conversion from half precision floating point format to single precision
8535 floating point format.
8540 The intrinsic function contains single argument - the value to be
8546 The '``llvm.convert.from.fp16``' intrinsic function performs a
8547 conversion from half single precision floating point format to single
8548 precision floating point format. The input half-float value is
8549 represented by an ``i16`` value.
8554 .. code-block:: llvm
8556 %a = load i16* @x, align 2
8557 %res = call f32 @llvm.convert.from.fp16(i16 %a)
8562 The LLVM debugger intrinsics (which all start with ``llvm.dbg.``
8563 prefix), are described in the `LLVM Source Level
8564 Debugging <SourceLevelDebugging.html#format_common_intrinsics>`_
8567 Exception Handling Intrinsics
8568 -----------------------------
8570 The LLVM exception handling intrinsics (which all start with
8571 ``llvm.eh.`` prefix), are described in the `LLVM Exception
8572 Handling <ExceptionHandling.html#format_common_intrinsics>`_ document.
8576 Trampoline Intrinsics
8577 ---------------------
8579 These intrinsics make it possible to excise one parameter, marked with
8580 the :ref:`nest <nest>` attribute, from a function. The result is a
8581 callable function pointer lacking the nest parameter - the caller does
8582 not need to provide a value for it. Instead, the value to use is stored
8583 in advance in a "trampoline", a block of memory usually allocated on the
8584 stack, which also contains code to splice the nest value into the
8585 argument list. This is used to implement the GCC nested function address
8588 For example, if the function is ``i32 f(i8* nest %c, i32 %x, i32 %y)``
8589 then the resulting function pointer has signature ``i32 (i32, i32)*``.
8590 It can be created as follows:
8592 .. code-block:: llvm
8594 %tramp = alloca [10 x i8], align 4 ; size and alignment only correct for X86
8595 %tramp1 = getelementptr [10 x i8]* %tramp, i32 0, i32 0
8596 call i8* @llvm.init.trampoline(i8* %tramp1, i8* bitcast (i32 (i8*, i32, i32)* @f to i8*), i8* %nval)
8597 %p = call i8* @llvm.adjust.trampoline(i8* %tramp1)
8598 %fp = bitcast i8* %p to i32 (i32, i32)*
8600 The call ``%val = call i32 %fp(i32 %x, i32 %y)`` is then equivalent to
8601 ``%val = call i32 %f(i8* %nval, i32 %x, i32 %y)``.
8605 '``llvm.init.trampoline``' Intrinsic
8606 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8613 declare void @llvm.init.trampoline(i8* <tramp>, i8* <func>, i8* <nval>)
8618 This fills the memory pointed to by ``tramp`` with executable code,
8619 turning it into a trampoline.
8624 The ``llvm.init.trampoline`` intrinsic takes three arguments, all
8625 pointers. The ``tramp`` argument must point to a sufficiently large and
8626 sufficiently aligned block of memory; this memory is written to by the
8627 intrinsic. Note that the size and the alignment are target-specific -
8628 LLVM currently provides no portable way of determining them, so a
8629 front-end that generates this intrinsic needs to have some
8630 target-specific knowledge. The ``func`` argument must hold a function
8631 bitcast to an ``i8*``.
8636 The block of memory pointed to by ``tramp`` is filled with target
8637 dependent code, turning it into a function. Then ``tramp`` needs to be
8638 passed to :ref:`llvm.adjust.trampoline <int_at>` to get a pointer which can
8639 be :ref:`bitcast (to a new function) and called <int_trampoline>`. The new
8640 function's signature is the same as that of ``func`` with any arguments
8641 marked with the ``nest`` attribute removed. At most one such ``nest``
8642 argument is allowed, and it must be of pointer type. Calling the new
8643 function is equivalent to calling ``func`` with the same argument list,
8644 but with ``nval`` used for the missing ``nest`` argument. If, after
8645 calling ``llvm.init.trampoline``, the memory pointed to by ``tramp`` is
8646 modified, then the effect of any later call to the returned function
8647 pointer is undefined.
