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).
133 It also shows a convention that we follow in this document. When
134 demonstrating instructions, we will follow an instruction with a comment
135 that defines the type and name of value produced.
143 LLVM programs are composed of ``Module``'s, each of which is a
144 translation unit of the input programs. Each module consists of
145 functions, global variables, and symbol table entries. Modules may be
146 combined together with the LLVM linker, which merges function (and
147 global variable) definitions, resolves forward declarations, and merges
148 symbol table entries. Here is an example of the "hello world" module:
152 ; Declare the string constant as a global constant.
153 @.str = private unnamed_addr constant [13 x i8] c"hello world\0A\00"
155 ; External declaration of the puts function
156 declare i32 @puts(i8* nocapture) nounwind
158 ; Definition of main function
159 define i32 @main() { ; i32()*
160 ; Convert [13 x i8]* to i8 *...
161 %cast210 = getelementptr [13 x i8]* @.str, i64 0, i64 0
163 ; Call puts function to write out the string to stdout.
164 call i32 @puts(i8* %cast210)
169 !1 = metadata !{i32 42}
172 This example is made up of a :ref:`global variable <globalvars>` named
173 "``.str``", an external declaration of the "``puts``" function, a
174 :ref:`function definition <functionstructure>` for "``main``" and
175 :ref:`named metadata <namedmetadatastructure>` "``foo``".
177 In general, a module is made up of a list of global values (where both
178 functions and global variables are global values). Global values are
179 represented by a pointer to a memory location (in this case, a pointer
180 to an array of char, and a pointer to a function), and have one of the
181 following :ref:`linkage types <linkage>`.
188 All Global Variables and Functions have one of the following types of
192 Global values with "``private``" linkage are only directly
193 accessible by objects in the current module. In particular, linking
194 code into a module with an private global value may cause the
195 private to be renamed as necessary to avoid collisions. Because the
196 symbol is private to the module, all references can be updated. This
197 doesn't show up in any symbol table in the object file.
199 Similar to ``private``, but the symbol is passed through the
200 assembler and evaluated by the linker. Unlike normal strong symbols,
201 they are removed by the linker from the final linked image
202 (executable or dynamic library).
203 ``linker_private_weak``
204 Similar to "``linker_private``", but the symbol is weak. Note that
205 ``linker_private_weak`` symbols are subject to coalescing by the
206 linker. The symbols are removed by the linker from the final linked
207 image (executable or dynamic library).
209 Similar to private, but the value shows as a local symbol
210 (``STB_LOCAL`` in the case of ELF) in the object file. This
211 corresponds to the notion of the '``static``' keyword in C.
212 ``available_externally``
213 Globals with "``available_externally``" linkage are never emitted
214 into the object file corresponding to the LLVM module. They exist to
215 allow inlining and other optimizations to take place given knowledge
216 of the definition of the global, which is known to be somewhere
217 outside the module. Globals with ``available_externally`` linkage
218 are allowed to be discarded at will, and are otherwise the same as
219 ``linkonce_odr``. This linkage type is only allowed on definitions,
222 Globals with "``linkonce``" linkage are merged with other globals of
223 the same name when linkage occurs. This can be used to implement
224 some forms of inline functions, templates, or other code which must
225 be generated in each translation unit that uses it, but where the
226 body may be overridden with a more definitive definition later.
227 Unreferenced ``linkonce`` globals are allowed to be discarded. Note
228 that ``linkonce`` linkage does not actually allow the optimizer to
229 inline the body of this function into callers because it doesn't
230 know if this definition of the function is the definitive definition
231 within the program or whether it will be overridden by a stronger
232 definition. To enable inlining and other optimizations, use
233 "``linkonce_odr``" linkage.
235 "``weak``" linkage has the same merging semantics as ``linkonce``
236 linkage, except that unreferenced globals with ``weak`` linkage may
237 not be discarded. This is used for globals that are declared "weak"
240 "``common``" linkage is most similar to "``weak``" linkage, but they
241 are used for tentative definitions in C, such as "``int X;``" at
242 global scope. Symbols with "``common``" linkage are merged in the
243 same way as ``weak symbols``, and they may not be deleted if
244 unreferenced. ``common`` symbols may not have an explicit section,
245 must have a zero initializer, and may not be marked
246 ':ref:`constant <globalvars>`'. Functions and aliases may not have
249 .. _linkage_appending:
252 "``appending``" linkage may only be applied to global variables of
253 pointer to array type. When two global variables with appending
254 linkage are linked together, the two global arrays are appended
255 together. This is the LLVM, typesafe, equivalent of having the
256 system linker append together "sections" with identical names when
259 The semantics of this linkage follow the ELF object file model: the
260 symbol is weak until linked, if not linked, the symbol becomes null
261 instead of being an undefined reference.
262 ``linkonce_odr``, ``weak_odr``
263 Some languages allow differing globals to be merged, such as two
264 functions with different semantics. Other languages, such as
265 ``C++``, ensure that only equivalent globals are ever merged (the
266 "one definition rule" --- "ODR"). Such languages can use the
267 ``linkonce_odr`` and ``weak_odr`` linkage types to indicate that the
268 global will only be merged with equivalent globals. These linkage
269 types are otherwise the same as their non-``odr`` versions.
270 ``linkonce_odr_auto_hide``
271 Similar to "``linkonce_odr``", but nothing in the translation unit
272 takes the address of this definition. For instance, functions that
273 had an inline definition, but the compiler decided not to inline it.
274 ``linkonce_odr_auto_hide`` may have only ``default`` visibility. The
275 symbols are removed by the linker from the final linked image
276 (executable or dynamic library).
278 If none of the above identifiers are used, the global is externally
279 visible, meaning that it participates in linkage and can be used to
280 resolve external symbol references.
282 The next two types of linkage are targeted for Microsoft Windows
283 platform only. They are designed to support importing (exporting)
284 symbols from (to) DLLs (Dynamic Link Libraries).
287 "``dllimport``" linkage causes the compiler to reference a function
288 or variable via a global pointer to a pointer that is set up by the
289 DLL exporting the symbol. On Microsoft Windows targets, the pointer
290 name is formed by combining ``__imp_`` and the function or variable
293 "``dllexport``" linkage causes the compiler to provide a global
294 pointer to a pointer in a DLL, so that it can be referenced with the
295 ``dllimport`` attribute. On Microsoft Windows targets, the pointer
296 name is formed by combining ``__imp_`` and the function or variable
299 For example, since the "``.LC0``" variable is defined to be internal, if
300 another module defined a "``.LC0``" variable and was linked with this
301 one, one of the two would be renamed, preventing a collision. Since
302 "``main``" and "``puts``" are external (i.e., lacking any linkage
303 declarations), they are accessible outside of the current module.
305 It is illegal for a function *declaration* to have any linkage type
306 other than ``external``, ``dllimport`` or ``extern_weak``.
308 Aliases can have only ``external``, ``internal``, ``weak`` or
309 ``weak_odr`` linkages.
316 LLVM :ref:`functions <functionstructure>`, :ref:`calls <i_call>` and
317 :ref:`invokes <i_invoke>` can all have an optional calling convention
318 specified for the call. The calling convention of any pair of dynamic
319 caller/callee must match, or the behavior of the program is undefined.
320 The following calling conventions are supported by LLVM, and more may be
323 "``ccc``" - The C calling convention
324 This calling convention (the default if no other calling convention
325 is specified) matches the target C calling conventions. This calling
326 convention supports varargs function calls and tolerates some
327 mismatch in the declared prototype and implemented declaration of
328 the function (as does normal C).
329 "``fastcc``" - The fast calling convention
330 This calling convention attempts to make calls as fast as possible
331 (e.g. by passing things in registers). This calling convention
332 allows the target to use whatever tricks it wants to produce fast
333 code for the target, without having to conform to an externally
334 specified ABI (Application Binary Interface). `Tail calls can only
335 be optimized when this, the GHC or the HiPE convention is
336 used. <CodeGenerator.html#id80>`_ This calling convention does not
337 support varargs and requires the prototype of all callees to exactly
338 match the prototype of the function definition.
339 "``coldcc``" - The cold calling convention
340 This calling convention attempts to make code in the caller as
341 efficient as possible under the assumption that the call is not
342 commonly executed. As such, these calls often preserve all registers
343 so that the call does not break any live ranges in the caller side.
344 This calling convention does not support varargs and requires the
345 prototype of all callees to exactly match the prototype of the
347 "``cc 10``" - GHC convention
348 This calling convention has been implemented specifically for use by
349 the `Glasgow Haskell Compiler (GHC) <http://www.haskell.org/ghc>`_.
350 It passes everything in registers, going to extremes to achieve this
351 by disabling callee save registers. This calling convention should
352 not be used lightly but only for specific situations such as an
353 alternative to the *register pinning* performance technique often
354 used when implementing functional programming languages. At the
355 moment only X86 supports this convention and it has the following
358 - On *X86-32* only supports up to 4 bit type parameters. No
359 floating point types are supported.
360 - On *X86-64* only supports up to 10 bit type parameters and 6
361 floating point parameters.
363 This calling convention supports `tail call
364 optimization <CodeGenerator.html#id80>`_ but requires both the
365 caller and callee are using it.
366 "``cc 11``" - The HiPE calling convention
367 This calling convention has been implemented specifically for use by
368 the `High-Performance Erlang
369 (HiPE) <http://www.it.uu.se/research/group/hipe/>`_ compiler, *the*
370 native code compiler of the `Ericsson's Open Source Erlang/OTP
371 system <http://www.erlang.org/download.shtml>`_. It uses more
372 registers for argument passing than the ordinary C calling
373 convention and defines no callee-saved registers. The calling
374 convention properly supports `tail call
375 optimization <CodeGenerator.html#id80>`_ but requires that both the
376 caller and the callee use it. It uses a *register pinning*
377 mechanism, similar to GHC's convention, for keeping frequently
378 accessed runtime components pinned to specific hardware registers.
379 At the moment only X86 supports this convention (both 32 and 64
381 "``cc <n>``" - Numbered convention
382 Any calling convention may be specified by number, allowing
383 target-specific calling conventions to be used. Target specific
384 calling conventions start at 64.
386 More calling conventions can be added/defined on an as-needed basis, to
387 support Pascal conventions or any other well-known target-independent
390 .. _visibilitystyles:
395 All Global Variables and Functions have one of the following visibility
398 "``default``" - Default style
399 On targets that use the ELF object file format, default visibility
400 means that the declaration is visible to other modules and, in
401 shared libraries, means that the declared entity may be overridden.
402 On Darwin, default visibility means that the declaration is visible
403 to other modules. Default visibility corresponds to "external
404 linkage" in the language.
405 "``hidden``" - Hidden style
406 Two declarations of an object with hidden visibility refer to the
407 same object if they are in the same shared object. Usually, hidden
408 visibility indicates that the symbol will not be placed into the
409 dynamic symbol table, so no other module (executable or shared
410 library) can reference it directly.
411 "``protected``" - Protected style
412 On ELF, protected visibility indicates that the symbol will be
413 placed in the dynamic symbol table, but that references within the
414 defining module will bind to the local symbol. That is, the symbol
415 cannot be overridden by another module.
422 LLVM IR allows you to specify name aliases for certain types. This can
423 make it easier to read the IR and make the IR more condensed
424 (particularly when recursive types are involved). An example of a name
429 %mytype = type { %mytype*, i32 }
431 You may give a name to any :ref:`type <typesystem>` except
432 ":ref:`void <t_void>`". Type name aliases may be used anywhere a type is
433 expected with the syntax "%mytype".
435 Note that type names are aliases for the structural type that they
436 indicate, and that you can therefore specify multiple names for the same
437 type. This often leads to confusing behavior when dumping out a .ll
438 file. Since LLVM IR uses structural typing, the name is not part of the
439 type. When printing out LLVM IR, the printer will pick *one name* to
440 render all types of a particular shape. This means that if you have code
441 where two different source types end up having the same LLVM type, that
442 the dumper will sometimes print the "wrong" or unexpected type. This is
443 an important design point and isn't going to change.
450 Global variables define regions of memory allocated at compilation time
451 instead of run-time. Global variables may optionally be initialized, may
452 have an explicit section to be placed in, and may have an optional
453 explicit alignment specified.
455 A variable may be defined as ``thread_local``, which means that it will
456 not be shared by threads (each thread will have a separated copy of the
457 variable). Not all targets support thread-local variables. Optionally, a
458 TLS model may be specified:
461 For variables that are only used within the current shared library.
463 For variables in modules that will not be loaded dynamically.
465 For variables defined in the executable and only used within it.
467 The models correspond to the ELF TLS models; see `ELF Handling For
468 Thread-Local Storage <http://people.redhat.com/drepper/tls.pdf>`_ for
469 more information on under which circumstances the different models may
470 be used. The target may choose a different TLS model if the specified
471 model is not supported, or if a better choice of model can be made.
473 A variable may be defined as a global ``constant``, which indicates that
474 the contents of the variable will **never** be modified (enabling better
475 optimization, allowing the global data to be placed in the read-only
476 section of an executable, etc). Note that variables that need runtime
477 initialization cannot be marked ``constant`` as there is a store to the
480 LLVM explicitly allows *declarations* of global variables to be marked
481 constant, even if the final definition of the global is not. This
482 capability can be used to enable slightly better optimization of the
483 program, but requires the language definition to guarantee that
484 optimizations based on the 'constantness' are valid for the translation
485 units that do not include the definition.
487 As SSA values, global variables define pointer values that are in scope
488 (i.e. they dominate) all basic blocks in the program. Global variables
489 always define a pointer to their "content" type because they describe a
490 region of memory, and all memory objects in LLVM are accessed through
493 Global variables can be marked with ``unnamed_addr`` which indicates
494 that the address is not significant, only the content. Constants marked
495 like this can be merged with other constants if they have the same
496 initializer. Note that a constant with significant address *can* be
497 merged with a ``unnamed_addr`` constant, the result being a constant
498 whose address is significant.
500 A global variable may be declared to reside in a target-specific
501 numbered address space. For targets that support them, address spaces
502 may affect how optimizations are performed and/or what target
503 instructions are used to access the variable. The default address space
504 is zero. The address space qualifier must precede any other attributes.
506 LLVM allows an explicit section to be specified for globals. If the
507 target supports it, it will emit globals to the section specified.
509 By default, global initializers are optimized by assuming that global
510 variables defined within the module are not modified from their
511 initial values before the start of the global initializer. This is
512 true even for variables potentially accessible from outside the
513 module, including those with external linkage or appearing in
514 ``@llvm.used``. This assumption may be suppressed by marking the
515 variable with ``externally_initialized``.
517 An explicit alignment may be specified for a global, which must be a
518 power of 2. If not present, or if the alignment is set to zero, the
519 alignment of the global is set by the target to whatever it feels
520 convenient. If an explicit alignment is specified, the global is forced
521 to have exactly that alignment. Targets and optimizers are not allowed
522 to over-align the global if the global has an assigned section. In this
523 case, the extra alignment could be observable: for example, code could
524 assume that the globals are densely packed in their section and try to
525 iterate over them as an array, alignment padding would break this
528 For example, the following defines a global in a numbered address space
529 with an initializer, section, and alignment:
533 @G = addrspace(5) constant float 1.0, section "foo", align 4
535 The following example defines a thread-local global with the
536 ``initialexec`` TLS model:
540 @G = thread_local(initialexec) global i32 0, align 4
542 .. _functionstructure:
547 LLVM function definitions consist of the "``define``" keyword, an
548 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
549 style <visibility>`, an optional :ref:`calling convention <callingconv>`,
550 an optional ``unnamed_addr`` attribute, a return type, an optional
551 :ref:`parameter attribute <paramattrs>` for the return type, a function
552 name, a (possibly empty) argument list (each with optional :ref:`parameter
553 attributes <paramattrs>`), optional :ref:`function attributes <fnattrs>`,
554 an optional section, an optional alignment, an optional :ref:`garbage
555 collector name <gc>`, an opening curly brace, a list of basic blocks,
556 and a closing curly brace.
558 LLVM function declarations consist of the "``declare``" keyword, an
559 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
560 style <visibility>`, an optional :ref:`calling convention <callingconv>`,
561 an optional ``unnamed_addr`` attribute, a return type, an optional
562 :ref:`parameter attribute <paramattrs>` for the return type, a function
563 name, a possibly empty list of arguments, an optional alignment, and an
564 optional :ref:`garbage collector name <gc>`.
566 A function definition contains a list of basic blocks, forming the CFG
567 (Control Flow Graph) for the function. Each basic block may optionally
568 start with a label (giving the basic block a symbol table entry),
569 contains a list of instructions, and ends with a
570 :ref:`terminator <terminators>` instruction (such as a branch or function
571 return). If explicit label is not provided, a block is assigned an
572 implicit numbered label, using a next value from the same counter as used
573 for unnamed temporaries (:ref:`see above<identifiers>`). For example, if a
574 function entry block does not have explicit label, it will be assigned
575 label "%0", then first unnamed temporary in that block will be "%1", etc.
577 The first basic block in a function is special in two ways: it is
578 immediately executed on entrance to the function, and it is not allowed
579 to have predecessor basic blocks (i.e. there can not be any branches to
580 the entry block of a function). Because the block can have no
581 predecessors, it also cannot have any :ref:`PHI nodes <i_phi>`.
583 LLVM allows an explicit section to be specified for functions. If the
584 target supports it, it will emit functions to the section specified.
586 An explicit alignment may be specified for a function. If not present,
587 or if the alignment is set to zero, the alignment of the function is set
588 by the target to whatever it feels convenient. If an explicit alignment
589 is specified, the function is forced to have at least that much
590 alignment. All alignments must be a power of 2.
592 If the ``unnamed_addr`` attribute is given, the address is know to not
593 be significant and two identical functions can be merged.
597 define [linkage] [visibility]
599 <ResultType> @<FunctionName> ([argument list])
600 [fn Attrs] [section "name"] [align N]
608 Aliases act as "second name" for the aliasee value (which can be either
609 function, global variable, another alias or bitcast of global value).
610 Aliases may have an optional :ref:`linkage type <linkage>`, and an optional
611 :ref:`visibility style <visibility>`.
615 @<Name> = alias [Linkage] [Visibility] <AliaseeTy> @<Aliasee>
617 .. _namedmetadatastructure:
622 Named metadata is a collection of metadata. :ref:`Metadata
623 nodes <metadata>` (but not metadata strings) are the only valid
624 operands for a named metadata.
628 ; Some unnamed metadata nodes, which are referenced by the named metadata.
629 !0 = metadata !{metadata !"zero"}
630 !1 = metadata !{metadata !"one"}
631 !2 = metadata !{metadata !"two"}
633 !name = !{!0, !1, !2}
640 The return type and each parameter of a function type may have a set of
641 *parameter attributes* associated with them. Parameter attributes are
642 used to communicate additional information about the result or
643 parameters of a function. Parameter attributes are considered to be part
644 of the function, not of the function type, so functions with different
645 parameter attributes can have the same function type.
647 Parameter attributes are simple keywords that follow the type specified.
648 If multiple parameter attributes are needed, they are space separated.
653 declare i32 @printf(i8* noalias nocapture, ...)
654 declare i32 @atoi(i8 zeroext)
655 declare signext i8 @returns_signed_char()
657 Note that any attributes for the function result (``nounwind``,
658 ``readonly``) come immediately after the argument list.
660 Currently, only the following parameter attributes are defined:
663 This indicates to the code generator that the parameter or return
664 value should be zero-extended to the extent required by the target's
665 ABI (which is usually 32-bits, but is 8-bits for a i1 on x86-64) by
666 the caller (for a parameter) or the callee (for a return value).
668 This indicates to the code generator that the parameter or return
669 value should be sign-extended to the extent required by the target's
670 ABI (which is usually 32-bits) by the caller (for a parameter) or
671 the callee (for a return value).
673 This indicates that this parameter or return value should be treated
674 in a special target-dependent fashion during while emitting code for
675 a function call or return (usually, by putting it in a register as
676 opposed to memory, though some targets use it to distinguish between
677 two different kinds of registers). Use of this attribute is
680 This indicates that the pointer parameter should really be passed by
681 value to the function. The attribute implies that a hidden copy of
682 the pointee is made between the caller and the callee, so the callee
683 is unable to modify the value in the caller. This attribute is only
684 valid on LLVM pointer arguments. It is generally used to pass
685 structs and arrays by value, but is also valid on pointers to
686 scalars. The copy is considered to belong to the caller not the
687 callee (for example, ``readonly`` functions should not write to
688 ``byval`` parameters). This is not a valid attribute for return
691 The byval attribute also supports specifying an alignment with the
692 align attribute. It indicates the alignment of the stack slot to
693 form and the known alignment of the pointer specified to the call
694 site. If the alignment is not specified, then the code generator
695 makes a target-specific assumption.
698 This indicates that the pointer parameter specifies the address of a
699 structure that is the return value of the function in the source
700 program. This pointer must be guaranteed by the caller to be valid:
701 loads and stores to the structure may be assumed by the callee
702 not to trap and to be properly aligned. This may only be applied to
703 the first parameter. This is not a valid attribute for return
706 This indicates that pointer values :ref:`based <pointeraliasing>` on
707 the argument or return value do not alias pointer values which are
708 not *based* on it, ignoring certain "irrelevant" dependencies. For a
709 call to the parent function, dependencies between memory references
710 from before or after the call and from those during the call are
711 "irrelevant" to the ``noalias`` keyword for the arguments and return
712 value used in that call. The caller shares the responsibility with
713 the callee for ensuring that these requirements are met. For further
714 details, please see the discussion of the NoAlias response in `alias
715 analysis <AliasAnalysis.html#MustMayNo>`_.
717 Note that this definition of ``noalias`` is intentionally similar
718 to the definition of ``restrict`` in C99 for function arguments,
719 though it is slightly weaker.
721 For function return values, C99's ``restrict`` is not meaningful,
722 while LLVM's ``noalias`` is.
724 This indicates that the callee does not make any copies of the
725 pointer that outlive the callee itself. This is not a valid
726 attribute for return values.
731 This indicates that the pointer parameter can be excised using the
732 :ref:`trampoline intrinsics <int_trampoline>`. This is not a valid
733 attribute for return values and can only be applied to one parameter.
736 This indicates that the function always returns the argument as its return
737 value. This is an optimization hint to the code generator when generating
738 the caller, allowing tail call optimization and omission of register saves
739 and restores in some cases; it is not checked or enforced when generating
740 the callee. The parameter and the function return type must be valid
741 operands for the :ref:`bitcast instruction <i_bitcast>`. This is not a
742 valid attribute for return values and can only be applied to one parameter.
746 Garbage Collector Names
747 -----------------------
749 Each function may specify a garbage collector name, which is simply a
754 define void @f() gc "name" { ... }
756 The compiler declares the supported values of *name*. Specifying a
757 collector which will cause the compiler to alter its output in order to
758 support the named garbage collection algorithm.
765 Attribute groups are groups of attributes that are referenced by objects within
766 the IR. They are important for keeping ``.ll`` files readable, because a lot of
767 functions will use the same set of attributes. In the degenerative case of a
768 ``.ll`` file that corresponds to a single ``.c`` file, the single attribute
769 group will capture the important command line flags used to build that file.
771 An attribute group is a module-level object. To use an attribute group, an
772 object references the attribute group's ID (e.g. ``#37``). An object may refer
773 to more than one attribute group. In that situation, the attributes from the
774 different groups are merged.
776 Here is an example of attribute groups for a function that should always be
777 inlined, has a stack alignment of 4, and which shouldn't use SSE instructions:
781 ; Target-independent attributes:
782 attributes #0 = { alwaysinline alignstack=4 }
784 ; Target-dependent attributes:
785 attributes #1 = { "no-sse" }
787 ; Function @f has attributes: alwaysinline, alignstack=4, and "no-sse".
788 define void @f() #0 #1 { ... }
795 Function attributes are set to communicate additional information about
796 a function. Function attributes are considered to be part of the
797 function, not of the function type, so functions with different function
798 attributes can have the same function type.
800 Function attributes are simple keywords that follow the type specified.
801 If multiple attributes are needed, they are space separated. For
806 define void @f() noinline { ... }
807 define void @f() alwaysinline { ... }
808 define void @f() alwaysinline optsize { ... }
809 define void @f() optsize { ... }
812 This attribute indicates that, when emitting the prologue and
813 epilogue, the backend should forcibly align the stack pointer.
814 Specify the desired alignment, which must be a power of two, in
817 This attribute indicates that the inliner should attempt to inline
818 this function into callers whenever possible, ignoring any active
819 inlining size threshold for this caller.
821 This attribute indicates that this function is rarely called. When
822 computing edge weights, basic blocks post-dominated by a cold
823 function call are also considered to be cold; and, thus, given low
826 This attribute suppresses lazy symbol binding for the function. This
827 may make calls to the function faster, at the cost of extra program
828 startup time if the function is not called during program startup.
830 This attribute indicates that the source code contained a hint that
831 inlining this function is desirable (such as the "inline" keyword in
832 C/C++). It is just a hint; it imposes no requirements on the
835 This attribute disables prologue / epilogue emission for the
836 function. This can have very system-specific consequences.
838 This indicates that the callee function at a call site is not
839 recognized as a built-in function. LLVM will retain the original call
840 and not replace it with equivalent code based on the semantics of the
841 built-in function. This is only valid at call sites, not on function
842 declarations or definitions.
844 This attribute indicates that calls to the function cannot be
845 duplicated. A call to a ``noduplicate`` function may be moved
846 within its parent function, but may not be duplicated within
849 A function containing a ``noduplicate`` call may still
850 be an inlining candidate, provided that the call is not
851 duplicated by inlining. That implies that the function has
852 internal linkage and only has one call site, so the original
853 call is dead after inlining.
855 This attributes disables implicit floating point instructions.
857 This attribute indicates that the inliner should never inline this
858 function in any situation. This attribute may not be used together
859 with the ``alwaysinline`` attribute.
861 This attribute indicates that the code generator should not use a
862 red zone, even if the target-specific ABI normally permits it.
864 This function attribute indicates that the function never returns
865 normally. This produces undefined behavior at runtime if the
866 function ever does dynamically return.
868 This function attribute indicates that the function never returns
869 with an unwind or exceptional control flow. If the function does
870 unwind, its runtime behavior is undefined.
872 This attribute suggests that optimization passes and code generator
873 passes make choices that keep the code size of this function low,
874 and otherwise do optimizations specifically to reduce code size.
876 This attribute indicates that the function computes its result (or
877 decides to unwind an exception) based strictly on its arguments,
878 without dereferencing any pointer arguments or otherwise accessing
879 any mutable state (e.g. memory, control registers, etc) visible to
880 caller functions. It does not write through any pointer arguments
881 (including ``byval`` arguments) and never changes any state visible
882 to callers. This means that it cannot unwind exceptions by calling
883 the ``C++`` exception throwing methods.
885 This attribute indicates that the function does not write through
886 any pointer arguments (including ``byval`` arguments) or otherwise
887 modify any state (e.g. memory, control registers, etc) visible to
888 caller functions. It may dereference pointer arguments and read
889 state that may be set in the caller. A readonly function always
890 returns the same value (or unwinds an exception identically) when
891 called with the same set of arguments and global state. It cannot
892 unwind an exception by calling the ``C++`` exception throwing
895 This attribute indicates that this function can return twice. The C
896 ``setjmp`` is an example of such a function. The compiler disables
897 some optimizations (like tail calls) in the caller of these
900 This attribute indicates that AddressSanitizer checks
901 (dynamic address safety analysis) are enabled for this function.
903 This attribute indicates that MemorySanitizer checks (dynamic detection
904 of accesses to uninitialized memory) are enabled for this function.
906 This attribute indicates that ThreadSanitizer checks
907 (dynamic thread safety analysis) are enabled for this function.
909 This attribute indicates that the function should emit a stack
910 smashing protector. It is in the form of a "canary" --- a random value
911 placed on the stack before the local variables that's checked upon
912 return from the function to see if it has been overwritten. A
913 heuristic is used to determine if a function needs stack protectors
914 or not. The heuristic used will enable protectors for functions with:
916 - Character arrays larger than ``ssp-buffer-size`` (default 8).
917 - Aggregates containing character arrays larger than ``ssp-buffer-size``.
918 - Calls to alloca() with variable sizes or constant sizes greater than
921 If a function that has an ``ssp`` attribute is inlined into a
922 function that doesn't have an ``ssp`` attribute, then the resulting
923 function will have an ``ssp`` attribute.
925 This attribute indicates that the function should *always* emit a
926 stack smashing protector. This overrides the ``ssp`` function
929 If a function that has an ``sspreq`` attribute is inlined into a
930 function that doesn't have an ``sspreq`` attribute or which has an
931 ``ssp`` or ``sspstrong`` attribute, then the resulting function will have
932 an ``sspreq`` attribute.
934 This attribute indicates that the function should emit a stack smashing
935 protector. This attribute causes a strong heuristic to be used when
936 determining if a function needs stack protectors. The strong heuristic
937 will enable protectors for functions with:
939 - Arrays of any size and type
940 - Aggregates containing an array of any size and type.
942 - Local variables that have had their address taken.
944 This overrides the ``ssp`` function attribute.
946 If a function that has an ``sspstrong`` attribute is inlined into a
947 function that doesn't have an ``sspstrong`` attribute, then the
948 resulting function will have an ``sspstrong`` attribute.
950 This attribute indicates that the ABI being targeted requires that
951 an unwind table entry be produce for this function even if we can
952 show that no exceptions passes by it. This is normally the case for
953 the ELF x86-64 abi, but it can be disabled for some compilation
958 Module-Level Inline Assembly
959 ----------------------------
961 Modules may contain "module-level inline asm" blocks, which corresponds
962 to the GCC "file scope inline asm" blocks. These blocks are internally
963 concatenated by LLVM and treated as a single unit, but may be separated
964 in the ``.ll`` file if desired. The syntax is very simple:
968 module asm "inline asm code goes here"
969 module asm "more can go here"
971 The strings can contain any character by escaping non-printable
972 characters. The escape sequence used is simply "\\xx" where "xx" is the
973 two digit hex code for the number.
975 The inline asm code is simply printed to the machine code .s file when
976 assembly code is generated.
978 .. _langref_datalayout:
983 A module may specify a target specific data layout string that specifies
984 how data is to be laid out in memory. The syntax for the data layout is
989 target datalayout = "layout specification"
991 The *layout specification* consists of a list of specifications
992 separated by the minus sign character ('-'). Each specification starts
993 with a letter and may include other information after the letter to
994 define some aspect of the data layout. The specifications accepted are
998 Specifies that the target lays out data in big-endian form. That is,
999 the bits with the most significance have the lowest address
1002 Specifies that the target lays out data in little-endian form. That
1003 is, the bits with the least significance have the lowest address
1006 Specifies the natural alignment of the stack in bits. Alignment
1007 promotion of stack variables is limited to the natural stack
1008 alignment to avoid dynamic stack realignment. The stack alignment
1009 must be a multiple of 8-bits. If omitted, the natural stack
1010 alignment defaults to "unspecified", which does not prevent any
1011 alignment promotions.
1012 ``p[n]:<size>:<abi>:<pref>``
1013 This specifies the *size* of a pointer and its ``<abi>`` and
1014 ``<pref>``\erred alignments for address space ``n``. All sizes are in
1015 bits. Specifying the ``<pref>`` alignment is optional. If omitted, the
1016 preceding ``:`` should be omitted too. The address space, ``n`` is
1017 optional, and if not specified, denotes the default address space 0.