8651 '``llvm.adjust.trampoline``' Intrinsic
8652 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8659 declare i8* @llvm.adjust.trampoline(i8* <tramp>)
8664 This performs any required machine-specific adjustment to the address of
8665 a trampoline (passed as ``tramp``).
8670 ``tramp`` must point to a block of memory which already has trampoline
8671 code filled in by a previous call to
8672 :ref:`llvm.init.trampoline <int_it>`.
8677 On some architectures the address of the code to be executed needs to be
8678 different to the address where the trampoline is actually stored. This
8679 intrinsic returns the executable address corresponding to ``tramp``
8680 after performing the required machine specific adjustments. The pointer
8681 returned can then be :ref:`bitcast and executed <int_trampoline>`.
8686 This class of intrinsics exists to information about the lifetime of
8687 memory objects and ranges where variables are immutable.
8691 '``llvm.lifetime.start``' Intrinsic
8692 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8699 declare void @llvm.lifetime.start(i64 <size>, i8* nocapture <ptr>)
8704 The '``llvm.lifetime.start``' intrinsic specifies the start of a memory
8710 The first argument is a constant integer representing the size of the
8711 object, or -1 if it is variable sized. The second argument is a pointer
8717 This intrinsic indicates that before this point in the code, the value
8718 of the memory pointed to by ``ptr`` is dead. This means that it is known
8719 to never be used and has an undefined value. A load from the pointer
8720 that precedes this intrinsic can be replaced with ``'undef'``.
8724 '``llvm.lifetime.end``' Intrinsic
8725 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8732 declare void @llvm.lifetime.end(i64 <size>, i8* nocapture <ptr>)
8737 The '``llvm.lifetime.end``' intrinsic specifies the end of a memory
8743 The first argument is a constant integer representing the size of the
8744 object, or -1 if it is variable sized. The second argument is a pointer
8750 This intrinsic indicates that after this point in the code, the value of
8751 the memory pointed to by ``ptr`` is dead. This means that it is known to
8752 never be used and has an undefined value. Any stores into the memory
8753 object following this intrinsic may be removed as dead.
8755 '``llvm.invariant.start``' Intrinsic
8756 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8763 declare {}* @llvm.invariant.start(i64 <size>, i8* nocapture <ptr>)
8768 The '``llvm.invariant.start``' intrinsic specifies that the contents of
8769 a memory object will not change.
8774 The first argument is a constant integer representing the size of the
8775 object, or -1 if it is variable sized. The second argument is a pointer
8781 This intrinsic indicates that until an ``llvm.invariant.end`` that uses
8782 the return value, the referenced memory location is constant and
8785 '``llvm.invariant.end``' Intrinsic
8786 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8793 declare void @llvm.invariant.end({}* <start>, i64 <size>, i8* nocapture <ptr>)
8798 The '``llvm.invariant.end``' intrinsic specifies that the contents of a
8799 memory object are mutable.
8804 The first argument is the matching ``llvm.invariant.start`` intrinsic.
8805 The second argument is a constant integer representing the size of the
8806 object, or -1 if it is variable sized and the third argument is a
8807 pointer to the object.
8812 This intrinsic indicates that the memory is mutable again.
8817 This class of intrinsics is designed to be generic and has no specific
8820 '``llvm.var.annotation``' Intrinsic
8821 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8828 declare void @llvm.var.annotation(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
8833 The '``llvm.var.annotation``' intrinsic.
8838 The first argument is a pointer to a value, the second is a pointer to a
8839 global string, the third is a pointer to a global string which is the
8840 source file name, and the last argument is the line number.
8845 This intrinsic allows annotation of local variables with arbitrary
8846 strings. This can be useful for special purpose optimizations that want
8847 to look for these annotations. These have no other defined use; they are
8848 ignored by code generation and optimization.