1018 The value of ``n`` must be in the range [1,2^23).
1019 ``i<size>:<abi>:<pref>``
1020 This specifies the alignment for an integer type of a given bit
1021 ``<size>``. The value of ``<size>`` must be in the range [1,2^23).
1022 ``v<size>:<abi>:<pref>``
1023 This specifies the alignment for a vector type of a given bit
1025 ``f<size>:<abi>:<pref>``
1026 This specifies the alignment for a floating point type of a given bit
1027 ``<size>``. Only values of ``<size>`` that are supported by the target
1028 will work. 32 (float) and 64 (double) are supported on all targets; 80
1029 or 128 (different flavors of long double) are also supported on some
1031 ``a<size>:<abi>:<pref>``
1032 This specifies the alignment for an aggregate type of a given bit
1034 ``s<size>:<abi>:<pref>``
1035 This specifies the alignment for a stack object of a given bit
1037 ``n<size1>:<size2>:<size3>...``
1038 This specifies a set of native integer widths for the target CPU in
1039 bits. For example, it might contain ``n32`` for 32-bit PowerPC,
1040 ``n32:64`` for PowerPC 64, or ``n8:16:32:64`` for X86-64. Elements of
1041 this set are considered to support most general arithmetic operations
1044 When constructing the data layout for a given target, LLVM starts with a
1045 default set of specifications which are then (possibly) overridden by
1046 the specifications in the ``datalayout`` keyword. The default
1047 specifications are given in this list:
1049 - ``E`` - big endian
1050 - ``p:64:64:64`` - 64-bit pointers with 64-bit alignment
1051 - ``S0`` - natural stack alignment is unspecified
1052 - ``i1:8:8`` - i1 is 8-bit (byte) aligned
1053 - ``i8:8:8`` - i8 is 8-bit (byte) aligned
1054 - ``i16:16:16`` - i16 is 16-bit aligned
1055 - ``i32:32:32`` - i32 is 32-bit aligned
1056 - ``i64:32:64`` - i64 has ABI alignment of 32-bits but preferred
1057 alignment of 64-bits
1058 - ``f16:16:16`` - half is 16-bit aligned
1059 - ``f32:32:32`` - float is 32-bit aligned
1060 - ``f64:64:64`` - double is 64-bit aligned
1061 - ``f128:128:128`` - quad is 128-bit aligned
1062 - ``v64:64:64`` - 64-bit vector is 64-bit aligned
1063 - ``v128:128:128`` - 128-bit vector is 128-bit aligned
1064 - ``a0:0:64`` - aggregates are 64-bit aligned
1066 When LLVM is determining the alignment for a given type, it uses the
1069 #. If the type sought is an exact match for one of the specifications,
1070 that specification is used.
1071 #. If no match is found, and the type sought is an integer type, then
1072 the smallest integer type that is larger than the bitwidth of the
1073 sought type is used. If none of the specifications are larger than
1074 the bitwidth then the largest integer type is used. For example,
1075 given the default specifications above, the i7 type will use the
1076 alignment of i8 (next largest) while both i65 and i256 will use the
1077 alignment of i64 (largest specified).
1078 #. If no match is found, and the type sought is a vector type, then the
1079 largest vector type that is smaller than the sought vector type will
1080 be used as a fall back. This happens because <128 x double> can be
1081 implemented in terms of 64 <2 x double>, for example.
1083 The function of the data layout string may not be what you expect.
1084 Notably, this is not a specification from the frontend of what alignment
1085 the code generator should use.
1087 Instead, if specified, the target data layout is required to match what
1088 the ultimate *code generator* expects. This string is used by the
1089 mid-level optimizers to improve code, and this only works if it matches
1090 what the ultimate code generator uses. If you would like to generate IR
1091 that does not embed this target-specific detail into the IR, then you
1092 don't have to specify the string. This will disable some optimizations
1093 that require precise layout information, but this also prevents those
1094 optimizations from introducing target specificity into the IR.
1096 .. _pointeraliasing:
1098 Pointer Aliasing Rules
1099 ----------------------
1101 Any memory access must be done through a pointer value associated with
1102 an address range of the memory access, otherwise the behavior is
1103 undefined. Pointer values are associated with address ranges according
1104 to the following rules:
1106 - A pointer value is associated with the addresses associated with any
1107 value it is *based* on.
1108 - An address of a global variable is associated with the address range
1109 of the variable's storage.
1110 - The result value of an allocation instruction is associated with the
1111 address range of the allocated storage.
1112 - A null pointer in the default address-space is associated with no
1114 - An integer constant other than zero or a pointer value returned from
1115 a function not defined within LLVM may be associated with address
1116 ranges allocated through mechanisms other than those provided by
1117 LLVM. Such ranges shall not overlap with any ranges of addresses
1118 allocated by mechanisms provided by LLVM.
1120 A pointer value is *based* on another pointer value according to the
1123 - A pointer value formed from a ``getelementptr`` operation is *based*
1124 on the first operand of the ``getelementptr``.
1125 - The result value of a ``bitcast`` is *based* on the operand of the
1127 - A pointer value formed by an ``inttoptr`` is *based* on all pointer
1128 values that contribute (directly or indirectly) to the computation of
1129 the pointer's value.
1130 - The "*based* on" relationship is transitive.
1132 Note that this definition of *"based"* is intentionally similar to the
1133 definition of *"based"* in C99, though it is slightly weaker.
1135 LLVM IR does not associate types with memory. The result type of a
1136 ``load`` merely indicates the size and alignment of the memory from
1137 which to load, as well as the interpretation of the value. The first
1138 operand type of a ``store`` similarly only indicates the size and
1139 alignment of the store.
1141 Consequently, type-based alias analysis, aka TBAA, aka
1142 ``-fstrict-aliasing``, is not applicable to general unadorned LLVM IR.
1143 :ref:`Metadata <metadata>` may be used to encode additional information
1144 which specialized optimization passes may use to implement type-based
1149 Volatile Memory Accesses
1150 ------------------------
1152 Certain memory accesses, such as :ref:`load <i_load>`'s,
1153 :ref:`store <i_store>`'s, and :ref:`llvm.memcpy <int_memcpy>`'s may be
1154 marked ``volatile``. The optimizers must not change the number of
1155 volatile operations or change their order of execution relative to other
1156 volatile operations. The optimizers *may* change the order of volatile
1157 operations relative to non-volatile operations. This is not Java's
1158 "volatile" and has no cross-thread synchronization behavior.
1160 IR-level volatile loads and stores cannot safely be optimized into
1161 llvm.memcpy or llvm.memmove intrinsics even when those intrinsics are
1162 flagged volatile. Likewise, the backend should never split or merge
1163 target-legal volatile load/store instructions.
1165 .. admonition:: Rationale
1167 Platforms may rely on volatile loads and stores of natively supported
1168 data width to be executed as single instruction. For example, in C
1169 this holds for an l-value of volatile primitive type with native
1170 hardware support, but not necessarily for aggregate types. The
1171 frontend upholds these expectations, which are intentionally
1172 unspecified in the IR. The rules above ensure that IR transformation
1173 do not violate the frontend's contract with the language.
1177 Memory Model for Concurrent Operations
1178 --------------------------------------
1180 The LLVM IR does not define any way to start parallel threads of
1181 execution or to register signal handlers. Nonetheless, there are
1182 platform-specific ways to create them, and we define LLVM IR's behavior
1183 in their presence. This model is inspired by the C++0x memory model.
1185 For a more informal introduction to this model, see the :doc:`Atomics`.
1187 We define a *happens-before* partial order as the least partial order
1190 - Is a superset of single-thread program order, and
1191 - When a *synchronizes-with* ``b``, includes an edge from ``a`` to
1192 ``b``. *Synchronizes-with* pairs are introduced by platform-specific
1193 techniques, like pthread locks, thread creation, thread joining,
1194 etc., and by atomic instructions. (See also :ref:`Atomic Memory Ordering
1195 Constraints <ordering>`).
1197 Note that program order does not introduce *happens-before* edges
1198 between a thread and signals executing inside that thread.
1200 Every (defined) read operation (load instructions, memcpy, atomic
1201 loads/read-modify-writes, etc.) R reads a series of bytes written by
1202 (defined) write operations (store instructions, atomic
1203 stores/read-modify-writes, memcpy, etc.). For the purposes of this
1204 section, initialized globals are considered to have a write of the
1205 initializer which is atomic and happens before any other read or write
1206 of the memory in question. For each byte of a read R, R\ :sub:`byte`
1207 may see any write to the same byte, except:
1209 - If write\ :sub:`1` happens before write\ :sub:`2`, and
1210 write\ :sub:`2` happens before R\ :sub:`byte`, then
1211 R\ :sub:`byte` does not see write\ :sub:`1`.
1212 - If R\ :sub:`byte` happens before write\ :sub:`3`, then
1213 R\ :sub:`byte` does not see write\ :sub:`3`.
1215 Given that definition, R\ :sub:`byte` is defined as follows:
1217 - If R is volatile, the result is target-dependent. (Volatile is
1218 supposed to give guarantees which can support ``sig_atomic_t`` in
1219 C/C++, and may be used for accesses to addresses which do not behave
1220 like normal memory. It does not generally provide cross-thread
1222 - Otherwise, if there is no write to the same byte that happens before
1223 R\ :sub:`byte`, R\ :sub:`byte` returns ``undef`` for that byte.
1224 - Otherwise, if R\ :sub:`byte` may see exactly one write,
1225 R\ :sub:`byte` returns the value written by that write.
1226 - Otherwise, if R is atomic, and all the writes R\ :sub:`byte` may
1227 see are atomic, it chooses one of the values written. See the :ref:`Atomic
1228 Memory Ordering Constraints <ordering>` section for additional
1229 constraints on how the choice is made.
1230 - Otherwise R\ :sub:`byte` returns ``undef``.
1232 R returns the value composed of the series of bytes it read. This
1233 implies that some bytes within the value may be ``undef`` **without**
1234 the entire value being ``undef``. Note that this only defines the
1235 semantics of the operation; it doesn't mean that targets will emit more
1236 than one instruction to read the series of bytes.
1238 Note that in cases where none of the atomic intrinsics are used, this
1239 model places only one restriction on IR transformations on top of what
1240 is required for single-threaded execution: introducing a store to a byte
1241 which might not otherwise be stored is not allowed in general.
1242 (Specifically, in the case where another thread might write to and read
1243 from an address, introducing a store can change a load that may see
1244 exactly one write into a load that may see multiple writes.)
1248 Atomic Memory Ordering Constraints
1249 ----------------------------------
1251 Atomic instructions (:ref:`cmpxchg <i_cmpxchg>`,
1252 :ref:`atomicrmw <i_atomicrmw>`, :ref:`fence <i_fence>`,
1253 :ref:`atomic load <i_load>`, and :ref:`atomic store <i_store>`) take
1254 an ordering parameter that determines which other atomic instructions on
1255 the same address they *synchronize with*. These semantics are borrowed
1256 from Java and C++0x, but are somewhat more colloquial. If these
1257 descriptions aren't precise enough, check those specs (see spec
1258 references in the :doc:`atomics guide <Atomics>`).
1259 :ref:`fence <i_fence>` instructions treat these orderings somewhat
1260 differently since they don't take an address. See that instruction's
1261 documentation for details.
1263 For a simpler introduction to the ordering constraints, see the
1267 The set of values that can be read is governed by the happens-before
1268 partial order. A value cannot be read unless some operation wrote
1269 it. This is intended to provide a guarantee strong enough to model
1270 Java's non-volatile shared variables. This ordering cannot be
1271 specified for read-modify-write operations; it is not strong enough
1272 to make them atomic in any interesting way.
1274 In addition to the guarantees of ``unordered``, there is a single
1275 total order for modifications by ``monotonic`` operations on each
1276 address. All modification orders must be compatible with the
1277 happens-before order. There is no guarantee that the modification
1278 orders can be combined to a global total order for the whole program
1279 (and this often will not be possible). The read in an atomic
1280 read-modify-write operation (:ref:`cmpxchg <i_cmpxchg>` and
1281 :ref:`atomicrmw <i_atomicrmw>`) reads the value in the modification
1282 order immediately before the value it writes. If one atomic read
1283 happens before another atomic read of the same address, the later
1284 read must see the same value or a later value in the address's
1285 modification order. This disallows reordering of ``monotonic`` (or
1286 stronger) operations on the same address. If an address is written
1287 ``monotonic``-ally by one thread, and other threads ``monotonic``-ally
1288 read that address repeatedly, the other threads must eventually see
1289 the write. This corresponds to the C++0x/C1x
1290 ``memory_order_relaxed``.
1292 In addition to the guarantees of ``monotonic``, a
1293 *synchronizes-with* edge may be formed with a ``release`` operation.
1294 This is intended to model C++'s ``memory_order_acquire``.
1296 In addition to the guarantees of ``monotonic``, if this operation
1297 writes a value which is subsequently read by an ``acquire``
1298 operation, it *synchronizes-with* that operation. (This isn't a
1299 complete description; see the C++0x definition of a release
1300 sequence.) This corresponds to the C++0x/C1x
1301 ``memory_order_release``.
1302 ``acq_rel`` (acquire+release)
1303 Acts as both an ``acquire`` and ``release`` operation on its
1304 address. This corresponds to the C++0x/C1x ``memory_order_acq_rel``.
1305 ``seq_cst`` (sequentially consistent)
1306 In addition to the guarantees of ``acq_rel`` (``acquire`` for an
1307 operation which only reads, ``release`` for an operation which only
1308 writes), there is a global total order on all
1309 sequentially-consistent operations on all addresses, which is
1310 consistent with the *happens-before* partial order and with the
1311 modification orders of all the affected addresses. Each
1312 sequentially-consistent read sees the last preceding write to the
1313 same address in this global order. This corresponds to the C++0x/C1x
1314 ``memory_order_seq_cst`` and Java volatile.
1318 If an atomic operation is marked ``singlethread``, it only *synchronizes
1319 with* or participates in modification and seq\_cst total orderings with
1320 other operations running in the same thread (for example, in signal
1328 LLVM IR floating-point binary ops (:ref:`fadd <i_fadd>`,
1329 :ref:`fsub <i_fsub>`, :ref:`fmul <i_fmul>`, :ref:`fdiv <i_fdiv>`,
1330 :ref:`frem <i_frem>`) have the following flags that can set to enable
1331 otherwise unsafe floating point operations
1334 No NaNs - Allow optimizations to assume the arguments and result are not
1335 NaN. Such optimizations are required to retain defined behavior over
1336 NaNs, but the value of the result is undefined.
1339 No Infs - Allow optimizations to assume the arguments and result are not
1340 +/-Inf. Such optimizations are required to retain defined behavior over
1341 +/-Inf, but the value of the result is undefined.
1344 No Signed Zeros - Allow optimizations to treat the sign of a zero
1345 argument or result as insignificant.
1348 Allow Reciprocal - Allow optimizations to use the reciprocal of an
1349 argument rather than perform division.
1352 Fast - Allow algebraically equivalent transformations that may
1353 dramatically change results in floating point (e.g. reassociate). This
1354 flag implies all the others.
1361 The LLVM type system is one of the most important features of the
1362 intermediate representation. Being typed enables a number of
1363 optimizations to be performed on the intermediate representation
1364 directly, without having to do extra analyses on the side before the
1365 transformation. A strong type system makes it easier to read the
1366 generated code and enables novel analyses and transformations that are
1367 not feasible to perform on normal three address code representations.
1369 .. _typeclassifications:
1371 Type Classifications
1372 --------------------
1374 The types fall into a few useful classifications:
1383 * - :ref:`integer <t_integer>`
1384 - ``i1``, ``i2``, ``i3``, ... ``i8``, ... ``i16``, ... ``i32``, ...
1387 * - :ref:`floating point <t_floating>`
1388 - ``half``, ``float``, ``double``, ``x86_fp80``, ``fp128``,
1396 - :ref:`integer <t_integer>`, :ref:`floating point <t_floating>`,
1397 :ref:`pointer <t_pointer>`, :ref:`vector <t_vector>`,
1398 :ref:`structure <t_struct>`, :ref:`array <t_array>`,
1399 :ref:`label <t_label>`, :ref:`metadata <t_metadata>`.
1401 * - :ref:`primitive <t_primitive>`
1402 - :ref:`label <t_label>`,
1403 :ref:`void <t_void>`,
1404 :ref:`integer <t_integer>`,
1405 :ref:`floating point <t_floating>`,
1406 :ref:`x86mmx <t_x86mmx>`,
1407 :ref:`metadata <t_metadata>`.
1409 * - :ref:`derived <t_derived>`
1410 - :ref:`array <t_array>`,
1411 :ref:`function <t_function>`,
1412 :ref:`pointer <t_pointer>`,
1413 :ref:`structure <t_struct>`,
1414 :ref:`vector <t_vector>`,
1415 :ref:`opaque <t_opaque>`.
1417 The :ref:`first class <t_firstclass>` types are perhaps the most important.
1418 Values of these types are the only ones which can be produced by
1426 The primitive types are the fundamental building blocks of the LLVM
1437 The integer type is a very simple type that simply specifies an
1438 arbitrary bit width for the integer type desired. Any bit width from 1
1439 bit to 2\ :sup:`23`\ -1 (about 8 million) can be specified.
1448 The number of bits the integer will occupy is specified by the ``N``
1454 +----------------+------------------------------------------------+
1455 | ``i1`` | a single-bit integer. |
1456 +----------------+------------------------------------------------+
1457 | ``i32`` | a 32-bit integer. |
1458 +----------------+------------------------------------------------+
1459 | ``i1942652`` | a really big integer of over 1 million bits. |
1460 +----------------+------------------------------------------------+
1464 Floating Point Types
1465 ^^^^^^^^^^^^^^^^^^^^
1474 - 16-bit floating point value
1477 - 32-bit floating point value
1480 - 64-bit floating point value
1483 - 128-bit floating point value (112-bit mantissa)
1486 - 80-bit floating point value (X87)
1489 - 128-bit floating point value (two 64-bits)
1499 The x86mmx type represents a value held in an MMX register on an x86
1500 machine. The operations allowed on it are quite limited: parameters and
1501 return values, load and store, and bitcast. User-specified MMX
1502 instructions are represented as intrinsic or asm calls with arguments
1503 and/or results of this type. There are no arrays, vectors or constants
1521 The void type does not represent any value and has no size.
1538 The label type represents code labels.
1555 The metadata type represents embedded metadata. No derived types may be
1556 created from metadata except for :ref:`function <t_function>` arguments.
1570 The real power in LLVM comes from the derived types in the system. This
1571 is what allows a programmer to represent arrays, functions, pointers,
1572 and other useful types. Each of these types contain one or more element
1573 types which may be a primitive type, or another derived type. For
1574 example, it is possible to have a two dimensional array, using an array
1575 as the element type of another array.
1582 Aggregate Types are a subset of derived types that can contain multiple
1583 member types. :ref:`Arrays <t_array>` and :ref:`structs <t_struct>` are
1584 aggregate types. :ref:`Vectors <t_vector>` are not considered to be
1595 The array type is a very simple derived type that arranges elements
1596 sequentially in memory. The array type requires a size (number of
1597 elements) and an underlying data type.
1604 [<# elements> x <elementtype>]
1606 The number of elements is a constant integer value; ``elementtype`` may
1607 be any type with a size.
1612 +------------------+--------------------------------------+
1613 | ``[40 x i32]`` | Array of 40 32-bit integer values. |
1614 +------------------+--------------------------------------+
1615 | ``[41 x i32]`` | Array of 41 32-bit integer values. |
1616 +------------------+--------------------------------------+
1617 | ``[4 x i8]`` | Array of 4 8-bit integer values. |
1618 +------------------+--------------------------------------+
1620 Here are some examples of multidimensional arrays:
1622 +-----------------------------+----------------------------------------------------------+
1623 | ``[3 x [4 x i32]]`` | 3x4 array of 32-bit integer values. |
1624 +-----------------------------+----------------------------------------------------------+
1625 | ``[12 x [10 x float]]`` | 12x10 array of single precision floating point values. |
1626 +-----------------------------+----------------------------------------------------------+
1627 | ``[2 x [3 x [4 x i16]]]`` | 2x3x4 array of 16-bit integer values. |
1628 +-----------------------------+----------------------------------------------------------+
1630 There is no restriction on indexing beyond the end of the array implied
1631 by a static type (though there are restrictions on indexing beyond the
1632 bounds of an allocated object in some cases). This means that
1633 single-dimension 'variable sized array' addressing can be implemented in
1634 LLVM with a zero length array type. An implementation of 'pascal style
1635 arrays' in LLVM could use the type "``{ i32, [0 x float]}``", for
1646 The function type can be thought of as a function signature. It consists
1647 of a return type and a list of formal parameter types. The return type
1648 of a function type is a first class type or a void type.
1655 <returntype> (<parameter list>)
1657 ...where '``<parameter list>``' is a comma-separated list of type
1658 specifiers. Optionally, the parameter list may include a type ``...``,
1659 which indicates that the function takes a variable number of arguments.
1660 Variable argument functions can access their arguments with the
1661 :ref:`variable argument handling intrinsic <int_varargs>` functions.
1662 '``<returntype>``' is any type except :ref:`label <t_label>`.
1667 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1668 | ``i32 (i32)`` | function taking an ``i32``, returning an ``i32`` |
1669 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1670 | ``float (i16, i32 *) *`` | :ref:`Pointer <t_pointer>` to a function that takes an ``i16`` and a :ref:`pointer <t_pointer>` to ``i32``, returning ``float``. |
1671 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1672 | ``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. |
1673 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1674 | ``{i32, i32} (i32)`` | A function taking an ``i32``, returning a :ref:`structure <t_struct>` containing two ``i32`` values |
1675 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1685 The structure type is used to represent a collection of data members
1686 together in memory. The elements of a structure may be any type that has
1689 Structures in memory are accessed using '``load``' and '``store``' by
1690 getting a pointer to a field with the '``getelementptr``' instruction.
1691 Structures in registers are accessed using the '``extractvalue``' and
1692 '``insertvalue``' instructions.
1694 Structures may optionally be "packed" structures, which indicate that
1695 the alignment of the struct is one byte, and that there is no padding
1696 between the elements. In non-packed structs, padding between field types
1697 is inserted as defined by the DataLayout string in the module, which is
1698 required to match what the underlying code generator expects.
1700 Structures can either be "literal" or "identified". A literal structure
1701 is defined inline with other types (e.g. ``{i32, i32}*``) whereas
1702 identified types are always defined at the top level with a name.
1703 Literal types are uniqued by their contents and can never be recursive
1704 or opaque since there is no way to write one. Identified types can be
1705 recursive, can be opaqued, and are never uniqued.
1712 %T1 = type { <type list> } ; Identified normal struct type
1713 %T2 = type <{ <type list> }> ; Identified packed struct type
1718 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1719 | ``{ i32, i32, i32 }`` | A triple of three ``i32`` values |
1720 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1721 | ``{ 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``. |
1722 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1723 | ``<{ i8, i32 }>`` | A packed struct known to be 5 bytes in size. |
1724 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1728 Opaque Structure Types
1729 ^^^^^^^^^^^^^^^^^^^^^^
1734 Opaque structure types are used to represent named structure types that
1735 do not have a body specified. This corresponds (for example) to the C
1736 notion of a forward declared structure.
1749 +--------------+-------------------+
1750 | ``opaque`` | An opaque type. |
1751 +--------------+-------------------+
1761 The pointer type is used to specify memory locations. Pointers are
1762 commonly used to reference objects in memory.
1764 Pointer types may have an optional address space attribute defining the
1765 numbered address space where the pointed-to object resides. The default
1766 address space is number zero. The semantics of non-zero address spaces
1767 are target-specific.
1769 Note that LLVM does not permit pointers to void (``void*``) nor does it
1770 permit pointers to labels (``label*``). Use ``i8*`` instead.
1782 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1783 | ``[4 x i32]*`` | A :ref:`pointer <t_pointer>` to :ref:`array <t_array>` of four ``i32`` values. |
1784 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1785 | ``i32 (i32*) *`` | A :ref:`pointer <t_pointer>` to a :ref:`function <t_function>` that takes an ``i32*``, returning an ``i32``. |
1786 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1787 | ``i32 addrspace(5)*`` | A :ref:`pointer <t_pointer>` to an ``i32`` value that resides in address space #5. |
1788 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1798 A vector type is a simple derived type that represents a vector of
1799 elements. Vector types are used when multiple primitive data are
1800 operated in parallel using a single instruction (SIMD). A vector type
1801 requires a size (number of elements) and an underlying primitive data
1802 type. Vector types are considered :ref:`first class <t_firstclass>`.
1809 < <# elements> x <elementtype> >
1811 The number of elements is a constant integer value larger than 0;
1812 elementtype may be any integer or floating point type, or a pointer to
1813 these types. Vectors of size zero are not allowed.
1818 +-------------------+--------------------------------------------------+
1819 | ``<4 x i32>`` | Vector of 4 32-bit integer values. |
1820 +-------------------+--------------------------------------------------+
1821 | ``<8 x float>`` | Vector of 8 32-bit floating-point values. |
1822 +-------------------+--------------------------------------------------+
1823 | ``<2 x i64>`` | Vector of 2 64-bit integer values. |
1824 +-------------------+--------------------------------------------------+
1825 | ``<4 x i64*>`` | Vector of 4 pointers to 64-bit integer values. |
1826 +-------------------+--------------------------------------------------+
1831 LLVM has several different basic types of constants. This section
1832 describes them all and their syntax.
1837 **Boolean constants**
1838 The two strings '``true``' and '``false``' are both valid constants
1840 **Integer constants**
1841 Standard integers (such as '4') are constants of the
1842 :ref:`integer <t_integer>` type. Negative numbers may be used with
1844 **Floating point constants**
1845 Floating point constants use standard decimal notation (e.g.
1846 123.421), exponential notation (e.g. 1.23421e+2), or a more precise
1847 hexadecimal notation (see below). The assembler requires the exact
1848 decimal value of a floating-point constant. For example, the
1849 assembler accepts 1.25 but rejects 1.3 because 1.3 is a repeating
1850 decimal in binary. Floating point constants must have a :ref:`floating
1851 point <t_floating>` type.
1852 **Null pointer constants**
1853 The identifier '``null``' is recognized as a null pointer constant
1854 and must be of :ref:`pointer type <t_pointer>`.
1856 The one non-intuitive notation for constants is the hexadecimal form of
1857 floating point constants. For example, the form
1858 '``double 0x432ff973cafa8000``' is equivalent to (but harder to read
1859 than) '``double 4.5e+15``'. The only time hexadecimal floating point
1860 constants are required (and the only time that they are generated by the
1861 disassembler) is when a floating point constant must be emitted but it
1862 cannot be represented as a decimal floating point number in a reasonable
1863 number of digits. For example, NaN's, infinities, and other special
1864 values are represented in their IEEE hexadecimal format so that assembly
1865 and disassembly do not cause any bits to change in the constants.
1867 When using the hexadecimal form, constants of types half, float, and
1868 double are represented using the 16-digit form shown above (which
1869 matches the IEEE754 representation for double); half and float values
1870 must, however, be exactly representable as IEEE 754 half and single
1871 precision, respectively. Hexadecimal format is always used for long
1872 double, and there are three forms of long double. The 80-bit format used
1873 by x86 is represented as ``0xK`` followed by 20 hexadecimal digits. The
1874 128-bit format used by PowerPC (two adjacent doubles) is represented by
1875 ``0xM`` followed by 32 hexadecimal digits. The IEEE 128-bit format is
1876 represented by ``0xL`` followed by 32 hexadecimal digits. Long doubles
1877 will only work if they match the long double format on your target.
1878 The IEEE 16-bit format (half precision) is represented by ``0xH``
1879 followed by 4 hexadecimal digits. All hexadecimal formats are big-endian
1880 (sign bit at the left).
1882 There are no constants of type x86mmx.
1884 .. _complexconstants:
1889 Complex constants are a (potentially recursive) combination of simple
1890 constants and smaller complex constants.
1892 **Structure constants**
1893 Structure constants are represented with notation similar to
1894 structure type definitions (a comma separated list of elements,
1895 surrounded by braces (``{}``)). For example:
1896 "``{ i32 4, float 17.0, i32* @G }``", where "``@G``" is declared as
1897 "``@G = external global i32``". Structure constants must have
1898 :ref:`structure type <t_struct>`, and the number and types of elements
1899 must match those specified by the type.
1901 Array constants are represented with notation similar to array type
1902 definitions (a comma separated list of elements, surrounded by
1903 square brackets (``[]``)). For example:
1904 "``[ i32 42, i32 11, i32 74 ]``". Array constants must have
1905 :ref:`array type <t_array>`, and the number and types of elements must
1906 match those specified by the type.
1907 **Vector constants**
1908 Vector constants are represented with notation similar to vector
1909 type definitions (a comma separated list of elements, surrounded by
1910 less-than/greater-than's (``<>``)). For example:
1911 "``< i32 42, i32 11, i32 74, i32 100 >``". Vector constants
1912 must have :ref:`vector type <t_vector>`, and the number and types of
1913 elements must match those specified by the type.
1914 **Zero initialization**
1915 The string '``zeroinitializer``' can be used to zero initialize a
1916 value to zero of *any* type, including scalar and
1917 :ref:`aggregate <t_aggregate>` types. This is often used to avoid
1918 having to print large zero initializers (e.g. for large arrays) and
1919 is always exactly equivalent to using explicit zero initializers.
1921 A metadata node is a structure-like constant with :ref:`metadata
1922 type <t_metadata>`. For example:
1923 "``metadata !{ i32 0, metadata !"test" }``". Unlike other
1924 constants that are meant to be interpreted as part of the
1925 instruction stream, metadata is a place to attach additional
1926 information such as debug info.
1928 Global Variable and Function Addresses
1929 --------------------------------------
1931 The addresses of :ref:`global variables <globalvars>` and
1932 :ref:`functions <functionstructure>` are always implicitly valid
1933 (link-time) constants. These constants are explicitly referenced when
1934 the :ref:`identifier for the global <identifiers>` is used and always have
1935 :ref:`pointer <t_pointer>` type. For example, the following is a legal LLVM
1938 .. code-block:: llvm
1942 @Z = global [2 x i32*] [ i32* @X, i32* @Y ]
1949 The string '``undef``' can be used anywhere a constant is expected, and
1950 indicates that the user of the value may receive an unspecified
1951 bit-pattern. Undefined values may be of any type (other than '``label``'
1952 or '``void``') and be used anywhere a constant is permitted.
1954 Undefined values are useful because they indicate to the compiler that
1955 the program is well defined no matter what value is used. This gives the
1956 compiler more freedom to optimize. Here are some examples of
1957 (potentially surprising) transformations that are valid (in pseudo IR):
1959 .. code-block:: llvm
1969 This is safe because all of the output bits are affected by the undef
1970 bits. Any output bit can have a zero or one depending on the input bits.
1972 .. code-block:: llvm
1983 These logical operations have bits that are not always affected by the
1984 input. For example, if ``%X`` has a zero bit, then the output of the
1985 '``and``' operation will always be a zero for that bit, no matter what
1986 the corresponding bit from the '``undef``' is. As such, it is unsafe to
1987 optimize or assume that the result of the '``and``' is '``undef``'.