8850 '``llvm.ptr.annotation.*``' Intrinsic
8851 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8856 This is an overloaded intrinsic. You can use '``llvm.ptr.annotation``' on a
8857 pointer to an integer of any width. *NOTE* you must specify an address space for
8858 the pointer. The identifier for the default address space is the integer
8863 declare i8* @llvm.ptr.annotation.p<address space>i8(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
8864 declare i16* @llvm.ptr.annotation.p<address space>i16(i16* <val>, i8* <str>, i8* <str>, i32 <int>)
8865 declare i32* @llvm.ptr.annotation.p<address space>i32(i32* <val>, i8* <str>, i8* <str>, i32 <int>)
8866 declare i64* @llvm.ptr.annotation.p<address space>i64(i64* <val>, i8* <str>, i8* <str>, i32 <int>)
8867 declare i256* @llvm.ptr.annotation.p<address space>i256(i256* <val>, i8* <str>, i8* <str>, i32 <int>)
8872 The '``llvm.ptr.annotation``' intrinsic.
8877 The first argument is a pointer to an integer value of arbitrary bitwidth
8878 (result of some expression), the second is a pointer to a global string, the
8879 third is a pointer to a global string which is the source file name, and the
8880 last argument is the line number. It returns the value of the first argument.
8885 This intrinsic allows annotation of a pointer to an integer with arbitrary
8886 strings. This can be useful for special purpose optimizations that want to look
8887 for these annotations. These have no other defined use; they are ignored by code
8888 generation and optimization.
8890 '``llvm.annotation.*``' Intrinsic
8891 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8896 This is an overloaded intrinsic. You can use '``llvm.annotation``' on
8897 any integer bit width.
8901 declare i8 @llvm.annotation.i8(i8 <val>, i8* <str>, i8* <str>, i32 <int>)
8902 declare i16 @llvm.annotation.i16(i16 <val>, i8* <str>, i8* <str>, i32 <int>)
8903 declare i32 @llvm.annotation.i32(i32 <val>, i8* <str>, i8* <str>, i32 <int>)
8904 declare i64 @llvm.annotation.i64(i64 <val>, i8* <str>, i8* <str>, i32 <int>)
8905 declare i256 @llvm.annotation.i256(i256 <val>, i8* <str>, i8* <str>, i32 <int>)
8910 The '``llvm.annotation``' intrinsic.
8915 The first argument is an integer value (result of some expression), the
8916 second is a pointer to a global string, the third is a pointer to a
8917 global string which is the source file name, and the last argument is
8918 the line number. It returns the value of the first argument.
8923 This intrinsic allows annotations to be put on arbitrary expressions
8924 with arbitrary strings. This can be useful for special purpose
8925 optimizations that want to look for these annotations. These have no
8926 other defined use; they are ignored by code generation and optimization.
8928 '``llvm.trap``' Intrinsic
8929 ^^^^^^^^^^^^^^^^^^^^^^^^^
8936 declare void @llvm.trap() noreturn nounwind
8941 The '``llvm.trap``' intrinsic.
8951 This intrinsic is lowered to the target dependent trap instruction. If
8952 the target does not have a trap instruction, this intrinsic will be
8953 lowered to a call of the ``abort()`` function.
8955 '``llvm.debugtrap``' Intrinsic
8956 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8963 declare void @llvm.debugtrap() nounwind
8968 The '``llvm.debugtrap``' intrinsic.
8978 This intrinsic is lowered to code which is intended to cause an
8979 execution trap with the intention of requesting the attention of a
8982 '``llvm.stackprotector``' Intrinsic
8983 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8990 declare void @llvm.stackprotector(i8* <guard>, i8** <slot>)
8995 The ``llvm.stackprotector`` intrinsic takes the ``guard`` and stores it
8996 onto the stack at ``slot``. The stack slot is adjusted to ensure that it
8997 is placed on the stack before local variables.
9002 The ``llvm.stackprotector`` intrinsic requires two pointer arguments.
9003 The first argument is the value loaded from the stack guard
9004 ``@__stack_chk_guard``. The second variable is an ``alloca`` that has
9005 enough space to hold the value of the guard.