1988 However, it is safe to assume that all bits of the '``undef``' could be
1989 0, and optimize the '``and``' to 0. Likewise, it is safe to assume that
1990 all the bits of the '``undef``' operand to the '``or``' could be set,
1991 allowing the '``or``' to be folded to -1.
1993 .. code-block:: llvm
1995 %A = select undef, %X, %Y
1996 %B = select undef, 42, %Y
1997 %C = select %X, %Y, undef
2007 This set of examples shows that undefined '``select``' (and conditional
2008 branch) conditions can go *either way*, but they have to come from one
2009 of the two operands. In the ``%A`` example, if ``%X`` and ``%Y`` were
2010 both known to have a clear low bit, then ``%A`` would have to have a
2011 cleared low bit. However, in the ``%C`` example, the optimizer is
2012 allowed to assume that the '``undef``' operand could be the same as
2013 ``%Y``, allowing the whole '``select``' to be eliminated.
2015 .. code-block:: llvm
2017 %A = xor undef, undef
2034 This example points out that two '``undef``' operands are not
2035 necessarily the same. This can be surprising to people (and also matches
2036 C semantics) where they assume that "``X^X``" is always zero, even if
2037 ``X`` is undefined. This isn't true for a number of reasons, but the
2038 short answer is that an '``undef``' "variable" can arbitrarily change
2039 its value over its "live range". This is true because the variable
2040 doesn't actually *have a live range*. Instead, the value is logically
2041 read from arbitrary registers that happen to be around when needed, so
2042 the value is not necessarily consistent over time. In fact, ``%A`` and
2043 ``%C`` need to have the same semantics or the core LLVM "replace all
2044 uses with" concept would not hold.
2046 .. code-block:: llvm
2054 These examples show the crucial difference between an *undefined value*
2055 and *undefined behavior*. An undefined value (like '``undef``') is
2056 allowed to have an arbitrary bit-pattern. This means that the ``%A``
2057 operation can be constant folded to '``undef``', because the '``undef``'
2058 could be an SNaN, and ``fdiv`` is not (currently) defined on SNaN's.
2059 However, in the second example, we can make a more aggressive
2060 assumption: because the ``undef`` is allowed to be an arbitrary value,
2061 we are allowed to assume that it could be zero. Since a divide by zero
2062 has *undefined behavior*, we are allowed to assume that the operation
2063 does not execute at all. This allows us to delete the divide and all
2064 code after it. Because the undefined operation "can't happen", the
2065 optimizer can assume that it occurs in dead code.
2067 .. code-block:: llvm
2069 a: store undef -> %X
2070 b: store %X -> undef
2075 These examples reiterate the ``fdiv`` example: a store *of* an undefined
2076 value can be assumed to not have any effect; we can assume that the
2077 value is overwritten with bits that happen to match what was already
2078 there. However, a store *to* an undefined location could clobber
2079 arbitrary memory, therefore, it has undefined behavior.
2086 Poison values are similar to :ref:`undef values <undefvalues>`, however
2087 they also represent the fact that an instruction or constant expression
2088 which cannot evoke side effects has nevertheless detected a condition
2089 which results in undefined behavior.
2091 There is currently no way of representing a poison value in the IR; they
2092 only exist when produced by operations such as :ref:`add <i_add>` with
2095 Poison value behavior is defined in terms of value *dependence*:
2097 - Values other than :ref:`phi <i_phi>` nodes depend on their operands.
2098 - :ref:`Phi <i_phi>` nodes depend on the operand corresponding to
2099 their dynamic predecessor basic block.
2100 - Function arguments depend on the corresponding actual argument values
2101 in the dynamic callers of their functions.
2102 - :ref:`Call <i_call>` instructions depend on the :ref:`ret <i_ret>`
2103 instructions that dynamically transfer control back to them.
2104 - :ref:`Invoke <i_invoke>` instructions depend on the
2105 :ref:`ret <i_ret>`, :ref:`resume <i_resume>`, or exception-throwing
2106 call instructions that dynamically transfer control back to them.
2107 - Non-volatile loads and stores depend on the most recent stores to all
2108 of the referenced memory addresses, following the order in the IR
2109 (including loads and stores implied by intrinsics such as
2110 :ref:`@llvm.memcpy <int_memcpy>`.)
2111 - An instruction with externally visible side effects depends on the
2112 most recent preceding instruction with externally visible side
2113 effects, following the order in the IR. (This includes :ref:`volatile
2114 operations <volatile>`.)
2115 - An instruction *control-depends* on a :ref:`terminator
2116 instruction <terminators>` if the terminator instruction has
2117 multiple successors and the instruction is always executed when
2118 control transfers to one of the successors, and may not be executed
2119 when control is transferred to another.
2120 - Additionally, an instruction also *control-depends* on a terminator
2121 instruction if the set of instructions it otherwise depends on would
2122 be different if the terminator had transferred control to a different
2124 - Dependence is transitive.
2126 Poison Values have the same behavior as :ref:`undef values <undefvalues>`,
2127 with the additional affect that any instruction which has a *dependence*
2128 on a poison value has undefined behavior.
2130 Here are some examples:
2132 .. code-block:: llvm
2135 %poison = sub nuw i32 0, 1 ; Results in a poison value.
2136 %still_poison = and i32 %poison, 0 ; 0, but also poison.
2137 %poison_yet_again = getelementptr i32* @h, i32 %still_poison
2138 store i32 0, i32* %poison_yet_again ; memory at @h[0] is poisoned
2140 store i32 %poison, i32* @g ; Poison value stored to memory.
2141 %poison2 = load i32* @g ; Poison value loaded back from memory.
2143 store volatile i32 %poison, i32* @g ; External observation; undefined behavior.
2145 %narrowaddr = bitcast i32* @g to i16*
2146 %wideaddr = bitcast i32* @g to i64*
2147 %poison3 = load i16* %narrowaddr ; Returns a poison value.
2148 %poison4 = load i64* %wideaddr ; Returns a poison value.
2150 %cmp = icmp slt i32 %poison, 0 ; Returns a poison value.
2151 br i1 %cmp, label %true, label %end ; Branch to either destination.
2154 store volatile i32 0, i32* @g ; This is control-dependent on %cmp, so
2155 ; it has undefined behavior.
2159 %p = phi i32 [ 0, %entry ], [ 1, %true ]
2160 ; Both edges into this PHI are
2161 ; control-dependent on %cmp, so this
2162 ; always results in a poison value.
2164 store volatile i32 0, i32* @g ; This would depend on the store in %true
2165 ; if %cmp is true, or the store in %entry
2166 ; otherwise, so this is undefined behavior.
2168 br i1 %cmp, label %second_true, label %second_end
2169 ; The same branch again, but this time the
2170 ; true block doesn't have side effects.
2177 store volatile i32 0, i32* @g ; This time, the instruction always depends
2178 ; on the store in %end. Also, it is
2179 ; control-equivalent to %end, so this is
2180 ; well-defined (ignoring earlier undefined
2181 ; behavior in this example).
2185 Addresses of Basic Blocks
2186 -------------------------
2188 ``blockaddress(@function, %block)``
2190 The '``blockaddress``' constant computes the address of the specified
2191 basic block in the specified function, and always has an ``i8*`` type.
2192 Taking the address of the entry block is illegal.
2194 This value only has defined behavior when used as an operand to the
2195 ':ref:`indirectbr <i_indirectbr>`' instruction, or for comparisons
2196 against null. Pointer equality tests between labels addresses results in
2197 undefined behavior --- though, again, comparison against null is ok, and
2198 no label is equal to the null pointer. This may be passed around as an
2199 opaque pointer sized value as long as the bits are not inspected. This
2200 allows ``ptrtoint`` and arithmetic to be performed on these values so
2201 long as the original value is reconstituted before the ``indirectbr``
2204 Finally, some targets may provide defined semantics when using the value
2205 as the operand to an inline assembly, but that is target specific.
2209 Constant Expressions
2210 --------------------
2212 Constant expressions are used to allow expressions involving other
2213 constants to be used as constants. Constant expressions may be of any
2214 :ref:`first class <t_firstclass>` type and may involve any LLVM operation
2215 that does not have side effects (e.g. load and call are not supported).
2216 The following is the syntax for constant expressions:
2218 ``trunc (CST to TYPE)``
2219 Truncate a constant to another type. The bit size of CST must be
2220 larger than the bit size of TYPE. Both types must be integers.
2221 ``zext (CST to TYPE)``
2222 Zero extend a constant to another type. The bit size of CST must be
2223 smaller than the bit size of TYPE. Both types must be integers.
2224 ``sext (CST to TYPE)``
2225 Sign extend a constant to another type. The bit size of CST must be
2226 smaller than the bit size of TYPE. Both types must be integers.
2227 ``fptrunc (CST to TYPE)``
2228 Truncate a floating point constant to another floating point type.
2229 The size of CST must be larger than the size of TYPE. Both types
2230 must be floating point.
2231 ``fpext (CST to TYPE)``
2232 Floating point extend a constant to another type. The size of CST
2233 must be smaller or equal to the size of TYPE. Both types must be
2235 ``fptoui (CST to TYPE)``
2236 Convert a floating point constant to the corresponding unsigned
2237 integer constant. TYPE must be a scalar or vector integer type. CST
2238 must be of scalar or vector floating point type. Both CST and TYPE
2239 must be scalars, or vectors of the same number of elements. If the
2240 value won't fit in the integer type, the results are undefined.
2241 ``fptosi (CST to TYPE)``
2242 Convert a floating point constant to the corresponding signed
2243 integer constant. TYPE must be a scalar or vector integer type. CST
2244 must be of scalar or vector floating point type. Both CST and TYPE
2245 must be scalars, or vectors of the same number of elements. If the
2246 value won't fit in the integer type, the results are undefined.
2247 ``uitofp (CST to TYPE)``
2248 Convert an unsigned integer constant to the corresponding floating
2249 point constant. TYPE must be a scalar or vector floating point type.
2250 CST must be of scalar or vector integer type. Both CST and TYPE must
2251 be scalars, or vectors of the same number of elements. If the value
2252 won't fit in the floating point type, the results are undefined.
2253 ``sitofp (CST to TYPE)``
2254 Convert a signed integer constant to the corresponding floating
2255 point constant. TYPE must be a scalar or vector floating point type.
2256 CST must be of scalar or vector integer type. Both CST and TYPE must
2257 be scalars, or vectors of the same number of elements. If the value
2258 won't fit in the floating point type, the results are undefined.
2259 ``ptrtoint (CST to TYPE)``
2260 Convert a pointer typed constant to the corresponding integer
2261 constant. ``TYPE`` must be an integer type. ``CST`` must be of
2262 pointer type. The ``CST`` value is zero extended, truncated, or
2263 unchanged to make it fit in ``TYPE``.
2264 ``inttoptr (CST to TYPE)``
2265 Convert an integer constant to a pointer constant. TYPE must be a
2266 pointer type. CST must be of integer type. The CST value is zero
2267 extended, truncated, or unchanged to make it fit in a pointer size.
2268 This one is *really* dangerous!
2269 ``bitcast (CST to TYPE)``
2270 Convert a constant, CST, to another TYPE. The constraints of the
2271 operands are the same as those for the :ref:`bitcast
2272 instruction <i_bitcast>`.
2273 ``getelementptr (CSTPTR, IDX0, IDX1, ...)``, ``getelementptr inbounds (CSTPTR, IDX0, IDX1, ...)``
2274 Perform the :ref:`getelementptr operation <i_getelementptr>` on
2275 constants. As with the :ref:`getelementptr <i_getelementptr>`
2276 instruction, the index list may have zero or more indexes, which are
2277 required to make sense for the type of "CSTPTR".
2278 ``select (COND, VAL1, VAL2)``
2279 Perform the :ref:`select operation <i_select>` on constants.
2280 ``icmp COND (VAL1, VAL2)``
2281 Performs the :ref:`icmp operation <i_icmp>` on constants.
2282 ``fcmp COND (VAL1, VAL2)``
2283 Performs the :ref:`fcmp operation <i_fcmp>` on constants.
2284 ``extractelement (VAL, IDX)``
2285 Perform the :ref:`extractelement operation <i_extractelement>` on
2287 ``insertelement (VAL, ELT, IDX)``
2288 Perform the :ref:`insertelement operation <i_insertelement>` on
2290 ``shufflevector (VEC1, VEC2, IDXMASK)``
2291 Perform the :ref:`shufflevector operation <i_shufflevector>` on
2293 ``extractvalue (VAL, IDX0, IDX1, ...)``
2294 Perform the :ref:`extractvalue operation <i_extractvalue>` on
2295 constants. The index list is interpreted in a similar manner as
2296 indices in a ':ref:`getelementptr <i_getelementptr>`' operation. At
2297 least one index value must be specified.
2298 ``insertvalue (VAL, ELT, IDX0, IDX1, ...)``
2299 Perform the :ref:`insertvalue operation <i_insertvalue>` on constants.
2300 The index list is interpreted in a similar manner as indices in a
2301 ':ref:`getelementptr <i_getelementptr>`' operation. At least one index
2302 value must be specified.
2303 ``OPCODE (LHS, RHS)``
2304 Perform the specified operation of the LHS and RHS constants. OPCODE
2305 may be any of the :ref:`binary <binaryops>` or :ref:`bitwise
2306 binary <bitwiseops>` operations. The constraints on operands are
2307 the same as those for the corresponding instruction (e.g. no bitwise
2308 operations on floating point values are allowed).
2315 Inline Assembler Expressions
2316 ----------------------------
2318 LLVM supports inline assembler expressions (as opposed to :ref:`Module-Level
2319 Inline Assembly <moduleasm>`) through the use of a special value. This
2320 value represents the inline assembler as a string (containing the
2321 instructions to emit), a list of operand constraints (stored as a
2322 string), a flag that indicates whether or not the inline asm expression
2323 has side effects, and a flag indicating whether the function containing
2324 the asm needs to align its stack conservatively. An example inline
2325 assembler expression is:
2327 .. code-block:: llvm
2329 i32 (i32) asm "bswap $0", "=r,r"
2331 Inline assembler expressions may **only** be used as the callee operand
2332 of a :ref:`call <i_call>` or an :ref:`invoke <i_invoke>` instruction.
2333 Thus, typically we have:
2335 .. code-block:: llvm
2337 %X = call i32 asm "bswap $0", "=r,r"(i32 %Y)
2339 Inline asms with side effects not visible in the constraint list must be
2340 marked as having side effects. This is done through the use of the
2341 '``sideeffect``' keyword, like so:
2343 .. code-block:: llvm
2345 call void asm sideeffect "eieio", ""()
2347 In some cases inline asms will contain code that will not work unless
2348 the stack is aligned in some way, such as calls or SSE instructions on
2349 x86, yet will not contain code that does that alignment within the asm.
2350 The compiler should make conservative assumptions about what the asm
2351 might contain and should generate its usual stack alignment code in the
2352 prologue if the '``alignstack``' keyword is present:
2354 .. code-block:: llvm
2356 call void asm alignstack "eieio", ""()
2358 Inline asms also support using non-standard assembly dialects. The
2359 assumed dialect is ATT. When the '``inteldialect``' keyword is present,
2360 the inline asm is using the Intel dialect. Currently, ATT and Intel are
2361 the only supported dialects. An example is:
2363 .. code-block:: llvm
2365 call void asm inteldialect "eieio", ""()
2367 If multiple keywords appear the '``sideeffect``' keyword must come
2368 first, the '``alignstack``' keyword second and the '``inteldialect``'
2374 The call instructions that wrap inline asm nodes may have a
2375 "``!srcloc``" MDNode attached to it that contains a list of constant
2376 integers. If present, the code generator will use the integer as the
2377 location cookie value when report errors through the ``LLVMContext``
2378 error reporting mechanisms. This allows a front-end to correlate backend
2379 errors that occur with inline asm back to the source code that produced
2382 .. code-block:: llvm
2384 call void asm sideeffect "something bad", ""(), !srcloc !42
2386 !42 = !{ i32 1234567 }
2388 It is up to the front-end to make sense of the magic numbers it places
2389 in the IR. If the MDNode contains multiple constants, the code generator
2390 will use the one that corresponds to the line of the asm that the error
2395 Metadata Nodes and Metadata Strings
2396 -----------------------------------
2398 LLVM IR allows metadata to be attached to instructions in the program
2399 that can convey extra information about the code to the optimizers and
2400 code generator. One example application of metadata is source-level
2401 debug information. There are two metadata primitives: strings and nodes.
2402 All metadata has the ``metadata`` type and is identified in syntax by a
2403 preceding exclamation point ('``!``').
2405 A metadata string is a string surrounded by double quotes. It can
2406 contain any character by escaping non-printable characters with
2407 "``\xx``" where "``xx``" is the two digit hex code. For example:
2410 Metadata nodes are represented with notation similar to structure
2411 constants (a comma separated list of elements, surrounded by braces and
2412 preceded by an exclamation point). Metadata nodes can have any values as
2413 their operand. For example:
2415 .. code-block:: llvm
2417 !{ metadata !"test\00", i32 10}
2419 A :ref:`named metadata <namedmetadatastructure>` is a collection of
2420 metadata nodes, which can be looked up in the module symbol table. For
2423 .. code-block:: llvm
2425 !foo = metadata !{!4, !3}
2427 Metadata can be used as function arguments. Here ``llvm.dbg.value``
2428 function is using two metadata arguments:
2430 .. code-block:: llvm
2432 call void @llvm.dbg.value(metadata !24, i64 0, metadata !25)
2434 Metadata can be attached with an instruction. Here metadata ``!21`` is
2435 attached to the ``add`` instruction using the ``!dbg`` identifier:
2437 .. code-block:: llvm
2439 %indvar.next = add i64 %indvar, 1, !dbg !21
2441 More information about specific metadata nodes recognized by the
2442 optimizers and code generator is found below.
2447 In LLVM IR, memory does not have types, so LLVM's own type system is not
2448 suitable for doing TBAA. Instead, metadata is added to the IR to
2449 describe a type system of a higher level language. This can be used to
2450 implement typical C/C++ TBAA, but it can also be used to implement
2451 custom alias analysis behavior for other languages.
2453 The current metadata format is very simple. TBAA metadata nodes have up
2454 to three fields, e.g.:
2456 .. code-block:: llvm
2458 !0 = metadata !{ metadata !"an example type tree" }
2459 !1 = metadata !{ metadata !"int", metadata !0 }
2460 !2 = metadata !{ metadata !"float", metadata !0 }
2461 !3 = metadata !{ metadata !"const float", metadata !2, i64 1 }
2463 The first field is an identity field. It can be any value, usually a
2464 metadata string, which uniquely identifies the type. The most important
2465 name in the tree is the name of the root node. Two trees with different
2466 root node names are entirely disjoint, even if they have leaves with
2469 The second field identifies the type's parent node in the tree, or is
2470 null or omitted for a root node. A type is considered to alias all of
2471 its descendants and all of its ancestors in the tree. Also, a type is
2472 considered to alias all types in other trees, so that bitcode produced
2473 from multiple front-ends is handled conservatively.
2475 If the third field is present, it's an integer which if equal to 1
2476 indicates that the type is "constant" (meaning
2477 ``pointsToConstantMemory`` should return true; see `other useful
2478 AliasAnalysis methods <AliasAnalysis.html#OtherItfs>`_).
2480 '``tbaa.struct``' Metadata
2481 ^^^^^^^^^^^^^^^^^^^^^^^^^^
2483 The :ref:`llvm.memcpy <int_memcpy>` is often used to implement
2484 aggregate assignment operations in C and similar languages, however it
2485 is defined to copy a contiguous region of memory, which is more than
2486 strictly necessary for aggregate types which contain holes due to
2487 padding. Also, it doesn't contain any TBAA information about the fields
2490 ``!tbaa.struct`` metadata can describe which memory subregions in a
2491 memcpy are padding and what the TBAA tags of the struct are.
2493 The current metadata format is very simple. ``!tbaa.struct`` metadata
2494 nodes are a list of operands which are in conceptual groups of three.
2495 For each group of three, the first operand gives the byte offset of a
2496 field in bytes, the second gives its size in bytes, and the third gives
2499 .. code-block:: llvm
2501 !4 = metadata !{ i64 0, i64 4, metadata !1, i64 8, i64 4, metadata !2 }
2503 This describes a struct with two fields. The first is at offset 0 bytes
2504 with size 4 bytes, and has tbaa tag !1. The second is at offset 8 bytes
2505 and has size 4 bytes and has tbaa tag !2.
2507 Note that the fields need not be contiguous. In this example, there is a
2508 4 byte gap between the two fields. This gap represents padding which
2509 does not carry useful data and need not be preserved.
2511 '``fpmath``' Metadata
2512 ^^^^^^^^^^^^^^^^^^^^^
2514 ``fpmath`` metadata may be attached to any instruction of floating point
2515 type. It can be used to express the maximum acceptable error in the
2516 result of that instruction, in ULPs, thus potentially allowing the
2517 compiler to use a more efficient but less accurate method of computing
2518 it. ULP is defined as follows:
2520 If ``x`` is a real number that lies between two finite consecutive
2521 floating-point numbers ``a`` and ``b``, without being equal to one
2522 of them, then ``ulp(x) = |b - a|``, otherwise ``ulp(x)`` is the
2523 distance between the two non-equal finite floating-point numbers
2524 nearest ``x``. Moreover, ``ulp(NaN)`` is ``NaN``.
2526 The metadata node shall consist of a single positive floating point
2527 number representing the maximum relative error, for example:
2529 .. code-block:: llvm
2531 !0 = metadata !{ float 2.5 } ; maximum acceptable inaccuracy is 2.5 ULPs
2533 '``range``' Metadata
2534 ^^^^^^^^^^^^^^^^^^^^
2536 ``range`` metadata may be attached only to loads of integer types. It
2537 expresses the possible ranges the loaded value is in. The ranges are
2538 represented with a flattened list of integers. The loaded value is known
2539 to be in the union of the ranges defined by each consecutive pair. Each
2540 pair has the following properties:
2542 - The type must match the type loaded by the instruction.
2543 - The pair ``a,b`` represents the range ``[a,b)``.
2544 - Both ``a`` and ``b`` are constants.
2545 - The range is allowed to wrap.
2546 - The range should not represent the full or empty set. That is,
2549 In addition, the pairs must be in signed order of the lower bound and
2550 they must be non-contiguous.
2554 .. code-block:: llvm
2556 %a = load i8* %x, align 1, !range !0 ; Can only be 0 or 1
2557 %b = load i8* %y, align 1, !range !1 ; Can only be 255 (-1), 0 or 1
2558 %c = load i8* %z, align 1, !range !2 ; Can only be 0, 1, 3, 4 or 5
2559 %d = load i8* %z, align 1, !range !3 ; Can only be -2, -1, 3, 4 or 5
2561 !0 = metadata !{ i8 0, i8 2 }
2562 !1 = metadata !{ i8 255, i8 2 }
2563 !2 = metadata !{ i8 0, i8 2, i8 3, i8 6 }
2564 !3 = metadata !{ i8 -2, i8 0, i8 3, i8 6 }
2569 It is sometimes useful to attach information to loop constructs. Currently,
2570 loop metadata is implemented as metadata attached to the branch instruction
2571 in the loop latch block. This type of metadata refer to a metadata node that is
2572 guaranteed to be separate for each loop. The loop identifier metadata is
2573 specified with the name ``llvm.loop``.
2575 The loop identifier metadata is implemented using a metadata that refers to
2576 itself to avoid merging it with any other identifier metadata, e.g.,
2577 during module linkage or function inlining. That is, each loop should refer
2578 to their own identification metadata even if they reside in separate functions.
2579 The following example contains loop identifier metadata for two separate loop
2582 .. code-block:: llvm
2584 !0 = metadata !{ metadata !0 }
2585 !1 = metadata !{ metadata !1 }
2587 The loop identifier metadata can be used to specify additional per-loop
2588 metadata. Any operands after the first operand can be treated as user-defined
2589 metadata. For example the ``llvm.vectorizer.unroll`` metadata is understood
2590 by the loop vectorizer to indicate how many times to unroll the loop:
2592 .. code-block:: llvm
2594 br i1 %exitcond, label %._crit_edge, label %.lr.ph, !llvm.loop !0
2596 !0 = metadata !{ metadata !0, metadata !1 }
2597 !1 = metadata !{ metadata !"llvm.vectorizer.unroll", i32 2 }
2602 Metadata types used to annotate memory accesses with information helpful
2603 for optimizations are prefixed with ``llvm.mem``.
2605 '``llvm.mem.parallel_loop_access``' Metadata
2606 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2608 For a loop to be parallel, in addition to using
2609 the ``llvm.loop`` metadata to mark the loop latch branch instruction,
2610 also all of the memory accessing instructions in the loop body need to be
2611 marked with the ``llvm.mem.parallel_loop_access`` metadata. If there
2612 is at least one memory accessing instruction not marked with the metadata,
2613 the loop must be considered a sequential loop. This causes parallel loops to be
2614 converted to sequential loops due to optimization passes that are unaware of
2615 the parallel semantics and that insert new memory instructions to the loop
2618 Example of a loop that is considered parallel due to its correct use of
2619 both ``llvm.loop`` and ``llvm.mem.parallel_loop_access``
2620 metadata types that refer to the same loop identifier metadata.
2622 .. code-block:: llvm
2626 %0 = load i32* %arrayidx, align 4, !llvm.mem.parallel_loop_access !0
2628 store i32 %0, i32* %arrayidx4, align 4, !llvm.mem.parallel_loop_access !0
2630 br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !0
2634 !0 = metadata !{ metadata !0 }
2636 It is also possible to have nested parallel loops. In that case the
2637 memory accesses refer to a list of loop identifier metadata nodes instead of
2638 the loop identifier metadata node directly:
2640 .. code-block:: llvm
2647 %0 = load i32* %arrayidx, align 4, !llvm.mem.parallel_loop_access !0
2649 store i32 %0, i32* %arrayidx4, align 4, !llvm.mem.parallel_loop_access !0
2651 br i1 %exitcond, label %inner.for.end, label %inner.for.body, !llvm.loop !1
2655 %0 = load i32* %arrayidx, align 4, !llvm.mem.parallel_loop_access !0
2657 store i32 %0, i32* %arrayidx4, align 4, !llvm.mem.parallel_loop_access !0
2659 br i1 %exitcond, label %outer.for.end, label %outer.for.body, !llvm.loop !2
2661 outer.for.end: ; preds = %for.body
2663 !0 = metadata !{ metadata !1, metadata !2 } ; a list of loop identifiers
2664 !1 = metadata !{ metadata !1 } ; an identifier for the inner loop
2665 !2 = metadata !{ metadata !2 } ; an identifier for the outer loop
2667 '``llvm.vectorizer``'
2668 ^^^^^^^^^^^^^^^^^^^^^
2670 Metadata prefixed with ``llvm.vectorizer`` is used to control per-loop
2671 vectorization parameters such as vectorization factor and unroll factor.
2673 ``llvm.vectorizer`` metadata should be used in conjunction with ``llvm.loop``
2674 loop identification metadata.
2676 '``llvm.vectorizer.unroll``' Metadata
2677 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2679 This metadata instructs the loop vectorizer to unroll the specified
2680 loop exactly ``N`` times.
2682 The first operand is the string ``llvm.vectorizer.unroll`` and the second
2683 operand is an integer specifying the unroll factor. For example:
2685 .. code-block:: llvm
2687 !0 = metadata !{ metadata !"llvm.vectorizer.unroll", i32 4 }
2689 Note that setting ``llvm.vectorizer.unroll`` to 1 disables unrolling of the
2692 If ``llvm.vectorizer.unroll`` is set to 0 then the amount of unrolling will be
2693 determined automatically.
2695 '``llvm.vectorizer.width``' Metadata
2696 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2698 This metadata sets the target width of the vectorizer to ``N``. Without
2699 this metadata, the vectorizer will choose a width automatically.
2700 Regardless of this metadata, the vectorizer will only vectorize loops if
2701 it believes it is valid to do so.
2703 The first operand is the string ``llvm.vectorizer.width`` and the second
2704 operand is an integer specifying the width. For example:
2706 .. code-block:: llvm
2708 !0 = metadata !{ metadata !"llvm.vectorizer.width", i32 4 }
2710 Note that setting ``llvm.vectorizer.width`` to 1 disables vectorization of the
2713 If ``llvm.vectorizer.width`` is set to 0 then the width will be determined
2716 Module Flags Metadata
2717 =====================
2719 Information about the module as a whole is difficult to convey to LLVM's
2720 subsystems. The LLVM IR isn't sufficient to transmit this information.
2721 The ``llvm.module.flags`` named metadata exists in order to facilitate
2722 this. These flags are in the form of key / value pairs --- much like a
2723 dictionary --- making it easy for any subsystem who cares about a flag to
2726 The ``llvm.module.flags`` metadata contains a list of metadata triplets.
2727 Each triplet has the following form:
2729 - The first element is a *behavior* flag, which specifies the behavior
2730 when two (or more) modules are merged together, and it encounters two
2731 (or more) metadata with the same ID. The supported behaviors are
2733 - The second element is a metadata string that is a unique ID for the
2734 metadata. Each module may only have one flag entry for each unique ID (not
2735 including entries with the **Require** behavior).
2736 - The third element is the value of the flag.
2738 When two (or more) modules are merged together, the resulting
2739 ``llvm.module.flags`` metadata is the union of the modules' flags. That is, for
2740 each unique metadata ID string, there will be exactly one entry in the merged
2741 modules ``llvm.module.flags`` metadata table, and the value for that entry will
2742 be determined by the merge behavior flag, as described below. The only exception
2743 is that entries with the *Require* behavior are always preserved.
2745 The following behaviors are supported:
2756 Emits an error if two values disagree, otherwise the resulting value
2757 is that of the operands.
2761 Emits a warning if two values disagree. The result value will be the
2762 operand for the flag from the first module being linked.
2766 Adds a requirement that another module flag be present and have a
2767 specified value after linking is performed. The value must be a
2768 metadata pair, where the first element of the pair is the ID of the
2769 module flag to be restricted, and the second element of the pair is
2770 the value the module flag should be restricted to. This behavior can
2771 be used to restrict the allowable results (via triggering of an
2772 error) of linking IDs with the **Override** behavior.
2776 Uses the specified value, regardless of the behavior or value of the
2777 other module. If both modules specify **Override**, but the values
2778 differ, an error will be emitted.
2782 Appends the two values, which are required to be metadata nodes.
2786 Appends the two values, which are required to be metadata
2787 nodes. However, duplicate entries in the second list are dropped
2788 during the append operation.
2790 It is an error for a particular unique flag ID to have multiple behaviors,
2791 except in the case of **Require** (which adds restrictions on another metadata
2792 value) or **Override**.
2794 An example of module flags:
2796 .. code-block:: llvm
2798 !0 = metadata !{ i32 1, metadata !"foo", i32 1 }
2799 !1 = metadata !{ i32 4, metadata !"bar", i32 37 }
2800 !2 = metadata !{ i32 2, metadata !"qux", i32 42 }
2801 !3 = metadata !{ i32 3, metadata !"qux",
2803 metadata !"foo", i32 1
2806 !llvm.module.flags = !{ !0, !1, !2, !3 }
2808 - Metadata ``!0`` has the ID ``!"foo"`` and the value '1'. The behavior
2809 if two or more ``!"foo"`` flags are seen is to emit an error if their
2810 values are not equal.
2812 - Metadata ``!1`` has the ID ``!"bar"`` and the value '37'. The
2813 behavior if two or more ``!"bar"`` flags are seen is to use the value
2816 - Metadata ``!2`` has the ID ``!"qux"`` and the value '42'. The
2817 behavior if two or more ``!"qux"`` flags are seen is to emit a
2818 warning if their values are not equal.
2820 - Metadata ``!3`` has the ID ``!"qux"`` and the value:
2824 metadata !{ metadata !"foo", i32 1 }
2826 The behavior is to emit an error if the ``llvm.module.flags`` does not
2827 contain a flag with the ID ``!"foo"`` that has the value '1' after linking is
2830 Objective-C Garbage Collection Module Flags Metadata
2831 ----------------------------------------------------
2833 On the Mach-O platform, Objective-C stores metadata about garbage
2834 collection in a special section called "image info". The metadata
2835 consists of a version number and a bitmask specifying what types of
2836 garbage collection are supported (if any) by the file. If two or more
2837 modules are linked together their garbage collection metadata needs to
2838 be merged rather than appended together.
2840 The Objective-C garbage collection module flags metadata consists of the
2841 following key-value pairs:
2850 * - ``Objective-C Version``
2851 - **[Required]** --- The Objective-C ABI version. Valid values are 1 and 2.
2853 * - ``Objective-C Image Info Version``
2854 - **[Required]** --- The version of the image info section. Currently
2857 * - ``Objective-C Image Info Section``
2858 - **[Required]** --- The section to place the metadata. Valid values are
2859 ``"__OBJC, __image_info, regular"`` for Objective-C ABI version 1, and
2860 ``"__DATA,__objc_imageinfo, regular, no_dead_strip"`` for
2861 Objective-C ABI version 2.
2863 * - ``Objective-C Garbage Collection``
2864 - **[Required]** --- Specifies whether garbage collection is supported or
2865 not. Valid values are 0, for no garbage collection, and 2, for garbage
2866 collection supported.
2868 * - ``Objective-C GC Only``
2869 - **[Optional]** --- Specifies that only garbage collection is supported.
2870 If present, its value must be 6. This flag requires that the
2871 ``Objective-C Garbage Collection`` flag have the value 2.
2873 Some important flag interactions:
2875 - If a module with ``Objective-C Garbage Collection`` set to 0 is
2876 merged with a module with ``Objective-C Garbage Collection`` set to
2877 2, then the resulting module has the
2878 ``Objective-C Garbage Collection`` flag set to 0.
2879 - A module with ``Objective-C Garbage Collection`` set to 0 cannot be
2880 merged with a module with ``Objective-C GC Only`` set to 6.
2882 Automatic Linker Flags Module Flags Metadata
2883 --------------------------------------------
2885 Some targets support embedding flags to the linker inside individual object
2886 files. Typically this is used in conjunction with language extensions which
2887 allow source files to explicitly declare the libraries they depend on, and have
2888 these automatically be transmitted to the linker via object files.
2890 These flags are encoded in the IR using metadata in the module flags section,
2891 using the ``Linker Options`` key. The merge behavior for this flag is required
2892 to be ``AppendUnique``, and the value for the key is expected to be a metadata
2893 node which should be a list of other metadata nodes, each of which should be a
2894 list of metadata strings defining linker options.
2896 For example, the following metadata section specifies two separate sets of
2897 linker options, presumably to link against ``libz`` and the ``Cocoa``
2900 !0 = metadata !{ i32 6, metadata !"Linker Options",
2902 metadata !{ metadata !"-lz" },
2903 metadata !{ metadata !"-framework", metadata !"Cocoa" } } }
2904 !llvm.module.flags = !{ !0 }
2906 The metadata encoding as lists of lists of options, as opposed to a collapsed
2907 list of options, is chosen so that the IR encoding can use multiple option
2908 strings to specify e.g., a single library, while still having that specifier be
2909 preserved as an atomic element that can be recognized by a target specific
2910 assembly writer or object file emitter.
2912 Each individual option is required to be either a valid option for the target's
2913 linker, or an option that is reserved by the target specific assembly writer or
2914 object file emitter. No other aspect of these options is defined by the IR.
2916 .. _intrinsicglobalvariables:
2918 Intrinsic Global Variables
2919 ==========================
2921 LLVM has a number of "magic" global variables that contain data that
2922 affect code generation or other IR semantics. These are documented here.
2923 All globals of this sort should have a section specified as
2924 "``llvm.metadata``". This section and all globals that start with
2925 "``llvm.``" are reserved for use by LLVM.
2929 The '``llvm.used``' Global Variable
2930 -----------------------------------
2932 The ``@llvm.used`` global is an array which has
2933 :ref:`appending linkage <linkage_appending>`. This array contains a list of
2934 pointers to named global variables, functions and aliases which may optionally
2935 have a pointer cast formed of bitcast or getelementptr. For example, a legal
2938 .. code-block:: llvm
2943 @llvm.used = appending global [2 x i8*] [
2945 i8* bitcast (i32* @Y to i8*)
2946 ], section "llvm.metadata"
2948 If a symbol appears in the ``@llvm.used`` list, then the compiler, assembler,
2949 and linker are required to treat the symbol as if there is a reference to the
2950 symbol that it cannot see (which is why they have to be named). For example, if
2951 a variable has internal linkage and no references other than that from the
2952 ``@llvm.used`` list, it cannot be deleted. This is commonly used to represent
2953 references from inline asms and other things the compiler cannot "see", and
2954 corresponds to "``attribute((used))``" in GNU C.
2956 On some targets, the code generator must emit a directive to the
2957 assembler or object file to prevent the assembler and linker from
2958 molesting the symbol.
2960 .. _gv_llvmcompilerused:
2962 The '``llvm.compiler.used``' Global Variable
2963 --------------------------------------------
2965 The ``@llvm.compiler.used`` directive is the same as the ``@llvm.used``
2966 directive, except that it only prevents the compiler from touching the
2967 symbol. On targets that support it, this allows an intelligent linker to
2968 optimize references to the symbol without being impeded as it would be
2971 This is a rare construct that should only be used in rare circumstances,
2972 and should not be exposed to source languages.
2974 .. _gv_llvmglobalctors:
2976 The '``llvm.global_ctors``' Global Variable
2977 -------------------------------------------
2979 .. code-block:: llvm
2981 %0 = type { i32, void ()* }
2982 @llvm.global_ctors = appending global [1 x %0] [%0 { i32 65535, void ()* @ctor }]
2984 The ``@llvm.global_ctors`` array contains a list of constructor
2985 functions and associated priorities. The functions referenced by this
2986 array will be called in ascending order of priority (i.e. lowest first)
2987 when the module is loaded. The order of functions with the same priority
2990 .. _llvmglobaldtors:
2992 The '``llvm.global_dtors``' Global Variable
2993 -------------------------------------------
2995 .. code-block:: llvm
2997 %0 = type { i32, void ()* }
2998 @llvm.global_dtors = appending global [1 x %0] [%0 { i32 65535, void ()* @dtor }]
3000 The ``@llvm.global_dtors`` array contains a list of destructor functions
3001 and associated priorities. The functions referenced by this array will
3002 be called in descending order of priority (i.e. highest first) when the
3003 module is loaded. The order of functions with the same priority is not
3006 Instruction Reference
3007 =====================
3009 The LLVM instruction set consists of several different classifications
3010 of instructions: :ref:`terminator instructions <terminators>`, :ref:`binary
3011 instructions <binaryops>`, :ref:`bitwise binary
3012 instructions <bitwiseops>`, :ref:`memory instructions <memoryops>`, and
3013 :ref:`other instructions <otherops>`.
3017 Terminator Instructions
3018 -----------------------
3020 As mentioned :ref:`previously <functionstructure>`, every basic block in a
3021 program ends with a "Terminator" instruction, which indicates which
3022 block should be executed after the current block is finished. These
3023 terminator instructions typically yield a '``void``' value: they produce
3024 control flow, not values (the one exception being the
3025 ':ref:`invoke <i_invoke>`' instruction).
3027 The terminator instructions are: ':ref:`ret <i_ret>`',
3028 ':ref:`br <i_br>`', ':ref:`switch <i_switch>`',
3029 ':ref:`indirectbr <i_indirectbr>`', ':ref:`invoke <i_invoke>`',
3030 ':ref:`resume <i_resume>`', and ':ref:`unreachable <i_unreachable>`'.
3034 '``ret``' Instruction
3035 ^^^^^^^^^^^^^^^^^^^^^
3042 ret <type> <value> ; Return a value from a non-void function
3043 ret void ; Return from void function
3048 The '``ret``' instruction is used to return control flow (and optionally
3049 a value) from a function back to the caller.
3051 There are two forms of the '``ret``' instruction: one that returns a
3052 value and then causes control flow, and one that just causes control
3058 The '``ret``' instruction optionally accepts a single argument, the
3059 return value. The type of the return value must be a ':ref:`first
3060 class <t_firstclass>`' type.
3062 A function is not :ref:`well formed <wellformed>` if it it has a non-void
3063 return type and contains a '``ret``' instruction with no return value or
3064 a return value with a type that does not match its type, or if it has a
3065 void return type and contains a '``ret``' instruction with a return
3071 When the '``ret``' instruction is executed, control flow returns back to
3072 the calling function's context. If the caller is a
3073 ":ref:`call <i_call>`" instruction, execution continues at the
3074 instruction after the call. If the caller was an
3075 ":ref:`invoke <i_invoke>`" instruction, execution continues at the
3076 beginning of the "normal" destination block. If the instruction returns
3077 a value, that value shall set the call or invoke instruction's return
3083 .. code-block:: llvm
3085 ret i32 5 ; Return an integer value of 5
3086 ret void ; Return from a void function
3087 ret { i32, i8 } { i32 4, i8 2 } ; Return a struct of values 4 and 2
3091 '``br``' Instruction
3092 ^^^^^^^^^^^^^^^^^^^^
3099 br i1 <cond>, label <iftrue>, label <iffalse>
3100 br label <dest> ; Unconditional branch
3105 The '``br``' instruction is used to cause control flow to transfer to a
3106 different basic block in the current function. There are two forms of
3107 this instruction, corresponding to a conditional branch and an
3108 unconditional branch.
3113 The conditional branch form of the '``br``' instruction takes a single
3114 '``i1``' value and two '``label``' values. The unconditional form of the
3115 '``br``' instruction takes a single '``label``' value as a target.
3120 Upon execution of a conditional '``br``' instruction, the '``i1``'
3121 argument is evaluated. If the value is ``true``, control flows to the
3122 '``iftrue``' ``label`` argument. If "cond" is ``false``, control flows
3123 to the '``iffalse``' ``label`` argument.
3128 .. code-block:: llvm
3131 %cond = icmp eq i32 %a, %b
3132 br i1 %cond, label %IfEqual, label %IfUnequal
3140 '``switch``' Instruction
3141 ^^^^^^^^^^^^^^^^^^^^^^^^
3148 switch <intty> <value>, label <defaultdest> [ <intty> <val>, label <dest> ... ]
3153 The '``switch``' instruction is used to transfer control flow to one of
3154 several different places. It is a generalization of the '``br``'
3155 instruction, allowing a branch to occur to one of many possible
3161 The '``switch``' instruction uses three parameters: an integer
3162 comparison value '``value``', a default '``label``' destination, and an
3163 array of pairs of comparison value constants and '``label``'s. The table
3164 is not allowed to contain duplicate constant entries.
3169 The ``switch`` instruction specifies a table of values and destinations.
3170 When the '``switch``' instruction is executed, this table is searched
3171 for the given value. If the value is found, control flow is transferred
3172 to the corresponding destination; otherwise, control flow is transferred
3173 to the default destination.
3178 Depending on properties of the target machine and the particular
3179 ``switch`` instruction, this instruction may be code generated in
3180 different ways. For example, it could be generated as a series of
3181 chained conditional branches or with a lookup table.
3186 .. code-block:: llvm
3188 ; Emulate a conditional br instruction
3189 %Val = zext i1 %value to i32
3190 switch i32 %Val, label %truedest [ i32 0, label %falsedest ]
3192 ; Emulate an unconditional br instruction
3193 switch i32 0, label %dest [ ]
3195 ; Implement a jump table:
3196 switch i32 %val, label %otherwise [ i32 0, label %onzero
3198 i32 2, label %ontwo ]
3202 '``indirectbr``' Instruction
3203 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3210 indirectbr <somety>* <address>, [ label <dest1>, label <dest2>, ... ]
3215 The '``indirectbr``' instruction implements an indirect branch to a
3216 label within the current function, whose address is specified by
3217 "``address``". Address must be derived from a
3218 :ref:`blockaddress <blockaddress>` constant.
3223 The '``address``' argument is the address of the label to jump to. The
3224 rest of the arguments indicate the full set of possible destinations
3225 that the address may point to. Blocks are allowed to occur multiple
3226 times in the destination list, though this isn't particularly useful.
3228 This destination list is required so that dataflow analysis has an
3229 accurate understanding of the CFG.
3234 Control transfers to the block specified in the address argument. All
3235 possible destination blocks must be listed in the label list, otherwise
3236 this instruction has undefined behavior. This implies that jumps to
3237 labels defined in other functions have undefined behavior as well.
3242 This is typically implemented with a jump through a register.
3247 .. code-block:: llvm
3249 indirectbr i8* %Addr, [ label %bb1, label %bb2, label %bb3 ]
3253 '``invoke``' Instruction
3254 ^^^^^^^^^^^^^^^^^^^^^^^^
3261 <result> = invoke [cconv] [ret attrs] <ptr to function ty> <function ptr val>(<function args>) [fn attrs]
3262 to label <normal label> unwind label <exception label>
3267 The '``invoke``' instruction causes control to transfer to a specified
3268 function, with the possibility of control flow transfer to either the
3269 '``normal``' label or the '``exception``' label. If the callee function
3270 returns with the "``ret``" instruction, control flow will return to the
3271 "normal" label. If the callee (or any indirect callees) returns via the
3272 ":ref:`resume <i_resume>`" instruction or other exception handling
3273 mechanism, control is interrupted and continued at the dynamically
3274 nearest "exception" label.
3276 The '``exception``' label is a `landing
3277 pad <ExceptionHandling.html#overview>`_ for the exception. As such,
3278 '``exception``' label is required to have the
3279 ":ref:`landingpad <i_landingpad>`" instruction, which contains the
3280 information about the behavior of the program after unwinding happens,
3281 as its first non-PHI instruction. The restrictions on the
3282 "``landingpad``" instruction's tightly couples it to the "``invoke``"
3283 instruction, so that the important information contained within the
3284 "``landingpad``" instruction can't be lost through normal code motion.
3289 This instruction requires several arguments:
3291 #. The optional "cconv" marker indicates which :ref:`calling
3292 convention <callingconv>` the call should use. If none is
3293 specified, the call defaults to using C calling conventions.
3294 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
3295 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
3297 #. '``ptr to function ty``': shall be the signature of the pointer to
3298 function value being invoked. In most cases, this is a direct
3299 function invocation, but indirect ``invoke``'s are just as possible,
3300 branching off an arbitrary pointer to function value.
3301 #. '``function ptr val``': An LLVM value containing a pointer to a
3302 function to be invoked.
3303 #. '``function args``': argument list whose types match the function
3304 signature argument types and parameter attributes. All arguments must
3305 be of :ref:`first class <t_firstclass>` type. If the function signature
3306 indicates the function accepts a variable number of arguments, the
3307 extra arguments can be specified.
3308 #. '``normal label``': the label reached when the called function
3309 executes a '``ret``' instruction.
3310 #. '``exception label``': the label reached when a callee returns via
3311 the :ref:`resume <i_resume>` instruction or other exception handling
3313 #. The optional :ref:`function attributes <fnattrs>` list. Only
3314 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
3315 attributes are valid here.
3320 This instruction is designed to operate as a standard '``call``'
3321 instruction in most regards. The primary difference is that it
3322 establishes an association with a label, which is used by the runtime
3323 library to unwind the stack.
3325 This instruction is used in languages with destructors to ensure that
3326 proper cleanup is performed in the case of either a ``longjmp`` or a
3327 thrown exception. Additionally, this is important for implementation of
3328 '``catch``' clauses in high-level languages that support them.
3330 For the purposes of the SSA form, the definition of the value returned
3331 by the '``invoke``' instruction is deemed to occur on the edge from the
3332 current block to the "normal" label. If the callee unwinds then no
3333 return value is available.
3338 .. code-block:: llvm
3340 %retval = invoke i32 @Test(i32 15) to label %Continue
3341 unwind label %TestCleanup ; {i32}:retval set
3342 %retval = invoke coldcc i32 %Testfnptr(i32 15) to label %Continue
3343 unwind label %TestCleanup ; {i32}:retval set
3347 '``resume``' Instruction
3348 ^^^^^^^^^^^^^^^^^^^^^^^^
3355 resume <type> <value>
3360 The '``resume``' instruction is a terminator instruction that has no
3366 The '``resume``' instruction requires one argument, which must have the
3367 same type as the result of any '``landingpad``' instruction in the same
3373 The '``resume``' instruction resumes propagation of an existing
3374 (in-flight) exception whose unwinding was interrupted with a
3375 :ref:`landingpad <i_landingpad>` instruction.
3380 .. code-block:: llvm
3382 resume { i8*, i32 } %exn
3386 '``unreachable``' Instruction
3387 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3399 The '``unreachable``' instruction has no defined semantics. This
3400 instruction is used to inform the optimizer that a particular portion of
3401 the code is not reachable. This can be used to indicate that the code
3402 after a no-return function cannot be reached, and other facts.
3407 The '``unreachable``' instruction has no defined semantics.
3414 Binary operators are used to do most of the computation in a program.
3415 They require two operands of the same type, execute an operation on
3416 them, and produce a single value. The operands might represent multiple
3417 data, as is the case with the :ref:`vector <t_vector>` data type. The
3418 result value has the same type as its operands.
3420 There are several different binary operators:
3424 '``add``' Instruction
3425 ^^^^^^^^^^^^^^^^^^^^^
3432 <result> = add <ty> <op1>, <op2> ; yields {ty}:result
3433 <result> = add nuw <ty> <op1>, <op2> ; yields {ty}:result
3434 <result> = add nsw <ty> <op1>, <op2> ; yields {ty}:result
3435 <result> = add nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3440 The '``add``' instruction returns the sum of its two operands.
3445 The two arguments to the '``add``' instruction must be
3446 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3447 arguments must have identical types.
3452 The value produced is the integer sum of the two operands.
3454 If the sum has unsigned overflow, the result returned is the
3455 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
3458 Because LLVM integers use a two's complement representation, this
3459 instruction is appropriate for both signed and unsigned integers.
3461 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3462 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3463 result value of the ``add`` is a :ref:`poison value <poisonvalues>` if
3464 unsigned and/or signed overflow, respectively, occurs.
3469 .. code-block:: llvm
3471 <result> = add i32 4, %var ; yields {i32}:result = 4 + %var
3475 '``fadd``' Instruction
3476 ^^^^^^^^^^^^^^^^^^^^^^
3483 <result> = fadd [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3488 The '``fadd``' instruction returns the sum of its two operands.
3493 The two arguments to the '``fadd``' instruction must be :ref:`floating
3494 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3495 Both arguments must have identical types.
3500 The value produced is the floating point sum of the two operands. This
3501 instruction can also take any number of :ref:`fast-math flags <fastmath>`,
3502 which are optimization hints to enable otherwise unsafe floating point
3508 .. code-block:: llvm
3510 <result> = fadd float 4.0, %var ; yields {float}:result = 4.0 + %var
3512 '``sub``' Instruction
3513 ^^^^^^^^^^^^^^^^^^^^^
3520 <result> = sub <ty> <op1>, <op2> ; yields {ty}:result
3521 <result> = sub nuw <ty> <op1>, <op2> ; yields {ty}:result
3522 <result> = sub nsw <ty> <op1>, <op2> ; yields {ty}:result
3523 <result> = sub nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3528 The '``sub``' instruction returns the difference of its two operands.
3530 Note that the '``sub``' instruction is used to represent the '``neg``'
3531 instruction present in most other intermediate representations.
3536 The two arguments to the '``sub``' instruction must be
3537 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3538 arguments must have identical types.
3543 The value produced is the integer difference of the two operands.
3545 If the difference has unsigned overflow, the result returned is the
3546 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
3549 Because LLVM integers use a two's complement representation, this
3550 instruction is appropriate for both signed and unsigned integers.
3552 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3553 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3554 result value of the ``sub`` is a :ref:`poison value <poisonvalues>` if
3555 unsigned and/or signed overflow, respectively, occurs.
3560 .. code-block:: llvm
3562 <result> = sub i32 4, %var ; yields {i32}:result = 4 - %var
3563 <result> = sub i32 0, %val ; yields {i32}:result = -%var
3567 '``fsub``' Instruction
3568 ^^^^^^^^^^^^^^^^^^^^^^
3575 <result> = fsub [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3580 The '``fsub``' instruction returns the difference of its two operands.
3582 Note that the '``fsub``' instruction is used to represent the '``fneg``'
3583 instruction present in most other intermediate representations.
3588 The two arguments to the '``fsub``' instruction must be :ref:`floating
3589 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3590 Both arguments must have identical types.
3595 The value produced is the floating point difference of the two operands.
3596 This instruction can also take any number of :ref:`fast-math
3597 flags <fastmath>`, which are optimization hints to enable otherwise
3598 unsafe floating point optimizations:
3603 .. code-block:: llvm
3605 <result> = fsub float 4.0, %var ; yields {float}:result = 4.0 - %var
3606 <result> = fsub float -0.0, %val ; yields {float}:result = -%var
3608 '``mul``' Instruction
3609 ^^^^^^^^^^^^^^^^^^^^^
3616 <result> = mul <ty> <op1>, <op2> ; yields {ty}:result
3617 <result> = mul nuw <ty> <op1>, <op2> ; yields {ty}:result
3618 <result> = mul nsw <ty> <op1>, <op2> ; yields {ty}:result
3619 <result> = mul nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3624 The '``mul``' instruction returns the product of its two operands.
3629 The two arguments to the '``mul``' instruction must be
3630 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3631 arguments must have identical types.
3636 The value produced is the integer product of the two operands.
3638 If the result of the multiplication has unsigned overflow, the result
3639 returned is the mathematical result modulo 2\ :sup:`n`\ , where n is the
3640 bit width of the result.
3642 Because LLVM integers use a two's complement representation, and the
3643 result is the same width as the operands, this instruction returns the
3644 correct result for both signed and unsigned integers. If a full product
3645 (e.g. ``i32`` * ``i32`` -> ``i64``) is needed, the operands should be
3646 sign-extended or zero-extended as appropriate to the width of the full
3649 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3650 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3651 result value of the ``mul`` is a :ref:`poison value <poisonvalues>` if
3652 unsigned and/or signed overflow, respectively, occurs.
3657 .. code-block:: llvm
3659 <result> = mul i32 4, %var ; yields {i32}:result = 4 * %var
3663 '``fmul``' Instruction
3664 ^^^^^^^^^^^^^^^^^^^^^^
3671 <result> = fmul [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3676 The '``fmul``' instruction returns the product of its two operands.
3681 The two arguments to the '``fmul``' instruction must be :ref:`floating
3682 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3683 Both arguments must have identical types.
3688 The value produced is the floating point product of the two operands.
3689 This instruction can also take any number of :ref:`fast-math
3690 flags <fastmath>`, which are optimization hints to enable otherwise
3691 unsafe floating point optimizations:
3696 .. code-block:: llvm
3698 <result> = fmul float 4.0, %var ; yields {float}:result = 4.0 * %var
3700 '``udiv``' Instruction
3701 ^^^^^^^^^^^^^^^^^^^^^^
3708 <result> = udiv <ty> <op1>, <op2> ; yields {ty}:result
3709 <result> = udiv exact <ty> <op1>, <op2> ; yields {ty}:result
3714 The '``udiv``' instruction returns the quotient of its two operands.
3719 The two arguments to the '``udiv``' instruction must be
3720 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3721 arguments must have identical types.
3726 The value produced is the unsigned integer quotient of the two operands.
3728 Note that unsigned integer division and signed integer division are
3729 distinct operations; for signed integer division, use '``sdiv``'.
3731 Division by zero leads to undefined behavior.
3733 If the ``exact`` keyword is present, the result value of the ``udiv`` is
3734 a :ref:`poison value <poisonvalues>` if %op1 is not a multiple of %op2 (as
3735 such, "((a udiv exact b) mul b) == a").
3740 .. code-block:: llvm
3742 <result> = udiv i32 4, %var ; yields {i32}:result = 4 / %var
3744 '``sdiv``' Instruction
3745 ^^^^^^^^^^^^^^^^^^^^^^
3752 <result> = sdiv <ty> <op1>, <op2> ; yields {ty}:result
3753 <result> = sdiv exact <ty> <op1>, <op2> ; yields {ty}:result
3758 The '``sdiv``' instruction returns the quotient of its two operands.
3763 The two arguments to the '``sdiv``' instruction must be
3764 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3765 arguments must have identical types.
3770 The value produced is the signed integer quotient of the two operands
3771 rounded towards zero.
3773 Note that signed integer division and unsigned integer division are
3774 distinct operations; for unsigned integer division, use '``udiv``'.
3776 Division by zero leads to undefined behavior. Overflow also leads to
3777 undefined behavior; this is a rare case, but can occur, for example, by
3778 doing a 32-bit division of -2147483648 by -1.
3780 If the ``exact`` keyword is present, the result value of the ``sdiv`` is
3781 a :ref:`poison value <poisonvalues>` if the result would be rounded.
3786 .. code-block:: llvm
3788 <result> = sdiv i32 4, %var ; yields {i32}:result = 4 / %var
3792 '``fdiv``' Instruction
3793 ^^^^^^^^^^^^^^^^^^^^^^
3800 <result> = fdiv [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3805 The '``fdiv``' instruction returns the quotient of its two operands.
3810 The two arguments to the '``fdiv``' instruction must be :ref:`floating
3811 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3812 Both arguments must have identical types.
3817 The value produced is the floating point quotient of the two operands.
3818 This instruction can also take any number of :ref:`fast-math
3819 flags <fastmath>`, which are optimization hints to enable otherwise
3820 unsafe floating point optimizations:
3825 .. code-block:: llvm
3827 <result> = fdiv float 4.0, %var ; yields {float}:result = 4.0 / %var
3829 '``urem``' Instruction
3830 ^^^^^^^^^^^^^^^^^^^^^^
3837 <result> = urem <ty> <op1>, <op2> ; yields {ty}:result
3842 The '``urem``' instruction returns the remainder from the unsigned
3843 division of its two arguments.
3848 The two arguments to the '``urem``' instruction must be
3849 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3850 arguments must have identical types.
3855 This instruction returns the unsigned integer *remainder* of a division.
3856 This instruction always performs an unsigned division to get the
3859 Note that unsigned integer remainder and signed integer remainder are
3860 distinct operations; for signed integer remainder, use '``srem``'.
3862 Taking the remainder of a division by zero leads to undefined behavior.
3867 .. code-block:: llvm
3869 <result> = urem i32 4, %var ; yields {i32}:result = 4 % %var
3871 '``srem``' Instruction
3872 ^^^^^^^^^^^^^^^^^^^^^^
3879 <result> = srem <ty> <op1>, <op2> ; yields {ty}:result
3884 The '``srem``' instruction returns the remainder from the signed
3885 division of its two operands. This instruction can also take
3886 :ref:`vector <t_vector>` versions of the values in which case the elements
3892 The two arguments to the '``srem``' instruction must be
3893 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3894 arguments must have identical types.
3899 This instruction returns the *remainder* of a division (where the result
3900 is either zero or has the same sign as the dividend, ``op1``), not the
3901 *modulo* operator (where the result is either zero or has the same sign
3902 as the divisor, ``op2``) of a value. For more information about the
3903 difference, see `The Math
3904 Forum <http://mathforum.org/dr.math/problems/anne.4.28.99.html>`_. For a
3905 table of how this is implemented in various languages, please see
3907 operation <http://en.wikipedia.org/wiki/Modulo_operation>`_.
3909 Note that signed integer remainder and unsigned integer remainder are
3910 distinct operations; for unsigned integer remainder, use '``urem``'.
3912 Taking the remainder of a division by zero leads to undefined behavior.
3913 Overflow also leads to undefined behavior; this is a rare case, but can
3914 occur, for example, by taking the remainder of a 32-bit division of
3915 -2147483648 by -1. (The remainder doesn't actually overflow, but this
3916 rule lets srem be implemented using instructions that return both the
3917 result of the division and the remainder.)
3922 .. code-block:: llvm
3924 <result> = srem i32 4, %var ; yields {i32}:result = 4 % %var
3928 '``frem``' Instruction
3929 ^^^^^^^^^^^^^^^^^^^^^^
3936 <result> = frem [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3941 The '``frem``' instruction returns the remainder from the division of
3947 The two arguments to the '``frem``' instruction must be :ref:`floating
3948 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3949 Both arguments must have identical types.
3954 This instruction returns the *remainder* of a division. The remainder
3955 has the same sign as the dividend. This instruction can also take any
3956 number of :ref:`fast-math flags <fastmath>`, which are optimization hints
3957 to enable otherwise unsafe floating point optimizations:
3962 .. code-block:: llvm
3964 <result> = frem float 4.0, %var ; yields {float}:result = 4.0 % %var
3968 Bitwise Binary Operations
3969 -------------------------
3971 Bitwise binary operators are used to do various forms of bit-twiddling
3972 in a program. They are generally very efficient instructions and can
3973 commonly be strength reduced from other instructions. They require two
3974 operands of the same type, execute an operation on them, and produce a
3975 single value. The resulting value is the same type as its operands.
3977 '``shl``' Instruction
3978 ^^^^^^^^^^^^^^^^^^^^^
3985 <result> = shl <ty> <op1>, <op2> ; yields {ty}:result
3986 <result> = shl nuw <ty> <op1>, <op2> ; yields {ty}:result
3987 <result> = shl nsw <ty> <op1>, <op2> ; yields {ty}:result
3988 <result> = shl nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3993 The '``shl``' instruction returns the first operand shifted to the left
3994 a specified number of bits.
3999 Both arguments to the '``shl``' instruction must be the same
4000 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4001 '``op2``' is treated as an unsigned value.
4006 The value produced is ``op1`` \* 2\ :sup:`op2` mod 2\ :sup:`n`,
4007 where ``n`` is the width of the result. If ``op2`` is (statically or
4008 dynamically) negative or equal to or larger than the number of bits in
4009 ``op1``, the result is undefined. If the arguments are vectors, each
4010 vector element of ``op1`` is shifted by the corresponding shift amount
4013 If the ``nuw`` keyword is present, then the shift produces a :ref:`poison
4014 value <poisonvalues>` if it shifts out any non-zero bits. If the
4015 ``nsw`` keyword is present, then the shift produces a :ref:`poison
4016 value <poisonvalues>` if it shifts out any bits that disagree with the
4017 resultant sign bit. As such, NUW/NSW have the same semantics as they
4018 would if the shift were expressed as a mul instruction with the same
4019 nsw/nuw bits in (mul %op1, (shl 1, %op2)).