9010 This intrinsic causes the prologue/epilogue inserter to force the position of
9011 the ``AllocaInst`` stack slot to be before local variables on the stack. This is
9012 to ensure that if a local variable on the stack is overwritten, it will destroy
9013 the value of the guard. When the function exits, the guard on the stack is
9014 checked against the original guard by ``llvm.stackprotectorcheck``. If they are
9015 different, then ``llvm.stackprotectorcheck`` causes the program to abort by
9016 calling the ``__stack_chk_fail()`` function.
9018 '``llvm.stackprotectorcheck``' Intrinsic
9019 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9026 declare void @llvm.stackprotectorcheck(i8** <guard>)
9031 The ``llvm.stackprotectorcheck`` intrinsic compares ``guard`` against an already
9032 created stack protector and if they are not equal calls the
9033 ``__stack_chk_fail()`` function.
9038 The ``llvm.stackprotectorcheck`` intrinsic requires one pointer argument, the
9039 the variable ``@__stack_chk_guard``.
9044 This intrinsic is provided to perform the stack protector check by comparing
9045 ``guard`` with the stack slot created by ``llvm.stackprotector`` and if the
9046 values do not match call the ``__stack_chk_fail()`` function.
9048 The reason to provide this as an IR level intrinsic instead of implementing it
9049 via other IR operations is that in order to perform this operation at the IR
9050 level without an intrinsic, one would need to create additional basic blocks to
9051 handle the success/failure cases. This makes it difficult to stop the stack
9052 protector check from disrupting sibling tail calls in Codegen. With this
9053 intrinsic, we are able to generate the stack protector basic blocks late in
9054 codegen after the tail call decision has occurred.
9056 '``llvm.objectsize``' Intrinsic
9057 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9064 declare i32 @llvm.objectsize.i32(i8* <object>, i1 <min>)
9065 declare i64 @llvm.objectsize.i64(i8* <object>, i1 <min>)
9070 The ``llvm.objectsize`` intrinsic is designed to provide information to
9071 the optimizers to determine at compile time whether a) an operation
9072 (like memcpy) will overflow a buffer that corresponds to an object, or
9073 b) that a runtime check for overflow isn't necessary. An object in this
9074 context means an allocation of a specific class, structure, array, or
9080 The ``llvm.objectsize`` intrinsic takes two arguments. The first
9081 argument is a pointer to or into the ``object``. The second argument is
9082 a boolean and determines whether ``llvm.objectsize`` returns 0 (if true)
9083 or -1 (if false) when the object size is unknown. The second argument
9084 only accepts constants.
9089 The ``llvm.objectsize`` intrinsic is lowered to a constant representing
9090 the size of the object concerned. If the size cannot be determined at
9091 compile time, ``llvm.objectsize`` returns ``i32/i64 -1 or 0`` (depending
9092 on the ``min`` argument).
9094 '``llvm.expect``' Intrinsic
9095 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
9100 This is an overloaded intrinsic. You can use ``llvm.expect`` on any
9105 declare i1 @llvm.expect.i1(i1 <val>, i1 <expected_val>)
9106 declare i32 @llvm.expect.i32(i32 <val>, i32 <expected_val>)
9107 declare i64 @llvm.expect.i64(i64 <val>, i64 <expected_val>)
9112 The ``llvm.expect`` intrinsic provides information about expected (the
9113 most probable) value of ``val``, which can be used by optimizers.
9118 The ``llvm.expect`` intrinsic takes two arguments. The first argument is
9119 a value. The second argument is an expected value, this needs to be a
9120 constant value, variables are not allowed.
9125 This intrinsic is lowered to the ``val``.
9127 '``llvm.donothing``' Intrinsic
9128 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
9135 declare void @llvm.donothing() nounwind readnone
9140 The ``llvm.donothing`` intrinsic doesn't perform any operation. It's the
9141 only intrinsic that can be called with an invoke instruction.
9151 This intrinsic does nothing, and it's removed by optimizers and ignored
9154 Stack Map Intrinsics
9155 --------------------
9157 LLVM provides experimental intrinsics to support runtime patching
9158 mechanisms commonly desired in dynamic language JITs. These intrinsics
9159 are described in :doc:`StackMaps`.