4024 .. code-block:: llvm
4026 <result> = shl i32 4, %var ; yields {i32}: 4 << %var
4027 <result> = shl i32 4, 2 ; yields {i32}: 16
4028 <result> = shl i32 1, 10 ; yields {i32}: 1024
4029 <result> = shl i32 1, 32 ; undefined
4030 <result> = shl <2 x i32> < i32 1, i32 1>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 2, i32 4>
4032 '``lshr``' Instruction
4033 ^^^^^^^^^^^^^^^^^^^^^^
4040 <result> = lshr <ty> <op1>, <op2> ; yields {ty}:result
4041 <result> = lshr exact <ty> <op1>, <op2> ; yields {ty}:result
4046 The '``lshr``' instruction (logical shift right) returns the first
4047 operand shifted to the right a specified number of bits with zero fill.
4052 Both arguments to the '``lshr``' instruction must be the same
4053 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4054 '``op2``' is treated as an unsigned value.
4059 This instruction always performs a logical shift right operation. The
4060 most significant bits of the result will be filled with zero bits after
4061 the shift. If ``op2`` is (statically or dynamically) equal to or larger
4062 than the number of bits in ``op1``, the result is undefined. If the
4063 arguments are vectors, each vector element of ``op1`` is shifted by the
4064 corresponding shift amount in ``op2``.
4066 If the ``exact`` keyword is present, the result value of the ``lshr`` is
4067 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
4073 .. code-block:: llvm
4075 <result> = lshr i32 4, 1 ; yields {i32}:result = 2
4076 <result> = lshr i32 4, 2 ; yields {i32}:result = 1
4077 <result> = lshr i8 4, 3 ; yields {i8}:result = 0
4078 <result> = lshr i8 -2, 1 ; yields {i8}:result = 0x7F
4079 <result> = lshr i32 1, 32 ; undefined
4080 <result> = lshr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 0x7FFFFFFF, i32 1>
4082 '``ashr``' Instruction
4083 ^^^^^^^^^^^^^^^^^^^^^^
4090 <result> = ashr <ty> <op1>, <op2> ; yields {ty}:result
4091 <result> = ashr exact <ty> <op1>, <op2> ; yields {ty}:result
4096 The '``ashr``' instruction (arithmetic shift right) returns the first
4097 operand shifted to the right a specified number of bits with sign
4103 Both arguments to the '``ashr``' instruction must be the same
4104 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4105 '``op2``' is treated as an unsigned value.
4110 This instruction always performs an arithmetic shift right operation,
4111 The most significant bits of the result will be filled with the sign bit
4112 of ``op1``. If ``op2`` is (statically or dynamically) equal to or larger
4113 than the number of bits in ``op1``, the result is undefined. If the
4114 arguments are vectors, each vector element of ``op1`` is shifted by the
4115 corresponding shift amount in ``op2``.
4117 If the ``exact`` keyword is present, the result value of the ``ashr`` is
4118 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
4124 .. code-block:: llvm
4126 <result> = ashr i32 4, 1 ; yields {i32}:result = 2
4127 <result> = ashr i32 4, 2 ; yields {i32}:result = 1
4128 <result> = ashr i8 4, 3 ; yields {i8}:result = 0
4129 <result> = ashr i8 -2, 1 ; yields {i8}:result = -1
4130 <result> = ashr i32 1, 32 ; undefined
4131 <result> = ashr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 3> ; yields: result=<2 x i32> < i32 -1, i32 0>
4133 '``and``' Instruction
4134 ^^^^^^^^^^^^^^^^^^^^^
4141 <result> = and <ty> <op1>, <op2> ; yields {ty}:result
4146 The '``and``' instruction returns the bitwise logical and of its two
4152 The two arguments to the '``and``' instruction must be
4153 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4154 arguments must have identical types.
4159 The truth table used for the '``and``' instruction is:
4176 .. code-block:: llvm
4178 <result> = and i32 4, %var ; yields {i32}:result = 4 & %var
4179 <result> = and i32 15, 40 ; yields {i32}:result = 8
4180 <result> = and i32 4, 8 ; yields {i32}:result = 0
4182 '``or``' Instruction
4183 ^^^^^^^^^^^^^^^^^^^^
4190 <result> = or <ty> <op1>, <op2> ; yields {ty}:result
4195 The '``or``' instruction returns the bitwise logical inclusive or of its
4201 The two arguments to the '``or``' instruction must be
4202 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4203 arguments must have identical types.
4208 The truth table used for the '``or``' instruction is:
4227 <result> = or i32 4, %var ; yields {i32}:result = 4 | %var
4228 <result> = or i32 15, 40 ; yields {i32}:result = 47
4229 <result> = or i32 4, 8 ; yields {i32}:result = 12
4231 '``xor``' Instruction
4232 ^^^^^^^^^^^^^^^^^^^^^
4239 <result> = xor <ty> <op1>, <op2> ; yields {ty}:result
4244 The '``xor``' instruction returns the bitwise logical exclusive or of
4245 its two operands. The ``xor`` is used to implement the "one's
4246 complement" operation, which is the "~" operator in C.
4251 The two arguments to the '``xor``' instruction must be
4252 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4253 arguments must have identical types.
4258 The truth table used for the '``xor``' instruction is:
4275 .. code-block:: llvm
4277 <result> = xor i32 4, %var ; yields {i32}:result = 4 ^ %var
4278 <result> = xor i32 15, 40 ; yields {i32}:result = 39
4279 <result> = xor i32 4, 8 ; yields {i32}:result = 12
4280 <result> = xor i32 %V, -1 ; yields {i32}:result = ~%V
4285 LLVM supports several instructions to represent vector operations in a
4286 target-independent manner. These instructions cover the element-access
4287 and vector-specific operations needed to process vectors effectively.
4288 While LLVM does directly support these vector operations, many
4289 sophisticated algorithms will want to use target-specific intrinsics to
4290 take full advantage of a specific target.
4292 .. _i_extractelement:
4294 '``extractelement``' Instruction
4295 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4302 <result> = extractelement <n x <ty>> <val>, i32 <idx> ; yields <ty>
4307 The '``extractelement``' instruction extracts a single scalar element
4308 from a vector at a specified index.
4313 The first operand of an '``extractelement``' instruction is a value of
4314 :ref:`vector <t_vector>` type. The second operand is an index indicating
4315 the position from which to extract the element. The index may be a
4321 The result is a scalar of the same type as the element type of ``val``.
4322 Its value is the value at position ``idx`` of ``val``. If ``idx``
4323 exceeds the length of ``val``, the results are undefined.
4328 .. code-block:: llvm
4330 <result> = extractelement <4 x i32> %vec, i32 0 ; yields i32
4332 .. _i_insertelement:
4334 '``insertelement``' Instruction
4335 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4342 <result> = insertelement <n x <ty>> <val>, <ty> <elt>, i32 <idx> ; yields <n x <ty>>
4347 The '``insertelement``' instruction inserts a scalar element into a
4348 vector at a specified index.
4353 The first operand of an '``insertelement``' instruction is a value of
4354 :ref:`vector <t_vector>` type. The second operand is a scalar value whose
4355 type must equal the element type of the first operand. The third operand
4356 is an index indicating the position at which to insert the value. The
4357 index may be a variable.
4362 The result is a vector of the same type as ``val``. Its element values
4363 are those of ``val`` except at position ``idx``, where it gets the value
4364 ``elt``. If ``idx`` exceeds the length of ``val``, the results are
4370 .. code-block:: llvm
4372 <result> = insertelement <4 x i32> %vec, i32 1, i32 0 ; yields <4 x i32>
4374 .. _i_shufflevector:
4376 '``shufflevector``' Instruction
4377 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4384 <result> = shufflevector <n x <ty>> <v1>, <n x <ty>> <v2>, <m x i32> <mask> ; yields <m x <ty>>
4389 The '``shufflevector``' instruction constructs a permutation of elements
4390 from two input vectors, returning a vector with the same element type as
4391 the input and length that is the same as the shuffle mask.
4396 The first two operands of a '``shufflevector``' instruction are vectors
4397 with the same type. The third argument is a shuffle mask whose element
4398 type is always 'i32'. The result of the instruction is a vector whose
4399 length is the same as the shuffle mask and whose element type is the
4400 same as the element type of the first two operands.
4402 The shuffle mask operand is required to be a constant vector with either
4403 constant integer or undef values.
4408 The elements of the two input vectors are numbered from left to right
4409 across both of the vectors. The shuffle mask operand specifies, for each
4410 element of the result vector, which element of the two input vectors the
4411 result element gets. The element selector may be undef (meaning "don't
4412 care") and the second operand may be undef if performing a shuffle from
4418 .. code-block:: llvm
4420 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4421 <4 x i32> <i32 0, i32 4, i32 1, i32 5> ; yields <4 x i32>
4422 <result> = shufflevector <4 x i32> %v1, <4 x i32> undef,
4423 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32> - Identity shuffle.
4424 <result> = shufflevector <8 x i32> %v1, <8 x i32> undef,
4425 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32>
4426 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4427 <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7 > ; yields <8 x i32>
4429 Aggregate Operations
4430 --------------------
4432 LLVM supports several instructions for working with
4433 :ref:`aggregate <t_aggregate>` values.
4437 '``extractvalue``' Instruction
4438 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4445 <result> = extractvalue <aggregate type> <val>, <idx>{, <idx>}*
4450 The '``extractvalue``' instruction extracts the value of a member field
4451 from an :ref:`aggregate <t_aggregate>` value.
4456 The first operand of an '``extractvalue``' instruction is a value of
4457 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The operands are
4458 constant indices to specify which value to extract in a similar manner
4459 as indices in a '``getelementptr``' instruction.
4461 The major differences to ``getelementptr`` indexing are:
4463 - Since the value being indexed is not a pointer, the first index is
4464 omitted and assumed to be zero.
4465 - At least one index must be specified.
4466 - Not only struct indices but also array indices must be in bounds.
4471 The result is the value at the position in the aggregate specified by
4477 .. code-block:: llvm
4479 <result> = extractvalue {i32, float} %agg, 0 ; yields i32
4483 '``insertvalue``' Instruction
4484 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4491 <result> = insertvalue <aggregate type> <val>, <ty> <elt>, <idx>{, <idx>}* ; yields <aggregate type>
4496 The '``insertvalue``' instruction inserts a value into a member field in
4497 an :ref:`aggregate <t_aggregate>` value.
4502 The first operand of an '``insertvalue``' instruction is a value of
4503 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The second operand is
4504 a first-class value to insert. The following operands are constant
4505 indices indicating the position at which to insert the value in a
4506 similar manner as indices in a '``extractvalue``' instruction. The value
4507 to insert must have the same type as the value identified by the
4513 The result is an aggregate of the same type as ``val``. Its value is
4514 that of ``val`` except that the value at the position specified by the
4515 indices is that of ``elt``.
4520 .. code-block:: llvm
4522 %agg1 = insertvalue {i32, float} undef, i32 1, 0 ; yields {i32 1, float undef}
4523 %agg2 = insertvalue {i32, float} %agg1, float %val, 1 ; yields {i32 1, float %val}
4524 %agg3 = insertvalue {i32, {float}} %agg1, float %val, 1, 0 ; yields {i32 1, float %val}
4528 Memory Access and Addressing Operations
4529 ---------------------------------------
4531 A key design point of an SSA-based representation is how it represents
4532 memory. In LLVM, no memory locations are in SSA form, which makes things
4533 very simple. This section describes how to read, write, and allocate
4538 '``alloca``' Instruction
4539 ^^^^^^^^^^^^^^^^^^^^^^^^
4546 <result> = alloca <type>[, <ty> <NumElements>][, align <alignment>] ; yields {type*}:result
4551 The '``alloca``' instruction allocates memory on the stack frame of the
4552 currently executing function, to be automatically released when this
4553 function returns to its caller. The object is always allocated in the
4554 generic address space (address space zero).
4559 The '``alloca``' instruction allocates ``sizeof(<type>)*NumElements``
4560 bytes of memory on the runtime stack, returning a pointer of the
4561 appropriate type to the program. If "NumElements" is specified, it is
4562 the number of elements allocated, otherwise "NumElements" is defaulted
4563 to be one. If a constant alignment is specified, the value result of the
4564 allocation is guaranteed to be aligned to at least that boundary. If not
4565 specified, or if zero, the target can choose to align the allocation on
4566 any convenient boundary compatible with the type.
4568 '``type``' may be any sized type.
4573 Memory is allocated; a pointer is returned. The operation is undefined
4574 if there is insufficient stack space for the allocation. '``alloca``'d
4575 memory is automatically released when the function returns. The
4576 '``alloca``' instruction is commonly used to represent automatic
4577 variables that must have an address available. When the function returns
4578 (either with the ``ret`` or ``resume`` instructions), the memory is
4579 reclaimed. Allocating zero bytes is legal, but the result is undefined.
4580 The order in which memory is allocated (ie., which way the stack grows)
4586 .. code-block:: llvm
4588 %ptr = alloca i32 ; yields {i32*}:ptr
4589 %ptr = alloca i32, i32 4 ; yields {i32*}:ptr
4590 %ptr = alloca i32, i32 4, align 1024 ; yields {i32*}:ptr
4591 %ptr = alloca i32, align 1024 ; yields {i32*}:ptr
4595 '``load``' Instruction
4596 ^^^^^^^^^^^^^^^^^^^^^^
4603 <result> = load [volatile] <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>][, !invariant.load !<index>]
4604 <result> = load atomic [volatile] <ty>* <pointer> [singlethread] <ordering>, align <alignment>
4605 !<index> = !{ i32 1 }
4610 The '``load``' instruction is used to read from memory.
4615 The argument to the ``load`` instruction specifies the memory address
4616 from which to load. The pointer must point to a :ref:`first
4617 class <t_firstclass>` type. If the ``load`` is marked as ``volatile``,
4618 then the optimizer is not allowed to modify the number or order of
4619 execution of this ``load`` with other :ref:`volatile
4620 operations <volatile>`.
4622 If the ``load`` is marked as ``atomic``, it takes an extra
4623 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
4624 ``release`` and ``acq_rel`` orderings are not valid on ``load``
4625 instructions. Atomic loads produce :ref:`defined <memmodel>` results
4626 when they may see multiple atomic stores. The type of the pointee must
4627 be an integer type whose bit width is a power of two greater than or
4628 equal to eight and less than or equal to a target-specific size limit.
4629 ``align`` must be explicitly specified on atomic loads, and the load has
4630 undefined behavior if the alignment is not set to a value which is at
4631 least the size in bytes of the pointee. ``!nontemporal`` does not have
4632 any defined semantics for atomic loads.
4634 The optional constant ``align`` argument specifies the alignment of the
4635 operation (that is, the alignment of the memory address). A value of 0
4636 or an omitted ``align`` argument means that the operation has the ABI
4637 alignment for the target. It is the responsibility of the code emitter
4638 to ensure that the alignment information is correct. Overestimating the
4639 alignment results in undefined behavior. Underestimating the alignment
4640 may produce less efficient code. An alignment of 1 is always safe.
4642 The optional ``!nontemporal`` metadata must reference a single
4643 metadata name ``<index>`` corresponding to a metadata node with one
4644 ``i32`` entry of value 1. The existence of the ``!nontemporal``
4645 metadata on the instruction tells the optimizer and code generator
4646 that this load is not expected to be reused in the cache. The code
4647 generator may select special instructions to save cache bandwidth, such
4648 as the ``MOVNT`` instruction on x86.
4650 The optional ``!invariant.load`` metadata must reference a single
4651 metadata name ``<index>`` corresponding to a metadata node with no
4652 entries. The existence of the ``!invariant.load`` metadata on the
4653 instruction tells the optimizer and code generator that this load
4654 address points to memory which does not change value during program
4655 execution. The optimizer may then move this load around, for example, by
4656 hoisting it out of loops using loop invariant code motion.
4661 The location of memory pointed to is loaded. If the value being loaded
4662 is of scalar type then the number of bytes read does not exceed the
4663 minimum number of bytes needed to hold all bits of the type. For
4664 example, loading an ``i24`` reads at most three bytes. When loading a
4665 value of a type like ``i20`` with a size that is not an integral number
4666 of bytes, the result is undefined if the value was not originally
4667 written using a store of the same type.
4672 .. code-block:: llvm
4674 %ptr = alloca i32 ; yields {i32*}:ptr
4675 store i32 3, i32* %ptr ; yields {void}
4676 %val = load i32* %ptr ; yields {i32}:val = i32 3
4680 '``store``' Instruction
4681 ^^^^^^^^^^^^^^^^^^^^^^^
4688 store [volatile] <ty> <value>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>] ; yields {void}
4689 store atomic [volatile] <ty> <value>, <ty>* <pointer> [singlethread] <ordering>, align <alignment> ; yields {void}
4694 The '``store``' instruction is used to write to memory.
4699 There are two arguments to the ``store`` instruction: a value to store
4700 and an address at which to store it. The type of the ``<pointer>``
4701 operand must be a pointer to the :ref:`first class <t_firstclass>` type of
4702 the ``<value>`` operand. If the ``store`` is marked as ``volatile``,
4703 then the optimizer is not allowed to modify the number or order of
4704 execution of this ``store`` with other :ref:`volatile
4705 operations <volatile>`.
4707 If the ``store`` is marked as ``atomic``, it takes an extra
4708 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
4709 ``acquire`` and ``acq_rel`` orderings aren't valid on ``store``
4710 instructions. Atomic loads produce :ref:`defined <memmodel>` results
4711 when they may see multiple atomic stores. The type of the pointee must
4712 be an integer type whose bit width is a power of two greater than or
4713 equal to eight and less than or equal to a target-specific size limit.
4714 ``align`` must be explicitly specified on atomic stores, and the store
4715 has undefined behavior if the alignment is not set to a value which is
4716 at least the size in bytes of the pointee. ``!nontemporal`` does not
4717 have any defined semantics for atomic stores.
4719 The optional constant ``align`` argument specifies the alignment of the
4720 operation (that is, the alignment of the memory address). A value of 0
4721 or an omitted ``align`` argument means that the operation has the ABI
4722 alignment for the target. It is the responsibility of the code emitter
4723 to ensure that the alignment information is correct. Overestimating the
4724 alignment results in undefined behavior. Underestimating the
4725 alignment may produce less efficient code. An alignment of 1 is always
4728 The optional ``!nontemporal`` metadata must reference a single metadata
4729 name ``<index>`` corresponding to a metadata node with one ``i32`` entry of
4730 value 1. The existence of the ``!nontemporal`` metadata on the instruction
4731 tells the optimizer and code generator that this load is not expected to
4732 be reused in the cache. The code generator may select special
4733 instructions to save cache bandwidth, such as the MOVNT instruction on
4739 The contents of memory are updated to contain ``<value>`` at the
4740 location specified by the ``<pointer>`` operand. If ``<value>`` is
4741 of scalar type then the number of bytes written does not exceed the
4742 minimum number of bytes needed to hold all bits of the type. For
4743 example, storing an ``i24`` writes at most three bytes. When writing a
4744 value of a type like ``i20`` with a size that is not an integral number
4745 of bytes, it is unspecified what happens to the extra bits that do not
4746 belong to the type, but they will typically be overwritten.
4751 .. code-block:: llvm
4753 %ptr = alloca i32 ; yields {i32*}:ptr
4754 store i32 3, i32* %ptr ; yields {void}
4755 %val = load i32* %ptr ; yields {i32}:val = i32 3
4759 '``fence``' Instruction
4760 ^^^^^^^^^^^^^^^^^^^^^^^
4767 fence [singlethread] <ordering> ; yields {void}
4772 The '``fence``' instruction is used to introduce happens-before edges
4778 '``fence``' instructions take an :ref:`ordering <ordering>` argument which
4779 defines what *synchronizes-with* edges they add. They can only be given
4780 ``acquire``, ``release``, ``acq_rel``, and ``seq_cst`` orderings.
4785 A fence A which has (at least) ``release`` ordering semantics
4786 *synchronizes with* a fence B with (at least) ``acquire`` ordering
4787 semantics if and only if there exist atomic operations X and Y, both
4788 operating on some atomic object M, such that A is sequenced before X, X
4789 modifies M (either directly or through some side effect of a sequence
4790 headed by X), Y is sequenced before B, and Y observes M. This provides a
4791 *happens-before* dependency between A and B. Rather than an explicit
4792 ``fence``, one (but not both) of the atomic operations X or Y might
4793 provide a ``release`` or ``acquire`` (resp.) ordering constraint and
4794 still *synchronize-with* the explicit ``fence`` and establish the
4795 *happens-before* edge.
4797 A ``fence`` which has ``seq_cst`` ordering, in addition to having both
4798 ``acquire`` and ``release`` semantics specified above, participates in
4799 the global program order of other ``seq_cst`` operations and/or fences.
4801 The optional ":ref:`singlethread <singlethread>`" argument specifies
4802 that the fence only synchronizes with other fences in the same thread.
4803 (This is useful for interacting with signal handlers.)
4808 .. code-block:: llvm
4810 fence acquire ; yields {void}
4811 fence singlethread seq_cst ; yields {void}
4815 '``cmpxchg``' Instruction
4816 ^^^^^^^^^^^^^^^^^^^^^^^^^
4823 cmpxchg [volatile] <ty>* <pointer>, <ty> <cmp>, <ty> <new> [singlethread] <ordering> ; yields {ty}
4828 The '``cmpxchg``' instruction is used to atomically modify memory. It
4829 loads a value in memory and compares it to a given value. If they are
4830 equal, it stores a new value into the memory.
4835 There are three arguments to the '``cmpxchg``' instruction: an address
4836 to operate on, a value to compare to the value currently be at that
4837 address, and a new value to place at that address if the compared values
4838 are equal. The type of '<cmp>' must be an integer type whose bit width
4839 is a power of two greater than or equal to eight and less than or equal
4840 to a target-specific size limit. '<cmp>' and '<new>' must have the same
4841 type, and the type of '<pointer>' must be a pointer to that type. If the
4842 ``cmpxchg`` is marked as ``volatile``, then the optimizer is not allowed
4843 to modify the number or order of execution of this ``cmpxchg`` with
4844 other :ref:`volatile operations <volatile>`.
4846 The :ref:`ordering <ordering>` argument specifies how this ``cmpxchg``
4847 synchronizes with other atomic operations.
4849 The optional "``singlethread``" argument declares that the ``cmpxchg``
4850 is only atomic with respect to code (usually signal handlers) running in
4851 the same thread as the ``cmpxchg``. Otherwise the cmpxchg is atomic with
4852 respect to all other code in the system.
4854 The pointer passed into cmpxchg must have alignment greater than or
4855 equal to the size in memory of the operand.
4860 The contents of memory at the location specified by the '``<pointer>``'
4861 operand is read and compared to '``<cmp>``'; if the read value is the
4862 equal, '``<new>``' is written. The original value at the location is
4865 A successful ``cmpxchg`` is a read-modify-write instruction for the purpose
4866 of identifying release sequences. A failed ``cmpxchg`` is equivalent to an
4867 atomic load with an ordering parameter determined by dropping any
4868 ``release`` part of the ``cmpxchg``'s ordering.
4873 .. code-block:: llvm
4876 %orig = atomic load i32* %ptr unordered ; yields {i32}
4880 %cmp = phi i32 [ %orig, %entry ], [%old, %loop]
4881 %squared = mul i32 %cmp, %cmp
4882 %old = cmpxchg i32* %ptr, i32 %cmp, i32 %squared ; yields {i32}
4883 %success = icmp eq i32 %cmp, %old
4884 br i1 %success, label %done, label %loop
4891 '``atomicrmw``' Instruction
4892 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
4899 atomicrmw [volatile] <operation> <ty>* <pointer>, <ty> <value> [singlethread] <ordering> ; yields {ty}
4904 The '``atomicrmw``' instruction is used to atomically modify memory.
4909 There are three arguments to the '``atomicrmw``' instruction: an
4910 operation to apply, an address whose value to modify, an argument to the
4911 operation. The operation must be one of the following keywords:
4925 The type of '<value>' must be an integer type whose bit width is a power
4926 of two greater than or equal to eight and less than or equal to a
4927 target-specific size limit. The type of the '``<pointer>``' operand must
4928 be a pointer to that type. If the ``atomicrmw`` is marked as
4929 ``volatile``, then the optimizer is not allowed to modify the number or
4930 order of execution of this ``atomicrmw`` with other :ref:`volatile
4931 operations <volatile>`.
4936 The contents of memory at the location specified by the '``<pointer>``'
4937 operand are atomically read, modified, and written back. The original
4938 value at the location is returned. The modification is specified by the
4941 - xchg: ``*ptr = val``
4942 - add: ``*ptr = *ptr + val``
4943 - sub: ``*ptr = *ptr - val``
4944 - and: ``*ptr = *ptr & val``
4945 - nand: ``*ptr = ~(*ptr & val)``
4946 - or: ``*ptr = *ptr | val``
4947 - xor: ``*ptr = *ptr ^ val``
4948 - max: ``*ptr = *ptr > val ? *ptr : val`` (using a signed comparison)
4949 - min: ``*ptr = *ptr < val ? *ptr : val`` (using a signed comparison)
4950 - umax: ``*ptr = *ptr > val ? *ptr : val`` (using an unsigned
4952 - umin: ``*ptr = *ptr < val ? *ptr : val`` (using an unsigned
4958 .. code-block:: llvm
4960 %old = atomicrmw add i32* %ptr, i32 1 acquire ; yields {i32}
4962 .. _i_getelementptr:
4964 '``getelementptr``' Instruction
4965 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4972 <result> = getelementptr <pty>* <ptrval>{, <ty> <idx>}*
4973 <result> = getelementptr inbounds <pty>* <ptrval>{, <ty> <idx>}*
4974 <result> = getelementptr <ptr vector> ptrval, <vector index type> idx
4979 The '``getelementptr``' instruction is used to get the address of a
4980 subelement of an :ref:`aggregate <t_aggregate>` data structure. It performs
4981 address calculation only and does not access memory.
4986 The first argument is always a pointer or a vector of pointers, and
4987 forms the basis of the calculation. The remaining arguments are indices
4988 that indicate which of the elements of the aggregate object are indexed.
4989 The interpretation of each index is dependent on the type being indexed
4990 into. The first index always indexes the pointer value given as the
4991 first argument, the second index indexes a value of the type pointed to
4992 (not necessarily the value directly pointed to, since the first index
4993 can be non-zero), etc. The first type indexed into must be a pointer
4994 value, subsequent types can be arrays, vectors, and structs. Note that
4995 subsequent types being indexed into can never be pointers, since that
4996 would require loading the pointer before continuing calculation.
4998 The type of each index argument depends on the type it is indexing into.
4999 When indexing into a (optionally packed) structure, only ``i32`` integer
5000 **constants** are allowed (when using a vector of indices they must all
5001 be the **same** ``i32`` integer constant). When indexing into an array,
5002 pointer or vector, integers of any width are allowed, and they are not
5003 required to be constant. These integers are treated as signed values
5006 For example, let's consider a C code fragment and how it gets compiled
5022 int *foo(struct ST *s) {
5023 return &s[1].Z.B[5][13];
5026 The LLVM code generated by Clang is:
5028 .. code-block:: llvm
5030 %struct.RT = type { i8, [10 x [20 x i32]], i8 }
5031 %struct.ST = type { i32, double, %struct.RT }
5033 define i32* @foo(%struct.ST* %s) nounwind uwtable readnone optsize ssp {
5035 %arrayidx = getelementptr inbounds %struct.ST* %s, i64 1, i32 2, i32 1, i64 5, i64 13
5042 In the example above, the first index is indexing into the
5043 '``%struct.ST*``' type, which is a pointer, yielding a '``%struct.ST``'
5044 = '``{ i32, double, %struct.RT }``' type, a structure. The second index
5045 indexes into the third element of the structure, yielding a
5046 '``%struct.RT``' = '``{ i8 , [10 x [20 x i32]], i8 }``' type, another
5047 structure. The third index indexes into the second element of the
5048 structure, yielding a '``[10 x [20 x i32]]``' type, an array. The two
5049 dimensions of the array are subscripted into, yielding an '``i32``'
5050 type. The '``getelementptr``' instruction returns a pointer to this
5051 element, thus computing a value of '``i32*``' type.
5053 Note that it is perfectly legal to index partially through a structure,
5054 returning a pointer to an inner element. Because of this, the LLVM code
5055 for the given testcase is equivalent to:
5057 .. code-block:: llvm
5059 define i32* @foo(%struct.ST* %s) {
5060 %t1 = getelementptr %struct.ST* %s, i32 1 ; yields %struct.ST*:%t1
5061 %t2 = getelementptr %struct.ST* %t1, i32 0, i32 2 ; yields %struct.RT*:%t2
5062 %t3 = getelementptr %struct.RT* %t2, i32 0, i32 1 ; yields [10 x [20 x i32]]*:%t3
5063 %t4 = getelementptr [10 x [20 x i32]]* %t3, i32 0, i32 5 ; yields [20 x i32]*:%t4
5064 %t5 = getelementptr [20 x i32]* %t4, i32 0, i32 13 ; yields i32*:%t5
5068 If the ``inbounds`` keyword is present, the result value of the
5069 ``getelementptr`` is a :ref:`poison value <poisonvalues>` if the base
5070 pointer is not an *in bounds* address of an allocated object, or if any
5071 of the addresses that would be formed by successive addition of the
5072 offsets implied by the indices to the base address with infinitely
5073 precise signed arithmetic are not an *in bounds* address of that
5074 allocated object. The *in bounds* addresses for an allocated object are
5075 all the addresses that point into the object, plus the address one byte
5076 past the end. In cases where the base is a vector of pointers the
5077 ``inbounds`` keyword applies to each of the computations element-wise.
5079 If the ``inbounds`` keyword is not present, the offsets are added to the
5080 base address with silently-wrapping two's complement arithmetic. If the
5081 offsets have a different width from the pointer, they are sign-extended
5082 or truncated to the width of the pointer. The result value of the
5083 ``getelementptr`` may be outside the object pointed to by the base
5084 pointer. The result value may not necessarily be used to access memory
5085 though, even if it happens to point into allocated storage. See the
5086 :ref:`Pointer Aliasing Rules <pointeraliasing>` section for more
5089 The getelementptr instruction is often confusing. For some more insight
5090 into how it works, see :doc:`the getelementptr FAQ <GetElementPtr>`.
5095 .. code-block:: llvm
5097 ; yields [12 x i8]*:aptr
5098 %aptr = getelementptr {i32, [12 x i8]}* %saptr, i64 0, i32 1
5100 %vptr = getelementptr {i32, <2 x i8>}* %svptr, i64 0, i32 1, i32 1
5102 %eptr = getelementptr [12 x i8]* %aptr, i64 0, i32 1
5104 %iptr = getelementptr [10 x i32]* @arr, i16 0, i16 0
5106 In cases where the pointer argument is a vector of pointers, each index
5107 must be a vector with the same number of elements. For example:
5109 .. code-block:: llvm
5111 %A = getelementptr <4 x i8*> %ptrs, <4 x i64> %offsets,
5113 Conversion Operations
5114 ---------------------
5116 The instructions in this category are the conversion instructions
5117 (casting) which all take a single operand and a type. They perform
5118 various bit conversions on the operand.
5120 '``trunc .. to``' Instruction
5121 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5128 <result> = trunc <ty> <value> to <ty2> ; yields ty2
5133 The '``trunc``' instruction truncates its operand to the type ``ty2``.
5138 The '``trunc``' instruction takes a value to trunc, and a type to trunc
5139 it to. Both types must be of :ref:`integer <t_integer>` types, or vectors
5140 of the same number of integers. The bit size of the ``value`` must be
5141 larger than the bit size of the destination type, ``ty2``. Equal sized
5142 types are not allowed.
5147 The '``trunc``' instruction truncates the high order bits in ``value``
5148 and converts the remaining bits to ``ty2``. Since the source size must
5149 be larger than the destination size, ``trunc`` cannot be a *no-op cast*.
5150 It will always truncate bits.
5155 .. code-block:: llvm
5157 %X = trunc i32 257 to i8 ; yields i8:1
5158 %Y = trunc i32 123 to i1 ; yields i1:true
5159 %Z = trunc i32 122 to i1 ; yields i1:false
5160 %W = trunc <2 x i16> <i16 8, i16 7> to <2 x i8> ; yields <i8 8, i8 7>
5162 '``zext .. to``' Instruction
5163 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5170 <result> = zext <ty> <value> to <ty2> ; yields ty2
5175 The '``zext``' instruction zero extends its operand to type ``ty2``.
5180 The '``zext``' instruction takes a value to cast, and a type to cast it
5181 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
5182 the same number of integers. The bit size of the ``value`` must be
5183 smaller than the bit size of the destination type, ``ty2``.
5188 The ``zext`` fills the high order bits of the ``value`` with zero bits
5189 until it reaches the size of the destination type, ``ty2``.
5191 When zero extending from i1, the result will always be either 0 or 1.
5196 .. code-block:: llvm
5198 %X = zext i32 257 to i64 ; yields i64:257
5199 %Y = zext i1 true to i32 ; yields i32:1
5200 %Z = zext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5202 '``sext .. to``' Instruction
5203 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5210 <result> = sext <ty> <value> to <ty2> ; yields ty2
5215 The '``sext``' sign extends ``value`` to the type ``ty2``.
5220 The '``sext``' instruction takes a value to cast, and a type to cast it
5221 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
5222 the same number of integers. The bit size of the ``value`` must be
5223 smaller than the bit size of the destination type, ``ty2``.
5228 The '``sext``' instruction performs a sign extension by copying the sign
5229 bit (highest order bit) of the ``value`` until it reaches the bit size
5230 of the type ``ty2``.
5232 When sign extending from i1, the extension always results in -1 or 0.
5237 .. code-block:: llvm
5239 %X = sext i8 -1 to i16 ; yields i16 :65535
5240 %Y = sext i1 true to i32 ; yields i32:-1
5241 %Z = sext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5243 '``fptrunc .. to``' Instruction
5244 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5251 <result> = fptrunc <ty> <value> to <ty2> ; yields ty2
5256 The '``fptrunc``' instruction truncates ``value`` to type ``ty2``.
5261 The '``fptrunc``' instruction takes a :ref:`floating point <t_floating>`
5262 value to cast and a :ref:`floating point <t_floating>` type to cast it to.
5263 The size of ``value`` must be larger than the size of ``ty2``. This
5264 implies that ``fptrunc`` cannot be used to make a *no-op cast*.
5269 The '``fptrunc``' instruction truncates a ``value`` from a larger
5270 :ref:`floating point <t_floating>` type to a smaller :ref:`floating
5271 point <t_floating>` type. If the value cannot fit within the
5272 destination type, ``ty2``, then the results are undefined.
5277 .. code-block:: llvm
5279 %X = fptrunc double 123.0 to float ; yields float:123.0
5280 %Y = fptrunc double 1.0E+300 to float ; yields undefined
5282 '``fpext .. to``' Instruction
5283 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5290 <result> = fpext <ty> <value> to <ty2> ; yields ty2
5295 The '``fpext``' extends a floating point ``value`` to a larger floating
5301 The '``fpext``' instruction takes a :ref:`floating point <t_floating>`
5302 ``value`` to cast, and a :ref:`floating point <t_floating>` type to cast it
5303 to. The source type must be smaller than the destination type.
5308 The '``fpext``' instruction extends the ``value`` from a smaller
5309 :ref:`floating point <t_floating>` type to a larger :ref:`floating
5310 point <t_floating>` type. The ``fpext`` cannot be used to make a
5311 *no-op cast* because it always changes bits. Use ``bitcast`` to make a
5312 *no-op cast* for a floating point cast.
5317 .. code-block:: llvm
5319 %X = fpext float 3.125 to double ; yields double:3.125000e+00
5320 %Y = fpext double %X to fp128 ; yields fp128:0xL00000000000000004000900000000000
5322 '``fptoui .. to``' Instruction
5323 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5330 <result> = fptoui <ty> <value> to <ty2> ; yields ty2
5335 The '``fptoui``' converts a floating point ``value`` to its unsigned
5336 integer equivalent of type ``ty2``.
5341 The '``fptoui``' instruction takes a value to cast, which must be a
5342 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5343 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5344 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5345 type with the same number of elements as ``ty``
5350 The '``fptoui``' instruction converts its :ref:`floating
5351 point <t_floating>` operand into the nearest (rounding towards zero)
5352 unsigned integer value. If the value cannot fit in ``ty2``, the results
5358 .. code-block:: llvm
5360 %X = fptoui double 123.0 to i32 ; yields i32:123
5361 %Y = fptoui float 1.0E+300 to i1 ; yields undefined:1
5362 %Z = fptoui float 1.04E+17 to i8 ; yields undefined:1
5364 '``fptosi .. to``' Instruction
5365 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5372 <result> = fptosi <ty> <value> to <ty2> ; yields ty2
5377 The '``fptosi``' instruction converts :ref:`floating point <t_floating>`
5378 ``value`` to type ``ty2``.
5383 The '``fptosi``' instruction takes a value to cast, which must be a
5384 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5385 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5386 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5387 type with the same number of elements as ``ty``
5392 The '``fptosi``' instruction converts its :ref:`floating
5393 point <t_floating>` operand into the nearest (rounding towards zero)
5394 signed integer value. If the value cannot fit in ``ty2``, the results
5400 .. code-block:: llvm
5402 %X = fptosi double -123.0 to i32 ; yields i32:-123
5403 %Y = fptosi float 1.0E-247 to i1 ; yields undefined:1
5404 %Z = fptosi float 1.04E+17 to i8 ; yields undefined:1
5406 '``uitofp .. to``' Instruction
5407 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5414 <result> = uitofp <ty> <value> to <ty2> ; yields ty2
5419 The '``uitofp``' instruction regards ``value`` as an unsigned integer
5420 and converts that value to the ``ty2`` type.
5425 The '``uitofp``' instruction takes a value to cast, which must be a
5426 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
5427 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
5428 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
5429 type with the same number of elements as ``ty``
5434 The '``uitofp``' instruction interprets its operand as an unsigned
5435 integer quantity and converts it to the corresponding floating point
5436 value. If the value cannot fit in the floating point value, the results
5442 .. code-block:: llvm
5444 %X = uitofp i32 257 to float ; yields float:257.0
5445 %Y = uitofp i8 -1 to double ; yields double:255.0
5447 '``sitofp .. to``' Instruction
5448 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5455 <result> = sitofp <ty> <value> to <ty2> ; yields ty2
5460 The '``sitofp``' instruction regards ``value`` as a signed integer and
5461 converts that value to the ``ty2`` type.
5466 The '``sitofp``' instruction takes a value to cast, which must be a
5467 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
5468 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
5469 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
5470 type with the same number of elements as ``ty``
5475 The '``sitofp``' instruction interprets its operand as a signed integer
5476 quantity and converts it to the corresponding floating point value. If
5477 the value cannot fit in the floating point value, the results are
5483 .. code-block:: llvm
5485 %X = sitofp i32 257 to float ; yields float:257.0
5486 %Y = sitofp i8 -1 to double ; yields double:-1.0
5490 '``ptrtoint .. to``' Instruction
5491 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5498 <result> = ptrtoint <ty> <value> to <ty2> ; yields ty2
5503 The '``ptrtoint``' instruction converts the pointer or a vector of
5504 pointers ``value`` to the integer (or vector of integers) type ``ty2``.
5509 The '``ptrtoint``' instruction takes a ``value`` to cast, which must be
5510 a a value of type :ref:`pointer <t_pointer>` or a vector of pointers, and a
5511 type to cast it to ``ty2``, which must be an :ref:`integer <t_integer>` or
5512 a vector of integers type.
5517 The '``ptrtoint``' instruction converts ``value`` to integer type
5518 ``ty2`` by interpreting the pointer value as an integer and either
5519 truncating or zero extending that value to the size of the integer type.
5520 If ``value`` is smaller than ``ty2`` then a zero extension is done. If
5521 ``value`` is larger than ``ty2`` then a truncation is done. If they are
5522 the same size, then nothing is done (*no-op cast*) other than a type
5528 .. code-block:: llvm
5530 %X = ptrtoint i32* %P to i8 ; yields truncation on 32-bit architecture
5531 %Y = ptrtoint i32* %P to i64 ; yields zero extension on 32-bit architecture
5532 %Z = ptrtoint <4 x i32*> %P to <4 x i64>; yields vector zero extension for a vector of addresses on 32-bit architecture
5536 '``inttoptr .. to``' Instruction
5537 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5544 <result> = inttoptr <ty> <value> to <ty2> ; yields ty2
5549 The '``inttoptr``' instruction converts an integer ``value`` to a
5550 pointer type, ``ty2``.
5555 The '``inttoptr``' instruction takes an :ref:`integer <t_integer>` value to
5556 cast, and a type to cast it to, which must be a :ref:`pointer <t_pointer>`
5562 The '``inttoptr``' instruction converts ``value`` to type ``ty2`` by
5563 applying either a zero extension or a truncation depending on the size
5564 of the integer ``value``. If ``value`` is larger than the size of a
5565 pointer then a truncation is done. If ``value`` is smaller than the size
5566 of a pointer then a zero extension is done. If they are the same size,
5567 nothing is done (*no-op cast*).
5572 .. code-block:: llvm
5574 %X = inttoptr i32 255 to i32* ; yields zero extension on 64-bit architecture
5575 %Y = inttoptr i32 255 to i32* ; yields no-op on 32-bit architecture
5576 %Z = inttoptr i64 0 to i32* ; yields truncation on 32-bit architecture
5577 %Z = inttoptr <4 x i32> %G to <4 x i8*>; yields truncation of vector G to four pointers
5581 '``bitcast .. to``' Instruction
5582 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5589 <result> = bitcast <ty> <value> to <ty2> ; yields ty2
5594 The '``bitcast``' instruction converts ``value`` to type ``ty2`` without
5600 The '``bitcast``' instruction takes a value to cast, which must be a
5601 non-aggregate first class value, and a type to cast it to, which must
5602 also be a non-aggregate :ref:`first class <t_firstclass>` type. The bit
5603 sizes of ``value`` and the destination type, ``ty2``, must be identical.
5604 If the source type is a pointer, the destination type must also be a
5605 pointer. This instruction supports bitwise conversion of vectors to
5606 integers and to vectors of other types (as long as they have the same
5612 The '``bitcast``' instruction converts ``value`` to type ``ty2``. It is
5613 always a *no-op cast* because no bits change with this conversion. The
5614 conversion is done as if the ``value`` had been stored to memory and
5615 read back as type ``ty2``. Pointer (or vector of pointers) types may
5616 only be converted to other pointer (or vector of pointers) types with
5617 this instruction. To convert pointers to other types, use the
5618 :ref:`inttoptr <i_inttoptr>` or :ref:`ptrtoint <i_ptrtoint>` instructions
5624 .. code-block:: llvm
5626 %X = bitcast i8 255 to i8 ; yields i8 :-1
5627 %Y = bitcast i32* %x to sint* ; yields sint*:%x
5628 %Z = bitcast <2 x int> %V to i64; ; yields i64: %V
5629 %Z = bitcast <2 x i32*> %V to <2 x i64*> ; yields <2 x i64*>
5636 The instructions in this category are the "miscellaneous" instructions,
5637 which defy better classification.
5641 '``icmp``' Instruction
5642 ^^^^^^^^^^^^^^^^^^^^^^
5649 <result> = icmp <cond> <ty> <op1>, <op2> ; yields {i1} or {<N x i1>}:result
5654 The '``icmp``' instruction returns a boolean value or a vector of
5655 boolean values based on comparison of its two integer, integer vector,
5656 pointer, or pointer vector operands.
5661 The '``icmp``' instruction takes three operands. The first operand is
5662 the condition code indicating the kind of comparison to perform. It is
5663 not a value, just a keyword. The possible condition code are:
5666 #. ``ne``: not equal
5667 #. ``ugt``: unsigned greater than
5668 #. ``uge``: unsigned greater or equal
5669 #. ``ult``: unsigned less than
5670 #. ``ule``: unsigned less or equal
5671 #. ``sgt``: signed greater than
5672 #. ``sge``: signed greater or equal
5673 #. ``slt``: signed less than
5674 #. ``sle``: signed less or equal
5676 The remaining two arguments must be :ref:`integer <t_integer>` or
5677 :ref:`pointer <t_pointer>` or integer :ref:`vector <t_vector>` typed. They
5678 must also be identical types.
5683 The '``icmp``' compares ``op1`` and ``op2`` according to the condition
5684 code given as ``cond``. The comparison performed always yields either an
5685 :ref:`i1 <t_integer>` or vector of ``i1`` result, as follows:
5687 #. ``eq``: yields ``true`` if the operands are equal, ``false``
5688 otherwise. No sign interpretation is necessary or performed.
5689 #. ``ne``: yields ``true`` if the operands are unequal, ``false``
5690 otherwise. No sign interpretation is necessary or performed.
5691 #. ``ugt``: interprets the operands as unsigned values and yields
5692 ``true`` if ``op1`` is greater than ``op2``.
5693 #. ``uge``: interprets the operands as unsigned values and yields
5694 ``true`` if ``op1`` is greater than or equal to ``op2``.
5695 #. ``ult``: interprets the operands as unsigned values and yields
5696 ``true`` if ``op1`` is less than ``op2``.
5697 #. ``ule``: interprets the operands as unsigned values and yields
5698 ``true`` if ``op1`` is less than or equal to ``op2``.
5699 #. ``sgt``: interprets the operands as signed values and yields ``true``
5700 if ``op1`` is greater than ``op2``.
5701 #. ``sge``: interprets the operands as signed values and yields ``true``
5702 if ``op1`` is greater than or equal to ``op2``.
5703 #. ``slt``: interprets the operands as signed values and yields ``true``
5704 if ``op1`` is less than ``op2``.
5705 #. ``sle``: interprets the operands as signed values and yields ``true``
5706 if ``op1`` is less than or equal to ``op2``.
5708 If the operands are :ref:`pointer <t_pointer>` typed, the pointer values
5709 are compared as if they were integers.
5711 If the operands are integer vectors, then they are compared element by
5712 element. The result is an ``i1`` vector with the same number of elements
5713 as the values being compared. Otherwise, the result is an ``i1``.
5718 .. code-block:: llvm
5720 <result> = icmp eq i32 4, 5 ; yields: result=false
5721 <result> = icmp ne float* %X, %X ; yields: result=false
5722 <result> = icmp ult i16 4, 5 ; yields: result=true
5723 <result> = icmp sgt i16 4, 5 ; yields: result=false
5724 <result> = icmp ule i16 -4, 5 ; yields: result=false
5725 <result> = icmp sge i16 4, 5 ; yields: result=false
5727 Note that the code generator does not yet support vector types with the
5728 ``icmp`` instruction.
5732 '``fcmp``' Instruction
5733 ^^^^^^^^^^^^^^^^^^^^^^
5740 <result> = fcmp <cond> <ty> <op1>, <op2> ; yields {i1} or {<N x i1>}:result
5745 The '``fcmp``' instruction returns a boolean value or vector of boolean
5746 values based on comparison of its operands.
5748 If the operands are floating point scalars, then the result type is a
5749 boolean (:ref:`i1 <t_integer>`).
5751 If the operands are floating point vectors, then the result type is a
5752 vector of boolean with the same number of elements as the operands being
5758 The '``fcmp``' instruction takes three operands. The first operand is
5759 the condition code indicating the kind of comparison to perform. It is
5760 not a value, just a keyword. The possible condition code are:
5762 #. ``false``: no comparison, always returns false
5763 #. ``oeq``: ordered and equal
5764 #. ``ogt``: ordered and greater than
5765 #. ``oge``: ordered and greater than or equal
5766 #. ``olt``: ordered and less than
5767 #. ``ole``: ordered and less than or equal
5768 #. ``one``: ordered and not equal
5769 #. ``ord``: ordered (no nans)
5770 #. ``ueq``: unordered or equal
5771 #. ``ugt``: unordered or greater than
5772 #. ``uge``: unordered or greater than or equal
5773 #. ``ult``: unordered or less than
5774 #. ``ule``: unordered or less than or equal
5775 #. ``une``: unordered or not equal
5776 #. ``uno``: unordered (either nans)
5777 #. ``true``: no comparison, always returns true
5779 *Ordered* means that neither operand is a QNAN while *unordered* means
5780 that either operand may be a QNAN.
5782 Each of ``val1`` and ``val2`` arguments must be either a :ref:`floating
5783 point <t_floating>` type or a :ref:`vector <t_vector>` of floating point
5784 type. They must have identical types.
5789 The '``fcmp``' instruction compares ``op1`` and ``op2`` according to the
5790 condition code given as ``cond``. If the operands are vectors, then the
5791 vectors are compared element by element. Each comparison performed
5792 always yields an :ref:`i1 <t_integer>` result, as follows:
5794 #. ``false``: always yields ``false``, regardless of operands.
5795 #. ``oeq``: yields ``true`` if both operands are not a QNAN and ``op1``
5796 is equal to ``op2``.
5797 #. ``ogt``: yields ``true`` if both operands are not a QNAN and ``op1``
5798 is greater than ``op2``.
5799 #. ``oge``: yields ``true`` if both operands are not a QNAN and ``op1``
5800 is greater than or equal to ``op2``.
5801 #. ``olt``: yields ``true`` if both operands are not a QNAN and ``op1``
5802 is less than ``op2``.
5803 #. ``ole``: yields ``true`` if both operands are not a QNAN and ``op1``
5804 is less than or equal to ``op2``.
5805 #. ``one``: yields ``true`` if both operands are not a QNAN and ``op1``
5806 is not equal to ``op2``.
5807 #. ``ord``: yields ``true`` if both operands are not a QNAN.
5808 #. ``ueq``: yields ``true`` if either operand is a QNAN or ``op1`` is
5810 #. ``ugt``: yields ``true`` if either operand is a QNAN or ``op1`` is
5811 greater than ``op2``.
5812 #. ``uge``: yields ``true`` if either operand is a QNAN or ``op1`` is
5813 greater than or equal to ``op2``.
5814 #. ``ult``: yields ``true`` if either operand is a QNAN or ``op1`` is
5816 #. ``ule``: yields ``true`` if either operand is a QNAN or ``op1`` is
5817 less than or equal to ``op2``.
5818 #. ``une``: yields ``true`` if either operand is a QNAN or ``op1`` is
5819 not equal to ``op2``.
5820 #. ``uno``: yields ``true`` if either operand is a QNAN.
5821 #. ``true``: always yields ``true``, regardless of operands.
5826 .. code-block:: llvm
5828 <result> = fcmp oeq float 4.0, 5.0 ; yields: result=false
5829 <result> = fcmp one float 4.0, 5.0 ; yields: result=true
5830 <result> = fcmp olt float 4.0, 5.0 ; yields: result=true
5831 <result> = fcmp ueq double 1.0, 2.0 ; yields: result=false
5833 Note that the code generator does not yet support vector types with the
5834 ``fcmp`` instruction.
5838 '``phi``' Instruction
5839 ^^^^^^^^^^^^^^^^^^^^^
5846 <result> = phi <ty> [ <val0>, <label0>], ...
5851 The '``phi``' instruction is used to implement the φ node in the SSA
5852 graph representing the function.
5857 The type of the incoming values is specified with the first type field.
5858 After this, the '``phi``' instruction takes a list of pairs as
5859 arguments, with one pair for each predecessor basic block of the current
5860 block. Only values of :ref:`first class <t_firstclass>` type may be used as
5861 the value arguments to the PHI node. Only labels may be used as the
5864 There must be no non-phi instructions between the start of a basic block
5865 and the PHI instructions: i.e. PHI instructions must be first in a basic
5868 For the purposes of the SSA form, the use of each incoming value is
5869 deemed to occur on the edge from the corresponding predecessor block to
5870 the current block (but after any definition of an '``invoke``'
5871 instruction's return value on the same edge).
5876 At runtime, the '``phi``' instruction logically takes on the value
5877 specified by the pair corresponding to the predecessor basic block that
5878 executed just prior to the current block.
5883 .. code-block:: llvm
5885 Loop: ; Infinite loop that counts from 0 on up...
5886 %indvar = phi i32 [ 0, %LoopHeader ], [ %nextindvar, %Loop ]
5887 %nextindvar = add i32 %indvar, 1
5892 '``select``' Instruction
5893 ^^^^^^^^^^^^^^^^^^^^^^^^
5900 <result> = select selty <cond>, <ty> <val1>, <ty> <val2> ; yields ty
5902 selty is either i1 or {<N x i1>}
5907 The '``select``' instruction is used to choose one value based on a
5908 condition, without branching.
5913 The '``select``' instruction requires an 'i1' value or a vector of 'i1'
5914 values indicating the condition, and two values of the same :ref:`first
5915 class <t_firstclass>` type. If the val1/val2 are vectors and the
5916 condition is a scalar, then entire vectors are selected, not individual
5922 If the condition is an i1 and it evaluates to 1, the instruction returns
5923 the first value argument; otherwise, it returns the second value
5926 If the condition is a vector of i1, then the value arguments must be
5927 vectors of the same size, and the selection is done element by element.
5932 .. code-block:: llvm
5934 %X = select i1 true, i8 17, i8 42 ; yields i8:17
5938 '``call``' Instruction
5939 ^^^^^^^^^^^^^^^^^^^^^^
5946 <result> = [tail] call [cconv] [ret attrs] <ty> [<fnty>*] <fnptrval>(<function args>) [fn attrs]
5951 The '``call``' instruction represents a simple function call.
5956 This instruction requires several arguments:
5958 #. The optional "tail" marker indicates that the callee function does
5959 not access any allocas or varargs in the caller. Note that calls may
5960 be marked "tail" even if they do not occur before a
5961 :ref:`ret <i_ret>` instruction. If the "tail" marker is present, the
5962 function call is eligible for tail call optimization, but `might not
5963 in fact be optimized into a jump <CodeGenerator.html#tailcallopt>`_.
5964 The code generator may optimize calls marked "tail" with either 1)
5965 automatic `sibling call
5966 optimization <CodeGenerator.html#sibcallopt>`_ when the caller and
5967 callee have matching signatures, or 2) forced tail call optimization
5968 when the following extra requirements are met:
5970 - Caller and callee both have the calling convention ``fastcc``.
5971 - The call is in tail position (ret immediately follows call and ret
5972 uses value of call or is void).
5973 - Option ``-tailcallopt`` is enabled, or
5974 ``llvm::GuaranteedTailCallOpt`` is ``true``.
5975 - `Platform specific constraints are
5976 met. <CodeGenerator.html#tailcallopt>`_
5978 #. The optional "cconv" marker indicates which :ref:`calling
5979 convention <callingconv>` the call should use. If none is
5980 specified, the call defaults to using C calling conventions. The
5981 calling convention of the call must match the calling convention of
5982 the target function, or else the behavior is undefined.
5983 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
5984 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
5986 #. '``ty``': the type of the call instruction itself which is also the
5987 type of the return value. Functions that return no value are marked
5989 #. '``fnty``': shall be the signature of the pointer to function value
5990 being invoked. The argument types must match the types implied by
5991 this signature. This type can be omitted if the function is not
5992 varargs and if the function type does not return a pointer to a
5994 #. '``fnptrval``': An LLVM value containing a pointer to a function to
5995 be invoked. In most cases, this is a direct function invocation, but
5996 indirect ``call``'s are just as possible, calling an arbitrary pointer
5998 #. '``function args``': argument list whose types match the function
5999 signature argument types and parameter attributes. All arguments must
6000 be of :ref:`first class <t_firstclass>` type. If the function signature
6001 indicates the function accepts a variable number of arguments, the
6002 extra arguments can be specified.
6003 #. The optional :ref:`function attributes <fnattrs>` list. Only
6004 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
6005 attributes are valid here.
6010 The '``call``' instruction is used to cause control flow to transfer to
6011 a specified function, with its incoming arguments bound to the specified
6012 values. Upon a '``ret``' instruction in the called function, control
6013 flow continues with the instruction after the function call, and the
6014 return value of the function is bound to the result argument.
6019 .. code-block:: llvm
6021 %retval = call i32 @test(i32 %argc)
6022 call i32 (i8*, ...)* @printf(i8* %msg, i32 12, i8 42) ; yields i32
6023 %X = tail call i32 @foo() ; yields i32
6024 %Y = tail call fastcc i32 @foo() ; yields i32
6025 call void %foo(i8 97 signext)
6027 %struct.A = type { i32, i8 }
6028 %r = call %struct.A @foo() ; yields { 32, i8 }
6029 %gr = extractvalue %struct.A %r, 0 ; yields i32
6030 %gr1 = extractvalue %struct.A %r, 1 ; yields i8
6031 %Z = call void @foo() noreturn ; indicates that %foo never returns normally
6032 %ZZ = call zeroext i32 @bar() ; Return value is %zero extended
6034 llvm treats calls to some functions with names and arguments that match
6035 the standard C99 library as being the C99 library functions, and may
6036 perform optimizations or generate code for them under that assumption.
6037 This is something we'd like to change in the future to provide better
6038 support for freestanding environments and non-C-based languages.
6042 '``va_arg``' Instruction
6043 ^^^^^^^^^^^^^^^^^^^^^^^^
6050 <resultval> = va_arg <va_list*> <arglist>, <argty>
6055 The '``va_arg``' instruction is used to access arguments passed through
6056 the "variable argument" area of a function call. It is used to implement
6057 the ``va_arg`` macro in C.
6062 This instruction takes a ``va_list*`` value and the type of the
6063 argument. It returns a value of the specified argument type and
6064 increments the ``va_list`` to point to the next argument. The actual
6065 type of ``va_list`` is target specific.
6070 The '``va_arg``' instruction loads an argument of the specified type
6071 from the specified ``va_list`` and causes the ``va_list`` to point to
6072 the next argument. For more information, see the variable argument
6073 handling :ref:`Intrinsic Functions <int_varargs>`.
6075 It is legal for this instruction to be called in a function which does
6076 not take a variable number of arguments, for example, the ``vfprintf``
6079 ``va_arg`` is an LLVM instruction instead of an :ref:`intrinsic
6080 function <intrinsics>` because it takes a type as an argument.
6085 See the :ref:`variable argument processing <int_varargs>` section.
6087 Note that the code generator does not yet fully support va\_arg on many
6088 targets. Also, it does not currently support va\_arg with aggregate
6089 types on any target.
6093 '``landingpad``' Instruction
6094 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6101 <resultval> = landingpad <resultty> personality <type> <pers_fn> <clause>+
6102 <resultval> = landingpad <resultty> personality <type> <pers_fn> cleanup <clause>*
6104 <clause> := catch <type> <value>
6105 <clause> := filter <array constant type> <array constant>
6110 The '``landingpad``' instruction is used by `LLVM's exception handling
6111 system <ExceptionHandling.html#overview>`_ to specify that a basic block
6112 is a landing pad --- one where the exception lands, and corresponds to the
6113 code found in the ``catch`` portion of a ``try``/``catch`` sequence. It
6114 defines values supplied by the personality function (``pers_fn``) upon
6115 re-entry to the function. The ``resultval`` has the type ``resultty``.
6120 This instruction takes a ``pers_fn`` value. This is the personality
6121 function associated with the unwinding mechanism. The optional
6122 ``cleanup`` flag indicates that the landing pad block is a cleanup.
6124 A ``clause`` begins with the clause type --- ``catch`` or ``filter`` --- and
6125 contains the global variable representing the "type" that may be caught
6126 or filtered respectively. Unlike the ``catch`` clause, the ``filter``
6127 clause takes an array constant as its argument. Use
6128 "``[0 x i8**] undef``" for a filter which cannot throw. The
6129 '``landingpad``' instruction must contain *at least* one ``clause`` or
6130 the ``cleanup`` flag.
6135 The '``landingpad``' instruction defines the values which are set by the
6136 personality function (``pers_fn``) upon re-entry to the function, and
6137 therefore the "result type" of the ``landingpad`` instruction. As with
6138 calling conventions, how the personality function results are
6139 represented in LLVM IR is target specific.
6141 The clauses are applied in order from top to bottom. If two
6142 ``landingpad`` instructions are merged together through inlining, the
6143 clauses from the calling function are appended to the list of clauses.
6144 When the call stack is being unwound due to an exception being thrown,
6145 the exception is compared against each ``clause`` in turn. If it doesn't
6146 match any of the clauses, and the ``cleanup`` flag is not set, then
6147 unwinding continues further up the call stack.
6149 The ``landingpad`` instruction has several restrictions:
6151 - A landing pad block is a basic block which is the unwind destination
6152 of an '``invoke``' instruction.
6153 - A landing pad block must have a '``landingpad``' instruction as its
6154 first non-PHI instruction.
6155 - There can be only one '``landingpad``' instruction within the landing
6157 - A basic block that is not a landing pad block may not include a
6158 '``landingpad``' instruction.
6159 - All '``landingpad``' instructions in a function must have the same
6160 personality function.
6165 .. code-block:: llvm
6167 ;; A landing pad which can catch an integer.
6168 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6170 ;; A landing pad that is a cleanup.
6171 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6173 ;; A landing pad which can catch an integer and can only throw a double.
6174 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6176 filter [1 x i8**] [@_ZTId]
6183 LLVM supports the notion of an "intrinsic function". These functions
6184 have well known names and semantics and are required to follow certain
6185 restrictions. Overall, these intrinsics represent an extension mechanism
6186 for the LLVM language that does not require changing all of the
6187 transformations in LLVM when adding to the language (or the bitcode
6188 reader/writer, the parser, etc...).
6190 Intrinsic function names must all start with an "``llvm.``" prefix. This
6191 prefix is reserved in LLVM for intrinsic names; thus, function names may
6192 not begin with this prefix. Intrinsic functions must always be external
6193 functions: you cannot define the body of intrinsic functions. Intrinsic
6194 functions may only be used in call or invoke instructions: it is illegal
6195 to take the address of an intrinsic function. Additionally, because
6196 intrinsic functions are part of the LLVM language, it is required if any
6197 are added that they be documented here.
6199 Some intrinsic functions can be overloaded, i.e., the intrinsic
6200 represents a family of functions that perform the same operation but on
6201 different data types. Because LLVM can represent over 8 million
6202 different integer types, overloading is used commonly to allow an
6203 intrinsic function to operate on any integer type. One or more of the
6204 argument types or the result type can be overloaded to accept any
6205 integer type. Argument types may also be defined as exactly matching a
6206 previous argument's type or the result type. This allows an intrinsic
6207 function which accepts multiple arguments, but needs all of them to be
6208 of the same type, to only be overloaded with respect to a single
6209 argument or the result.
6211 Overloaded intrinsics will have the names of its overloaded argument
6212 types encoded into its function name, each preceded by a period. Only
6213 those types which are overloaded result in a name suffix. Arguments
6214 whose type is matched against another type do not. For example, the
6215 ``llvm.ctpop`` function can take an integer of any width and returns an
6216 integer of exactly the same integer width. This leads to a family of
6217 functions such as ``i8 @llvm.ctpop.i8(i8 %val)`` and
6218 ``i29 @llvm.ctpop.i29(i29 %val)``. Only one type, the return type, is
6219 overloaded, and only one type suffix is required. Because the argument's
6220 type is matched against the return type, it does not require its own
6223 To learn how to add an intrinsic function, please see the `Extending
6224 LLVM Guide <ExtendingLLVM.html>`_.
6228 Variable Argument Handling Intrinsics
6229 -------------------------------------
6231 Variable argument support is defined in LLVM with the
6232 :ref:`va_arg <i_va_arg>` instruction and these three intrinsic
6233 functions. These functions are related to the similarly named macros
6234 defined in the ``<stdarg.h>`` header file.
6236 All of these functions operate on arguments that use a target-specific
6237 value type "``va_list``". The LLVM assembly language reference manual
6238 does not define what this type is, so all transformations should be
6239 prepared to handle these functions regardless of the type used.
6241 This example shows how the :ref:`va_arg <i_va_arg>` instruction and the
6242 variable argument handling intrinsic functions are used.
6244 .. code-block:: llvm
6246 define i32 @test(i32 %X, ...) {
6247 ; Initialize variable argument processing
6249 %ap2 = bitcast i8** %ap to i8*
6250 call void @llvm.va_start(i8* %ap2)
6252 ; Read a single integer argument
6253 %tmp = va_arg i8** %ap, i32
6255 ; Demonstrate usage of llvm.va_copy and llvm.va_end
6257 %aq2 = bitcast i8** %aq to i8*
6258 call void @llvm.va_copy(i8* %aq2, i8* %ap2)
6259 call void @llvm.va_end(i8* %aq2)
6261 ; Stop processing of arguments.
6262 call void @llvm.va_end(i8* %ap2)
6266 declare void @llvm.va_start(i8*)
6267 declare void @llvm.va_copy(i8*, i8*)
6268 declare void @llvm.va_end(i8*)
6272 '``llvm.va_start``' Intrinsic
6273 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6280 declare void %llvm.va_start(i8* <arglist>)
6285 The '``llvm.va_start``' intrinsic initializes ``*<arglist>`` for
6286 subsequent use by ``va_arg``.
6291 The argument is a pointer to a ``va_list`` element to initialize.
6296 The '``llvm.va_start``' intrinsic works just like the ``va_start`` macro
6297 available in C. In a target-dependent way, it initializes the
6298 ``va_list`` element to which the argument points, so that the next call
6299 to ``va_arg`` will produce the first variable argument passed to the
6300 function. Unlike the C ``va_start`` macro, this intrinsic does not need
6301 to know the last argument of the function as the compiler can figure
6304 '``llvm.va_end``' Intrinsic
6305 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6312 declare void @llvm.va_end(i8* <arglist>)
6317 The '``llvm.va_end``' intrinsic destroys ``*<arglist>``, which has been
6318 initialized previously with ``llvm.va_start`` or ``llvm.va_copy``.
6323 The argument is a pointer to a ``va_list`` to destroy.
6328 The '``llvm.va_end``' intrinsic works just like the ``va_end`` macro
6329 available in C. In a target-dependent way, it destroys the ``va_list``
6330 element to which the argument points. Calls to
6331 :ref:`llvm.va_start <int_va_start>` and
6332 :ref:`llvm.va_copy <int_va_copy>` must be matched exactly with calls to
6337 '``llvm.va_copy``' Intrinsic
6338 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6345 declare void @llvm.va_copy(i8* <destarglist>, i8* <srcarglist>)
6350 The '``llvm.va_copy``' intrinsic copies the current argument position
6351 from the source argument list to the destination argument list.
6356 The first argument is a pointer to a ``va_list`` element to initialize.
6357 The second argument is a pointer to a ``va_list`` element to copy from.
6362 The '``llvm.va_copy``' intrinsic works just like the ``va_copy`` macro
6363 available in C. In a target-dependent way, it copies the source
6364 ``va_list`` element into the destination ``va_list`` element. This
6365 intrinsic is necessary because the `` llvm.va_start`` intrinsic may be
6366 arbitrarily complex and require, for example, memory allocation.
6368 Accurate Garbage Collection Intrinsics
6369 --------------------------------------
6371 LLVM support for `Accurate Garbage Collection <GarbageCollection.html>`_
6372 (GC) requires the implementation and generation of these intrinsics.
6373 These intrinsics allow identification of :ref:`GC roots on the
6374 stack <int_gcroot>`, as well as garbage collector implementations that
6375 require :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers.
6376 Front-ends for type-safe garbage collected languages should generate
6377 these intrinsics to make use of the LLVM garbage collectors. For more
6378 details, see `Accurate Garbage Collection with
6379 LLVM <GarbageCollection.html>`_.
6381 The garbage collection intrinsics only operate on objects in the generic
6382 address space (address space zero).
6386 '``llvm.gcroot``' Intrinsic
6387 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6394 declare void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
6399 The '``llvm.gcroot``' intrinsic declares the existence of a GC root to
6400 the code generator, and allows some metadata to be associated with it.
6405 The first argument specifies the address of a stack object that contains
6406 the root pointer. The second pointer (which must be either a constant or
6407 a global value address) contains the meta-data to be associated with the
6413 At runtime, a call to this intrinsic stores a null pointer into the
6414 "ptrloc" location. At compile-time, the code generator generates
6415 information to allow the runtime to find the pointer at GC safe points.
6416 The '``llvm.gcroot``' intrinsic may only be used in a function which
6417 :ref:`specifies a GC algorithm <gc>`.
6421 '``llvm.gcread``' Intrinsic
6422 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6429 declare i8* @llvm.gcread(i8* %ObjPtr, i8** %Ptr)
6434 The '``llvm.gcread``' intrinsic identifies reads of references from heap
6435 locations, allowing garbage collector implementations that require read
6441 The second argument is the address to read from, which should be an
6442 address allocated from the garbage collector. The first object is a
6443 pointer to the start of the referenced object, if needed by the language
6444 runtime (otherwise null).
6449 The '``llvm.gcread``' intrinsic has the same semantics as a load
6450 instruction, but may be replaced with substantially more complex code by
6451 the garbage collector runtime, as needed. The '``llvm.gcread``'
6452 intrinsic may only be used in a function which :ref:`specifies a GC
6457 '``llvm.gcwrite``' Intrinsic
6458 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6465 declare void @llvm.gcwrite(i8* %P1, i8* %Obj, i8** %P2)
6470 The '``llvm.gcwrite``' intrinsic identifies writes of references to heap
6471 locations, allowing garbage collector implementations that require write
6472 barriers (such as generational or reference counting collectors).
6477 The first argument is the reference to store, the second is the start of
6478 the object to store it to, and the third is the address of the field of
6479 Obj to store to. If the runtime does not require a pointer to the
6480 object, Obj may be null.
6485 The '``llvm.gcwrite``' intrinsic has the same semantics as a store
6486 instruction, but may be replaced with substantially more complex code by
6487 the garbage collector runtime, as needed. The '``llvm.gcwrite``'
6488 intrinsic may only be used in a function which :ref:`specifies a GC
6491 Code Generator Intrinsics
6492 -------------------------
6494 These intrinsics are provided by LLVM to expose special features that
6495 may only be implemented with code generator support.
6497 '``llvm.returnaddress``' Intrinsic
6498 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6505 declare i8 *@llvm.returnaddress(i32 <level>)
6510 The '``llvm.returnaddress``' intrinsic attempts to compute a
6511 target-specific value indicating the return address of the current
6512 function or one of its callers.
6517 The argument to this intrinsic indicates which function to return the
6518 address for. Zero indicates the calling function, one indicates its
6519 caller, etc. The argument is **required** to be a constant integer
6525 The '``llvm.returnaddress``' intrinsic either returns a pointer
6526 indicating the return address of the specified call frame, or zero if it
6527 cannot be identified. The value returned by this intrinsic is likely to
6528 be incorrect or 0 for arguments other than zero, so it should only be
6529 used for debugging purposes.
6531 Note that calling this intrinsic does not prevent function inlining or
6532 other aggressive transformations, so the value returned may not be that
6533 of the obvious source-language caller.
6535 '``llvm.frameaddress``' Intrinsic
6536 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6543 declare i8* @llvm.frameaddress(i32 <level>)
6548 The '``llvm.frameaddress``' intrinsic attempts to return the
6549 target-specific frame pointer value for the specified stack frame.
6554 The argument to this intrinsic indicates which function to return the
6555 frame pointer for. Zero indicates the calling function, one indicates
6556 its caller, etc. The argument is **required** to be a constant integer
6562 The '``llvm.frameaddress``' intrinsic either returns a pointer
6563 indicating the frame address of the specified call frame, or zero if it
6564 cannot be identified. The value returned by this intrinsic is likely to
6565 be incorrect or 0 for arguments other than zero, so it should only be
6566 used for debugging purposes.
6568 Note that calling this intrinsic does not prevent function inlining or
6569 other aggressive transformations, so the value returned may not be that
6570 of the obvious source-language caller.
6574 '``llvm.stacksave``' Intrinsic
6575 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6582 declare i8* @llvm.stacksave()
6587 The '``llvm.stacksave``' intrinsic is used to remember the current state
6588 of the function stack, for use with
6589 :ref:`llvm.stackrestore <int_stackrestore>`. This is useful for
6590 implementing language features like scoped automatic variable sized
6596 This intrinsic returns a opaque pointer value that can be passed to
6597 :ref:`llvm.stackrestore <int_stackrestore>`. When an
6598 ``llvm.stackrestore`` intrinsic is executed with a value saved from
6599 ``llvm.stacksave``, it effectively restores the state of the stack to
6600 the state it was in when the ``llvm.stacksave`` intrinsic executed. In
6601 practice, this pops any :ref:`alloca <i_alloca>` blocks from the stack that
6602 were allocated after the ``llvm.stacksave`` was executed.
6604 .. _int_stackrestore:
6606 '``llvm.stackrestore``' Intrinsic
6607 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6614 declare void @llvm.stackrestore(i8* %ptr)
6619 The '``llvm.stackrestore``' intrinsic is used to restore the state of
6620 the function stack to the state it was in when the corresponding
6621 :ref:`llvm.stacksave <int_stacksave>` intrinsic executed. This is
6622 useful for implementing language features like scoped automatic variable
6623 sized arrays in C99.
6628 See the description for :ref:`llvm.stacksave <int_stacksave>`.
6630 '``llvm.prefetch``' Intrinsic
6631 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6638 declare void @llvm.prefetch(i8* <address>, i32 <rw>, i32 <locality>, i32 <cache type>)
6643 The '``llvm.prefetch``' intrinsic is a hint to the code generator to
6644 insert a prefetch instruction if supported; otherwise, it is a noop.
6645 Prefetches have no effect on the behavior of the program but can change
6646 its performance characteristics.
6651 ``address`` is the address to be prefetched, ``rw`` is the specifier
6652 determining if the fetch should be for a read (0) or write (1), and
6653 ``locality`` is a temporal locality specifier ranging from (0) - no
6654 locality, to (3) - extremely local keep in cache. The ``cache type``
6655 specifies whether the prefetch is performed on the data (1) or
6656 instruction (0) cache. The ``rw``, ``locality`` and ``cache type``
6657 arguments must be constant integers.
6662 This intrinsic does not modify the behavior of the program. In
6663 particular, prefetches cannot trap and do not produce a value. On
6664 targets that support this intrinsic, the prefetch can provide hints to
6665 the processor cache for better performance.
6667 '``llvm.pcmarker``' Intrinsic
6668 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6675 declare void @llvm.pcmarker(i32 <id>)
6680 The '``llvm.pcmarker``' intrinsic is a method to export a Program
6681 Counter (PC) in a region of code to simulators and other tools. The
6682 method is target specific, but it is expected that the marker will use
6683 exported symbols to transmit the PC of the marker. The marker makes no
6684 guarantees that it will remain with any specific instruction after
6685 optimizations. It is possible that the presence of a marker will inhibit
6686 optimizations. The intended use is to be inserted after optimizations to
6687 allow correlations of simulation runs.
6692 ``id`` is a numerical id identifying the marker.
6697 This intrinsic does not modify the behavior of the program. Backends
6698 that do not support this intrinsic may ignore it.
6700 '``llvm.readcyclecounter``' Intrinsic
6701 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6708 declare i64 @llvm.readcyclecounter()
6713 The '``llvm.readcyclecounter``' intrinsic provides access to the cycle
6714 counter register (or similar low latency, high accuracy clocks) on those
6715 targets that support it. On X86, it should map to RDTSC. On Alpha, it
6716 should map to RPCC. As the backing counters overflow quickly (on the
6717 order of 9 seconds on alpha), this should only be used for small
6723 When directly supported, reading the cycle counter should not modify any
6724 memory. Implementations are allowed to either return a application
6725 specific value or a system wide value. On backends without support, this
6726 is lowered to a constant 0.
6728 Note that runtime support may be conditional on the privilege-level code is
6729 running at and the host platform.
6731 Standard C Library Intrinsics
6732 -----------------------------
6734 LLVM provides intrinsics for a few important standard C library
6735 functions. These intrinsics allow source-language front-ends to pass
6736 information about the alignment of the pointer arguments to the code
6737 generator, providing opportunity for more efficient code generation.
6741 '``llvm.memcpy``' Intrinsic
6742 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6747 This is an overloaded intrinsic. You can use ``llvm.memcpy`` on any
6748 integer bit width and for different address spaces. Not all targets
6749 support all bit widths however.
6753 declare void @llvm.memcpy.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
6754 i32 <len>, i32 <align>, i1 <isvolatile>)
6755 declare void @llvm.memcpy.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
6756 i64 <len>, i32 <align>, i1 <isvolatile>)
6761 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
6762 source location to the destination location.
6764 Note that, unlike the standard libc function, the ``llvm.memcpy.*``
6765 intrinsics do not return a value, takes extra alignment/isvolatile
6766 arguments and the pointers can be in specified address spaces.
6771 The first argument is a pointer to the destination, the second is a
6772 pointer to the source. The third argument is an integer argument
6773 specifying the number of bytes to copy, the fourth argument is the
6774 alignment of the source and destination locations, and the fifth is a
6775 boolean indicating a volatile access.
6777 If the call to this intrinsic has an alignment value that is not 0 or 1,
6778 then the caller guarantees that both the source and destination pointers
6779 are aligned to that boundary.
6781 If the ``isvolatile`` parameter is ``true``, the ``llvm.memcpy`` call is
6782 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
6783 very cleanly specified and it is unwise to depend on it.
6788 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
6789 source location to the destination location, which are not allowed to
6790 overlap. It copies "len" bytes of memory over. If the argument is known
6791 to be aligned to some boundary, this can be specified as the fourth
6792 argument, otherwise it should be set to 0 or 1.
6794 '``llvm.memmove``' Intrinsic
6795 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6800 This is an overloaded intrinsic. You can use llvm.memmove on any integer
6801 bit width and for different address space. Not all targets support all
6806 declare void @llvm.memmove.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
6807 i32 <len>, i32 <align>, i1 <isvolatile>)
6808 declare void @llvm.memmove.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
6809 i64 <len>, i32 <align>, i1 <isvolatile>)
6814 The '``llvm.memmove.*``' intrinsics move a block of memory from the
6815 source location to the destination location. It is similar to the
6816 '``llvm.memcpy``' intrinsic but allows the two memory locations to
6819 Note that, unlike the standard libc function, the ``llvm.memmove.*``
6820 intrinsics do not return a value, takes extra alignment/isvolatile
6821 arguments and the pointers can be in specified address spaces.
6826 The first argument is a pointer to the destination, the second is a
6827 pointer to the source. The third argument is an integer argument
6828 specifying the number of bytes to copy, the fourth argument is the
6829 alignment of the source and destination locations, and the fifth is a
6830 boolean indicating a volatile access.
6832 If the call to this intrinsic has an alignment value that is not 0 or 1,
6833 then the caller guarantees that the source and destination pointers are
6834 aligned to that boundary.
6836 If the ``isvolatile`` parameter is ``true``, the ``llvm.memmove`` call
6837 is a :ref:`volatile operation <volatile>`. The detailed access behavior is
6838 not very cleanly specified and it is unwise to depend on it.
6843 The '``llvm.memmove.*``' intrinsics copy a block of memory from the
6844 source location to the destination location, which may overlap. It
6845 copies "len" bytes of memory over. If the argument is known to be
6846 aligned to some boundary, this can be specified as the fourth argument,
6847 otherwise it should be set to 0 or 1.
6849 '``llvm.memset.*``' Intrinsics
6850 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6855 This is an overloaded intrinsic. You can use llvm.memset on any integer
6856 bit width and for different address spaces. However, not all targets
6857 support all bit widths.
6861 declare void @llvm.memset.p0i8.i32(i8* <dest>, i8 <val>,
6862 i32 <len>, i32 <align>, i1 <isvolatile>)
6863 declare void @llvm.memset.p0i8.i64(i8* <dest>, i8 <val>,
6864 i64 <len>, i32 <align>, i1 <isvolatile>)
6869 The '``llvm.memset.*``' intrinsics fill a block of memory with a
6870 particular byte value.
6872 Note that, unlike the standard libc function, the ``llvm.memset``
6873 intrinsic does not return a value and takes extra alignment/volatile
6874 arguments. Also, the destination can be in an arbitrary address space.
6879 The first argument is a pointer to the destination to fill, the second
6880 is the byte value with which to fill it, the third argument is an
6881 integer argument specifying the number of bytes to fill, and the fourth
6882 argument is the known alignment of the destination location.
6884 If the call to this intrinsic has an alignment value that is not 0 or 1,
6885 then the caller guarantees that the destination pointer is aligned to
6888 If the ``isvolatile`` parameter is ``true``, the ``llvm.memset`` call is
6889 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
6890 very cleanly specified and it is unwise to depend on it.
6895 The '``llvm.memset.*``' intrinsics fill "len" bytes of memory starting
6896 at the destination location. If the argument is known to be aligned to
6897 some boundary, this can be specified as the fourth argument, otherwise
6898 it should be set to 0 or 1.
6900 '``llvm.sqrt.*``' Intrinsic
6901 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6906 This is an overloaded intrinsic. You can use ``llvm.sqrt`` on any
6907 floating point or vector of floating point type. Not all targets support
6912 declare float @llvm.sqrt.f32(float %Val)
6913 declare double @llvm.sqrt.f64(double %Val)
6914 declare x86_fp80 @llvm.sqrt.f80(x86_fp80 %Val)
6915 declare fp128 @llvm.sqrt.f128(fp128 %Val)
6916 declare ppc_fp128 @llvm.sqrt.ppcf128(ppc_fp128 %Val)
6921 The '``llvm.sqrt``' intrinsics return the sqrt of the specified operand,
6922 returning the same value as the libm '``sqrt``' functions would. Unlike
6923 ``sqrt`` in libm, however, ``llvm.sqrt`` has undefined behavior for
6924 negative numbers other than -0.0 (which allows for better optimization,
6925 because there is no need to worry about errno being set).
6926 ``llvm.sqrt(-0.0)`` is defined to return -0.0 like IEEE sqrt.
6931 The argument and return value are floating point numbers of the same
6937 This function returns the sqrt of the specified operand if it is a
6938 nonnegative floating point number.
6940 '``llvm.powi.*``' Intrinsic
6941 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6946 This is an overloaded intrinsic. You can use ``llvm.powi`` on any
6947 floating point or vector of floating point type. Not all targets support
6952 declare float @llvm.powi.f32(float %Val, i32 %power)
6953 declare double @llvm.powi.f64(double %Val, i32 %power)
6954 declare x86_fp80 @llvm.powi.f80(x86_fp80 %Val, i32 %power)
6955 declare fp128 @llvm.powi.f128(fp128 %Val, i32 %power)
6956 declare ppc_fp128 @llvm.powi.ppcf128(ppc_fp128 %Val, i32 %power)
6961 The '``llvm.powi.*``' intrinsics return the first operand raised to the
6962 specified (positive or negative) power. The order of evaluation of
6963 multiplications is not defined. When a vector of floating point type is
6964 used, the second argument remains a scalar integer value.
6969 The second argument is an integer power, and the first is a value to
6970 raise to that power.
6975 This function returns the first value raised to the second power with an
6976 unspecified sequence of rounding operations.
6978 '``llvm.sin.*``' Intrinsic
6979 ^^^^^^^^^^^^^^^^^^^^^^^^^^
6984 This is an overloaded intrinsic. You can use ``llvm.sin`` on any
6985 floating point or vector of floating point type. Not all targets support
6990 declare float @llvm.sin.f32(float %Val)
6991 declare double @llvm.sin.f64(double %Val)
6992 declare x86_fp80 @llvm.sin.f80(x86_fp80 %Val)
6993 declare fp128 @llvm.sin.f128(fp128 %Val)
6994 declare ppc_fp128 @llvm.sin.ppcf128(ppc_fp128 %Val)
6999 The '``llvm.sin.*``' intrinsics return the sine of the operand.
7004 The argument and return value are floating point numbers of the same
7010 This function returns the sine of the specified operand, returning the
7011 same values as the libm ``sin`` functions would, and handles error
7012 conditions in the same way.
7014 '``llvm.cos.*``' Intrinsic
7015 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7020 This is an overloaded intrinsic. You can use ``llvm.cos`` on any
7021 floating point or vector of floating point type. Not all targets support
7026 declare float @llvm.cos.f32(float %Val)
7027 declare double @llvm.cos.f64(double %Val)
7028 declare x86_fp80 @llvm.cos.f80(x86_fp80 %Val)
7029 declare fp128 @llvm.cos.f128(fp128 %Val)
7030 declare ppc_fp128 @llvm.cos.ppcf128(ppc_fp128 %Val)
7035 The '``llvm.cos.*``' intrinsics return the cosine of the operand.
7040 The argument and return value are floating point numbers of the same
7046 This function returns the cosine of the specified operand, returning the
7047 same values as the libm ``cos`` functions would, and handles error
7048 conditions in the same way.
7050 '``llvm.pow.*``' Intrinsic
7051 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7056 This is an overloaded intrinsic. You can use ``llvm.pow`` on any
7057 floating point or vector of floating point type. Not all targets support
7062 declare float @llvm.pow.f32(float %Val, float %Power)
7063 declare double @llvm.pow.f64(double %Val, double %Power)
7064 declare x86_fp80 @llvm.pow.f80(x86_fp80 %Val, x86_fp80 %Power)
7065 declare fp128 @llvm.pow.f128(fp128 %Val, fp128 %Power)
7066 declare ppc_fp128 @llvm.pow.ppcf128(ppc_fp128 %Val, ppc_fp128 Power)
7071 The '``llvm.pow.*``' intrinsics return the first operand raised to the
7072 specified (positive or negative) power.
7077 The second argument is a floating point power, and the first is a value
7078 to raise to that power.
7083 This function returns the first value raised to the second power,
7084 returning the same values as the libm ``pow`` functions would, and
7085 handles error conditions in the same way.
7087 '``llvm.exp.*``' Intrinsic
7088 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7093 This is an overloaded intrinsic. You can use ``llvm.exp`` on any
7094 floating point or vector of floating point type. Not all targets support
7099 declare float @llvm.exp.f32(float %Val)
7100 declare double @llvm.exp.f64(double %Val)
7101 declare x86_fp80 @llvm.exp.f80(x86_fp80 %Val)
7102 declare fp128 @llvm.exp.f128(fp128 %Val)
7103 declare ppc_fp128 @llvm.exp.ppcf128(ppc_fp128 %Val)
7108 The '``llvm.exp.*``' intrinsics perform the exp function.
7113 The argument and return value are floating point numbers of the same
7119 This function returns the same values as the libm ``exp`` functions
7120 would, and handles error conditions in the same way.
7122 '``llvm.exp2.*``' Intrinsic
7123 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7128 This is an overloaded intrinsic. You can use ``llvm.exp2`` on any
7129 floating point or vector of floating point type. Not all targets support
7134 declare float @llvm.exp2.f32(float %Val)
7135 declare double @llvm.exp2.f64(double %Val)
7136 declare x86_fp80 @llvm.exp2.f80(x86_fp80 %Val)
7137 declare fp128 @llvm.exp2.f128(fp128 %Val)
7138 declare ppc_fp128 @llvm.exp2.ppcf128(ppc_fp128 %Val)
7143 The '``llvm.exp2.*``' intrinsics perform the exp2 function.
7148 The argument and return value are floating point numbers of the same
7154 This function returns the same values as the libm ``exp2`` functions
7155 would, and handles error conditions in the same way.
7157 '``llvm.log.*``' Intrinsic
7158 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7163 This is an overloaded intrinsic. You can use ``llvm.log`` on any
7164 floating point or vector of floating point type. Not all targets support
7169 declare float @llvm.log.f32(float %Val)
7170 declare double @llvm.log.f64(double %Val)
7171 declare x86_fp80 @llvm.log.f80(x86_fp80 %Val)
7172 declare fp128 @llvm.log.f128(fp128 %Val)
7173 declare ppc_fp128 @llvm.log.ppcf128(ppc_fp128 %Val)
7178 The '``llvm.log.*``' intrinsics perform the log function.
7183 The argument and return value are floating point numbers of the same
7189 This function returns the same values as the libm ``log`` functions
7190 would, and handles error conditions in the same way.
7192 '``llvm.log10.*``' Intrinsic
7193 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7198 This is an overloaded intrinsic. You can use ``llvm.log10`` on any
7199 floating point or vector of floating point type. Not all targets support
7204 declare float @llvm.log10.f32(float %Val)
7205 declare double @llvm.log10.f64(double %Val)
7206 declare x86_fp80 @llvm.log10.f80(x86_fp80 %Val)
7207 declare fp128 @llvm.log10.f128(fp128 %Val)
7208 declare ppc_fp128 @llvm.log10.ppcf128(ppc_fp128 %Val)
7213 The '``llvm.log10.*``' intrinsics perform the log10 function.
7218 The argument and return value are floating point numbers of the same
7224 This function returns the same values as the libm ``log10`` functions
7225 would, and handles error conditions in the same way.
7227 '``llvm.log2.*``' Intrinsic
7228 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7233 This is an overloaded intrinsic. You can use ``llvm.log2`` on any
7234 floating point or vector of floating point type. Not all targets support
7239 declare float @llvm.log2.f32(float %Val)
7240 declare double @llvm.log2.f64(double %Val)
7241 declare x86_fp80 @llvm.log2.f80(x86_fp80 %Val)
7242 declare fp128 @llvm.log2.f128(fp128 %Val)
7243 declare ppc_fp128 @llvm.log2.ppcf128(ppc_fp128 %Val)
7248 The '``llvm.log2.*``' intrinsics perform the log2 function.
7253 The argument and return value are floating point numbers of the same
7259 This function returns the same values as the libm ``log2`` functions
7260 would, and handles error conditions in the same way.
7262 '``llvm.fma.*``' Intrinsic
7263 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7268 This is an overloaded intrinsic. You can use ``llvm.fma`` on any
7269 floating point or vector of floating point type. Not all targets support
7274 declare float @llvm.fma.f32(float %a, float %b, float %c)
7275 declare double @llvm.fma.f64(double %a, double %b, double %c)
7276 declare x86_fp80 @llvm.fma.f80(x86_fp80 %a, x86_fp80 %b, x86_fp80 %c)
7277 declare fp128 @llvm.fma.f128(fp128 %a, fp128 %b, fp128 %c)
7278 declare ppc_fp128 @llvm.fma.ppcf128(ppc_fp128 %a, ppc_fp128 %b, ppc_fp128 %c)
7283 The '``llvm.fma.*``' intrinsics perform the fused multiply-add
7289 The argument and return value are floating point numbers of the same
7295 This function returns the same values as the libm ``fma`` functions
7298 '``llvm.fabs.*``' Intrinsic
7299 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7304 This is an overloaded intrinsic. You can use ``llvm.fabs`` on any
7305 floating point or vector of floating point type. Not all targets support
7310 declare float @llvm.fabs.f32(float %Val)
7311 declare double @llvm.fabs.f64(double %Val)
7312 declare x86_fp80 @llvm.fabs.f80(x86_fp80 %Val)
7313 declare fp128 @llvm.fabs.f128(fp128 %Val)
7314 declare ppc_fp128 @llvm.fabs.ppcf128(ppc_fp128 %Val)
7319 The '``llvm.fabs.*``' intrinsics return the absolute value of the
7325 The argument and return value are floating point numbers of the same
7331 This function returns the same values as the libm ``fabs`` functions
7332 would, and handles error conditions in the same way.
7334 '``llvm.floor.*``' Intrinsic
7335 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7340 This is an overloaded intrinsic. You can use ``llvm.floor`` on any
7341 floating point or vector of floating point type. Not all targets support
7346 declare float @llvm.floor.f32(float %Val)
7347 declare double @llvm.floor.f64(double %Val)
7348 declare x86_fp80 @llvm.floor.f80(x86_fp80 %Val)
7349 declare fp128 @llvm.floor.f128(fp128 %Val)
7350 declare ppc_fp128 @llvm.floor.ppcf128(ppc_fp128 %Val)
7355 The '``llvm.floor.*``' intrinsics return the floor of the operand.
7360 The argument and return value are floating point numbers of the same
7366 This function returns the same values as the libm ``floor`` functions
7367 would, and handles error conditions in the same way.
7369 '``llvm.ceil.*``' Intrinsic
7370 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7375 This is an overloaded intrinsic. You can use ``llvm.ceil`` on any
7376 floating point or vector of floating point type. Not all targets support
7381 declare float @llvm.ceil.f32(float %Val)
7382 declare double @llvm.ceil.f64(double %Val)
7383 declare x86_fp80 @llvm.ceil.f80(x86_fp80 %Val)
7384 declare fp128 @llvm.ceil.f128(fp128 %Val)
7385 declare ppc_fp128 @llvm.ceil.ppcf128(ppc_fp128 %Val)
7390 The '``llvm.ceil.*``' intrinsics return the ceiling of the operand.
7395 The argument and return value are floating point numbers of the same
7401 This function returns the same values as the libm ``ceil`` functions
7402 would, and handles error conditions in the same way.
7404 '``llvm.trunc.*``' Intrinsic
7405 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7410 This is an overloaded intrinsic. You can use ``llvm.trunc`` on any
7411 floating point or vector of floating point type. Not all targets support
7416 declare float @llvm.trunc.f32(float %Val)
7417 declare double @llvm.trunc.f64(double %Val)
7418 declare x86_fp80 @llvm.trunc.f80(x86_fp80 %Val)
7419 declare fp128 @llvm.trunc.f128(fp128 %Val)
7420 declare ppc_fp128 @llvm.trunc.ppcf128(ppc_fp128 %Val)
7425 The '``llvm.trunc.*``' intrinsics returns the operand rounded to the
7426 nearest integer not larger in magnitude than the operand.
7431 The argument and return value are floating point numbers of the same
7437 This function returns the same values as the libm ``trunc`` functions
7438 would, and handles error conditions in the same way.
7440 '``llvm.rint.*``' Intrinsic
7441 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7446 This is an overloaded intrinsic. You can use ``llvm.rint`` on any
7447 floating point or vector of floating point type. Not all targets support
7452 declare float @llvm.rint.f32(float %Val)
7453 declare double @llvm.rint.f64(double %Val)
7454 declare x86_fp80 @llvm.rint.f80(x86_fp80 %Val)
7455 declare fp128 @llvm.rint.f128(fp128 %Val)
7456 declare ppc_fp128 @llvm.rint.ppcf128(ppc_fp128 %Val)
7461 The '``llvm.rint.*``' intrinsics returns the operand rounded to the
7462 nearest integer. It may raise an inexact floating-point exception if the
7463 operand isn't an integer.
7468 The argument and return value are floating point numbers of the same
7474 This function returns the same values as the libm ``rint`` functions
7475 would, and handles error conditions in the same way.
7477 '``llvm.nearbyint.*``' Intrinsic
7478 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7483 This is an overloaded intrinsic. You can use ``llvm.nearbyint`` on any
7484 floating point or vector of floating point type. Not all targets support
7489 declare float @llvm.nearbyint.f32(float %Val)
7490 declare double @llvm.nearbyint.f64(double %Val)
7491 declare x86_fp80 @llvm.nearbyint.f80(x86_fp80 %Val)
7492 declare fp128 @llvm.nearbyint.f128(fp128 %Val)
7493 declare ppc_fp128 @llvm.nearbyint.ppcf128(ppc_fp128 %Val)
7498 The '``llvm.nearbyint.*``' intrinsics returns the operand rounded to the
7504 The argument and return value are floating point numbers of the same
7510 This function returns the same values as the libm ``nearbyint``
7511 functions would, and handles error conditions in the same way.
7513 Bit Manipulation Intrinsics
7514 ---------------------------
7516 LLVM provides intrinsics for a few important bit manipulation
7517 operations. These allow efficient code generation for some algorithms.
7519 '``llvm.bswap.*``' Intrinsics
7520 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7525 This is an overloaded intrinsic function. You can use bswap on any
7526 integer type that is an even number of bytes (i.e. BitWidth % 16 == 0).
7530 declare i16 @llvm.bswap.i16(i16 <id>)
7531 declare i32 @llvm.bswap.i32(i32 <id>)
7532 declare i64 @llvm.bswap.i64(i64 <id>)
7537 The '``llvm.bswap``' family of intrinsics is used to byte swap integer
7538 values with an even number of bytes (positive multiple of 16 bits).
7539 These are useful for performing operations on data that is not in the
7540 target's native byte order.
7545 The ``llvm.bswap.i16`` intrinsic returns an i16 value that has the high
7546 and low byte of the input i16 swapped. Similarly, the ``llvm.bswap.i32``
7547 intrinsic returns an i32 value that has the four bytes of the input i32
7548 swapped, so that if the input bytes are numbered 0, 1, 2, 3 then the
7549 returned i32 will have its bytes in 3, 2, 1, 0 order. The
7550 ``llvm.bswap.i48``, ``llvm.bswap.i64`` and other intrinsics extend this
7551 concept to additional even-byte lengths (6 bytes, 8 bytes and more,
7554 '``llvm.ctpop.*``' Intrinsic
7555 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7560 This is an overloaded intrinsic. You can use llvm.ctpop on any integer
7561 bit width, or on any vector with integer elements. Not all targets
7562 support all bit widths or vector types, however.
7566 declare i8 @llvm.ctpop.i8(i8 <src>)
7567 declare i16 @llvm.ctpop.i16(i16 <src>)
7568 declare i32 @llvm.ctpop.i32(i32 <src>)
7569 declare i64 @llvm.ctpop.i64(i64 <src>)
7570 declare i256 @llvm.ctpop.i256(i256 <src>)
7571 declare <2 x i32> @llvm.ctpop.v2i32(<2 x i32> <src>)
7576 The '``llvm.ctpop``' family of intrinsics counts the number of bits set
7582 The only argument is the value to be counted. The argument may be of any
7583 integer type, or a vector with integer elements. The return type must
7584 match the argument type.
7589 The '``llvm.ctpop``' intrinsic counts the 1's in a variable, or within
7590 each element of a vector.
7592 '``llvm.ctlz.*``' Intrinsic
7593 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7598 This is an overloaded intrinsic. You can use ``llvm.ctlz`` on any
7599 integer bit width, or any vector whose elements are integers. Not all
7600 targets support all bit widths or vector types, however.
7604 declare i8 @llvm.ctlz.i8 (i8 <src>, i1 <is_zero_undef>)
7605 declare i16 @llvm.ctlz.i16 (i16 <src>, i1 <is_zero_undef>)
7606 declare i32 @llvm.ctlz.i32 (i32 <src>, i1 <is_zero_undef>)
7607 declare i64 @llvm.ctlz.i64 (i64 <src>, i1 <is_zero_undef>)
7608 declare i256 @llvm.ctlz.i256(i256 <src>, i1 <is_zero_undef>)
7609 declase <2 x i32> @llvm.ctlz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
7614 The '``llvm.ctlz``' family of intrinsic functions counts the number of
7615 leading zeros in a variable.
7620 The first argument is the value to be counted. This argument may be of
7621 any integer type, or a vectory with integer element type. The return
7622 type must match the first argument type.
7624 The second argument must be a constant and is a flag to indicate whether
7625 the intrinsic should ensure that a zero as the first argument produces a
7626 defined result. Historically some architectures did not provide a
7627 defined result for zero values as efficiently, and many algorithms are
7628 now predicated on avoiding zero-value inputs.
7633 The '``llvm.ctlz``' intrinsic counts the leading (most significant)
7634 zeros in a variable, or within each element of the vector. If
7635 ``src == 0`` then the result is the size in bits of the type of ``src``
7636 if ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
7637 ``llvm.ctlz(i32 2) = 30``.
7639 '``llvm.cttz.*``' Intrinsic
7640 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7645 This is an overloaded intrinsic. You can use ``llvm.cttz`` on any
7646 integer bit width, or any vector of integer elements. Not all targets
7647 support all bit widths or vector types, however.
7651 declare i8 @llvm.cttz.i8 (i8 <src>, i1 <is_zero_undef>)
7652 declare i16 @llvm.cttz.i16 (i16 <src>, i1 <is_zero_undef>)
7653 declare i32 @llvm.cttz.i32 (i32 <src>, i1 <is_zero_undef>)
7654 declare i64 @llvm.cttz.i64 (i64 <src>, i1 <is_zero_undef>)
7655 declare i256 @llvm.cttz.i256(i256 <src>, i1 <is_zero_undef>)
7656 declase <2 x i32> @llvm.cttz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
7661 The '``llvm.cttz``' family of intrinsic functions counts the number of
7667 The first argument is the value to be counted. This argument may be of
7668 any integer type, or a vectory with integer element type. The return
7669 type must match the first argument type.
7671 The second argument must be a constant and is a flag to indicate whether
7672 the intrinsic should ensure that a zero as the first argument produces a
7673 defined result. Historically some architectures did not provide a
7674 defined result for zero values as efficiently, and many algorithms are
7675 now predicated on avoiding zero-value inputs.
7680 The '``llvm.cttz``' intrinsic counts the trailing (least significant)
7681 zeros in a variable, or within each element of a vector. If ``src == 0``
7682 then the result is the size in bits of the type of ``src`` if
7683 ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
7684 ``llvm.cttz(2) = 1``.
7686 Arithmetic with Overflow Intrinsics
7687 -----------------------------------
7689 LLVM provides intrinsics for some arithmetic with overflow operations.
7691 '``llvm.sadd.with.overflow.*``' Intrinsics
7692 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7697 This is an overloaded intrinsic. You can use ``llvm.sadd.with.overflow``
7698 on any integer bit width.
7702 declare {i16, i1} @llvm.sadd.with.overflow.i16(i16 %a, i16 %b)
7703 declare {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
7704 declare {i64, i1} @llvm.sadd.with.overflow.i64(i64 %a, i64 %b)
7709 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
7710 a signed addition of the two arguments, and indicate whether an overflow
7711 occurred during the signed summation.
7716 The arguments (%a and %b) and the first element of the result structure
7717 may be of integer types of any bit width, but they must have the same
7718 bit width. The second element of the result structure must be of type
7719 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
7725 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
7726 a signed addition of the two variables. They return a structure --- the
7727 first element of which is the signed summation, and the second element
7728 of which is a bit specifying if the signed summation resulted in an
7734 .. code-block:: llvm
7736 %res = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
7737 %sum = extractvalue {i32, i1} %res, 0
7738 %obit = extractvalue {i32, i1} %res, 1
7739 br i1 %obit, label %overflow, label %normal
7741 '``llvm.uadd.with.overflow.*``' Intrinsics
7742 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7747 This is an overloaded intrinsic. You can use ``llvm.uadd.with.overflow``
7748 on any integer bit width.
7752 declare {i16, i1} @llvm.uadd.with.overflow.i16(i16 %a, i16 %b)
7753 declare {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
7754 declare {i64, i1} @llvm.uadd.with.overflow.i64(i64 %a, i64 %b)
7759 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
7760 an unsigned addition of the two arguments, and indicate whether a carry
7761 occurred during the unsigned summation.
7766 The arguments (%a and %b) and the first element of the result structure
7767 may be of integer types of any bit width, but they must have the same
7768 bit width. The second element of the result structure must be of type
7769 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
7775 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
7776 an unsigned addition of the two arguments. They return a structure --- the
7777 first element of which is the sum, and the second element of which is a
7778 bit specifying if the unsigned summation resulted in a carry.
7783 .. code-block:: llvm
7785 %res = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
7786 %sum = extractvalue {i32, i1} %res, 0
7787 %obit = extractvalue {i32, i1} %res, 1
7788 br i1 %obit, label %carry, label %normal
7790 '``llvm.ssub.with.overflow.*``' Intrinsics
7791 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7796 This is an overloaded intrinsic. You can use ``llvm.ssub.with.overflow``
7797 on any integer bit width.
7801 declare {i16, i1} @llvm.ssub.with.overflow.i16(i16 %a, i16 %b)
7802 declare {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
7803 declare {i64, i1} @llvm.ssub.with.overflow.i64(i64 %a, i64 %b)
7808 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
7809 a signed subtraction of the two arguments, and indicate whether an
7810 overflow occurred during the signed subtraction.
7815 The arguments (%a and %b) and the first element of the result structure
7816 may be of integer types of any bit width, but they must have the same
7817 bit width. The second element of the result structure must be of type
7818 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
7824 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
7825 a signed subtraction of the two arguments. They return a structure --- the
7826 first element of which is the subtraction, and the second element of
7827 which is a bit specifying if the signed subtraction resulted in an
7833 .. code-block:: llvm
7835 %res = call {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
7836 %sum = extractvalue {i32, i1} %res, 0
7837 %obit = extractvalue {i32, i1} %res, 1
7838 br i1 %obit, label %overflow, label %normal
7840 '``llvm.usub.with.overflow.*``' Intrinsics
7841 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7846 This is an overloaded intrinsic. You can use ``llvm.usub.with.overflow``
7847 on any integer bit width.
7851 declare {i16, i1} @llvm.usub.with.overflow.i16(i16 %a, i16 %b)
7852 declare {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
7853 declare {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
7858 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
7859 an unsigned subtraction of the two arguments, and indicate whether an
7860 overflow occurred during the unsigned subtraction.
7865 The arguments (%a and %b) and the first element of the result structure
7866 may be of integer types of any bit width, but they must have the same
7867 bit width. The second element of the result structure must be of type
7868 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
7874 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
7875 an unsigned subtraction of the two arguments. They return a structure ---
7876 the first element of which is the subtraction, and the second element of
7877 which is a bit specifying if the unsigned subtraction resulted in an
7883 .. code-block:: llvm
7885 %res = call {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
7886 %sum = extractvalue {i32, i1} %res, 0
7887 %obit = extractvalue {i32, i1} %res, 1
7888 br i1 %obit, label %overflow, label %normal
7890 '``llvm.smul.with.overflow.*``' Intrinsics
7891 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7896 This is an overloaded intrinsic. You can use ``llvm.smul.with.overflow``
7897 on any integer bit width.
7901 declare {i16, i1} @llvm.smul.with.overflow.i16(i16 %a, i16 %b)
7902 declare {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
7903 declare {i64, i1} @llvm.smul.with.overflow.i64(i64 %a, i64 %b)
7908 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
7909 a signed multiplication of the two arguments, and indicate whether an
7910 overflow occurred during the signed multiplication.
7915 The arguments (%a and %b) and the first element of the result structure
7916 may be of integer types of any bit width, but they must have the same
7917 bit width. The second element of the result structure must be of type
7918 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
7924 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
7925 a signed multiplication of the two arguments. They return a structure ---
7926 the first element of which is the multiplication, and the second element
7927 of which is a bit specifying if the signed multiplication resulted in an
7933 .. code-block:: llvm
7935 %res = call {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
7936 %sum = extractvalue {i32, i1} %res, 0
7937 %obit = extractvalue {i32, i1} %res, 1
7938 br i1 %obit, label %overflow, label %normal
7940 '``llvm.umul.with.overflow.*``' Intrinsics
7941 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7946 This is an overloaded intrinsic. You can use ``llvm.umul.with.overflow``
7947 on any integer bit width.
7951 declare {i16, i1} @llvm.umul.with.overflow.i16(i16 %a, i16 %b)
7952 declare {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
7953 declare {i64, i1} @llvm.umul.with.overflow.i64(i64 %a, i64 %b)
7958 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
7959 a unsigned multiplication of the two arguments, and indicate whether an
7960 overflow occurred during the unsigned multiplication.
7965 The arguments (%a and %b) and the first element of the result structure
7966 may be of integer types of any bit width, but they must have the same
7967 bit width. The second element of the result structure must be of type
7968 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
7974 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
7975 an unsigned multiplication of the two arguments. They return a structure ---
7976 the first element of which is the multiplication, and the second
7977 element of which is a bit specifying if the unsigned multiplication
7978 resulted in an overflow.
7983 .. code-block:: llvm
7985 %res = call {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
7986 %sum = extractvalue {i32, i1} %res, 0
7987 %obit = extractvalue {i32, i1} %res, 1
7988 br i1 %obit, label %overflow, label %normal
7990 Specialised Arithmetic Intrinsics
7991 ---------------------------------
7993 '``llvm.fmuladd.*``' Intrinsic
7994 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8001 declare float @llvm.fmuladd.f32(float %a, float %b, float %c)
8002 declare double @llvm.fmuladd.f64(double %a, double %b, double %c)
8007 The '``llvm.fmuladd.*``' intrinsic functions represent multiply-add
8008 expressions that can be fused if the code generator determines that (a) the
8009 target instruction set has support for a fused operation, and (b) that the
8010 fused operation is more efficient than the equivalent, separate pair of mul
8011 and add instructions.
8016 The '``llvm.fmuladd.*``' intrinsics each take three arguments: two
8017 multiplicands, a and b, and an addend c.
8026 %0 = call float @llvm.fmuladd.f32(%a, %b, %c)
8028 is equivalent to the expression a \* b + c, except that rounding will
8029 not be performed between the multiplication and addition steps if the
8030 code generator fuses the operations. Fusion is not guaranteed, even if
8031 the target platform supports it. If a fused multiply-add is required the
8032 corresponding llvm.fma.\* intrinsic function should be used instead.
8037 .. code-block:: llvm
8039 %r2 = call float @llvm.fmuladd.f32(float %a, float %b, float %c) ; yields {float}:r2 = (a * b) + c
8041 Half Precision Floating Point Intrinsics
8042 ----------------------------------------
8044 For most target platforms, half precision floating point is a
8045 storage-only format. This means that it is a dense encoding (in memory)
8046 but does not support computation in the format.
8048 This means that code must first load the half-precision floating point
8049 value as an i16, then convert it to float with
8050 :ref:`llvm.convert.from.fp16 <int_convert_from_fp16>`. Computation can
8051 then be performed on the float value (including extending to double
8052 etc). To store the value back to memory, it is first converted to float
8053 if needed, then converted to i16 with
8054 :ref:`llvm.convert.to.fp16 <int_convert_to_fp16>`, then storing as an
8057 .. _int_convert_to_fp16:
8059 '``llvm.convert.to.fp16``' Intrinsic
8060 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8067 declare i16 @llvm.convert.to.fp16(f32 %a)
8072 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion
8073 from single precision floating point format to half precision floating
8079 The intrinsic function contains single argument - the value to be
8085 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion
8086 from single precision floating point format to half precision floating
8087 point format. The return value is an ``i16`` which contains the
8093 .. code-block:: llvm
8095 %res = call i16 @llvm.convert.to.fp16(f32 %a)
8096 store i16 %res, i16* @x, align 2
8098 .. _int_convert_from_fp16:
8100 '``llvm.convert.from.fp16``' Intrinsic
8101 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8108 declare f32 @llvm.convert.from.fp16(i16 %a)
8113 The '``llvm.convert.from.fp16``' intrinsic function performs a
8114 conversion from half precision floating point format to single precision
8115 floating point format.
8120 The intrinsic function contains single argument - the value to be
8126 The '``llvm.convert.from.fp16``' intrinsic function performs a
8127 conversion from half single precision floating point format to single
8128 precision floating point format. The input half-float value is
8129 represented by an ``i16`` value.
8134 .. code-block:: llvm
8136 %a = load i16* @x, align 2
8137 %res = call f32 @llvm.convert.from.fp16(i16 %a)
8142 The LLVM debugger intrinsics (which all start with ``llvm.dbg.``
8143 prefix), are described in the `LLVM Source Level
8144 Debugging <SourceLevelDebugging.html#format_common_intrinsics>`_
8147 Exception Handling Intrinsics
8148 -----------------------------
8150 The LLVM exception handling intrinsics (which all start with
8151 ``llvm.eh.`` prefix), are described in the `LLVM Exception
8152 Handling <ExceptionHandling.html#format_common_intrinsics>`_ document.
8156 Trampoline Intrinsics
8157 ---------------------
8159 These intrinsics make it possible to excise one parameter, marked with
8160 the :ref:`nest <nest>` attribute, from a function. The result is a
8161 callable function pointer lacking the nest parameter - the caller does
8162 not need to provide a value for it. Instead, the value to use is stored
8163 in advance in a "trampoline", a block of memory usually allocated on the
8164 stack, which also contains code to splice the nest value into the
8165 argument list. This is used to implement the GCC nested function address
8168 For example, if the function is ``i32 f(i8* nest %c, i32 %x, i32 %y)``
8169 then the resulting function pointer has signature ``i32 (i32, i32)*``.
8170 It can be created as follows:
8172 .. code-block:: llvm
8174 %tramp = alloca [10 x i8], align 4 ; size and alignment only correct for X86
8175 %tramp1 = getelementptr [10 x i8]* %tramp, i32 0, i32 0
8176 call i8* @llvm.init.trampoline(i8* %tramp1, i8* bitcast (i32 (i8*, i32, i32)* @f to i8*), i8* %nval)
8177 %p = call i8* @llvm.adjust.trampoline(i8* %tramp1)
8178 %fp = bitcast i8* %p to i32 (i32, i32)*
8180 The call ``%val = call i32 %fp(i32 %x, i32 %y)`` is then equivalent to
8181 ``%val = call i32 %f(i8* %nval, i32 %x, i32 %y)``.
8185 '``llvm.init.trampoline``' Intrinsic
8186 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8193 declare void @llvm.init.trampoline(i8* <tramp>, i8* <func>, i8* <nval>)
8198 This fills the memory pointed to by ``tramp`` with executable code,
8199 turning it into a trampoline.
8204 The ``llvm.init.trampoline`` intrinsic takes three arguments, all
8205 pointers. The ``tramp`` argument must point to a sufficiently large and
8206 sufficiently aligned block of memory; this memory is written to by the
8207 intrinsic. Note that the size and the alignment are target-specific -
8208 LLVM currently provides no portable way of determining them, so a
8209 front-end that generates this intrinsic needs to have some
8210 target-specific knowledge. The ``func`` argument must hold a function
8211 bitcast to an ``i8*``.
8216 The block of memory pointed to by ``tramp`` is filled with target
8217 dependent code, turning it into a function. Then ``tramp`` needs to be
8218 passed to :ref:`llvm.adjust.trampoline <int_at>` to get a pointer which can
8219 be :ref:`bitcast (to a new function) and called <int_trampoline>`. The new
8220 function's signature is the same as that of ``func`` with any arguments
8221 marked with the ``nest`` attribute removed. At most one such ``nest``
8222 argument is allowed, and it must be of pointer type. Calling the new
8223 function is equivalent to calling ``func`` with the same argument list,
8224 but with ``nval`` used for the missing ``nest`` argument. If, after
8225 calling ``llvm.init.trampoline``, the memory pointed to by ``tramp`` is
8226 modified, then the effect of any later call to the returned function
8227 pointer is undefined.
8231 '``llvm.adjust.trampoline``' Intrinsic
8232 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8239 declare i8* @llvm.adjust.trampoline(i8* <tramp>)
8244 This performs any required machine-specific adjustment to the address of
8245 a trampoline (passed as ``tramp``).
8250 ``tramp`` must point to a block of memory which already has trampoline
8251 code filled in by a previous call to
8252 :ref:`llvm.init.trampoline <int_it>`.
8257 On some architectures the address of the code to be executed needs to be
8258 different to the address where the trampoline is actually stored. This
8259 intrinsic returns the executable address corresponding to ``tramp``
8260 after performing the required machine specific adjustments. The pointer
8261 returned can then be :ref:`bitcast and executed <int_trampoline>`.
8266 This class of intrinsics exists to information about the lifetime of
8267 memory objects and ranges where variables are immutable.
8269 '``llvm.lifetime.start``' Intrinsic
8270 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8277 declare void @llvm.lifetime.start(i64 <size>, i8* nocapture <ptr>)
8282 The '``llvm.lifetime.start``' intrinsic specifies the start of a memory
8288 The first argument is a constant integer representing the size of the
8289 object, or -1 if it is variable sized. The second argument is a pointer
8295 This intrinsic indicates that before this point in the code, the value
8296 of the memory pointed to by ``ptr`` is dead. This means that it is known
8297 to never be used and has an undefined value. A load from the pointer
8298 that precedes this intrinsic can be replaced with ``'undef'``.
8300 '``llvm.lifetime.end``' Intrinsic
8301 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8308 declare void @llvm.lifetime.end(i64 <size>, i8* nocapture <ptr>)
8313 The '``llvm.lifetime.end``' intrinsic specifies the end of a memory
8319 The first argument is a constant integer representing the size of the
8320 object, or -1 if it is variable sized. The second argument is a pointer
8326 This intrinsic indicates that after this point in the code, the value of
8327 the memory pointed to by ``ptr`` is dead. This means that it is known to
8328 never be used and has an undefined value. Any stores into the memory
8329 object following this intrinsic may be removed as dead.
8331 '``llvm.invariant.start``' Intrinsic
8332 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8339 declare {}* @llvm.invariant.start(i64 <size>, i8* nocapture <ptr>)
8344 The '``llvm.invariant.start``' intrinsic specifies that the contents of
8345 a memory object will not change.
8350 The first argument is a constant integer representing the size of the
8351 object, or -1 if it is variable sized. The second argument is a pointer
8357 This intrinsic indicates that until an ``llvm.invariant.end`` that uses
8358 the return value, the referenced memory location is constant and
8361 '``llvm.invariant.end``' Intrinsic
8362 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8369 declare void @llvm.invariant.end({}* <start>, i64 <size>, i8* nocapture <ptr>)
8374 The '``llvm.invariant.end``' intrinsic specifies that the contents of a
8375 memory object are mutable.
8380 The first argument is the matching ``llvm.invariant.start`` intrinsic.
8381 The second argument is a constant integer representing the size of the
8382 object, or -1 if it is variable sized and the third argument is a
8383 pointer to the object.
8388 This intrinsic indicates that the memory is mutable again.
8393 This class of intrinsics is designed to be generic and has no specific
8396 '``llvm.var.annotation``' Intrinsic
8397 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8404 declare void @llvm.var.annotation(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
8409 The '``llvm.var.annotation``' intrinsic.
8414 The first argument is a pointer to a value, the second is a pointer to a
8415 global string, the third is a pointer to a global string which is the
8416 source file name, and the last argument is the line number.
8421 This intrinsic allows annotation of local variables with arbitrary
8422 strings. This can be useful for special purpose optimizations that want
8423 to look for these annotations. These have no other defined use; they are
8424 ignored by code generation and optimization.
8426 '``llvm.ptr.annotation.*``' Intrinsic
8427 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8432 This is an overloaded intrinsic. You can use '``llvm.ptr.annotation``' on a
8433 pointer to an integer of any width. *NOTE* you must specify an address space for
8434 the pointer. The identifier for the default address space is the integer
8439 declare i8* @llvm.ptr.annotation.p<address space>i8(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
8440 declare i16* @llvm.ptr.annotation.p<address space>i16(i16* <val>, i8* <str>, i8* <str>, i32 <int>)
8441 declare i32* @llvm.ptr.annotation.p<address space>i32(i32* <val>, i8* <str>, i8* <str>, i32 <int>)
8442 declare i64* @llvm.ptr.annotation.p<address space>i64(i64* <val>, i8* <str>, i8* <str>, i32 <int>)
8443 declare i256* @llvm.ptr.annotation.p<address space>i256(i256* <val>, i8* <str>, i8* <str>, i32 <int>)
8448 The '``llvm.ptr.annotation``' intrinsic.
8453 The first argument is a pointer to an integer value of arbitrary bitwidth
8454 (result of some expression), the second is a pointer to a global string, the
8455 third is a pointer to a global string which is the source file name, and the
8456 last argument is the line number. It returns the value of the first argument.
8461 This intrinsic allows annotation of a pointer to an integer with arbitrary
8462 strings. This can be useful for special purpose optimizations that want to look
8463 for these annotations. These have no other defined use; they are ignored by code
8464 generation and optimization.
8466 '``llvm.annotation.*``' Intrinsic
8467 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8472 This is an overloaded intrinsic. You can use '``llvm.annotation``' on
8473 any integer bit width.
8477 declare i8 @llvm.annotation.i8(i8 <val>, i8* <str>, i8* <str>, i32 <int>)
8478 declare i16 @llvm.annotation.i16(i16 <val>, i8* <str>, i8* <str>, i32 <int>)
8479 declare i32 @llvm.annotation.i32(i32 <val>, i8* <str>, i8* <str>, i32 <int>)
8480 declare i64 @llvm.annotation.i64(i64 <val>, i8* <str>, i8* <str>, i32 <int>)
8481 declare i256 @llvm.annotation.i256(i256 <val>, i8* <str>, i8* <str>, i32 <int>)
8486 The '``llvm.annotation``' intrinsic.
8491 The first argument is an integer value (result of some expression), the
8492 second is a pointer to a global string, the third is a pointer to a
8493 global string which is the source file name, and the last argument is
8494 the line number. It returns the value of the first argument.
8499 This intrinsic allows annotations to be put on arbitrary expressions
8500 with arbitrary strings. This can be useful for special purpose
8501 optimizations that want to look for these annotations. These have no
8502 other defined use; they are ignored by code generation and optimization.
8504 '``llvm.trap``' Intrinsic
8505 ^^^^^^^^^^^^^^^^^^^^^^^^^
8512 declare void @llvm.trap() noreturn nounwind
8517 The '``llvm.trap``' intrinsic.
8527 This intrinsic is lowered to the target dependent trap instruction. If
8528 the target does not have a trap instruction, this intrinsic will be
8529 lowered to a call of the ``abort()`` function.
8531 '``llvm.debugtrap``' Intrinsic
8532 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8539 declare void @llvm.debugtrap() nounwind
8544 The '``llvm.debugtrap``' intrinsic.
8554 This intrinsic is lowered to code which is intended to cause an
8555 execution trap with the intention of requesting the attention of a
8558 '``llvm.stackprotector``' Intrinsic
8559 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8566 declare void @llvm.stackprotector(i8* <guard>, i8** <slot>)
8571 The ``llvm.stackprotector`` intrinsic takes the ``guard`` and stores it
8572 onto the stack at ``slot``. The stack slot is adjusted to ensure that it
8573 is placed on the stack before local variables.
8578 The ``llvm.stackprotector`` intrinsic requires two pointer arguments.
8579 The first argument is the value loaded from the stack guard
8580 ``@__stack_chk_guard``. The second variable is an ``alloca`` that has
8581 enough space to hold the value of the guard.
8586 This intrinsic causes the prologue/epilogue inserter to force the
8587 position of the ``AllocaInst`` stack slot to be before local variables
8588 on the stack. This is to ensure that if a local variable on the stack is
8589 overwritten, it will destroy the value of the guard. When the function
8590 exits, the guard on the stack is checked against the original guard. If
8591 they are different, then the program aborts by calling the
8592 ``__stack_chk_fail()`` function.
8594 '``llvm.objectsize``' Intrinsic
8595 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8602 declare i32 @llvm.objectsize.i32(i8* <object>, i1 <min>)
8603 declare i64 @llvm.objectsize.i64(i8* <object>, i1 <min>)
8608 The ``llvm.objectsize`` intrinsic is designed to provide information to
8609 the optimizers to determine at compile time whether a) an operation
8610 (like memcpy) will overflow a buffer that corresponds to an object, or
8611 b) that a runtime check for overflow isn't necessary. An object in this
8612 context means an allocation of a specific class, structure, array, or
8618 The ``llvm.objectsize`` intrinsic takes two arguments. The first
8619 argument is a pointer to or into the ``object``. The second argument is
8620 a boolean and determines whether ``llvm.objectsize`` returns 0 (if true)
8621 or -1 (if false) when the object size is unknown. The second argument
8622 only accepts constants.
8627 The ``llvm.objectsize`` intrinsic is lowered to a constant representing
8628 the size of the object concerned. If the size cannot be determined at
8629 compile time, ``llvm.objectsize`` returns ``i32/i64 -1 or 0`` (depending
8630 on the ``min`` argument).
8632 '``llvm.expect``' Intrinsic
8633 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8640 declare i32 @llvm.expect.i32(i32 <val>, i32 <expected_val>)
8641 declare i64 @llvm.expect.i64(i64 <val>, i64 <expected_val>)
8646 The ``llvm.expect`` intrinsic provides information about expected (the
8647 most probable) value of ``val``, which can be used by optimizers.
8652 The ``llvm.expect`` intrinsic takes two arguments. The first argument is
8653 a value. The second argument is an expected value, this needs to be a
8654 constant value, variables are not allowed.
8659 This intrinsic is lowered to the ``val``.
8661 '``llvm.donothing``' Intrinsic
8662 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8669 declare void @llvm.donothing() nounwind readnone
8674 The ``llvm.donothing`` intrinsic doesn't perform any operation. It's the
8675 only intrinsic that can be called with an invoke instruction.
8685 This intrinsic does nothing, and it's removed by optimizers and ignored